projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: assign VS/TCS/TES/GS shader input parameter locations dynamically
[mesa.git] / src / gallium / drivers / radeonsi / si_shader.c
1 /*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 * Tom Stellard <thomas.stellard@amd.com>
25 * Michel Dänzer <michel.daenzer@amd.com>
26 * Christian König <christian.koenig@amd.com>
27 */
28
29 #include "gallivm/lp_bld_const.h"
30 #include "gallivm/lp_bld_gather.h"
31 #include "gallivm/lp_bld_intr.h"
32 #include "gallivm/lp_bld_logic.h"
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_flow.h"
35 #include "gallivm/lp_bld_misc.h"
36 #include "util/u_memory.h"
37 #include "util/u_string.h"
38 #include "tgsi/tgsi_build.h"
39 #include "tgsi/tgsi_util.h"
40 #include "tgsi/tgsi_dump.h"
41
42 #include "ac_binary.h"
43 #include "ac_llvm_util.h"
44 #include "ac_exp_param.h"
45 #include "si_shader_internal.h"
46 #include "si_pipe.h"
47 #include "sid.h"
48
49
50 static const char *scratch_rsrc_dword0_symbol =
51 "SCRATCH_RSRC_DWORD0";
52
53 static const char *scratch_rsrc_dword1_symbol =
54 "SCRATCH_RSRC_DWORD1";
55
56 struct si_shader_output_values
57 {
58 LLVMValueRef values[4];
59 unsigned semantic_name;
60 unsigned semantic_index;
61 ubyte vertex_stream[4];
62 };
63
64 static void si_init_shader_ctx(struct si_shader_context *ctx,
65 struct si_screen *sscreen,
66 LLVMTargetMachineRef tm);
67
68 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
69 struct lp_build_tgsi_context *bld_base,
70 struct lp_build_emit_data *emit_data);
71
72 static void si_dump_shader_key(unsigned processor, struct si_shader *shader,
73 FILE *f);
74
75 static unsigned llvm_get_type_size(LLVMTypeRef type);
76
77 static void si_build_vs_prolog_function(struct si_shader_context *ctx,
78 union si_shader_part_key *key);
79 static void si_build_vs_epilog_function(struct si_shader_context *ctx,
80 union si_shader_part_key *key);
81 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
82 union si_shader_part_key *key);
83 static void si_build_ps_prolog_function(struct si_shader_context *ctx,
84 union si_shader_part_key *key);
85 static void si_build_ps_epilog_function(struct si_shader_context *ctx,
86 union si_shader_part_key *key);
87
88 /* Ideally pass the sample mask input to the PS epilog as v13, which
89 * is its usual location, so that the shader doesn't have to add v_mov.
90 */
91 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 13
92
93 /* The VS location of the PrimitiveID input is the same in the epilog,
94 * so that the main shader part doesn't have to move it.
95 */
96 #define VS_EPILOG_PRIMID_LOC 2
97
98 enum {
99 CONST_ADDR_SPACE = 2,
100 LOCAL_ADDR_SPACE = 3,
101 };
102
103 /**
104 * Returns a unique index for a semantic name and index. The index must be
105 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
106 * calculated.
107 */
108 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index)
109 {
110 switch (semantic_name) {
111 case TGSI_SEMANTIC_POSITION:
112 return 0;
113 case TGSI_SEMANTIC_PSIZE:
114 return 1;
115 case TGSI_SEMANTIC_CLIPDIST:
116 assert(index <= 1);
117 return 2 + index;
118 case TGSI_SEMANTIC_GENERIC:
119 if (index <= 63-4)
120 return 4 + index;
121
122 assert(!"invalid generic index");
123 return 0;
124
125 /* patch indices are completely separate and thus start from 0 */
126 case TGSI_SEMANTIC_TESSOUTER:
127 return 0;
128 case TGSI_SEMANTIC_TESSINNER:
129 return 1;
130 case TGSI_SEMANTIC_PATCH:
131 return 2 + index;
132
133 default:
134 assert(!"invalid semantic name");
135 return 0;
136 }
137 }
138
139 unsigned si_shader_io_get_unique_index2(unsigned name, unsigned index)
140 {
141 switch (name) {
142 case TGSI_SEMANTIC_FOG:
143 return 0;
144 case TGSI_SEMANTIC_LAYER:
145 return 1;
146 case TGSI_SEMANTIC_VIEWPORT_INDEX:
147 return 2;
148 case TGSI_SEMANTIC_PRIMID:
149 return 3;
150 case TGSI_SEMANTIC_COLOR: /* these alias */
151 case TGSI_SEMANTIC_BCOLOR:
152 return 4 + index;
153 case TGSI_SEMANTIC_TEXCOORD:
154 return 6 + index;
155 default:
156 assert(!"invalid semantic name");
157 return 0;
158 }
159 }
160
161 /**
162 * Get the value of a shader input parameter and extract a bitfield.
163 */
164 static LLVMValueRef unpack_param(struct si_shader_context *ctx,
165 unsigned param, unsigned rshift,
166 unsigned bitwidth)
167 {
168 struct gallivm_state *gallivm = &ctx->gallivm;
169 LLVMValueRef value = LLVMGetParam(ctx->main_fn,
170 param);
171
172 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
173 value = bitcast(&ctx->bld_base,
174 TGSI_TYPE_UNSIGNED, value);
175
176 if (rshift)
177 value = LLVMBuildLShr(gallivm->builder, value,
178 LLVMConstInt(ctx->i32, rshift, 0), "");
179
180 if (rshift + bitwidth < 32) {
181 unsigned mask = (1 << bitwidth) - 1;
182 value = LLVMBuildAnd(gallivm->builder, value,
183 LLVMConstInt(ctx->i32, mask, 0), "");
184 }
185
186 return value;
187 }
188
189 static LLVMValueRef get_rel_patch_id(struct si_shader_context *ctx)
190 {
191 switch (ctx->type) {
192 case PIPE_SHADER_TESS_CTRL:
193 return unpack_param(ctx, ctx->param_tcs_rel_ids, 0, 8);
194
195 case PIPE_SHADER_TESS_EVAL:
196 return LLVMGetParam(ctx->main_fn,
197 ctx->param_tes_rel_patch_id);
198
199 default:
200 assert(0);
201 return NULL;
202 }
203 }
204
205 /* Tessellation shaders pass outputs to the next shader using LDS.
206 *
207 * LS outputs = TCS inputs
208 * TCS outputs = TES inputs
209 *
210 * The LDS layout is:
211 * - TCS inputs for patch 0
212 * - TCS inputs for patch 1
213 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
214 * - ...
215 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
216 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
217 * - TCS outputs for patch 1
218 * - Per-patch TCS outputs for patch 1
219 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
220 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
221 * - ...
222 *
223 * All three shaders VS(LS), TCS, TES share the same LDS space.
224 */
225
226 static LLVMValueRef
227 get_tcs_in_patch_stride(struct si_shader_context *ctx)
228 {
229 return unpack_param(ctx, ctx->param_vs_state_bits, 8, 13);
230 }
231
232 static LLVMValueRef
233 get_tcs_out_patch_stride(struct si_shader_context *ctx)
234 {
235 return unpack_param(ctx, ctx->param_tcs_out_lds_layout, 0, 13);
236 }
237
238 static LLVMValueRef
239 get_tcs_out_patch0_offset(struct si_shader_context *ctx)
240 {
241 return lp_build_mul_imm(&ctx->bld_base.uint_bld,
242 unpack_param(ctx,
243 ctx->param_tcs_out_lds_offsets,
244 0, 16),
245 4);
246 }
247
248 static LLVMValueRef
249 get_tcs_out_patch0_patch_data_offset(struct si_shader_context *ctx)
250 {
251 return lp_build_mul_imm(&ctx->bld_base.uint_bld,
252 unpack_param(ctx,
253 ctx->param_tcs_out_lds_offsets,
254 16, 16),
255 4);
256 }
257
258 static LLVMValueRef
259 get_tcs_in_current_patch_offset(struct si_shader_context *ctx)
260 {
261 struct gallivm_state *gallivm = &ctx->gallivm;
262 LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
263 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
264
265 return LLVMBuildMul(gallivm->builder, patch_stride, rel_patch_id, "");
266 }
267
268 static LLVMValueRef
269 get_tcs_out_current_patch_offset(struct si_shader_context *ctx)
270 {
271 struct gallivm_state *gallivm = &ctx->gallivm;
272 LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
273 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
274 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
275
276 return LLVMBuildAdd(gallivm->builder, patch0_offset,
277 LLVMBuildMul(gallivm->builder, patch_stride,
278 rel_patch_id, ""),
279 "");
280 }
281
282 static LLVMValueRef
283 get_tcs_out_current_patch_data_offset(struct si_shader_context *ctx)
284 {
285 struct gallivm_state *gallivm = &ctx->gallivm;
286 LLVMValueRef patch0_patch_data_offset =
287 get_tcs_out_patch0_patch_data_offset(ctx);
288 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
289 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
290
291 return LLVMBuildAdd(gallivm->builder, patch0_patch_data_offset,
292 LLVMBuildMul(gallivm->builder, patch_stride,
293 rel_patch_id, ""),
294 "");
295 }
296
297 static LLVMValueRef get_instance_index_for_fetch(
298 struct si_shader_context *ctx,
299 unsigned param_start_instance, unsigned divisor)
300 {
301 struct gallivm_state *gallivm = &ctx->gallivm;
302
303 LLVMValueRef result = LLVMGetParam(ctx->main_fn,
304 ctx->param_instance_id);
305
306 /* The division must be done before START_INSTANCE is added. */
307 if (divisor > 1)
308 result = LLVMBuildUDiv(gallivm->builder, result,
309 LLVMConstInt(ctx->i32, divisor, 0), "");
310
311 return LLVMBuildAdd(gallivm->builder, result,
312 LLVMGetParam(ctx->main_fn, param_start_instance), "");
313 }
314
315 /* Bitcast <4 x float> to <2 x double>, extract the component, and convert
316 * to float. */
317 static LLVMValueRef extract_double_to_float(struct si_shader_context *ctx,
318 LLVMValueRef vec4,
319 unsigned double_index)
320 {
321 LLVMBuilderRef builder = ctx->gallivm.builder;
322 LLVMTypeRef f64 = LLVMDoubleTypeInContext(ctx->gallivm.context);
323 LLVMValueRef dvec2 = LLVMBuildBitCast(builder, vec4,
324 LLVMVectorType(f64, 2), "");
325 LLVMValueRef index = LLVMConstInt(ctx->i32, double_index, 0);
326 LLVMValueRef value = LLVMBuildExtractElement(builder, dvec2, index, "");
327 return LLVMBuildFPTrunc(builder, value, ctx->f32, "");
328 }
329
330 static void declare_input_vs(
331 struct si_shader_context *ctx,
332 unsigned input_index,
333 const struct tgsi_full_declaration *decl,
334 LLVMValueRef out[4])
335 {
336 struct gallivm_state *gallivm = &ctx->gallivm;
337
338 unsigned chan;
339 unsigned fix_fetch;
340 unsigned num_fetches;
341 unsigned fetch_stride;
342
343 LLVMValueRef t_list_ptr;
344 LLVMValueRef t_offset;
345 LLVMValueRef t_list;
346 LLVMValueRef vertex_index;
347 LLVMValueRef input[3];
348
349 /* Load the T list */
350 t_list_ptr = LLVMGetParam(ctx->main_fn, ctx->param_vertex_buffers);
351
352 t_offset = LLVMConstInt(ctx->i32, input_index, 0);
353
354 t_list = ac_build_indexed_load_const(&ctx->ac, t_list_ptr, t_offset);
355
356 vertex_index = LLVMGetParam(ctx->main_fn,
357 ctx->param_vertex_index0 +
358 input_index);
359
360 fix_fetch = ctx->shader->key.mono.vs_fix_fetch[input_index];
361
362 /* Do multiple loads for special formats. */
363 switch (fix_fetch) {
364 case SI_FIX_FETCH_RGB_64_FLOAT:
365 num_fetches = 3; /* 3 2-dword loads */
366 fetch_stride = 8;
367 break;
368 case SI_FIX_FETCH_RGBA_64_FLOAT:
369 num_fetches = 2; /* 2 4-dword loads */
370 fetch_stride = 16;
371 break;
372 case SI_FIX_FETCH_RGB_8:
373 case SI_FIX_FETCH_RGB_8_INT:
374 num_fetches = 3;
375 fetch_stride = 1;
376 break;
377 case SI_FIX_FETCH_RGB_16:
378 case SI_FIX_FETCH_RGB_16_INT:
379 num_fetches = 3;
380 fetch_stride = 2;
381 break;
382 default:
383 num_fetches = 1;
384 fetch_stride = 0;
385 }
386
387 for (unsigned i = 0; i < num_fetches; i++) {
388 LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
389
390 input[i] = ac_build_buffer_load_format(&ctx->ac, t_list,
391 vertex_index, voffset,
392 true);
393 }
394
395 /* Break up the vec4 into individual components */
396 for (chan = 0; chan < 4; chan++) {
397 LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, 0);
398 out[chan] = LLVMBuildExtractElement(gallivm->builder,
399 input[0], llvm_chan, "");
400 }
401
402 switch (fix_fetch) {
403 case SI_FIX_FETCH_A2_SNORM:
404 case SI_FIX_FETCH_A2_SSCALED:
405 case SI_FIX_FETCH_A2_SINT: {
406 /* The hardware returns an unsigned value; convert it to a
407 * signed one.
408 */
409 LLVMValueRef tmp = out[3];
410 LLVMValueRef c30 = LLVMConstInt(ctx->i32, 30, 0);
411
412 /* First, recover the sign-extended signed integer value. */
413 if (fix_fetch == SI_FIX_FETCH_A2_SSCALED)
414 tmp = LLVMBuildFPToUI(gallivm->builder, tmp, ctx->i32, "");
415 else
416 tmp = LLVMBuildBitCast(gallivm->builder, tmp, ctx->i32, "");
417
418 /* For the integer-like cases, do a natural sign extension.
419 *
420 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
421 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
422 * exponent.
423 */
424 tmp = LLVMBuildShl(gallivm->builder, tmp,
425 fix_fetch == SI_FIX_FETCH_A2_SNORM ?
426 LLVMConstInt(ctx->i32, 7, 0) : c30, "");
427 tmp = LLVMBuildAShr(gallivm->builder, tmp, c30, "");
428
429 /* Convert back to the right type. */
430 if (fix_fetch == SI_FIX_FETCH_A2_SNORM) {
431 LLVMValueRef clamp;
432 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
433 tmp = LLVMBuildSIToFP(gallivm->builder, tmp, ctx->f32, "");
434 clamp = LLVMBuildFCmp(gallivm->builder, LLVMRealULT, tmp, neg_one, "");
435 tmp = LLVMBuildSelect(gallivm->builder, clamp, neg_one, tmp, "");
436 } else if (fix_fetch == SI_FIX_FETCH_A2_SSCALED) {
437 tmp = LLVMBuildSIToFP(gallivm->builder, tmp, ctx->f32, "");
438 }
439
440 out[3] = tmp;
441 break;
442 }
443 case SI_FIX_FETCH_RGBA_32_UNORM:
444 case SI_FIX_FETCH_RGBX_32_UNORM:
445 for (chan = 0; chan < 4; chan++) {
446 out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
447 ctx->i32, "");
448 out[chan] = LLVMBuildUIToFP(gallivm->builder,
449 out[chan], ctx->f32, "");
450 out[chan] = LLVMBuildFMul(gallivm->builder, out[chan],
451 LLVMConstReal(ctx->f32, 1.0 / UINT_MAX), "");
452 }
453 /* RGBX UINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
454 if (fix_fetch == SI_FIX_FETCH_RGBX_32_UNORM)
455 out[3] = LLVMConstReal(ctx->f32, 1);
456 break;
457 case SI_FIX_FETCH_RGBA_32_SNORM:
458 case SI_FIX_FETCH_RGBX_32_SNORM:
459 case SI_FIX_FETCH_RGBA_32_FIXED:
460 case SI_FIX_FETCH_RGBX_32_FIXED: {
461 double scale;
462 if (fix_fetch >= SI_FIX_FETCH_RGBA_32_FIXED)
463 scale = 1.0 / 0x10000;
464 else
465 scale = 1.0 / INT_MAX;
466
467 for (chan = 0; chan < 4; chan++) {
468 out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
469 ctx->i32, "");
470 out[chan] = LLVMBuildSIToFP(gallivm->builder,
471 out[chan], ctx->f32, "");
472 out[chan] = LLVMBuildFMul(gallivm->builder, out[chan],
473 LLVMConstReal(ctx->f32, scale), "");
474 }
475 /* RGBX SINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
476 if (fix_fetch == SI_FIX_FETCH_RGBX_32_SNORM ||
477 fix_fetch == SI_FIX_FETCH_RGBX_32_FIXED)
478 out[3] = LLVMConstReal(ctx->f32, 1);
479 break;
480 }
481 case SI_FIX_FETCH_RGBA_32_USCALED:
482 for (chan = 0; chan < 4; chan++) {
483 out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
484 ctx->i32, "");
485 out[chan] = LLVMBuildUIToFP(gallivm->builder,
486 out[chan], ctx->f32, "");
487 }
488 break;
489 case SI_FIX_FETCH_RGBA_32_SSCALED:
490 for (chan = 0; chan < 4; chan++) {
491 out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
492 ctx->i32, "");
493 out[chan] = LLVMBuildSIToFP(gallivm->builder,
494 out[chan], ctx->f32, "");
495 }
496 break;
497 case SI_FIX_FETCH_RG_64_FLOAT:
498 for (chan = 0; chan < 2; chan++)
499 out[chan] = extract_double_to_float(ctx, input[0], chan);
500
501 out[2] = LLVMConstReal(ctx->f32, 0);
502 out[3] = LLVMConstReal(ctx->f32, 1);
503 break;
504 case SI_FIX_FETCH_RGB_64_FLOAT:
505 for (chan = 0; chan < 3; chan++)
506 out[chan] = extract_double_to_float(ctx, input[chan], 0);
507
508 out[3] = LLVMConstReal(ctx->f32, 1);
509 break;
510 case SI_FIX_FETCH_RGBA_64_FLOAT:
511 for (chan = 0; chan < 4; chan++) {
512 out[chan] = extract_double_to_float(ctx, input[chan / 2],
513 chan % 2);
514 }
515 break;
516 case SI_FIX_FETCH_RGB_8:
517 case SI_FIX_FETCH_RGB_8_INT:
518 case SI_FIX_FETCH_RGB_16:
519 case SI_FIX_FETCH_RGB_16_INT:
520 for (chan = 0; chan < 3; chan++) {
521 out[chan] = LLVMBuildExtractElement(gallivm->builder,
522 input[chan],
523 ctx->i32_0, "");
524 }
525 if (fix_fetch == SI_FIX_FETCH_RGB_8 ||
526 fix_fetch == SI_FIX_FETCH_RGB_16) {
527 out[3] = LLVMConstReal(ctx->f32, 1);
528 } else {
529 out[3] = LLVMBuildBitCast(gallivm->builder, ctx->i32_1,
530 ctx->f32, "");
531 }
532 break;
533 }
534 }
535
536 static LLVMValueRef get_primitive_id(struct lp_build_tgsi_context *bld_base,
537 unsigned swizzle)
538 {
539 struct si_shader_context *ctx = si_shader_context(bld_base);
540
541 if (swizzle > 0)
542 return ctx->i32_0;
543
544 switch (ctx->type) {
545 case PIPE_SHADER_VERTEX:
546 return LLVMGetParam(ctx->main_fn,
547 ctx->param_vs_prim_id);
548 case PIPE_SHADER_TESS_CTRL:
549 return LLVMGetParam(ctx->main_fn,
550 ctx->param_tcs_patch_id);
551 case PIPE_SHADER_TESS_EVAL:
552 return LLVMGetParam(ctx->main_fn,
553 ctx->param_tes_patch_id);
554 case PIPE_SHADER_GEOMETRY:
555 return LLVMGetParam(ctx->main_fn,
556 ctx->param_gs_prim_id);
557 default:
558 assert(0);
559 return ctx->i32_0;
560 }
561 }
562
563 /**
564 * Return the value of tgsi_ind_register for indexing.
565 * This is the indirect index with the constant offset added to it.
566 */
567 static LLVMValueRef get_indirect_index(struct si_shader_context *ctx,
568 const struct tgsi_ind_register *ind,
569 int rel_index)
570 {
571 struct gallivm_state *gallivm = &ctx->gallivm;
572 LLVMValueRef result;
573
574 result = ctx->addrs[ind->Index][ind->Swizzle];
575 result = LLVMBuildLoad(gallivm->builder, result, "");
576 result = LLVMBuildAdd(gallivm->builder, result,
577 LLVMConstInt(ctx->i32, rel_index, 0), "");
578 return result;
579 }
580
581 /**
582 * Like get_indirect_index, but restricts the return value to a (possibly
583 * undefined) value inside [0..num).
584 */
585 static LLVMValueRef get_bounded_indirect_index(struct si_shader_context *ctx,
586 const struct tgsi_ind_register *ind,
587 int rel_index, unsigned num)
588 {
589 LLVMValueRef result = get_indirect_index(ctx, ind, rel_index);
590
591 /* LLVM 3.8: If indirect resource indexing is used:
592 * - SI & CIK hang
593 * - VI crashes
594 */
595 if (HAVE_LLVM == 0x0308)
596 return LLVMGetUndef(ctx->i32);
597
598 return si_llvm_bound_index(ctx, result, num);
599 }
600
601
602 /**
603 * Calculate a dword address given an input or output register and a stride.
604 */
605 static LLVMValueRef get_dw_address(struct si_shader_context *ctx,
606 const struct tgsi_full_dst_register *dst,
607 const struct tgsi_full_src_register *src,
608 LLVMValueRef vertex_dw_stride,
609 LLVMValueRef base_addr)
610 {
611 struct gallivm_state *gallivm = &ctx->gallivm;
612 struct tgsi_shader_info *info = &ctx->shader->selector->info;
613 ubyte *name, *index, *array_first;
614 int first, param;
615 struct tgsi_full_dst_register reg;
616
617 /* Set the register description. The address computation is the same
618 * for sources and destinations. */
619 if (src) {
620 reg.Register.File = src->Register.File;
621 reg.Register.Index = src->Register.Index;
622 reg.Register.Indirect = src->Register.Indirect;
623 reg.Register.Dimension = src->Register.Dimension;
624 reg.Indirect = src->Indirect;
625 reg.Dimension = src->Dimension;
626 reg.DimIndirect = src->DimIndirect;
627 } else
628 reg = *dst;
629
630 /* If the register is 2-dimensional (e.g. an array of vertices
631 * in a primitive), calculate the base address of the vertex. */
632 if (reg.Register.Dimension) {
633 LLVMValueRef index;
634
635 if (reg.Dimension.Indirect)
636 index = get_indirect_index(ctx, &reg.DimIndirect,
637 reg.Dimension.Index);
638 else
639 index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0);
640
641 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
642 LLVMBuildMul(gallivm->builder, index,
643 vertex_dw_stride, ""), "");
644 }
645
646 /* Get information about the register. */
647 if (reg.Register.File == TGSI_FILE_INPUT) {
648 name = info->input_semantic_name;
649 index = info->input_semantic_index;
650 array_first = info->input_array_first;
651 } else if (reg.Register.File == TGSI_FILE_OUTPUT) {
652 name = info->output_semantic_name;
653 index = info->output_semantic_index;
654 array_first = info->output_array_first;
655 } else {
656 assert(0);
657 return NULL;
658 }
659
660 if (reg.Register.Indirect) {
661 /* Add the relative address of the element. */
662 LLVMValueRef ind_index;
663
664 if (reg.Indirect.ArrayID)
665 first = array_first[reg.Indirect.ArrayID];
666 else
667 first = reg.Register.Index;
668
669 ind_index = get_indirect_index(ctx, &reg.Indirect,
670 reg.Register.Index - first);
671
672 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
673 LLVMBuildMul(gallivm->builder, ind_index,
674 LLVMConstInt(ctx->i32, 4, 0), ""), "");
675
676 param = si_shader_io_get_unique_index(name[first], index[first]);
677 } else {
678 param = si_shader_io_get_unique_index(name[reg.Register.Index],
679 index[reg.Register.Index]);
680 }
681
682 /* Add the base address of the element. */
683 return LLVMBuildAdd(gallivm->builder, base_addr,
684 LLVMConstInt(ctx->i32, param * 4, 0), "");
685 }
686
687 /* The offchip buffer layout for TCS->TES is
688 *
689 * - attribute 0 of patch 0 vertex 0
690 * - attribute 0 of patch 0 vertex 1
691 * - attribute 0 of patch 0 vertex 2
692 * ...
693 * - attribute 0 of patch 1 vertex 0
694 * - attribute 0 of patch 1 vertex 1
695 * ...
696 * - attribute 1 of patch 0 vertex 0
697 * - attribute 1 of patch 0 vertex 1
698 * ...
699 * - per patch attribute 0 of patch 0
700 * - per patch attribute 0 of patch 1
701 * ...
702 *
703 * Note that every attribute has 4 components.
704 */
705 static LLVMValueRef get_tcs_tes_buffer_address(struct si_shader_context *ctx,
706 LLVMValueRef rel_patch_id,
707 LLVMValueRef vertex_index,
708 LLVMValueRef param_index)
709 {
710 struct gallivm_state *gallivm = &ctx->gallivm;
711 LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
712 LLVMValueRef param_stride, constant16;
713
714 vertices_per_patch = unpack_param(ctx, ctx->param_tcs_offchip_layout, 9, 6);
715 num_patches = unpack_param(ctx, ctx->param_tcs_offchip_layout, 0, 9);
716 total_vertices = LLVMBuildMul(gallivm->builder, vertices_per_patch,
717 num_patches, "");
718
719 constant16 = LLVMConstInt(ctx->i32, 16, 0);
720 if (vertex_index) {
721 base_addr = LLVMBuildMul(gallivm->builder, rel_patch_id,
722 vertices_per_patch, "");
723
724 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
725 vertex_index, "");
726
727 param_stride = total_vertices;
728 } else {
729 base_addr = rel_patch_id;
730 param_stride = num_patches;
731 }
732
733 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
734 LLVMBuildMul(gallivm->builder, param_index,
735 param_stride, ""), "");
736
737 base_addr = LLVMBuildMul(gallivm->builder, base_addr, constant16, "");
738
739 if (!vertex_index) {
740 LLVMValueRef patch_data_offset =
741 unpack_param(ctx, ctx->param_tcs_offchip_layout, 16, 16);
742
743 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
744 patch_data_offset, "");
745 }
746 return base_addr;
747 }
748
749 static LLVMValueRef get_tcs_tes_buffer_address_from_reg(
750 struct si_shader_context *ctx,
751 const struct tgsi_full_dst_register *dst,
752 const struct tgsi_full_src_register *src)
753 {
754 struct gallivm_state *gallivm = &ctx->gallivm;
755 struct tgsi_shader_info *info = &ctx->shader->selector->info;
756 ubyte *name, *index, *array_first;
757 struct tgsi_full_src_register reg;
758 LLVMValueRef vertex_index = NULL;
759 LLVMValueRef param_index = NULL;
760 unsigned param_index_base, param_base;
761
762 reg = src ? *src : tgsi_full_src_register_from_dst(dst);
763
764 if (reg.Register.Dimension) {
765
766 if (reg.Dimension.Indirect)
767 vertex_index = get_indirect_index(ctx, &reg.DimIndirect,
768 reg.Dimension.Index);
769 else
770 vertex_index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0);
771 }
772
773 /* Get information about the register. */
774 if (reg.Register.File == TGSI_FILE_INPUT) {
775 name = info->input_semantic_name;
776 index = info->input_semantic_index;
777 array_first = info->input_array_first;
778 } else if (reg.Register.File == TGSI_FILE_OUTPUT) {
779 name = info->output_semantic_name;
780 index = info->output_semantic_index;
781 array_first = info->output_array_first;
782 } else {
783 assert(0);
784 return NULL;
785 }
786
787 if (reg.Register.Indirect) {
788 if (reg.Indirect.ArrayID)
789 param_base = array_first[reg.Indirect.ArrayID];
790 else
791 param_base = reg.Register.Index;
792
793 param_index = get_indirect_index(ctx, &reg.Indirect,
794 reg.Register.Index - param_base);
795
796 } else {
797 param_base = reg.Register.Index;
798 param_index = ctx->i32_0;
799 }
800
801 param_index_base = si_shader_io_get_unique_index(name[param_base],
802 index[param_base]);
803
804 param_index = LLVMBuildAdd(gallivm->builder, param_index,
805 LLVMConstInt(ctx->i32, param_index_base, 0),
806 "");
807
808 return get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx),
809 vertex_index, param_index);
810 }
811
812 static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base,
813 enum tgsi_opcode_type type, unsigned swizzle,
814 LLVMValueRef buffer, LLVMValueRef offset,
815 LLVMValueRef base, bool readonly_memory)
816 {
817 struct si_shader_context *ctx = si_shader_context(bld_base);
818 struct gallivm_state *gallivm = &ctx->gallivm;
819 LLVMValueRef value, value2;
820 LLVMTypeRef llvm_type = tgsi2llvmtype(bld_base, type);
821 LLVMTypeRef vec_type = LLVMVectorType(llvm_type, 4);
822
823 if (swizzle == ~0) {
824 value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
825 0, 1, 0, readonly_memory);
826
827 return LLVMBuildBitCast(gallivm->builder, value, vec_type, "");
828 }
829
830 if (!tgsi_type_is_64bit(type)) {
831 value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
832 0, 1, 0, readonly_memory);
833
834 value = LLVMBuildBitCast(gallivm->builder, value, vec_type, "");
835 return LLVMBuildExtractElement(gallivm->builder, value,
836 LLVMConstInt(ctx->i32, swizzle, 0), "");
837 }
838
839 value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
840 swizzle * 4, 1, 0, readonly_memory);
841
842 value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
843 swizzle * 4 + 4, 1, 0, readonly_memory);
844
845 return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
846 }
847
848 /**
849 * Load from LDS.
850 *
851 * \param type output value type
852 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
853 * \param dw_addr address in dwords
854 */
855 static LLVMValueRef lds_load(struct lp_build_tgsi_context *bld_base,
856 enum tgsi_opcode_type type, unsigned swizzle,
857 LLVMValueRef dw_addr)
858 {
859 struct si_shader_context *ctx = si_shader_context(bld_base);
860 struct gallivm_state *gallivm = &ctx->gallivm;
861 LLVMValueRef value;
862
863 if (swizzle == ~0) {
864 LLVMValueRef values[TGSI_NUM_CHANNELS];
865
866 for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; chan++)
867 values[chan] = lds_load(bld_base, type, chan, dw_addr);
868
869 return lp_build_gather_values(gallivm, values,
870 TGSI_NUM_CHANNELS);
871 }
872
873 dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
874 LLVMConstInt(ctx->i32, swizzle, 0));
875
876 value = ac_build_indexed_load(&ctx->ac, ctx->lds, dw_addr, false);
877 if (tgsi_type_is_64bit(type)) {
878 LLVMValueRef value2;
879 dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
880 ctx->i32_1);
881 value2 = ac_build_indexed_load(&ctx->ac, ctx->lds, dw_addr, false);
882 return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
883 }
884
885 return LLVMBuildBitCast(gallivm->builder, value,
886 tgsi2llvmtype(bld_base, type), "");
887 }
888
889 /**
890 * Store to LDS.
891 *
892 * \param swizzle offset (typically 0..3)
893 * \param dw_addr address in dwords
894 * \param value value to store
895 */
896 static void lds_store(struct lp_build_tgsi_context *bld_base,
897 unsigned swizzle, LLVMValueRef dw_addr,
898 LLVMValueRef value)
899 {
900 struct si_shader_context *ctx = si_shader_context(bld_base);
901 struct gallivm_state *gallivm = &ctx->gallivm;
902
903 dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
904 LLVMConstInt(ctx->i32, swizzle, 0));
905
906 value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
907 ac_build_indexed_store(&ctx->ac, ctx->lds,
908 dw_addr, value);
909 }
910
911 static LLVMValueRef fetch_input_tcs(
912 struct lp_build_tgsi_context *bld_base,
913 const struct tgsi_full_src_register *reg,
914 enum tgsi_opcode_type type, unsigned swizzle)
915 {
916 struct si_shader_context *ctx = si_shader_context(bld_base);
917 LLVMValueRef dw_addr, stride;
918
919 stride = unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);
920 dw_addr = get_tcs_in_current_patch_offset(ctx);
921 dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
922
923 return lds_load(bld_base, type, swizzle, dw_addr);
924 }
925
926 static LLVMValueRef fetch_output_tcs(
927 struct lp_build_tgsi_context *bld_base,
928 const struct tgsi_full_src_register *reg,
929 enum tgsi_opcode_type type, unsigned swizzle)
930 {
931 struct si_shader_context *ctx = si_shader_context(bld_base);
932 LLVMValueRef dw_addr, stride;
933
934 if (reg->Register.Dimension) {
935 stride = unpack_param(ctx, ctx->param_tcs_out_lds_layout, 13, 8);
936 dw_addr = get_tcs_out_current_patch_offset(ctx);
937 dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
938 } else {
939 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
940 dw_addr = get_dw_address(ctx, NULL, reg, NULL, dw_addr);
941 }
942
943 return lds_load(bld_base, type, swizzle, dw_addr);
944 }
945
946 static LLVMValueRef fetch_input_tes(
947 struct lp_build_tgsi_context *bld_base,
948 const struct tgsi_full_src_register *reg,
949 enum tgsi_opcode_type type, unsigned swizzle)
950 {
951 struct si_shader_context *ctx = si_shader_context(bld_base);
952 LLVMValueRef rw_buffers, buffer, base, addr;
953
954 rw_buffers = LLVMGetParam(ctx->main_fn,
955 ctx->param_rw_buffers);
956 buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
957 LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
958
959 base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
960 addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg);
961
962 return buffer_load(bld_base, type, swizzle, buffer, base, addr, true);
963 }
964
965 static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
966 const struct tgsi_full_instruction *inst,
967 const struct tgsi_opcode_info *info,
968 LLVMValueRef dst[4])
969 {
970 struct si_shader_context *ctx = si_shader_context(bld_base);
971 struct gallivm_state *gallivm = &ctx->gallivm;
972 const struct tgsi_full_dst_register *reg = &inst->Dst[0];
973 const struct tgsi_shader_info *sh_info = &ctx->shader->selector->info;
974 unsigned chan_index;
975 LLVMValueRef dw_addr, stride;
976 LLVMValueRef rw_buffers, buffer, base, buf_addr;
977 LLVMValueRef values[4];
978 bool skip_lds_store;
979 bool is_tess_factor = false;
980
981 /* Only handle per-patch and per-vertex outputs here.
982 * Vectors will be lowered to scalars and this function will be called again.
983 */
984 if (reg->Register.File != TGSI_FILE_OUTPUT ||
985 (dst[0] && LLVMGetTypeKind(LLVMTypeOf(dst[0])) == LLVMVectorTypeKind)) {
986 si_llvm_emit_store(bld_base, inst, info, dst);
987 return;
988 }
989
990 if (reg->Register.Dimension) {
991 stride = unpack_param(ctx, ctx->param_tcs_out_lds_layout, 13, 8);
992 dw_addr = get_tcs_out_current_patch_offset(ctx);
993 dw_addr = get_dw_address(ctx, reg, NULL, stride, dw_addr);
994 skip_lds_store = !sh_info->reads_pervertex_outputs;
995 } else {
996 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
997 dw_addr = get_dw_address(ctx, reg, NULL, NULL, dw_addr);
998 skip_lds_store = !sh_info->reads_perpatch_outputs;
999
1000 if (!reg->Register.Indirect) {
1001 int name = sh_info->output_semantic_name[reg->Register.Index];
1002
1003 /* Always write tess factors into LDS for the TCS epilog. */
1004 if (name == TGSI_SEMANTIC_TESSINNER ||
1005 name == TGSI_SEMANTIC_TESSOUTER) {
1006 skip_lds_store = false;
1007 is_tess_factor = true;
1008 }
1009 }
1010 }
1011
1012 rw_buffers = LLVMGetParam(ctx->main_fn,
1013 ctx->param_rw_buffers);
1014 buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
1015 LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
1016
1017 base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
1018 buf_addr = get_tcs_tes_buffer_address_from_reg(ctx, reg, NULL);
1019
1020
1021 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan_index) {
1022 LLVMValueRef value = dst[chan_index];
1023
1024 if (inst->Instruction.Saturate)
1025 value = ac_build_clamp(&ctx->ac, value);
1026
1027 /* Skip LDS stores if there is no LDS read of this output. */
1028 if (!skip_lds_store)
1029 lds_store(bld_base, chan_index, dw_addr, value);
1030
1031 value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
1032 values[chan_index] = value;
1033
1034 if (inst->Dst[0].Register.WriteMask != 0xF && !is_tess_factor) {
1035 ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
1036 buf_addr, base,
1037 4 * chan_index, 1, 0, true, false);
1038 }
1039 }
1040
1041 if (inst->Dst[0].Register.WriteMask == 0xF && !is_tess_factor) {
1042 LLVMValueRef value = lp_build_gather_values(gallivm,
1043 values, 4);
1044 ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buf_addr,
1045 base, 0, 1, 0, true, false);
1046 }
1047 }
1048
1049 static LLVMValueRef fetch_input_gs(
1050 struct lp_build_tgsi_context *bld_base,
1051 const struct tgsi_full_src_register *reg,
1052 enum tgsi_opcode_type type,
1053 unsigned swizzle)
1054 {
1055 struct si_shader_context *ctx = si_shader_context(bld_base);
1056 struct si_shader *shader = ctx->shader;
1057 struct lp_build_context *uint = &ctx->bld_base.uint_bld;
1058 struct gallivm_state *gallivm = &ctx->gallivm;
1059 LLVMValueRef vtx_offset, soffset;
1060 unsigned vtx_offset_param;
1061 struct tgsi_shader_info *info = &shader->selector->info;
1062 unsigned semantic_name = info->input_semantic_name[reg->Register.Index];
1063 unsigned semantic_index = info->input_semantic_index[reg->Register.Index];
1064 unsigned param;
1065 LLVMValueRef value;
1066
1067 if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID)
1068 return get_primitive_id(bld_base, swizzle);
1069
1070 if (!reg->Register.Dimension)
1071 return NULL;
1072
1073 if (swizzle == ~0) {
1074 LLVMValueRef values[TGSI_NUM_CHANNELS];
1075 unsigned chan;
1076 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
1077 values[chan] = fetch_input_gs(bld_base, reg, type, chan);
1078 }
1079 return lp_build_gather_values(gallivm, values,
1080 TGSI_NUM_CHANNELS);
1081 }
1082
1083 /* Get the vertex offset parameter */
1084 vtx_offset_param = reg->Dimension.Index;
1085 if (vtx_offset_param < 2) {
1086 vtx_offset_param += ctx->param_gs_vtx0_offset;
1087 } else {
1088 assert(vtx_offset_param < 6);
1089 vtx_offset_param += ctx->param_gs_vtx2_offset - 2;
1090 }
1091 vtx_offset = lp_build_mul_imm(uint,
1092 LLVMGetParam(ctx->main_fn,
1093 vtx_offset_param),
1094 4);
1095
1096 param = si_shader_io_get_unique_index(semantic_name, semantic_index);
1097 soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0);
1098
1099 value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0,
1100 vtx_offset, soffset, 0, 1, 0, true);
1101 if (tgsi_type_is_64bit(type)) {
1102 LLVMValueRef value2;
1103 soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0);
1104
1105 value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1,
1106 ctx->i32_0, vtx_offset, soffset,
1107 0, 1, 0, true);
1108 return si_llvm_emit_fetch_64bit(bld_base, type,
1109 value, value2);
1110 }
1111 return LLVMBuildBitCast(gallivm->builder,
1112 value,
1113 tgsi2llvmtype(bld_base, type), "");
1114 }
1115
1116 static int lookup_interp_param_index(unsigned interpolate, unsigned location)
1117 {
1118 switch (interpolate) {
1119 case TGSI_INTERPOLATE_CONSTANT:
1120 return 0;
1121
1122 case TGSI_INTERPOLATE_LINEAR:
1123 if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
1124 return SI_PARAM_LINEAR_SAMPLE;
1125 else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
1126 return SI_PARAM_LINEAR_CENTROID;
1127 else
1128 return SI_PARAM_LINEAR_CENTER;
1129 break;
1130 case TGSI_INTERPOLATE_COLOR:
1131 case TGSI_INTERPOLATE_PERSPECTIVE:
1132 if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
1133 return SI_PARAM_PERSP_SAMPLE;
1134 else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
1135 return SI_PARAM_PERSP_CENTROID;
1136 else
1137 return SI_PARAM_PERSP_CENTER;
1138 break;
1139 default:
1140 fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
1141 return -1;
1142 }
1143 }
1144
1145 /**
1146 * Interpolate a fragment shader input.
1147 *
1148 * @param ctx context
1149 * @param input_index index of the input in hardware
1150 * @param semantic_name TGSI_SEMANTIC_*
1151 * @param semantic_index semantic index
1152 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1153 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1154 * @param interp_param interpolation weights (i,j)
1155 * @param prim_mask SI_PARAM_PRIM_MASK
1156 * @param face SI_PARAM_FRONT_FACE
1157 * @param result the return value (4 components)
1158 */
1159 static void interp_fs_input(struct si_shader_context *ctx,
1160 unsigned input_index,
1161 unsigned semantic_name,
1162 unsigned semantic_index,
1163 unsigned num_interp_inputs,
1164 unsigned colors_read_mask,
1165 LLVMValueRef interp_param,
1166 LLVMValueRef prim_mask,
1167 LLVMValueRef face,
1168 LLVMValueRef result[4])
1169 {
1170 struct gallivm_state *gallivm = &ctx->gallivm;
1171 LLVMValueRef attr_number;
1172 LLVMValueRef i, j;
1173
1174 unsigned chan;
1175
1176 /* fs.constant returns the param from the middle vertex, so it's not
1177 * really useful for flat shading. It's meant to be used for custom
1178 * interpolation (but the intrinsic can't fetch from the other two
1179 * vertices).
1180 *
1181 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1182 * to do the right thing. The only reason we use fs.constant is that
1183 * fs.interp cannot be used on integers, because they can be equal
1184 * to NaN.
1185 *
1186 * When interp is false we will use fs.constant or for newer llvm,
1187 * amdgcn.interp.mov.
1188 */
1189 bool interp = interp_param != NULL;
1190
1191 attr_number = LLVMConstInt(ctx->i32, input_index, 0);
1192
1193 if (interp) {
1194 interp_param = LLVMBuildBitCast(gallivm->builder, interp_param,
1195 LLVMVectorType(ctx->f32, 2), "");
1196
1197 i = LLVMBuildExtractElement(gallivm->builder, interp_param,
1198 ctx->i32_0, "");
1199 j = LLVMBuildExtractElement(gallivm->builder, interp_param,
1200 ctx->i32_1, "");
1201 }
1202
1203 if (semantic_name == TGSI_SEMANTIC_COLOR &&
1204 ctx->shader->key.part.ps.prolog.color_two_side) {
1205 LLVMValueRef is_face_positive;
1206 LLVMValueRef back_attr_number;
1207
1208 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1209 * otherwise it's at offset "num_inputs".
1210 */
1211 unsigned back_attr_offset = num_interp_inputs;
1212 if (semantic_index == 1 && colors_read_mask & 0xf)
1213 back_attr_offset += 1;
1214
1215 back_attr_number = LLVMConstInt(ctx->i32, back_attr_offset, 0);
1216
1217 is_face_positive = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
1218 face, ctx->i32_0, "");
1219
1220 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
1221 LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, 0);
1222 LLVMValueRef front, back;
1223
1224 if (interp) {
1225 front = ac_build_fs_interp(&ctx->ac, llvm_chan,
1226 attr_number, prim_mask,
1227 i, j);
1228 back = ac_build_fs_interp(&ctx->ac, llvm_chan,
1229 back_attr_number, prim_mask,
1230 i, j);
1231 } else {
1232 front = ac_build_fs_interp_mov(&ctx->ac,
1233 LLVMConstInt(ctx->i32, 2, 0), /* P0 */
1234 llvm_chan, attr_number, prim_mask);
1235 back = ac_build_fs_interp_mov(&ctx->ac,
1236 LLVMConstInt(ctx->i32, 2, 0), /* P0 */
1237 llvm_chan, back_attr_number, prim_mask);
1238 }
1239
1240 result[chan] = LLVMBuildSelect(gallivm->builder,
1241 is_face_positive,
1242 front,
1243 back,
1244 "");
1245 }
1246 } else if (semantic_name == TGSI_SEMANTIC_FOG) {
1247 if (interp) {
1248 result[0] = ac_build_fs_interp(&ctx->ac, ctx->i32_0,
1249 attr_number, prim_mask, i, j);
1250 } else {
1251 result[0] = ac_build_fs_interp_mov(&ctx->ac, ctx->i32_0,
1252 LLVMConstInt(ctx->i32, 2, 0), /* P0 */
1253 attr_number, prim_mask);
1254 }
1255 result[1] =
1256 result[2] = LLVMConstReal(ctx->f32, 0.0f);
1257 result[3] = LLVMConstReal(ctx->f32, 1.0f);
1258 } else {
1259 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
1260 LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, 0);
1261
1262 if (interp) {
1263 result[chan] = ac_build_fs_interp(&ctx->ac,
1264 llvm_chan, attr_number, prim_mask, i, j);
1265 } else {
1266 result[chan] = ac_build_fs_interp_mov(&ctx->ac,
1267 LLVMConstInt(ctx->i32, 2, 0), /* P0 */
1268 llvm_chan, attr_number, prim_mask);
1269 }
1270 }
1271 }
1272 }
1273
1274 static void declare_input_fs(
1275 struct si_shader_context *ctx,
1276 unsigned input_index,
1277 const struct tgsi_full_declaration *decl,
1278 LLVMValueRef out[4])
1279 {
1280 struct lp_build_context *base = &ctx->bld_base.base;
1281 struct si_shader *shader = ctx->shader;
1282 LLVMValueRef main_fn = ctx->main_fn;
1283 LLVMValueRef interp_param = NULL;
1284 int interp_param_idx;
1285
1286 /* Get colors from input VGPRs (set by the prolog). */
1287 if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR) {
1288 unsigned i = decl->Semantic.Index;
1289 unsigned colors_read = shader->selector->info.colors_read;
1290 unsigned mask = colors_read >> (i * 4);
1291 unsigned offset = SI_PARAM_POS_FIXED_PT + 1 +
1292 (i ? util_bitcount(colors_read & 0xf) : 0);
1293
1294 out[0] = mask & 0x1 ? LLVMGetParam(main_fn, offset++) : base->undef;
1295 out[1] = mask & 0x2 ? LLVMGetParam(main_fn, offset++) : base->undef;
1296 out[2] = mask & 0x4 ? LLVMGetParam(main_fn, offset++) : base->undef;
1297 out[3] = mask & 0x8 ? LLVMGetParam(main_fn, offset++) : base->undef;
1298 return;
1299 }
1300
1301 interp_param_idx = lookup_interp_param_index(decl->Interp.Interpolate,
1302 decl->Interp.Location);
1303 if (interp_param_idx == -1)
1304 return;
1305 else if (interp_param_idx) {
1306 interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
1307 }
1308
1309 if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
1310 decl->Interp.Interpolate == TGSI_INTERPOLATE_COLOR &&
1311 ctx->shader->key.part.ps.prolog.flatshade_colors)
1312 interp_param = NULL; /* load the constant color */
1313
1314 interp_fs_input(ctx, input_index, decl->Semantic.Name,
1315 decl->Semantic.Index, shader->selector->info.num_inputs,
1316 shader->selector->info.colors_read, interp_param,
1317 LLVMGetParam(main_fn, SI_PARAM_PRIM_MASK),
1318 LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE),
1319 &out[0]);
1320 }
1321
1322 static LLVMValueRef get_sample_id(struct si_shader_context *ctx)
1323 {
1324 return unpack_param(ctx, SI_PARAM_ANCILLARY, 8, 4);
1325 }
1326
1327
1328 /**
1329 * Load a dword from a constant buffer.
1330 */
1331 static LLVMValueRef buffer_load_const(struct si_shader_context *ctx,
1332 LLVMValueRef resource,
1333 LLVMValueRef offset)
1334 {
1335 LLVMBuilderRef builder = ctx->gallivm.builder;
1336 LLVMValueRef args[2] = {resource, offset};
1337
1338 return lp_build_intrinsic(builder, "llvm.SI.load.const", ctx->f32, args, 2,
1339 LP_FUNC_ATTR_READNONE |
1340 LP_FUNC_ATTR_LEGACY);
1341 }
1342
1343 static LLVMValueRef load_sample_position(struct si_shader_context *ctx, LLVMValueRef sample_id)
1344 {
1345 struct lp_build_context *uint_bld = &ctx->bld_base.uint_bld;
1346 struct gallivm_state *gallivm = &ctx->gallivm;
1347 LLVMBuilderRef builder = gallivm->builder;
1348 LLVMValueRef desc = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
1349 LLVMValueRef buf_index = LLVMConstInt(ctx->i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
1350 LLVMValueRef resource = ac_build_indexed_load_const(&ctx->ac, desc, buf_index);
1351
1352 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1353 LLVMValueRef offset0 = lp_build_mul_imm(uint_bld, sample_id, 8);
1354 LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, LLVMConstInt(ctx->i32, 4, 0), "");
1355
1356 LLVMValueRef pos[4] = {
1357 buffer_load_const(ctx, resource, offset0),
1358 buffer_load_const(ctx, resource, offset1),
1359 LLVMConstReal(ctx->f32, 0),
1360 LLVMConstReal(ctx->f32, 0)
1361 };
1362
1363 return lp_build_gather_values(gallivm, pos, 4);
1364 }
1365
1366 static void declare_system_value(struct si_shader_context *ctx,
1367 unsigned index,
1368 const struct tgsi_full_declaration *decl)
1369 {
1370 struct lp_build_context *bld = &ctx->bld_base.base;
1371 struct gallivm_state *gallivm = &ctx->gallivm;
1372 LLVMValueRef value = 0;
1373
1374 assert(index < RADEON_LLVM_MAX_SYSTEM_VALUES);
1375
1376 switch (decl->Semantic.Name) {
1377 case TGSI_SEMANTIC_INSTANCEID:
1378 value = LLVMGetParam(ctx->main_fn,
1379 ctx->param_instance_id);
1380 break;
1381
1382 case TGSI_SEMANTIC_VERTEXID:
1383 value = LLVMBuildAdd(gallivm->builder,
1384 LLVMGetParam(ctx->main_fn,
1385 ctx->param_vertex_id),
1386 LLVMGetParam(ctx->main_fn,
1387 ctx->param_base_vertex), "");
1388 break;
1389
1390 case TGSI_SEMANTIC_VERTEXID_NOBASE:
1391 /* Unused. Clarify the meaning in indexed vs. non-indexed
1392 * draws if this is ever used again. */
1393 assert(false);
1394 break;
1395
1396 case TGSI_SEMANTIC_BASEVERTEX:
1397 {
1398 /* For non-indexed draws, the base vertex set by the driver
1399 * (for direct draws) or the CP (for indirect draws) is the
1400 * first vertex ID, but GLSL expects 0 to be returned.
1401 */
1402 LLVMValueRef vs_state = LLVMGetParam(ctx->main_fn, ctx->param_vs_state_bits);
1403 LLVMValueRef indexed;
1404
1405 indexed = LLVMBuildLShr(gallivm->builder, vs_state, ctx->i32_1, "");
1406 indexed = LLVMBuildTrunc(gallivm->builder, indexed, ctx->i1, "");
1407
1408 value = LLVMBuildSelect(gallivm->builder, indexed,
1409 LLVMGetParam(ctx->main_fn, ctx->param_base_vertex),
1410 ctx->i32_0, "");
1411 break;
1412 }
1413
1414 case TGSI_SEMANTIC_BASEINSTANCE:
1415 value = LLVMGetParam(ctx->main_fn, ctx->param_start_instance);
1416 break;
1417
1418 case TGSI_SEMANTIC_DRAWID:
1419 value = LLVMGetParam(ctx->main_fn, ctx->param_draw_id);
1420 break;
1421
1422 case TGSI_SEMANTIC_INVOCATIONID:
1423 if (ctx->type == PIPE_SHADER_TESS_CTRL)
1424 value = unpack_param(ctx, ctx->param_tcs_rel_ids, 8, 5);
1425 else if (ctx->type == PIPE_SHADER_GEOMETRY)
1426 value = LLVMGetParam(ctx->main_fn,
1427 ctx->param_gs_instance_id);
1428 else
1429 assert(!"INVOCATIONID not implemented");
1430 break;
1431
1432 case TGSI_SEMANTIC_POSITION:
1433 {
1434 LLVMValueRef pos[4] = {
1435 LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
1436 LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
1437 LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Z_FLOAT),
1438 lp_build_emit_llvm_unary(&ctx->bld_base, TGSI_OPCODE_RCP,
1439 LLVMGetParam(ctx->main_fn,
1440 SI_PARAM_POS_W_FLOAT)),
1441 };
1442 value = lp_build_gather_values(gallivm, pos, 4);
1443 break;
1444 }
1445
1446 case TGSI_SEMANTIC_FACE:
1447 value = LLVMGetParam(ctx->main_fn, SI_PARAM_FRONT_FACE);
1448 break;
1449
1450 case TGSI_SEMANTIC_SAMPLEID:
1451 value = get_sample_id(ctx);
1452 break;
1453
1454 case TGSI_SEMANTIC_SAMPLEPOS: {
1455 LLVMValueRef pos[4] = {
1456 LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
1457 LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
1458 LLVMConstReal(ctx->f32, 0),
1459 LLVMConstReal(ctx->f32, 0)
1460 };
1461 pos[0] = lp_build_emit_llvm_unary(&ctx->bld_base,
1462 TGSI_OPCODE_FRC, pos[0]);
1463 pos[1] = lp_build_emit_llvm_unary(&ctx->bld_base,
1464 TGSI_OPCODE_FRC, pos[1]);
1465 value = lp_build_gather_values(gallivm, pos, 4);
1466 break;
1467 }
1468
1469 case TGSI_SEMANTIC_SAMPLEMASK:
1470 /* This can only occur with the OpenGL Core profile, which
1471 * doesn't support smoothing.
1472 */
1473 value = LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLE_COVERAGE);
1474 break;
1475
1476 case TGSI_SEMANTIC_TESSCOORD:
1477 {
1478 LLVMValueRef coord[4] = {
1479 LLVMGetParam(ctx->main_fn, ctx->param_tes_u),
1480 LLVMGetParam(ctx->main_fn, ctx->param_tes_v),
1481 bld->zero,
1482 bld->zero
1483 };
1484
1485 /* For triangles, the vector should be (u, v, 1-u-v). */
1486 if (ctx->shader->selector->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] ==
1487 PIPE_PRIM_TRIANGLES)
1488 coord[2] = lp_build_sub(bld, bld->one,
1489 lp_build_add(bld, coord[0], coord[1]));
1490
1491 value = lp_build_gather_values(gallivm, coord, 4);
1492 break;
1493 }
1494
1495 case TGSI_SEMANTIC_VERTICESIN:
1496 if (ctx->type == PIPE_SHADER_TESS_CTRL)
1497 value = unpack_param(ctx, ctx->param_tcs_out_lds_layout, 26, 6);
1498 else if (ctx->type == PIPE_SHADER_TESS_EVAL)
1499 value = unpack_param(ctx, ctx->param_tcs_offchip_layout, 9, 7);
1500 else
1501 assert(!"invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1502 break;
1503
1504 case TGSI_SEMANTIC_TESSINNER:
1505 case TGSI_SEMANTIC_TESSOUTER:
1506 {
1507 LLVMValueRef rw_buffers, buffer, base, addr;
1508 int param = si_shader_io_get_unique_index(decl->Semantic.Name, 0);
1509
1510 rw_buffers = LLVMGetParam(ctx->main_fn,
1511 ctx->param_rw_buffers);
1512 buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
1513 LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
1514
1515 base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
1516 addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL,
1517 LLVMConstInt(ctx->i32, param, 0));
1518
1519 value = buffer_load(&ctx->bld_base, TGSI_TYPE_FLOAT,
1520 ~0, buffer, base, addr, true);
1521
1522 break;
1523 }
1524
1525 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI:
1526 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI:
1527 {
1528 LLVMValueRef buf, slot, val[4];
1529 int i, offset;
1530
1531 slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
1532 buf = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
1533 buf = ac_build_indexed_load_const(&ctx->ac, buf, slot);
1534 offset = decl->Semantic.Name == TGSI_SEMANTIC_DEFAULT_TESSINNER_SI ? 4 : 0;
1535
1536 for (i = 0; i < 4; i++)
1537 val[i] = buffer_load_const(ctx, buf,
1538 LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
1539 value = lp_build_gather_values(gallivm, val, 4);
1540 break;
1541 }
1542
1543 case TGSI_SEMANTIC_PRIMID:
1544 value = get_primitive_id(&ctx->bld_base, 0);
1545 break;
1546
1547 case TGSI_SEMANTIC_GRID_SIZE:
1548 value = LLVMGetParam(ctx->main_fn, SI_PARAM_GRID_SIZE);
1549 break;
1550
1551 case TGSI_SEMANTIC_BLOCK_SIZE:
1552 {
1553 LLVMValueRef values[3];
1554 unsigned i;
1555 unsigned *properties = ctx->shader->selector->info.properties;
1556
1557 if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
1558 unsigned sizes[3] = {
1559 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
1560 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
1561 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
1562 };
1563
1564 for (i = 0; i < 3; ++i)
1565 values[i] = LLVMConstInt(ctx->i32, sizes[i], 0);
1566
1567 value = lp_build_gather_values(gallivm, values, 3);
1568 } else {
1569 value = LLVMGetParam(ctx->main_fn, SI_PARAM_BLOCK_SIZE);
1570 }
1571 break;
1572 }
1573
1574 case TGSI_SEMANTIC_BLOCK_ID:
1575 value = LLVMGetParam(ctx->main_fn, SI_PARAM_BLOCK_ID);
1576 break;
1577
1578 case TGSI_SEMANTIC_THREAD_ID:
1579 value = LLVMGetParam(ctx->main_fn, SI_PARAM_THREAD_ID);
1580 break;
1581
1582 case TGSI_SEMANTIC_HELPER_INVOCATION:
1583 if (HAVE_LLVM >= 0x0309) {
1584 value = lp_build_intrinsic(gallivm->builder,
1585 "llvm.amdgcn.ps.live",
1586 ctx->i1, NULL, 0,
1587 LP_FUNC_ATTR_READNONE);
1588 value = LLVMBuildNot(gallivm->builder, value, "");
1589 value = LLVMBuildSExt(gallivm->builder, value, ctx->i32, "");
1590 } else {
1591 assert(!"TGSI_SEMANTIC_HELPER_INVOCATION unsupported");
1592 return;
1593 }
1594 break;
1595
1596 case TGSI_SEMANTIC_SUBGROUP_SIZE:
1597 value = LLVMConstInt(ctx->i32, 64, 0);
1598 break;
1599
1600 case TGSI_SEMANTIC_SUBGROUP_INVOCATION:
1601 value = ac_get_thread_id(&ctx->ac);
1602 break;
1603
1604 case TGSI_SEMANTIC_SUBGROUP_EQ_MASK:
1605 {
1606 LLVMValueRef id = ac_get_thread_id(&ctx->ac);
1607 id = LLVMBuildZExt(gallivm->builder, id, ctx->i64, "");
1608 value = LLVMBuildShl(gallivm->builder, LLVMConstInt(ctx->i64, 1, 0), id, "");
1609 value = LLVMBuildBitCast(gallivm->builder, value, ctx->v2i32, "");
1610 break;
1611 }
1612
1613 case TGSI_SEMANTIC_SUBGROUP_GE_MASK:
1614 case TGSI_SEMANTIC_SUBGROUP_GT_MASK:
1615 case TGSI_SEMANTIC_SUBGROUP_LE_MASK:
1616 case TGSI_SEMANTIC_SUBGROUP_LT_MASK:
1617 {
1618 LLVMValueRef id = ac_get_thread_id(&ctx->ac);
1619 if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_GT_MASK ||
1620 decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK) {
1621 /* All bits set except LSB */
1622 value = LLVMConstInt(ctx->i64, -2, 0);
1623 } else {
1624 /* All bits set */
1625 value = LLVMConstInt(ctx->i64, -1, 0);
1626 }
1627 id = LLVMBuildZExt(gallivm->builder, id, ctx->i64, "");
1628 value = LLVMBuildShl(gallivm->builder, value, id, "");
1629 if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK ||
1630 decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LT_MASK)
1631 value = LLVMBuildNot(gallivm->builder, value, "");
1632 value = LLVMBuildBitCast(gallivm->builder, value, ctx->v2i32, "");
1633 break;
1634 }
1635
1636 default:
1637 assert(!"unknown system value");
1638 return;
1639 }
1640
1641 ctx->system_values[index] = value;
1642 }
1643
1644 static void declare_compute_memory(struct si_shader_context *ctx,
1645 const struct tgsi_full_declaration *decl)
1646 {
1647 struct si_shader_selector *sel = ctx->shader->selector;
1648 struct gallivm_state *gallivm = &ctx->gallivm;
1649
1650 LLVMTypeRef i8p = LLVMPointerType(ctx->i8, LOCAL_ADDR_SPACE);
1651 LLVMValueRef var;
1652
1653 assert(decl->Declaration.MemType == TGSI_MEMORY_TYPE_SHARED);
1654 assert(decl->Range.First == decl->Range.Last);
1655 assert(!ctx->shared_memory);
1656
1657 var = LLVMAddGlobalInAddressSpace(gallivm->module,
1658 LLVMArrayType(ctx->i8, sel->local_size),
1659 "compute_lds",
1660 LOCAL_ADDR_SPACE);
1661 LLVMSetAlignment(var, 4);
1662
1663 ctx->shared_memory = LLVMBuildBitCast(gallivm->builder, var, i8p, "");
1664 }
1665
1666 static LLVMValueRef load_const_buffer_desc(struct si_shader_context *ctx, int i)
1667 {
1668 LLVMValueRef list_ptr = LLVMGetParam(ctx->main_fn,
1669 ctx->param_const_buffers);
1670
1671 return ac_build_indexed_load_const(&ctx->ac, list_ptr,
1672 LLVMConstInt(ctx->i32, i, 0));
1673 }
1674
1675 static LLVMValueRef fetch_constant(
1676 struct lp_build_tgsi_context *bld_base,
1677 const struct tgsi_full_src_register *reg,
1678 enum tgsi_opcode_type type,
1679 unsigned swizzle)
1680 {
1681 struct si_shader_context *ctx = si_shader_context(bld_base);
1682 struct lp_build_context *base = &bld_base->base;
1683 const struct tgsi_ind_register *ireg = &reg->Indirect;
1684 unsigned buf, idx;
1685
1686 LLVMValueRef addr, bufp;
1687 LLVMValueRef result;
1688
1689 if (swizzle == LP_CHAN_ALL) {
1690 unsigned chan;
1691 LLVMValueRef values[4];
1692 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
1693 values[chan] = fetch_constant(bld_base, reg, type, chan);
1694
1695 return lp_build_gather_values(&ctx->gallivm, values, 4);
1696 }
1697
1698 buf = reg->Register.Dimension ? reg->Dimension.Index : 0;
1699 idx = reg->Register.Index * 4 + swizzle;
1700
1701 if (reg->Register.Dimension && reg->Dimension.Indirect) {
1702 LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_const_buffers);
1703 LLVMValueRef index;
1704 index = get_bounded_indirect_index(ctx, &reg->DimIndirect,
1705 reg->Dimension.Index,
1706 SI_NUM_CONST_BUFFERS);
1707 bufp = ac_build_indexed_load_const(&ctx->ac, ptr, index);
1708 } else
1709 bufp = load_const_buffer_desc(ctx, buf);
1710
1711 if (reg->Register.Indirect) {
1712 addr = ctx->addrs[ireg->Index][ireg->Swizzle];
1713 addr = LLVMBuildLoad(base->gallivm->builder, addr, "load addr reg");
1714 addr = lp_build_mul_imm(&bld_base->uint_bld, addr, 16);
1715 addr = lp_build_add(&bld_base->uint_bld, addr,
1716 LLVMConstInt(ctx->i32, idx * 4, 0));
1717 } else {
1718 addr = LLVMConstInt(ctx->i32, idx * 4, 0);
1719 }
1720
1721 result = buffer_load_const(ctx, bufp, addr);
1722
1723 if (!tgsi_type_is_64bit(type))
1724 result = bitcast(bld_base, type, result);
1725 else {
1726 LLVMValueRef addr2, result2;
1727
1728 addr2 = lp_build_add(&bld_base->uint_bld, addr,
1729 LLVMConstInt(ctx->i32, 4, 0));
1730 result2 = buffer_load_const(ctx, bufp, addr2);
1731
1732 result = si_llvm_emit_fetch_64bit(bld_base, type,
1733 result, result2);
1734 }
1735 return result;
1736 }
1737
1738 /* Upper 16 bits must be zero. */
1739 static LLVMValueRef si_llvm_pack_two_int16(struct si_shader_context *ctx,
1740 LLVMValueRef val[2])
1741 {
1742 return LLVMBuildOr(ctx->gallivm.builder, val[0],
1743 LLVMBuildShl(ctx->gallivm.builder, val[1],
1744 LLVMConstInt(ctx->i32, 16, 0),
1745 ""), "");
1746 }
1747
1748 /* Upper 16 bits are ignored and will be dropped. */
1749 static LLVMValueRef si_llvm_pack_two_int32_as_int16(struct si_shader_context *ctx,
1750 LLVMValueRef val[2])
1751 {
1752 LLVMValueRef v[2] = {
1753 LLVMBuildAnd(ctx->gallivm.builder, val[0],
1754 LLVMConstInt(ctx->i32, 0xffff, 0), ""),
1755 val[1],
1756 };
1757 return si_llvm_pack_two_int16(ctx, v);
1758 }
1759
1760 /* Initialize arguments for the shader export intrinsic */
1761 static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
1762 LLVMValueRef *values,
1763 unsigned target,
1764 struct ac_export_args *args)
1765 {
1766 struct si_shader_context *ctx = si_shader_context(bld_base);
1767 struct lp_build_context *base = &bld_base->base;
1768 LLVMBuilderRef builder = ctx->gallivm.builder;
1769 LLVMValueRef val[4];
1770 unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
1771 unsigned chan;
1772 bool is_int8, is_int10;
1773
1774 /* Default is 0xf. Adjusted below depending on the format. */
1775 args->enabled_channels = 0xf; /* writemask */
1776
1777 /* Specify whether the EXEC mask represents the valid mask */
1778 args->valid_mask = 0;
1779
1780 /* Specify whether this is the last export */
1781 args->done = 0;
1782
1783 /* Specify the target we are exporting */
1784 args->target = target;
1785
1786 if (ctx->type == PIPE_SHADER_FRAGMENT) {
1787 const struct si_shader_key *key = &ctx->shader->key;
1788 unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
1789 int cbuf = target - V_008DFC_SQ_EXP_MRT;
1790
1791 assert(cbuf >= 0 && cbuf < 8);
1792 spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
1793 is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
1794 is_int10 = (key->part.ps.epilog.color_is_int10 >> cbuf) & 0x1;
1795 }
1796
1797 args->compr = false;
1798 args->out[0] = base->undef;
1799 args->out[1] = base->undef;
1800 args->out[2] = base->undef;
1801 args->out[3] = base->undef;
1802
1803 switch (spi_shader_col_format) {
1804 case V_028714_SPI_SHADER_ZERO:
1805 args->enabled_channels = 0; /* writemask */
1806 args->target = V_008DFC_SQ_EXP_NULL;
1807 break;
1808
1809 case V_028714_SPI_SHADER_32_R:
1810 args->enabled_channels = 1; /* writemask */
1811 args->out[0] = values[0];
1812 break;
1813
1814 case V_028714_SPI_SHADER_32_GR:
1815 args->enabled_channels = 0x3; /* writemask */
1816 args->out[0] = values[0];
1817 args->out[1] = values[1];
1818 break;
1819
1820 case V_028714_SPI_SHADER_32_AR:
1821 args->enabled_channels = 0x9; /* writemask */
1822 args->out[0] = values[0];
1823 args->out[3] = values[3];
1824 break;
1825
1826 case V_028714_SPI_SHADER_FP16_ABGR:
1827 args->compr = 1; /* COMPR flag */
1828
1829 for (chan = 0; chan < 2; chan++) {
1830 LLVMValueRef pack_args[2] = {
1831 values[2 * chan],
1832 values[2 * chan + 1]
1833 };
1834 LLVMValueRef packed;
1835
1836 packed = ac_build_cvt_pkrtz_f16(&ctx->ac, pack_args);
1837 args->out[chan] =
1838 LLVMBuildBitCast(ctx->gallivm.builder,
1839 packed, ctx->f32, "");
1840 }
1841 break;
1842
1843 case V_028714_SPI_SHADER_UNORM16_ABGR:
1844 for (chan = 0; chan < 4; chan++) {
1845 val[chan] = ac_build_clamp(&ctx->ac, values[chan]);
1846 val[chan] = LLVMBuildFMul(builder, val[chan],
1847 LLVMConstReal(ctx->f32, 65535), "");
1848 val[chan] = LLVMBuildFAdd(builder, val[chan],
1849 LLVMConstReal(ctx->f32, 0.5), "");
1850 val[chan] = LLVMBuildFPToUI(builder, val[chan],
1851 ctx->i32, "");
1852 }
1853
1854 args->compr = 1; /* COMPR flag */
1855 args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1856 si_llvm_pack_two_int16(ctx, val));
1857 args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1858 si_llvm_pack_two_int16(ctx, val+2));
1859 break;
1860
1861 case V_028714_SPI_SHADER_SNORM16_ABGR:
1862 for (chan = 0; chan < 4; chan++) {
1863 /* Clamp between [-1, 1]. */
1864 val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MIN,
1865 values[chan],
1866 LLVMConstReal(ctx->f32, 1));
1867 val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MAX,
1868 val[chan],
1869 LLVMConstReal(ctx->f32, -1));
1870 /* Convert to a signed integer in [-32767, 32767]. */
1871 val[chan] = LLVMBuildFMul(builder, val[chan],
1872 LLVMConstReal(ctx->f32, 32767), "");
1873 /* If positive, add 0.5, else add -0.5. */
1874 val[chan] = LLVMBuildFAdd(builder, val[chan],
1875 LLVMBuildSelect(builder,
1876 LLVMBuildFCmp(builder, LLVMRealOGE,
1877 val[chan], base->zero, ""),
1878 LLVMConstReal(ctx->f32, 0.5),
1879 LLVMConstReal(ctx->f32, -0.5), ""), "");
1880 val[chan] = LLVMBuildFPToSI(builder, val[chan], ctx->i32, "");
1881 }
1882
1883 args->compr = 1; /* COMPR flag */
1884 args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1885 si_llvm_pack_two_int32_as_int16(ctx, val));
1886 args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1887 si_llvm_pack_two_int32_as_int16(ctx, val+2));
1888 break;
1889
1890 case V_028714_SPI_SHADER_UINT16_ABGR: {
1891 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1892 is_int8 ? 255 : is_int10 ? 1023 : 65535, 0);
1893 LLVMValueRef max_alpha =
1894 !is_int10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
1895
1896 /* Clamp. */
1897 for (chan = 0; chan < 4; chan++) {
1898 val[chan] = bitcast(bld_base, TGSI_TYPE_UNSIGNED, values[chan]);
1899 val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_UMIN,
1900 val[chan],
1901 chan == 3 ? max_alpha : max_rgb);
1902 }
1903
1904 args->compr = 1; /* COMPR flag */
1905 args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1906 si_llvm_pack_two_int16(ctx, val));
1907 args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1908 si_llvm_pack_two_int16(ctx, val+2));
1909 break;
1910 }
1911
1912 case V_028714_SPI_SHADER_SINT16_ABGR: {
1913 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1914 is_int8 ? 127 : is_int10 ? 511 : 32767, 0);
1915 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
1916 is_int8 ? -128 : is_int10 ? -512 : -32768, 0);
1917 LLVMValueRef max_alpha =
1918 !is_int10 ? max_rgb : ctx->i32_1;
1919 LLVMValueRef min_alpha =
1920 !is_int10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
1921
1922 /* Clamp. */
1923 for (chan = 0; chan < 4; chan++) {
1924 val[chan] = bitcast(bld_base, TGSI_TYPE_UNSIGNED, values[chan]);
1925 val[chan] = lp_build_emit_llvm_binary(bld_base,
1926 TGSI_OPCODE_IMIN,
1927 val[chan], chan == 3 ? max_alpha : max_rgb);
1928 val[chan] = lp_build_emit_llvm_binary(bld_base,
1929 TGSI_OPCODE_IMAX,
1930 val[chan], chan == 3 ? min_alpha : min_rgb);
1931 }
1932
1933 args->compr = 1; /* COMPR flag */
1934 args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1935 si_llvm_pack_two_int32_as_int16(ctx, val));
1936 args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
1937 si_llvm_pack_two_int32_as_int16(ctx, val+2));
1938 break;
1939 }
1940
1941 case V_028714_SPI_SHADER_32_ABGR:
1942 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
1943 break;
1944 }
1945 }
1946
1947 static void si_alpha_test(struct lp_build_tgsi_context *bld_base,
1948 LLVMValueRef alpha)
1949 {
1950 struct si_shader_context *ctx = si_shader_context(bld_base);
1951
1952 if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
1953 LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
1954 SI_PARAM_ALPHA_REF);
1955
1956 LLVMValueRef alpha_pass =
1957 lp_build_cmp(&bld_base->base,
1958 ctx->shader->key.part.ps.epilog.alpha_func,
1959 alpha, alpha_ref);
1960 LLVMValueRef arg =
1961 lp_build_select(&bld_base->base,
1962 alpha_pass,
1963 LLVMConstReal(ctx->f32, 1.0f),
1964 LLVMConstReal(ctx->f32, -1.0f));
1965
1966 ac_build_kill(&ctx->ac, arg);
1967 } else {
1968 ac_build_kill(&ctx->ac, NULL);
1969 }
1970 }
1971
1972 static LLVMValueRef si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context *bld_base,
1973 LLVMValueRef alpha,
1974 unsigned samplemask_param)
1975 {
1976 struct si_shader_context *ctx = si_shader_context(bld_base);
1977 struct gallivm_state *gallivm = &ctx->gallivm;
1978 LLVMValueRef coverage;
1979
1980 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
1981 coverage = LLVMGetParam(ctx->main_fn,
1982 samplemask_param);
1983 coverage = bitcast(bld_base, TGSI_TYPE_SIGNED, coverage);
1984
1985 coverage = lp_build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
1986 ctx->i32,
1987 &coverage, 1, LP_FUNC_ATTR_READNONE);
1988
1989 coverage = LLVMBuildUIToFP(gallivm->builder, coverage,
1990 ctx->f32, "");
1991
1992 coverage = LLVMBuildFMul(gallivm->builder, coverage,
1993 LLVMConstReal(ctx->f32,
1994 1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");
1995
1996 return LLVMBuildFMul(gallivm->builder, alpha, coverage, "");
1997 }
1998
1999 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context *bld_base,
2000 struct ac_export_args *pos, LLVMValueRef *out_elts)
2001 {
2002 struct si_shader_context *ctx = si_shader_context(bld_base);
2003 struct lp_build_context *base = &bld_base->base;
2004 unsigned reg_index;
2005 unsigned chan;
2006 unsigned const_chan;
2007 LLVMValueRef base_elt;
2008 LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
2009 LLVMValueRef constbuf_index = LLVMConstInt(ctx->i32,
2010 SI_VS_CONST_CLIP_PLANES, 0);
2011 LLVMValueRef const_resource = ac_build_indexed_load_const(&ctx->ac, ptr, constbuf_index);
2012
2013 for (reg_index = 0; reg_index < 2; reg_index ++) {
2014 struct ac_export_args *args = &pos[2 + reg_index];
2015
2016 args->out[0] =
2017 args->out[1] =
2018 args->out[2] =
2019 args->out[3] = LLVMConstReal(ctx->f32, 0.0f);
2020
2021 /* Compute dot products of position and user clip plane vectors */
2022 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
2023 for (const_chan = 0; const_chan < TGSI_NUM_CHANNELS; const_chan++) {
2024 LLVMValueRef addr =
2025 LLVMConstInt(ctx->i32, ((reg_index * 4 + chan) * 4 +
2026 const_chan) * 4, 0);
2027 base_elt = buffer_load_const(ctx, const_resource,
2028 addr);
2029 args->out[chan] =
2030 lp_build_add(base, args->out[chan],
2031 lp_build_mul(base, base_elt,
2032 out_elts[const_chan]));
2033 }
2034 }
2035
2036 args->enabled_channels = 0xf;
2037 args->valid_mask = 0;
2038 args->done = 0;
2039 args->target = V_008DFC_SQ_EXP_POS + 2 + reg_index;
2040 args->compr = 0;
2041 }
2042 }
2043
2044 static void si_dump_streamout(struct pipe_stream_output_info *so)
2045 {
2046 unsigned i;
2047
2048 if (so->num_outputs)
2049 fprintf(stderr, "STREAMOUT\n");
2050
2051 for (i = 0; i < so->num_outputs; i++) {
2052 unsigned mask = ((1 << so->output[i].num_components) - 1) <<
2053 so->output[i].start_component;
2054 fprintf(stderr, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2055 i, so->output[i].output_buffer,
2056 so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1,
2057 so->output[i].register_index,
2058 mask & 1 ? "x" : "",
2059 mask & 2 ? "y" : "",
2060 mask & 4 ? "z" : "",
2061 mask & 8 ? "w" : "");
2062 }
2063 }
2064
2065 static void emit_streamout_output(struct si_shader_context *ctx,
2066 LLVMValueRef const *so_buffers,
2067 LLVMValueRef const *so_write_offsets,
2068 struct pipe_stream_output *stream_out,
2069 struct si_shader_output_values *shader_out)
2070 {
2071 struct gallivm_state *gallivm = &ctx->gallivm;
2072 LLVMBuilderRef builder = gallivm->builder;
2073 unsigned buf_idx = stream_out->output_buffer;
2074 unsigned start = stream_out->start_component;
2075 unsigned num_comps = stream_out->num_components;
2076 LLVMValueRef out[4];
2077
2078 assert(num_comps && num_comps <= 4);
2079 if (!num_comps || num_comps > 4)
2080 return;
2081
2082 /* Load the output as int. */
2083 for (int j = 0; j < num_comps; j++) {
2084 assert(stream_out->stream == shader_out->vertex_stream[start + j]);
2085
2086 out[j] = LLVMBuildBitCast(builder,
2087 shader_out->values[start + j],
2088 ctx->i32, "");
2089 }
2090
2091 /* Pack the output. */
2092 LLVMValueRef vdata = NULL;
2093
2094 switch (num_comps) {
2095 case 1: /* as i32 */
2096 vdata = out[0];
2097 break;
2098 case 2: /* as v2i32 */
2099 case 3: /* as v4i32 (aligned to 4) */
2100 case 4: /* as v4i32 */
2101 vdata = LLVMGetUndef(LLVMVectorType(ctx->i32, util_next_power_of_two(num_comps)));
2102 for (int j = 0; j < num_comps; j++) {
2103 vdata = LLVMBuildInsertElement(builder, vdata, out[j],
2104 LLVMConstInt(ctx->i32, j, 0), "");
2105 }
2106 break;
2107 }
2108
2109 ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf_idx],
2110 vdata, num_comps,
2111 so_write_offsets[buf_idx],
2112 ctx->i32_0,
2113 stream_out->dst_offset * 4, 1, 1, true, false);
2114 }
2115
2116 /**
2117 * Write streamout data to buffers for vertex stream @p stream (different
2118 * vertex streams can occur for GS copy shaders).
2119 */
2120 static void si_llvm_emit_streamout(struct si_shader_context *ctx,
2121 struct si_shader_output_values *outputs,
2122 unsigned noutput, unsigned stream)
2123 {
2124 struct si_shader_selector *sel = ctx->shader->selector;
2125 struct pipe_stream_output_info *so = &sel->so;
2126 struct gallivm_state *gallivm = &ctx->gallivm;
2127 LLVMBuilderRef builder = gallivm->builder;
2128 int i;
2129 struct lp_build_if_state if_ctx;
2130
2131 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2132 LLVMValueRef so_vtx_count =
2133 unpack_param(ctx, ctx->param_streamout_config, 16, 7);
2134
2135 LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
2136
2137 /* can_emit = tid < so_vtx_count; */
2138 LLVMValueRef can_emit =
2139 LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
2140
2141 /* Emit the streamout code conditionally. This actually avoids
2142 * out-of-bounds buffer access. The hw tells us via the SGPR
2143 * (so_vtx_count) which threads are allowed to emit streamout data. */
2144 lp_build_if(&if_ctx, gallivm, can_emit);
2145 {
2146 /* The buffer offset is computed as follows:
2147 * ByteOffset = streamout_offset[buffer_id]*4 +
2148 * (streamout_write_index + thread_id)*stride[buffer_id] +
2149 * attrib_offset
2150 */
2151
2152 LLVMValueRef so_write_index =
2153 LLVMGetParam(ctx->main_fn,
2154 ctx->param_streamout_write_index);
2155
2156 /* Compute (streamout_write_index + thread_id). */
2157 so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
2158
2159 /* Load the descriptor and compute the write offset for each
2160 * enabled buffer. */
2161 LLVMValueRef so_write_offset[4] = {};
2162 LLVMValueRef so_buffers[4];
2163 LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
2164 ctx->param_rw_buffers);
2165
2166 for (i = 0; i < 4; i++) {
2167 if (!so->stride[i])
2168 continue;
2169
2170 LLVMValueRef offset = LLVMConstInt(ctx->i32,
2171 SI_VS_STREAMOUT_BUF0 + i, 0);
2172
2173 so_buffers[i] = ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
2174
2175 LLVMValueRef so_offset = LLVMGetParam(ctx->main_fn,
2176 ctx->param_streamout_offset[i]);
2177 so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
2178
2179 so_write_offset[i] = LLVMBuildMul(builder, so_write_index,
2180 LLVMConstInt(ctx->i32, so->stride[i]*4, 0), "");
2181 so_write_offset[i] = LLVMBuildAdd(builder, so_write_offset[i], so_offset, "");
2182 }
2183
2184 /* Write streamout data. */
2185 for (i = 0; i < so->num_outputs; i++) {
2186 unsigned reg = so->output[i].register_index;
2187
2188 if (reg >= noutput)
2189 continue;
2190
2191 if (stream != so->output[i].stream)
2192 continue;
2193
2194 emit_streamout_output(ctx, so_buffers, so_write_offset,
2195 &so->output[i], &outputs[reg]);
2196 }
2197 }
2198 lp_build_endif(&if_ctx);
2199 }
2200
2201
2202 /* Generate export instructions for hardware VS shader stage */
2203 static void si_llvm_export_vs(struct lp_build_tgsi_context *bld_base,
2204 struct si_shader_output_values *outputs,
2205 unsigned noutput)
2206 {
2207 struct si_shader_context *ctx = si_shader_context(bld_base);
2208 struct si_shader *shader = ctx->shader;
2209 struct lp_build_context *base = &bld_base->base;
2210 struct ac_export_args args, pos_args[4] = {};
2211 LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL, viewport_index_value = NULL;
2212 unsigned semantic_name, semantic_index;
2213 unsigned target;
2214 unsigned param_count = 0;
2215 unsigned pos_idx;
2216 int i;
2217
2218 for (i = 0; i < noutput; i++) {
2219 semantic_name = outputs[i].semantic_name;
2220 semantic_index = outputs[i].semantic_index;
2221 bool export_param = true;
2222
2223 switch (semantic_name) {
2224 case TGSI_SEMANTIC_POSITION: /* ignore these */
2225 case TGSI_SEMANTIC_PSIZE:
2226 case TGSI_SEMANTIC_CLIPVERTEX:
2227 case TGSI_SEMANTIC_EDGEFLAG:
2228 break;
2229 case TGSI_SEMANTIC_GENERIC:
2230 case TGSI_SEMANTIC_CLIPDIST:
2231 if (shader->key.opt.hw_vs.kill_outputs &
2232 (1ull << si_shader_io_get_unique_index(semantic_name, semantic_index)))
2233 export_param = false;
2234 break;
2235 default:
2236 if (shader->key.opt.hw_vs.kill_outputs2 &
2237 (1u << si_shader_io_get_unique_index2(semantic_name, semantic_index)))
2238 export_param = false;
2239 break;
2240 }
2241
2242 if (outputs[i].vertex_stream[0] != 0 &&
2243 outputs[i].vertex_stream[1] != 0 &&
2244 outputs[i].vertex_stream[2] != 0 &&
2245 outputs[i].vertex_stream[3] != 0)
2246 export_param = false;
2247
2248 handle_semantic:
2249 /* Select the correct target */
2250 switch(semantic_name) {
2251 case TGSI_SEMANTIC_PSIZE:
2252 psize_value = outputs[i].values[0];
2253 continue;
2254 case TGSI_SEMANTIC_EDGEFLAG:
2255 edgeflag_value = outputs[i].values[0];
2256 continue;
2257 case TGSI_SEMANTIC_LAYER:
2258 layer_value = outputs[i].values[0];
2259 semantic_name = TGSI_SEMANTIC_GENERIC;
2260 goto handle_semantic;
2261 case TGSI_SEMANTIC_VIEWPORT_INDEX:
2262 viewport_index_value = outputs[i].values[0];
2263 semantic_name = TGSI_SEMANTIC_GENERIC;
2264 goto handle_semantic;
2265 case TGSI_SEMANTIC_POSITION:
2266 target = V_008DFC_SQ_EXP_POS;
2267 break;
2268 case TGSI_SEMANTIC_CLIPDIST:
2269 if (shader->key.opt.hw_vs.clip_disable) {
2270 semantic_name = TGSI_SEMANTIC_GENERIC;
2271 goto handle_semantic;
2272 }
2273 target = V_008DFC_SQ_EXP_POS + 2 + semantic_index;
2274 break;
2275 case TGSI_SEMANTIC_CLIPVERTEX:
2276 if (shader->key.opt.hw_vs.clip_disable)
2277 continue;
2278 si_llvm_emit_clipvertex(bld_base, pos_args, outputs[i].values);
2279 continue;
2280 case TGSI_SEMANTIC_COLOR:
2281 case TGSI_SEMANTIC_BCOLOR:
2282 case TGSI_SEMANTIC_PRIMID:
2283 case TGSI_SEMANTIC_FOG:
2284 case TGSI_SEMANTIC_TEXCOORD:
2285 case TGSI_SEMANTIC_GENERIC:
2286 if (!export_param)
2287 continue;
2288 target = V_008DFC_SQ_EXP_PARAM + param_count;
2289 assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
2290 shader->info.vs_output_param_offset[i] = param_count;
2291 param_count++;
2292 break;
2293 default:
2294 target = 0;
2295 fprintf(stderr,
2296 "Warning: SI unhandled vs output type:%d\n",
2297 semantic_name);
2298 }
2299
2300 si_llvm_init_export_args(bld_base, outputs[i].values, target, &args);
2301
2302 if (target >= V_008DFC_SQ_EXP_POS &&
2303 target <= (V_008DFC_SQ_EXP_POS + 3)) {
2304 memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
2305 &args, sizeof(args));
2306 } else {
2307 ac_build_export(&ctx->ac, &args);
2308 }
2309
2310 if (semantic_name == TGSI_SEMANTIC_CLIPDIST) {
2311 semantic_name = TGSI_SEMANTIC_GENERIC;
2312 goto handle_semantic;
2313 }
2314 }
2315
2316 shader->info.nr_param_exports = param_count;
2317
2318 /* We need to add the position output manually if it's missing. */
2319 if (!pos_args[0].out[0]) {
2320 pos_args[0].enabled_channels = 0xf; /* writemask */
2321 pos_args[0].valid_mask = 0; /* EXEC mask */
2322 pos_args[0].done = 0; /* last export? */
2323 pos_args[0].target = V_008DFC_SQ_EXP_POS;
2324 pos_args[0].compr = 0; /* COMPR flag */
2325 pos_args[0].out[0] = base->zero; /* X */
2326 pos_args[0].out[1] = base->zero; /* Y */
2327 pos_args[0].out[2] = base->zero; /* Z */
2328 pos_args[0].out[3] = base->one; /* W */
2329 }
2330
2331 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2332 if (shader->selector->info.writes_psize ||
2333 shader->selector->info.writes_edgeflag ||
2334 shader->selector->info.writes_viewport_index ||
2335 shader->selector->info.writes_layer) {
2336 pos_args[1].enabled_channels = shader->selector->info.writes_psize |
2337 (shader->selector->info.writes_edgeflag << 1) |
2338 (shader->selector->info.writes_layer << 2) |
2339 (shader->selector->info.writes_viewport_index << 3);
2340 pos_args[1].valid_mask = 0; /* EXEC mask */
2341 pos_args[1].done = 0; /* last export? */
2342 pos_args[1].target = V_008DFC_SQ_EXP_POS + 1;
2343 pos_args[1].compr = 0; /* COMPR flag */
2344 pos_args[1].out[0] = base->zero; /* X */
2345 pos_args[1].out[1] = base->zero; /* Y */
2346 pos_args[1].out[2] = base->zero; /* Z */
2347 pos_args[1].out[3] = base->zero; /* W */
2348
2349 if (shader->selector->info.writes_psize)
2350 pos_args[1].out[0] = psize_value;
2351
2352 if (shader->selector->info.writes_edgeflag) {
2353 /* The output is a float, but the hw expects an integer
2354 * with the first bit containing the edge flag. */
2355 edgeflag_value = LLVMBuildFPToUI(ctx->gallivm.builder,
2356 edgeflag_value,
2357 ctx->i32, "");
2358 edgeflag_value = lp_build_min(&bld_base->int_bld,
2359 edgeflag_value,
2360 ctx->i32_1);
2361
2362 /* The LLVM intrinsic expects a float. */
2363 pos_args[1].out[1] = LLVMBuildBitCast(ctx->gallivm.builder,
2364 edgeflag_value,
2365 ctx->f32, "");
2366 }
2367
2368 if (shader->selector->info.writes_layer)
2369 pos_args[1].out[2] = layer_value;
2370
2371 if (shader->selector->info.writes_viewport_index)
2372 pos_args[1].out[3] = viewport_index_value;
2373 }
2374
2375 for (i = 0; i < 4; i++)
2376 if (pos_args[i].out[0])
2377 shader->info.nr_pos_exports++;
2378
2379 pos_idx = 0;
2380 for (i = 0; i < 4; i++) {
2381 if (!pos_args[i].out[0])
2382 continue;
2383
2384 /* Specify the target we are exporting */
2385 pos_args[i].target = V_008DFC_SQ_EXP_POS + pos_idx++;
2386
2387 if (pos_idx == shader->info.nr_pos_exports)
2388 /* Specify that this is the last export */
2389 pos_args[i].done = 1;
2390
2391 ac_build_export(&ctx->ac, &pos_args[i]);
2392 }
2393 }
2394
2395 /**
2396 * Forward all outputs from the vertex shader to the TES. This is only used
2397 * for the fixed function TCS.
2398 */
2399 static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
2400 {
2401 struct si_shader_context *ctx = si_shader_context(bld_base);
2402 struct gallivm_state *gallivm = &ctx->gallivm;
2403 LLVMValueRef invocation_id, rw_buffers, buffer, buffer_offset;
2404 LLVMValueRef lds_vertex_stride, lds_vertex_offset, lds_base;
2405 uint64_t inputs;
2406
2407 invocation_id = unpack_param(ctx, ctx->param_tcs_rel_ids, 8, 5);
2408
2409 rw_buffers = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
2410 buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
2411 LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
2412
2413 buffer_offset = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
2414
2415 lds_vertex_stride = unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);
2416 lds_vertex_offset = LLVMBuildMul(gallivm->builder, invocation_id,
2417 lds_vertex_stride, "");
2418 lds_base = get_tcs_in_current_patch_offset(ctx);
2419 lds_base = LLVMBuildAdd(gallivm->builder, lds_base, lds_vertex_offset, "");
2420
2421 inputs = ctx->shader->key.mono.ff_tcs_inputs_to_copy;
2422 while (inputs) {
2423 unsigned i = u_bit_scan64(&inputs);
2424
2425 LLVMValueRef lds_ptr = LLVMBuildAdd(gallivm->builder, lds_base,
2426 LLVMConstInt(ctx->i32, 4 * i, 0),
2427 "");
2428
2429 LLVMValueRef buffer_addr = get_tcs_tes_buffer_address(ctx,
2430 get_rel_patch_id(ctx),
2431 invocation_id,
2432 LLVMConstInt(ctx->i32, i, 0));
2433
2434 LLVMValueRef value = lds_load(bld_base, TGSI_TYPE_SIGNED, ~0,
2435 lds_ptr);
2436
2437 ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buffer_addr,
2438 buffer_offset, 0, 1, 0, true, false);
2439 }
2440 }
2441
2442 static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
2443 LLVMValueRef rel_patch_id,
2444 LLVMValueRef invocation_id,
2445 LLVMValueRef tcs_out_current_patch_data_offset)
2446 {
2447 struct si_shader_context *ctx = si_shader_context(bld_base);
2448 struct gallivm_state *gallivm = &ctx->gallivm;
2449 struct si_shader *shader = ctx->shader;
2450 unsigned tess_inner_index, tess_outer_index;
2451 LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
2452 LLVMValueRef out[6], vec0, vec1, rw_buffers, tf_base, inner[4], outer[4];
2453 unsigned stride, outer_comps, inner_comps, i;
2454 struct lp_build_if_state if_ctx, inner_if_ctx;
2455
2456 si_llvm_emit_barrier(NULL, bld_base, NULL);
2457
2458 /* Do this only for invocation 0, because the tess levels are per-patch,
2459 * not per-vertex.
2460 *
2461 * This can't jump, because invocation 0 executes this. It should
2462 * at least mask out the loads and stores for other invocations.
2463 */
2464 lp_build_if(&if_ctx, gallivm,
2465 LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
2466 invocation_id, ctx->i32_0, ""));
2467
2468 /* Determine the layout of one tess factor element in the buffer. */
2469 switch (shader->key.part.tcs.epilog.prim_mode) {
2470 case PIPE_PRIM_LINES:
2471 stride = 2; /* 2 dwords, 1 vec2 store */
2472 outer_comps = 2;
2473 inner_comps = 0;
2474 break;
2475 case PIPE_PRIM_TRIANGLES:
2476 stride = 4; /* 4 dwords, 1 vec4 store */
2477 outer_comps = 3;
2478 inner_comps = 1;
2479 break;
2480 case PIPE_PRIM_QUADS:
2481 stride = 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2482 outer_comps = 4;
2483 inner_comps = 2;
2484 break;
2485 default:
2486 assert(0);
2487 return;
2488 }
2489
2490 /* Load tess_inner and tess_outer from LDS.
2491 * Any invocation can write them, so we can't get them from a temporary.
2492 */
2493 tess_inner_index = si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER, 0);
2494 tess_outer_index = si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER, 0);
2495
2496 lds_base = tcs_out_current_patch_data_offset;
2497 lds_inner = LLVMBuildAdd(gallivm->builder, lds_base,
2498 LLVMConstInt(ctx->i32,
2499 tess_inner_index * 4, 0), "");
2500 lds_outer = LLVMBuildAdd(gallivm->builder, lds_base,
2501 LLVMConstInt(ctx->i32,
2502 tess_outer_index * 4, 0), "");
2503
2504 for (i = 0; i < 4; i++) {
2505 inner[i] = LLVMGetUndef(ctx->i32);
2506 outer[i] = LLVMGetUndef(ctx->i32);
2507 }
2508
2509 if (shader->key.part.tcs.epilog.prim_mode == PIPE_PRIM_LINES) {
2510 /* For isolines, the hardware expects tess factors in the
2511 * reverse order from what GLSL / TGSI specify.
2512 */
2513 outer[0] = out[1] = lds_load(bld_base, TGSI_TYPE_SIGNED, 0, lds_outer);
2514 outer[1] = out[0] = lds_load(bld_base, TGSI_TYPE_SIGNED, 1, lds_outer);
2515 } else {
2516 for (i = 0; i < outer_comps; i++) {
2517 outer[i] = out[i] =
2518 lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_outer);
2519 }
2520 for (i = 0; i < inner_comps; i++) {
2521 inner[i] = out[outer_comps+i] =
2522 lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_inner);
2523 }
2524 }
2525
2526 /* Convert the outputs to vectors for stores. */
2527 vec0 = lp_build_gather_values(gallivm, out, MIN2(stride, 4));
2528 vec1 = NULL;
2529
2530 if (stride > 4)
2531 vec1 = lp_build_gather_values(gallivm, out+4, stride - 4);
2532
2533 /* Get the buffer. */
2534 rw_buffers = LLVMGetParam(ctx->main_fn,
2535 ctx->param_rw_buffers);
2536 buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
2537 LLVMConstInt(ctx->i32, SI_HS_RING_TESS_FACTOR, 0));
2538
2539 /* Get the offset. */
2540 tf_base = LLVMGetParam(ctx->main_fn,
2541 ctx->param_tcs_factor_offset);
2542 byteoffset = LLVMBuildMul(gallivm->builder, rel_patch_id,
2543 LLVMConstInt(ctx->i32, 4 * stride, 0), "");
2544
2545 lp_build_if(&inner_if_ctx, gallivm,
2546 LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
2547 rel_patch_id, ctx->i32_0, ""));
2548
2549 /* Store the dynamic HS control word. */
2550 ac_build_buffer_store_dword(&ctx->ac, buffer,
2551 LLVMConstInt(ctx->i32, 0x80000000, 0),
2552 1, ctx->i32_0, tf_base,
2553 0, 1, 0, true, false);
2554
2555 lp_build_endif(&inner_if_ctx);
2556
2557 /* Store the tessellation factors. */
2558 ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
2559 MIN2(stride, 4), byteoffset, tf_base,
2560 4, 1, 0, true, false);
2561 if (vec1)
2562 ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
2563 stride - 4, byteoffset, tf_base,
2564 20, 1, 0, true, false);
2565
2566 /* Store the tess factors into the offchip buffer if TES reads them. */
2567 if (shader->key.part.tcs.epilog.tes_reads_tess_factors) {
2568 LLVMValueRef buf, base, inner_vec, outer_vec, tf_outer_offset;
2569 LLVMValueRef tf_inner_offset;
2570 unsigned param_outer, param_inner;
2571
2572 buf = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
2573 LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
2574 base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
2575
2576 param_outer = si_shader_io_get_unique_index(
2577 TGSI_SEMANTIC_TESSOUTER, 0);
2578 tf_outer_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
2579 LLVMConstInt(ctx->i32, param_outer, 0));
2580
2581 outer_vec = lp_build_gather_values(gallivm, outer,
2582 util_next_power_of_two(outer_comps));
2583
2584 ac_build_buffer_store_dword(&ctx->ac, buf, outer_vec,
2585 outer_comps, tf_outer_offset,
2586 base, 0, 1, 0, true, false);
2587 if (inner_comps) {
2588 param_inner = si_shader_io_get_unique_index(
2589 TGSI_SEMANTIC_TESSINNER, 0);
2590 tf_inner_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
2591 LLVMConstInt(ctx->i32, param_inner, 0));
2592
2593 inner_vec = inner_comps == 1 ? inner[0] :
2594 lp_build_gather_values(gallivm, inner, inner_comps);
2595 ac_build_buffer_store_dword(&ctx->ac, buf, inner_vec,
2596 inner_comps, tf_inner_offset,
2597 base, 0, 1, 0, true, false);
2598 }
2599 }
2600
2601 lp_build_endif(&if_ctx);
2602 }
2603
2604 /* This only writes the tessellation factor levels. */
2605 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context *bld_base)
2606 {
2607 struct si_shader_context *ctx = si_shader_context(bld_base);
2608 LLVMValueRef rel_patch_id, invocation_id, tf_lds_offset;
2609 LLVMValueRef offchip_soffset, offchip_layout;
2610
2611 si_copy_tcs_inputs(bld_base);
2612
2613 rel_patch_id = get_rel_patch_id(ctx);
2614 invocation_id = unpack_param(ctx, ctx->param_tcs_rel_ids, 8, 5);
2615 tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
2616
2617 /* Return epilog parameters from this function. */
2618 LLVMBuilderRef builder = ctx->gallivm.builder;
2619 LLVMValueRef ret = ctx->return_value;
2620 LLVMValueRef rw_buffers, rw0, rw1, tf_soffset;
2621 unsigned vgpr;
2622
2623 /* RW_BUFFERS pointer */
2624 rw_buffers = LLVMGetParam(ctx->main_fn,
2625 ctx->param_rw_buffers);
2626 rw_buffers = LLVMBuildPtrToInt(builder, rw_buffers, ctx->i64, "");
2627 rw_buffers = LLVMBuildBitCast(builder, rw_buffers, ctx->v2i32, "");
2628 rw0 = LLVMBuildExtractElement(builder, rw_buffers,
2629 ctx->i32_0, "");
2630 rw1 = LLVMBuildExtractElement(builder, rw_buffers,
2631 ctx->i32_1, "");
2632 ret = LLVMBuildInsertValue(builder, ret, rw0, 0, "");
2633 ret = LLVMBuildInsertValue(builder, ret, rw1, 1, "");
2634
2635 /* Tess offchip and factor buffer soffset are after user SGPRs. */
2636 offchip_layout = LLVMGetParam(ctx->main_fn,
2637 ctx->param_tcs_offchip_layout);
2638 offchip_soffset = LLVMGetParam(ctx->main_fn,
2639 ctx->param_tcs_offchip_offset);
2640 tf_soffset = LLVMGetParam(ctx->main_fn,
2641 ctx->param_tcs_factor_offset);
2642 ret = LLVMBuildInsertValue(builder, ret, offchip_layout,
2643 GFX6_SGPR_TCS_OFFCHIP_LAYOUT, "");
2644 ret = LLVMBuildInsertValue(builder, ret, offchip_soffset,
2645 GFX6_TCS_NUM_USER_SGPR, "");
2646 ret = LLVMBuildInsertValue(builder, ret, tf_soffset,
2647 GFX6_TCS_NUM_USER_SGPR + 1, "");
2648
2649 /* VGPRs */
2650 rel_patch_id = bitcast(bld_base, TGSI_TYPE_FLOAT, rel_patch_id);
2651 invocation_id = bitcast(bld_base, TGSI_TYPE_FLOAT, invocation_id);
2652 tf_lds_offset = bitcast(bld_base, TGSI_TYPE_FLOAT, tf_lds_offset);
2653
2654 vgpr = GFX6_TCS_NUM_USER_SGPR + 2;
2655 ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
2656 ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");
2657 ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
2658 ctx->return_value = ret;
2659 }
2660
2661 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context *bld_base)
2662 {
2663 struct si_shader_context *ctx = si_shader_context(bld_base);
2664 struct si_shader *shader = ctx->shader;
2665 struct tgsi_shader_info *info = &shader->selector->info;
2666 struct gallivm_state *gallivm = &ctx->gallivm;
2667 unsigned i, chan;
2668 LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
2669 ctx->param_rel_auto_id);
2670 LLVMValueRef vertex_dw_stride =
2671 unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);
2672 LLVMValueRef base_dw_addr = LLVMBuildMul(gallivm->builder, vertex_id,
2673 vertex_dw_stride, "");
2674
2675 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2676 * its inputs from it. */
2677 for (i = 0; i < info->num_outputs; i++) {
2678 LLVMValueRef *out_ptr = ctx->outputs[i];
2679 unsigned name = info->output_semantic_name[i];
2680 unsigned index = info->output_semantic_index[i];
2681
2682 /* The ARB_shader_viewport_layer_array spec contains the
2683 * following issue:
2684 *
2685 * 2) What happens if gl_ViewportIndex or gl_Layer is
2686 * written in the vertex shader and a geometry shader is
2687 * present?
2688 *
2689 * RESOLVED: The value written by the last vertex processing
2690 * stage is used. If the last vertex processing stage
2691 * (vertex, tessellation evaluation or geometry) does not
2692 * statically assign to gl_ViewportIndex or gl_Layer, index
2693 * or layer zero is assumed.
2694 *
2695 * So writes to those outputs in VS-as-LS are simply ignored.
2696 */
2697 if (name == TGSI_SEMANTIC_LAYER ||
2698 name == TGSI_SEMANTIC_VIEWPORT_INDEX)
2699 continue;
2700
2701 int param = si_shader_io_get_unique_index(name, index);
2702 LLVMValueRef dw_addr = LLVMBuildAdd(gallivm->builder, base_dw_addr,
2703 LLVMConstInt(ctx->i32, param * 4, 0), "");
2704
2705 for (chan = 0; chan < 4; chan++) {
2706 lds_store(bld_base, chan, dw_addr,
2707 LLVMBuildLoad(gallivm->builder, out_ptr[chan], ""));
2708 }
2709 }
2710 }
2711
2712 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context *bld_base)
2713 {
2714 struct si_shader_context *ctx = si_shader_context(bld_base);
2715 struct gallivm_state *gallivm = &ctx->gallivm;
2716 struct si_shader *es = ctx->shader;
2717 struct tgsi_shader_info *info = &es->selector->info;
2718 LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
2719 ctx->param_es2gs_offset);
2720 unsigned chan;
2721 int i;
2722
2723 for (i = 0; i < info->num_outputs; i++) {
2724 LLVMValueRef *out_ptr = ctx->outputs[i];
2725 int param_index;
2726
2727 if (info->output_semantic_name[i] == TGSI_SEMANTIC_VIEWPORT_INDEX ||
2728 info->output_semantic_name[i] == TGSI_SEMANTIC_LAYER)
2729 continue;
2730
2731 param_index = si_shader_io_get_unique_index(info->output_semantic_name[i],
2732 info->output_semantic_index[i]);
2733
2734 for (chan = 0; chan < 4; chan++) {
2735 LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
2736 out_val = LLVMBuildBitCast(gallivm->builder, out_val, ctx->i32, "");
2737
2738 ac_build_buffer_store_dword(&ctx->ac,
2739 ctx->esgs_ring,
2740 out_val, 1, NULL, soffset,
2741 (4 * param_index + chan) * 4,
2742 1, 1, true, true);
2743 }
2744 }
2745 }
2746
2747 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context *bld_base)
2748 {
2749 struct si_shader_context *ctx = si_shader_context(bld_base);
2750
2751 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
2752 LLVMGetParam(ctx->main_fn, ctx->param_gs_wave_id));
2753 }
2754
2755 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context *bld_base)
2756 {
2757 struct si_shader_context *ctx = si_shader_context(bld_base);
2758 struct gallivm_state *gallivm = &ctx->gallivm;
2759 struct tgsi_shader_info *info = &ctx->shader->selector->info;
2760 struct si_shader_output_values *outputs = NULL;
2761 int i,j;
2762
2763 assert(!ctx->shader->is_gs_copy_shader);
2764
2765 outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
2766
2767 /* Vertex color clamping.
2768 *
2769 * This uses a state constant loaded in a user data SGPR and
2770 * an IF statement is added that clamps all colors if the constant
2771 * is true.
2772 */
2773 if (ctx->type == PIPE_SHADER_VERTEX) {
2774 struct lp_build_if_state if_ctx;
2775 LLVMValueRef cond = NULL;
2776 LLVMValueRef addr, val;
2777
2778 for (i = 0; i < info->num_outputs; i++) {
2779 if (info->output_semantic_name[i] != TGSI_SEMANTIC_COLOR &&
2780 info->output_semantic_name[i] != TGSI_SEMANTIC_BCOLOR)
2781 continue;
2782
2783 /* We've found a color. */
2784 if (!cond) {
2785 /* The state is in the first bit of the user SGPR. */
2786 cond = LLVMGetParam(ctx->main_fn,
2787 ctx->param_vs_state_bits);
2788 cond = LLVMBuildTrunc(gallivm->builder, cond,
2789 ctx->i1, "");
2790 lp_build_if(&if_ctx, gallivm, cond);
2791 }
2792
2793 for (j = 0; j < 4; j++) {
2794 addr = ctx->outputs[i][j];
2795 val = LLVMBuildLoad(gallivm->builder, addr, "");
2796 val = ac_build_clamp(&ctx->ac, val);
2797 LLVMBuildStore(gallivm->builder, val, addr);
2798 }
2799 }
2800
2801 if (cond)
2802 lp_build_endif(&if_ctx);
2803 }
2804
2805 for (i = 0; i < info->num_outputs; i++) {
2806 outputs[i].semantic_name = info->output_semantic_name[i];
2807 outputs[i].semantic_index = info->output_semantic_index[i];
2808
2809 for (j = 0; j < 4; j++) {
2810 outputs[i].values[j] =
2811 LLVMBuildLoad(gallivm->builder,
2812 ctx->outputs[i][j],
2813 "");
2814 outputs[i].vertex_stream[j] =
2815 (info->output_streams[i] >> (2 * j)) & 3;
2816 }
2817
2818 }
2819
2820 /* Return the primitive ID from the LLVM function. */
2821 ctx->return_value =
2822 LLVMBuildInsertValue(gallivm->builder,
2823 ctx->return_value,
2824 bitcast(bld_base, TGSI_TYPE_FLOAT,
2825 get_primitive_id(bld_base, 0)),
2826 VS_EPILOG_PRIMID_LOC, "");
2827
2828 if (ctx->shader->selector->so.num_outputs)
2829 si_llvm_emit_streamout(ctx, outputs, i, 0);
2830 si_llvm_export_vs(bld_base, outputs, i);
2831 FREE(outputs);
2832 }
2833
2834 struct si_ps_exports {
2835 unsigned num;
2836 struct ac_export_args args[10];
2837 };
2838
2839 unsigned si_get_spi_shader_z_format(bool writes_z, bool writes_stencil,
2840 bool writes_samplemask)
2841 {
2842 if (writes_z) {
2843 /* Z needs 32 bits. */
2844 if (writes_samplemask)
2845 return V_028710_SPI_SHADER_32_ABGR;
2846 else if (writes_stencil)
2847 return V_028710_SPI_SHADER_32_GR;
2848 else
2849 return V_028710_SPI_SHADER_32_R;
2850 } else if (writes_stencil || writes_samplemask) {
2851 /* Both stencil and sample mask need only 16 bits. */
2852 return V_028710_SPI_SHADER_UINT16_ABGR;
2853 } else {
2854 return V_028710_SPI_SHADER_ZERO;
2855 }
2856 }
2857
2858 static void si_export_mrt_z(struct lp_build_tgsi_context *bld_base,
2859 LLVMValueRef depth, LLVMValueRef stencil,
2860 LLVMValueRef samplemask, struct si_ps_exports *exp)
2861 {
2862 struct si_shader_context *ctx = si_shader_context(bld_base);
2863 struct lp_build_context *base = &bld_base->base;
2864 struct ac_export_args args;
2865 unsigned mask = 0;
2866 unsigned format = si_get_spi_shader_z_format(depth != NULL,
2867 stencil != NULL,
2868 samplemask != NULL);
2869
2870 assert(depth || stencil || samplemask);
2871
2872 args.valid_mask = 1; /* whether the EXEC mask is valid */
2873 args.done = 1; /* DONE bit */
2874
2875 /* Specify the target we are exporting */
2876 args.target = V_008DFC_SQ_EXP_MRTZ;
2877
2878 args.compr = 0; /* COMP flag */
2879 args.out[0] = base->undef; /* R, depth */
2880 args.out[1] = base->undef; /* G, stencil test value[0:7], stencil op value[8:15] */
2881 args.out[2] = base->undef; /* B, sample mask */
2882 args.out[3] = base->undef; /* A, alpha to mask */
2883
2884 if (format == V_028710_SPI_SHADER_UINT16_ABGR) {
2885 assert(!depth);
2886 args.compr = 1; /* COMPR flag */
2887
2888 if (stencil) {
2889 /* Stencil should be in X[23:16]. */
2890 stencil = bitcast(bld_base, TGSI_TYPE_UNSIGNED, stencil);
2891 stencil = LLVMBuildShl(ctx->gallivm.builder, stencil,
2892 LLVMConstInt(ctx->i32, 16, 0), "");
2893 args.out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT, stencil);
2894 mask |= 0x3;
2895 }
2896 if (samplemask) {
2897 /* SampleMask should be in Y[15:0]. */
2898 args.out[1] = samplemask;
2899 mask |= 0xc;
2900 }
2901 } else {
2902 if (depth) {
2903 args.out[0] = depth;
2904 mask |= 0x1;
2905 }
2906 if (stencil) {
2907 args.out[1] = stencil;
2908 mask |= 0x2;
2909 }
2910 if (samplemask) {
2911 args.out[2] = samplemask;
2912 mask |= 0x4;
2913 }
2914 }
2915
2916 /* SI (except OLAND and HAINAN) has a bug that it only looks
2917 * at the X writemask component. */
2918 if (ctx->screen->b.chip_class == SI &&
2919 ctx->screen->b.family != CHIP_OLAND &&
2920 ctx->screen->b.family != CHIP_HAINAN)
2921 mask |= 0x1;
2922
2923 /* Specify which components to enable */
2924 args.enabled_channels = mask;
2925
2926 memcpy(&exp->args[exp->num++], &args, sizeof(args));
2927 }
2928
2929 static void si_export_mrt_color(struct lp_build_tgsi_context *bld_base,
2930 LLVMValueRef *color, unsigned index,
2931 unsigned samplemask_param,
2932 bool is_last, struct si_ps_exports *exp)
2933 {
2934 struct si_shader_context *ctx = si_shader_context(bld_base);
2935 struct lp_build_context *base = &bld_base->base;
2936 int i;
2937
2938 /* Clamp color */
2939 if (ctx->shader->key.part.ps.epilog.clamp_color)
2940 for (i = 0; i < 4; i++)
2941 color[i] = ac_build_clamp(&ctx->ac, color[i]);
2942
2943 /* Alpha to one */
2944 if (ctx->shader->key.part.ps.epilog.alpha_to_one)
2945 color[3] = base->one;
2946
2947 /* Alpha test */
2948 if (index == 0 &&
2949 ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
2950 si_alpha_test(bld_base, color[3]);
2951
2952 /* Line & polygon smoothing */
2953 if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
2954 color[3] = si_scale_alpha_by_sample_mask(bld_base, color[3],
2955 samplemask_param);
2956
2957 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2958 if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
2959 struct ac_export_args args[8];
2960 int c, last = -1;
2961
2962 /* Get the export arguments, also find out what the last one is. */
2963 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
2964 si_llvm_init_export_args(bld_base, color,
2965 V_008DFC_SQ_EXP_MRT + c, &args[c]);
2966 if (args[c].enabled_channels)
2967 last = c;
2968 }
2969
2970 /* Emit all exports. */
2971 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
2972 if (is_last && last == c) {
2973 args[c].valid_mask = 1; /* whether the EXEC mask is valid */
2974 args[c].done = 1; /* DONE bit */
2975 } else if (!args[c].enabled_channels)
2976 continue; /* unnecessary NULL export */
2977
2978 memcpy(&exp->args[exp->num++], &args[c], sizeof(args[c]));
2979 }
2980 } else {
2981 struct ac_export_args args;
2982
2983 /* Export */
2984 si_llvm_init_export_args(bld_base, color, V_008DFC_SQ_EXP_MRT + index,
2985 &args);
2986 if (is_last) {
2987 args.valid_mask = 1; /* whether the EXEC mask is valid */
2988 args.done = 1; /* DONE bit */
2989 } else if (!args.enabled_channels)
2990 return; /* unnecessary NULL export */
2991
2992 memcpy(&exp->args[exp->num++], &args, sizeof(args));
2993 }
2994 }
2995
2996 static void si_emit_ps_exports(struct si_shader_context *ctx,
2997 struct si_ps_exports *exp)
2998 {
2999 for (unsigned i = 0; i < exp->num; i++)
3000 ac_build_export(&ctx->ac, &exp->args[i]);
3001 }
3002
3003 static void si_export_null(struct lp_build_tgsi_context *bld_base)
3004 {
3005 struct si_shader_context *ctx = si_shader_context(bld_base);
3006 struct lp_build_context *base = &bld_base->base;
3007 struct ac_export_args args;
3008
3009 args.enabled_channels = 0x0; /* enabled channels */
3010 args.valid_mask = 1; /* whether the EXEC mask is valid */
3011 args.done = 1; /* DONE bit */
3012 args.target = V_008DFC_SQ_EXP_NULL;
3013 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
3014 args.out[0] = base->undef; /* R */
3015 args.out[1] = base->undef; /* G */
3016 args.out[2] = base->undef; /* B */
3017 args.out[3] = base->undef; /* A */
3018
3019 ac_build_export(&ctx->ac, &args);
3020 }
3021
3022 /**
3023 * Return PS outputs in this order:
3024 *
3025 * v[0:3] = color0.xyzw
3026 * v[4:7] = color1.xyzw
3027 * ...
3028 * vN+0 = Depth
3029 * vN+1 = Stencil
3030 * vN+2 = SampleMask
3031 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3032 *
3033 * The alpha-ref SGPR is returned via its original location.
3034 */
3035 static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context *bld_base)
3036 {
3037 struct si_shader_context *ctx = si_shader_context(bld_base);
3038 struct si_shader *shader = ctx->shader;
3039 struct tgsi_shader_info *info = &shader->selector->info;
3040 LLVMBuilderRef builder = ctx->gallivm.builder;
3041 unsigned i, j, first_vgpr, vgpr;
3042
3043 LLVMValueRef color[8][4] = {};
3044 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
3045 LLVMValueRef ret;
3046
3047 /* Read the output values. */
3048 for (i = 0; i < info->num_outputs; i++) {
3049 unsigned semantic_name = info->output_semantic_name[i];
3050 unsigned semantic_index = info->output_semantic_index[i];
3051
3052 switch (semantic_name) {
3053 case TGSI_SEMANTIC_COLOR:
3054 assert(semantic_index < 8);
3055 for (j = 0; j < 4; j++) {
3056 LLVMValueRef ptr = ctx->outputs[i][j];
3057 LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
3058 color[semantic_index][j] = result;
3059 }
3060 break;
3061 case TGSI_SEMANTIC_POSITION:
3062 depth = LLVMBuildLoad(builder,
3063 ctx->outputs[i][2], "");
3064 break;
3065 case TGSI_SEMANTIC_STENCIL:
3066 stencil = LLVMBuildLoad(builder,
3067 ctx->outputs[i][1], "");
3068 break;
3069 case TGSI_SEMANTIC_SAMPLEMASK:
3070 samplemask = LLVMBuildLoad(builder,
3071 ctx->outputs[i][0], "");
3072 break;
3073 default:
3074 fprintf(stderr, "Warning: SI unhandled fs output type:%d\n",
3075 semantic_name);
3076 }
3077 }
3078
3079 /* Fill the return structure. */
3080 ret = ctx->return_value;
3081
3082 /* Set SGPRs. */
3083 ret = LLVMBuildInsertValue(builder, ret,
3084 bitcast(bld_base, TGSI_TYPE_SIGNED,
3085 LLVMGetParam(ctx->main_fn,
3086 SI_PARAM_ALPHA_REF)),
3087 SI_SGPR_ALPHA_REF, "");
3088
3089 /* Set VGPRs */
3090 first_vgpr = vgpr = SI_SGPR_ALPHA_REF + 1;
3091 for (i = 0; i < ARRAY_SIZE(color); i++) {
3092 if (!color[i][0])
3093 continue;
3094
3095 for (j = 0; j < 4; j++)
3096 ret = LLVMBuildInsertValue(builder, ret, color[i][j], vgpr++, "");
3097 }
3098 if (depth)
3099 ret = LLVMBuildInsertValue(builder, ret, depth, vgpr++, "");
3100 if (stencil)
3101 ret = LLVMBuildInsertValue(builder, ret, stencil, vgpr++, "");
3102 if (samplemask)
3103 ret = LLVMBuildInsertValue(builder, ret, samplemask, vgpr++, "");
3104
3105 /* Add the input sample mask for smoothing at the end. */
3106 if (vgpr < first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC)
3107 vgpr = first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC;
3108 ret = LLVMBuildInsertValue(builder, ret,
3109 LLVMGetParam(ctx->main_fn,
3110 SI_PARAM_SAMPLE_COVERAGE), vgpr++, "");
3111
3112 ctx->return_value = ret;
3113 }
3114
3115 /**
3116 * Given a v8i32 resource descriptor for a buffer, extract the size of the
3117 * buffer in number of elements and return it as an i32.
3118 */
3119 static LLVMValueRef get_buffer_size(
3120 struct lp_build_tgsi_context *bld_base,
3121 LLVMValueRef descriptor)
3122 {
3123 struct si_shader_context *ctx = si_shader_context(bld_base);
3124 struct gallivm_state *gallivm = &ctx->gallivm;
3125 LLVMBuilderRef builder = gallivm->builder;
3126 LLVMValueRef size =
3127 LLVMBuildExtractElement(builder, descriptor,
3128 LLVMConstInt(ctx->i32, 2, 0), "");
3129
3130 if (ctx->screen->b.chip_class == VI) {
3131 /* On VI, the descriptor contains the size in bytes,
3132 * but TXQ must return the size in elements.
3133 * The stride is always non-zero for resources using TXQ.
3134 */
3135 LLVMValueRef stride =
3136 LLVMBuildExtractElement(builder, descriptor,
3137 ctx->i32_1, "");
3138 stride = LLVMBuildLShr(builder, stride,
3139 LLVMConstInt(ctx->i32, 16, 0), "");
3140 stride = LLVMBuildAnd(builder, stride,
3141 LLVMConstInt(ctx->i32, 0x3FFF, 0), "");
3142
3143 size = LLVMBuildUDiv(builder, size, stride, "");
3144 }
3145
3146 return size;
3147 }
3148
3149 static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
3150 struct lp_build_tgsi_context *bld_base,
3151 struct lp_build_emit_data *emit_data);
3152
3153 /* Prevent optimizations (at least of memory accesses) across the current
3154 * point in the program by emitting empty inline assembly that is marked as
3155 * having side effects.
3156 *
3157 * Optionally, a value can be passed through the inline assembly to prevent
3158 * LLVM from hoisting calls to ReadNone functions.
3159 */
3160 static void emit_optimization_barrier(struct si_shader_context *ctx,
3161 LLVMValueRef *pvgpr)
3162 {
3163 static int counter = 0;
3164
3165 LLVMBuilderRef builder = ctx->gallivm.builder;
3166 char code[16];
3167
3168 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
3169
3170 if (!pvgpr) {
3171 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
3172 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
3173 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
3174 } else {
3175 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
3176 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
3177 LLVMValueRef vgpr = *pvgpr;
3178 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
3179 unsigned vgpr_size = llvm_get_type_size(vgpr_type);
3180 LLVMValueRef vgpr0;
3181
3182 assert(vgpr_size % 4 == 0);
3183
3184 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
3185 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
3186 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
3187 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
3188 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
3189
3190 *pvgpr = vgpr;
3191 }
3192 }
3193
3194 /* Combine these with & instead of |. */
3195 #define NOOP_WAITCNT 0xf7f
3196 #define LGKM_CNT 0x07f
3197 #define VM_CNT 0xf70
3198
3199 static void emit_waitcnt(struct si_shader_context *ctx, unsigned simm16)
3200 {
3201 struct gallivm_state *gallivm = &ctx->gallivm;
3202 LLVMBuilderRef builder = gallivm->builder;
3203 LLVMValueRef args[1] = {
3204 LLVMConstInt(ctx->i32, simm16, 0)
3205 };
3206 lp_build_intrinsic(builder, "llvm.amdgcn.s.waitcnt",
3207 ctx->voidt, args, 1, 0);
3208 }
3209
3210 static void membar_emit(
3211 const struct lp_build_tgsi_action *action,
3212 struct lp_build_tgsi_context *bld_base,
3213 struct lp_build_emit_data *emit_data)
3214 {
3215 struct si_shader_context *ctx = si_shader_context(bld_base);
3216 LLVMValueRef src0 = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0);
3217 unsigned flags = LLVMConstIntGetZExtValue(src0);
3218 unsigned waitcnt = NOOP_WAITCNT;
3219
3220 if (flags & TGSI_MEMBAR_THREAD_GROUP)
3221 waitcnt &= VM_CNT & LGKM_CNT;
3222
3223 if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER |
3224 TGSI_MEMBAR_SHADER_BUFFER |
3225 TGSI_MEMBAR_SHADER_IMAGE))
3226 waitcnt &= VM_CNT;
3227
3228 if (flags & TGSI_MEMBAR_SHARED)
3229 waitcnt &= LGKM_CNT;
3230
3231 if (waitcnt != NOOP_WAITCNT)
3232 emit_waitcnt(ctx, waitcnt);
3233 }
3234
3235 static void clock_emit(
3236 const struct lp_build_tgsi_action *action,
3237 struct lp_build_tgsi_context *bld_base,
3238 struct lp_build_emit_data *emit_data)
3239 {
3240 struct si_shader_context *ctx = si_shader_context(bld_base);
3241 struct gallivm_state *gallivm = &ctx->gallivm;
3242 LLVMValueRef tmp;
3243
3244 tmp = lp_build_intrinsic(gallivm->builder, "llvm.readcyclecounter",
3245 ctx->i64, NULL, 0, 0);
3246 tmp = LLVMBuildBitCast(gallivm->builder, tmp, ctx->v2i32, "");
3247
3248 emit_data->output[0] =
3249 LLVMBuildExtractElement(gallivm->builder, tmp, ctx->i32_0, "");
3250 emit_data->output[1] =
3251 LLVMBuildExtractElement(gallivm->builder, tmp, ctx->i32_1, "");
3252 }
3253
3254 static LLVMValueRef
3255 shader_buffer_fetch_rsrc(struct si_shader_context *ctx,
3256 const struct tgsi_full_src_register *reg)
3257 {
3258 LLVMValueRef index;
3259 LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
3260 ctx->param_shader_buffers);
3261
3262 if (!reg->Register.Indirect)
3263 index = LLVMConstInt(ctx->i32, reg->Register.Index, 0);
3264 else
3265 index = get_bounded_indirect_index(ctx, &reg->Indirect,
3266 reg->Register.Index,
3267 SI_NUM_SHADER_BUFFERS);
3268
3269 return ac_build_indexed_load_const(&ctx->ac, rsrc_ptr, index);
3270 }
3271
3272 static bool tgsi_is_array_sampler(unsigned target)
3273 {
3274 return target == TGSI_TEXTURE_1D_ARRAY ||
3275 target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
3276 target == TGSI_TEXTURE_2D_ARRAY ||
3277 target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
3278 target == TGSI_TEXTURE_CUBE_ARRAY ||
3279 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY ||
3280 target == TGSI_TEXTURE_2D_ARRAY_MSAA;
3281 }
3282
3283 static bool tgsi_is_array_image(unsigned target)
3284 {
3285 return target == TGSI_TEXTURE_3D ||
3286 target == TGSI_TEXTURE_CUBE ||
3287 target == TGSI_TEXTURE_1D_ARRAY ||
3288 target == TGSI_TEXTURE_2D_ARRAY ||
3289 target == TGSI_TEXTURE_CUBE_ARRAY ||
3290 target == TGSI_TEXTURE_2D_ARRAY_MSAA;
3291 }
3292
3293 /**
3294 * Given a 256-bit resource descriptor, force the DCC enable bit to off.
3295 *
3296 * At least on Tonga, executing image stores on images with DCC enabled and
3297 * non-trivial can eventually lead to lockups. This can occur when an
3298 * application binds an image as read-only but then uses a shader that writes
3299 * to it. The OpenGL spec allows almost arbitrarily bad behavior (including
3300 * program termination) in this case, but it doesn't cost much to be a bit
3301 * nicer: disabling DCC in the shader still leads to undefined results but
3302 * avoids the lockup.
3303 */
3304 static LLVMValueRef force_dcc_off(struct si_shader_context *ctx,
3305 LLVMValueRef rsrc)
3306 {
3307 if (ctx->screen->b.chip_class <= CIK) {
3308 return rsrc;
3309 } else {
3310 LLVMBuilderRef builder = ctx->gallivm.builder;
3311 LLVMValueRef i32_6 = LLVMConstInt(ctx->i32, 6, 0);
3312 LLVMValueRef i32_C = LLVMConstInt(ctx->i32, C_008F28_COMPRESSION_EN, 0);
3313 LLVMValueRef tmp;
3314
3315 tmp = LLVMBuildExtractElement(builder, rsrc, i32_6, "");
3316 tmp = LLVMBuildAnd(builder, tmp, i32_C, "");
3317 return LLVMBuildInsertElement(builder, rsrc, tmp, i32_6, "");
3318 }
3319 }
3320
3321 static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
3322 {
3323 return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
3324 CONST_ADDR_SPACE);
3325 }
3326
3327 static LLVMValueRef load_image_desc(struct si_shader_context *ctx,
3328 LLVMValueRef list, LLVMValueRef index,
3329 unsigned target)
3330 {
3331 LLVMBuilderRef builder = ctx->gallivm.builder;
3332
3333 if (target == TGSI_TEXTURE_BUFFER) {
3334 index = LLVMBuildMul(builder, index,
3335 LLVMConstInt(ctx->i32, 2, 0), "");
3336 index = LLVMBuildAdd(builder, index,
3337 ctx->i32_1, "");
3338 list = LLVMBuildPointerCast(builder, list,
3339 const_array(ctx->v4i32, 0), "");
3340 }
3341
3342 return ac_build_indexed_load_const(&ctx->ac, list, index);
3343 }
3344
3345 /**
3346 * Load the resource descriptor for \p image.
3347 */
3348 static void
3349 image_fetch_rsrc(
3350 struct lp_build_tgsi_context *bld_base,
3351 const struct tgsi_full_src_register *image,
3352 bool is_store, unsigned target,
3353 LLVMValueRef *rsrc)
3354 {
3355 struct si_shader_context *ctx = si_shader_context(bld_base);
3356 LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
3357 ctx->param_images);
3358 LLVMValueRef index;
3359 bool dcc_off = is_store;
3360
3361 assert(image->Register.File == TGSI_FILE_IMAGE);
3362
3363 if (!image->Register.Indirect) {
3364 const struct tgsi_shader_info *info = bld_base->info;
3365 unsigned images_writemask = info->images_store |
3366 info->images_atomic;
3367
3368 index = LLVMConstInt(ctx->i32, image->Register.Index, 0);
3369
3370 if (images_writemask & (1 << image->Register.Index))
3371 dcc_off = true;
3372 } else {
3373 /* From the GL_ARB_shader_image_load_store extension spec:
3374 *
3375 * If a shader performs an image load, store, or atomic
3376 * operation using an image variable declared as an array,
3377 * and if the index used to select an individual element is
3378 * negative or greater than or equal to the size of the
3379 * array, the results of the operation are undefined but may
3380 * not lead to termination.
3381 */
3382 index = get_bounded_indirect_index(ctx, &image->Indirect,
3383 image->Register.Index,
3384 SI_NUM_IMAGES);
3385 }
3386
3387 *rsrc = load_image_desc(ctx, rsrc_ptr, index, target);
3388 if (dcc_off && target != TGSI_TEXTURE_BUFFER)
3389 *rsrc = force_dcc_off(ctx, *rsrc);
3390 }
3391
3392 static LLVMValueRef image_fetch_coords(
3393 struct lp_build_tgsi_context *bld_base,
3394 const struct tgsi_full_instruction *inst,
3395 unsigned src, LLVMValueRef desc)
3396 {
3397 struct si_shader_context *ctx = si_shader_context(bld_base);
3398 struct gallivm_state *gallivm = &ctx->gallivm;
3399 LLVMBuilderRef builder = gallivm->builder;
3400 unsigned target = inst->Memory.Texture;
3401 unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
3402 LLVMValueRef coords[4];
3403 LLVMValueRef tmp;
3404 int chan;
3405
3406 for (chan = 0; chan < num_coords; ++chan) {
3407 tmp = lp_build_emit_fetch(bld_base, inst, src, chan);
3408 tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
3409 coords[chan] = tmp;
3410 }
3411
3412 if (ctx->screen->b.chip_class >= GFX9) {
3413 /* 1D textures are allocated and used as 2D on GFX9. */
3414 if (target == TGSI_TEXTURE_1D) {
3415 coords[1] = ctx->i32_0;
3416 num_coords++;
3417 } else if (target == TGSI_TEXTURE_1D_ARRAY) {
3418 coords[2] = coords[1];
3419 coords[1] = ctx->i32_0;
3420 num_coords++;
3421 } else if (target == TGSI_TEXTURE_2D) {
3422 /* The hw can't bind a slice of a 3D image as a 2D
3423 * image, because it ignores BASE_ARRAY if the target
3424 * is 3D. The workaround is to read BASE_ARRAY and set
3425 * it as the 3rd address operand for all 2D images.
3426 */
3427 LLVMValueRef first_layer, const5, mask;
3428
3429 const5 = LLVMConstInt(ctx->i32, 5, 0);
3430 mask = LLVMConstInt(ctx->i32, S_008F24_BASE_ARRAY(~0), 0);
3431 first_layer = LLVMBuildExtractElement(builder, desc, const5, "");
3432 first_layer = LLVMBuildAnd(builder, first_layer, mask, "");
3433
3434 coords[2] = first_layer;
3435 num_coords++;
3436 }
3437 }
3438
3439 if (num_coords == 1)
3440 return coords[0];
3441
3442 if (num_coords == 3) {
3443 /* LLVM has difficulties lowering 3-element vectors. */
3444 coords[3] = bld_base->uint_bld.undef;
3445 num_coords = 4;
3446 }
3447
3448 return lp_build_gather_values(gallivm, coords, num_coords);
3449 }
3450
3451 /**
3452 * Append the extra mode bits that are used by image load and store.
3453 */
3454 static void image_append_args(
3455 struct si_shader_context *ctx,
3456 struct lp_build_emit_data * emit_data,
3457 unsigned target,
3458 bool atomic,
3459 bool force_glc)
3460 {
3461 const struct tgsi_full_instruction *inst = emit_data->inst;
3462 LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
3463 LLVMValueRef i1true = LLVMConstInt(ctx->i1, 1, 0);
3464 LLVMValueRef r128 = i1false;
3465 LLVMValueRef da = tgsi_is_array_image(target) ? i1true : i1false;
3466 LLVMValueRef glc =
3467 force_glc ||
3468 inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE) ?
3469 i1true : i1false;
3470 LLVMValueRef slc = i1false;
3471 LLVMValueRef lwe = i1false;
3472
3473 if (atomic || (HAVE_LLVM <= 0x0309)) {
3474 emit_data->args[emit_data->arg_count++] = r128;
3475 emit_data->args[emit_data->arg_count++] = da;
3476 if (!atomic) {
3477 emit_data->args[emit_data->arg_count++] = glc;
3478 }
3479 emit_data->args[emit_data->arg_count++] = slc;
3480 return;
3481 }
3482
3483 /* HAVE_LLVM >= 0x0400 */
3484 emit_data->args[emit_data->arg_count++] = glc;
3485 emit_data->args[emit_data->arg_count++] = slc;
3486 emit_data->args[emit_data->arg_count++] = lwe;
3487 emit_data->args[emit_data->arg_count++] = da;
3488 }
3489
3490 /**
3491 * Append the resource and indexing arguments for buffer intrinsics.
3492 *
3493 * \param rsrc the v4i32 buffer resource
3494 * \param index index into the buffer (stride-based)
3495 * \param offset byte offset into the buffer
3496 */
3497 static void buffer_append_args(
3498 struct si_shader_context *ctx,
3499 struct lp_build_emit_data *emit_data,
3500 LLVMValueRef rsrc,
3501 LLVMValueRef index,
3502 LLVMValueRef offset,
3503 bool atomic,
3504 bool force_glc)
3505 {
3506 const struct tgsi_full_instruction *inst = emit_data->inst;
3507 LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
3508 LLVMValueRef i1true = LLVMConstInt(ctx->i1, 1, 0);
3509
3510 emit_data->args[emit_data->arg_count++] = rsrc;
3511 emit_data->args[emit_data->arg_count++] = index; /* vindex */
3512 emit_data->args[emit_data->arg_count++] = offset; /* voffset */
3513 if (!atomic) {
3514 emit_data->args[emit_data->arg_count++] =
3515 force_glc ||
3516 inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE) ?
3517 i1true : i1false; /* glc */
3518 }
3519 emit_data->args[emit_data->arg_count++] = i1false; /* slc */
3520 }
3521
3522 static void load_fetch_args(
3523 struct lp_build_tgsi_context * bld_base,
3524 struct lp_build_emit_data * emit_data)
3525 {
3526 struct si_shader_context *ctx = si_shader_context(bld_base);
3527 struct gallivm_state *gallivm = &ctx->gallivm;
3528 const struct tgsi_full_instruction * inst = emit_data->inst;
3529 unsigned target = inst->Memory.Texture;
3530 LLVMValueRef rsrc;
3531
3532 emit_data->dst_type = ctx->v4f32;
3533
3534 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
3535 LLVMBuilderRef builder = gallivm->builder;
3536 LLVMValueRef offset;
3537 LLVMValueRef tmp;
3538
3539 rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
3540
3541 tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
3542 offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
3543
3544 buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
3545 offset, false, false);
3546 } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
3547 LLVMValueRef coords;
3548
3549 image_fetch_rsrc(bld_base, &inst->Src[0], false, target, &rsrc);
3550 coords = image_fetch_coords(bld_base, inst, 1, rsrc);
3551
3552 if (target == TGSI_TEXTURE_BUFFER) {
3553 buffer_append_args(ctx, emit_data, rsrc, coords,
3554 ctx->i32_0, false, false);
3555 } else {
3556 emit_data->args[0] = coords;
3557 emit_data->args[1] = rsrc;
3558 emit_data->args[2] = LLVMConstInt(ctx->i32, 15, 0); /* dmask */
3559 emit_data->arg_count = 3;
3560
3561 image_append_args(ctx, emit_data, target, false, false);
3562 }
3563 }
3564 }
3565
3566 static unsigned get_load_intr_attribs(bool readonly_memory)
3567 {
3568 /* READNONE means writes can't affect it, while READONLY means that
3569 * writes can affect it. */
3570 return readonly_memory && HAVE_LLVM >= 0x0400 ?
3571 LP_FUNC_ATTR_READNONE :
3572 LP_FUNC_ATTR_READONLY;
3573 }
3574
3575 static unsigned get_store_intr_attribs(bool writeonly_memory)
3576 {
3577 return writeonly_memory && HAVE_LLVM >= 0x0400 ?
3578 LP_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
3579 LP_FUNC_ATTR_WRITEONLY;
3580 }
3581
3582 static void load_emit_buffer(struct si_shader_context *ctx,
3583 struct lp_build_emit_data *emit_data,
3584 bool readonly_memory)
3585 {
3586 const struct tgsi_full_instruction *inst = emit_data->inst;
3587 struct gallivm_state *gallivm = &ctx->gallivm;
3588 LLVMBuilderRef builder = gallivm->builder;
3589 uint writemask = inst->Dst[0].Register.WriteMask;
3590 uint count = util_last_bit(writemask);
3591 const char *intrinsic_name;
3592 LLVMTypeRef dst_type;
3593
3594 switch (count) {
3595 case 1:
3596 intrinsic_name = "llvm.amdgcn.buffer.load.f32";
3597 dst_type = ctx->f32;
3598 break;
3599 case 2:
3600 intrinsic_name = "llvm.amdgcn.buffer.load.v2f32";
3601 dst_type = LLVMVectorType(ctx->f32, 2);
3602 break;
3603 default: // 3 & 4
3604 intrinsic_name = "llvm.amdgcn.buffer.load.v4f32";
3605 dst_type = ctx->v4f32;
3606 count = 4;
3607 }
3608
3609 emit_data->output[emit_data->chan] = lp_build_intrinsic(
3610 builder, intrinsic_name, dst_type,
3611 emit_data->args, emit_data->arg_count,
3612 get_load_intr_attribs(readonly_memory));
3613 }
3614
3615 static LLVMValueRef get_memory_ptr(struct si_shader_context *ctx,
3616 const struct tgsi_full_instruction *inst,
3617 LLVMTypeRef type, int arg)
3618 {
3619 struct gallivm_state *gallivm = &ctx->gallivm;
3620 LLVMBuilderRef builder = gallivm->builder;
3621 LLVMValueRef offset, ptr;
3622 int addr_space;
3623
3624 offset = lp_build_emit_fetch(&ctx->bld_base, inst, arg, 0);
3625 offset = LLVMBuildBitCast(builder, offset, ctx->i32, "");
3626
3627 ptr = ctx->shared_memory;
3628 ptr = LLVMBuildGEP(builder, ptr, &offset, 1, "");
3629 addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3630 ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, addr_space), "");
3631
3632 return ptr;
3633 }
3634
3635 static void load_emit_memory(
3636 struct si_shader_context *ctx,
3637 struct lp_build_emit_data *emit_data)
3638 {
3639 const struct tgsi_full_instruction *inst = emit_data->inst;
3640 struct gallivm_state *gallivm = &ctx->gallivm;
3641 LLVMBuilderRef builder = gallivm->builder;
3642 unsigned writemask = inst->Dst[0].Register.WriteMask;
3643 LLVMValueRef channels[4], ptr, derived_ptr, index;
3644 int chan;
3645
3646 ptr = get_memory_ptr(ctx, inst, ctx->f32, 1);
3647
3648 for (chan = 0; chan < 4; ++chan) {
3649 if (!(writemask & (1 << chan))) {
3650 channels[chan] = LLVMGetUndef(ctx->f32);
3651 continue;
3652 }
3653
3654 index = LLVMConstInt(ctx->i32, chan, 0);
3655 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
3656 channels[chan] = LLVMBuildLoad(builder, derived_ptr, "");
3657 }
3658 emit_data->output[emit_data->chan] = lp_build_gather_values(gallivm, channels, 4);
3659 }
3660
3661 /**
3662 * Return true if the memory accessed by a LOAD or STORE instruction is
3663 * read-only or write-only, respectively.
3664 *
3665 * \param shader_buffers_reverse_access_mask
3666 * For LOAD, set this to (store | atomic) slot usage in the shader.
3667 * For STORE, set this to (load | atomic) slot usage in the shader.
3668 * \param images_reverse_access_mask Same as above, but for images.
3669 */
3670 static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
3671 const struct tgsi_shader_info *info,
3672 unsigned shader_buffers_reverse_access_mask,
3673 unsigned images_reverse_access_mask)
3674 {
3675 /* RESTRICT means NOALIAS.
3676 * If there are no writes, we can assume the accessed memory is read-only.
3677 * If there are no reads, we can assume the accessed memory is write-only.
3678 */
3679 if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT) {
3680 unsigned reverse_access_mask;
3681
3682 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
3683 reverse_access_mask = shader_buffers_reverse_access_mask;
3684 } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
3685 reverse_access_mask = info->images_buffers &
3686 images_reverse_access_mask;
3687 } else {
3688 reverse_access_mask = ~info->images_buffers &
3689 images_reverse_access_mask;
3690 }
3691
3692 if (inst->Src[0].Register.Indirect) {
3693 if (!reverse_access_mask)
3694 return true;
3695 } else {
3696 if (!(reverse_access_mask &
3697 (1u << inst->Src[0].Register.Index)))
3698 return true;
3699 }
3700 }
3701
3702 /* If there are no buffer writes (for both shader buffers & image
3703 * buffers), it implies that buffer memory is read-only.
3704 * If there are no buffer reads (for both shader buffers & image
3705 * buffers), it implies that buffer memory is write-only.
3706 *
3707 * Same for the case when there are no writes/reads for non-buffer
3708 * images.
3709 */
3710 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
3711 (inst->Src[0].Register.File == TGSI_FILE_IMAGE &&
3712 inst->Memory.Texture == TGSI_TEXTURE_BUFFER)) {
3713 if (!shader_buffers_reverse_access_mask &&
3714 !(info->images_buffers & images_reverse_access_mask))
3715 return true;
3716 } else {
3717 if (!(~info->images_buffers & images_reverse_access_mask))
3718 return true;
3719 }
3720 return false;
3721 }
3722
3723 static void load_emit(
3724 const struct lp_build_tgsi_action *action,
3725 struct lp_build_tgsi_context *bld_base,
3726 struct lp_build_emit_data *emit_data)
3727 {
3728 struct si_shader_context *ctx = si_shader_context(bld_base);
3729 struct gallivm_state *gallivm = &ctx->gallivm;
3730 LLVMBuilderRef builder = gallivm->builder;
3731 const struct tgsi_full_instruction * inst = emit_data->inst;
3732 const struct tgsi_shader_info *info = &ctx->shader->selector->info;
3733 char intrinsic_name[64];
3734 bool readonly_memory = false;
3735
3736 if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
3737 load_emit_memory(ctx, emit_data);
3738 return;
3739 }
3740
3741 if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
3742 emit_waitcnt(ctx, VM_CNT);
3743
3744 readonly_memory = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
3745 is_oneway_access_only(inst, info,
3746 info->shader_buffers_store |
3747 info->shader_buffers_atomic,
3748 info->images_store |
3749 info->images_atomic);
3750
3751 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
3752 load_emit_buffer(ctx, emit_data, readonly_memory);
3753 return;
3754 }
3755
3756 if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
3757 emit_data->output[emit_data->chan] =
3758 lp_build_intrinsic(
3759 builder, "llvm.amdgcn.buffer.load.format.v4f32", emit_data->dst_type,
3760 emit_data->args, emit_data->arg_count,
3761 get_load_intr_attribs(readonly_memory));
3762 } else {
3763 ac_get_image_intr_name("llvm.amdgcn.image.load",
3764 emit_data->dst_type, /* vdata */
3765 LLVMTypeOf(emit_data->args[0]), /* coords */
3766 LLVMTypeOf(emit_data->args[1]), /* rsrc */
3767 intrinsic_name, sizeof(intrinsic_name));
3768
3769 emit_data->output[emit_data->chan] =
3770 lp_build_intrinsic(
3771 builder, intrinsic_name, emit_data->dst_type,
3772 emit_data->args, emit_data->arg_count,
3773 get_load_intr_attribs(readonly_memory));
3774 }
3775 }
3776
3777 static void store_fetch_args(
3778 struct lp_build_tgsi_context * bld_base,
3779 struct lp_build_emit_data * emit_data)
3780 {
3781 struct si_shader_context *ctx = si_shader_context(bld_base);
3782 struct gallivm_state *gallivm = &ctx->gallivm;
3783 LLVMBuilderRef builder = gallivm->builder;
3784 const struct tgsi_full_instruction * inst = emit_data->inst;
3785 struct tgsi_full_src_register memory;
3786 LLVMValueRef chans[4];
3787 LLVMValueRef data;
3788 LLVMValueRef rsrc;
3789 unsigned chan;
3790
3791 emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
3792
3793 for (chan = 0; chan < 4; ++chan) {
3794 chans[chan] = lp_build_emit_fetch(bld_base, inst, 1, chan);
3795 }
3796 data = lp_build_gather_values(gallivm, chans, 4);
3797
3798 emit_data->args[emit_data->arg_count++] = data;
3799
3800 memory = tgsi_full_src_register_from_dst(&inst->Dst[0]);
3801
3802 if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
3803 LLVMValueRef offset;
3804 LLVMValueRef tmp;
3805
3806 rsrc = shader_buffer_fetch_rsrc(ctx, &memory);
3807
3808 tmp = lp_build_emit_fetch(bld_base, inst, 0, 0);
3809 offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
3810
3811 buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
3812 offset, false, false);
3813 } else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE) {
3814 unsigned target = inst->Memory.Texture;
3815 LLVMValueRef coords;
3816
3817 /* 8bit/16bit TC L1 write corruption bug on SI.
3818 * All store opcodes not aligned to a dword are affected.
3819 *
3820 * The only way to get unaligned stores in radeonsi is through
3821 * shader images.
3822 */
3823 bool force_glc = ctx->screen->b.chip_class == SI;
3824
3825 image_fetch_rsrc(bld_base, &memory, true, target, &rsrc);
3826 coords = image_fetch_coords(bld_base, inst, 0, rsrc);
3827
3828 if (target == TGSI_TEXTURE_BUFFER) {
3829 buffer_append_args(ctx, emit_data, rsrc, coords,
3830 ctx->i32_0, false, force_glc);
3831 } else {
3832 emit_data->args[1] = coords;
3833 emit_data->args[2] = rsrc;
3834 emit_data->args[3] = LLVMConstInt(ctx->i32, 15, 0); /* dmask */
3835 emit_data->arg_count = 4;
3836
3837 image_append_args(ctx, emit_data, target, false, force_glc);
3838 }
3839 }
3840 }
3841
3842 static void store_emit_buffer(
3843 struct si_shader_context *ctx,
3844 struct lp_build_emit_data *emit_data,
3845 bool writeonly_memory)
3846 {
3847 const struct tgsi_full_instruction *inst = emit_data->inst;
3848 struct gallivm_state *gallivm = &ctx->gallivm;
3849 LLVMBuilderRef builder = gallivm->builder;
3850 LLVMValueRef base_data = emit_data->args[0];
3851 LLVMValueRef base_offset = emit_data->args[3];
3852 unsigned writemask = inst->Dst[0].Register.WriteMask;
3853
3854 while (writemask) {
3855 int start, count;
3856 const char *intrinsic_name;
3857 LLVMValueRef data;
3858 LLVMValueRef offset;
3859 LLVMValueRef tmp;
3860
3861 u_bit_scan_consecutive_range(&writemask, &start, &count);
3862
3863 /* Due to an LLVM limitation, split 3-element writes
3864 * into a 2-element and a 1-element write. */
3865 if (count == 3) {
3866 writemask |= 1 << (start + 2);
3867 count = 2;
3868 }
3869
3870 if (count == 4) {
3871 data = base_data;
3872 intrinsic_name = "llvm.amdgcn.buffer.store.v4f32";
3873 } else if (count == 2) {
3874 LLVMTypeRef v2f32 = LLVMVectorType(ctx->f32, 2);
3875
3876 tmp = LLVMBuildExtractElement(
3877 builder, base_data,
3878 LLVMConstInt(ctx->i32, start, 0), "");
3879 data = LLVMBuildInsertElement(
3880 builder, LLVMGetUndef(v2f32), tmp,
3881 ctx->i32_0, "");
3882
3883 tmp = LLVMBuildExtractElement(
3884 builder, base_data,
3885 LLVMConstInt(ctx->i32, start + 1, 0), "");
3886 data = LLVMBuildInsertElement(
3887 builder, data, tmp, ctx->i32_1, "");
3888
3889 intrinsic_name = "llvm.amdgcn.buffer.store.v2f32";
3890 } else {
3891 assert(count == 1);
3892 data = LLVMBuildExtractElement(
3893 builder, base_data,
3894 LLVMConstInt(ctx->i32, start, 0), "");
3895 intrinsic_name = "llvm.amdgcn.buffer.store.f32";
3896 }
3897
3898 offset = base_offset;
3899 if (start != 0) {
3900 offset = LLVMBuildAdd(
3901 builder, offset,
3902 LLVMConstInt(ctx->i32, start * 4, 0), "");
3903 }
3904
3905 emit_data->args[0] = data;
3906 emit_data->args[3] = offset;
3907
3908 lp_build_intrinsic(
3909 builder, intrinsic_name, emit_data->dst_type,
3910 emit_data->args, emit_data->arg_count,
3911 get_store_intr_attribs(writeonly_memory));
3912 }
3913 }
3914
3915 static void store_emit_memory(
3916 struct si_shader_context *ctx,
3917 struct lp_build_emit_data *emit_data)
3918 {
3919 const struct tgsi_full_instruction *inst = emit_data->inst;
3920 struct gallivm_state *gallivm = &ctx->gallivm;
3921 LLVMBuilderRef builder = gallivm->builder;
3922 unsigned writemask = inst->Dst[0].Register.WriteMask;
3923 LLVMValueRef ptr, derived_ptr, data, index;
3924 int chan;
3925
3926 ptr = get_memory_ptr(ctx, inst, ctx->f32, 0);
3927
3928 for (chan = 0; chan < 4; ++chan) {
3929 if (!(writemask & (1 << chan))) {
3930 continue;
3931 }
3932 data = lp_build_emit_fetch(&ctx->bld_base, inst, 1, chan);
3933 index = LLVMConstInt(ctx->i32, chan, 0);
3934 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
3935 LLVMBuildStore(builder, data, derived_ptr);
3936 }
3937 }
3938
3939 static void store_emit(
3940 const struct lp_build_tgsi_action *action,
3941 struct lp_build_tgsi_context *bld_base,
3942 struct lp_build_emit_data *emit_data)
3943 {
3944 struct si_shader_context *ctx = si_shader_context(bld_base);
3945 struct gallivm_state *gallivm = &ctx->gallivm;
3946 LLVMBuilderRef builder = gallivm->builder;
3947 const struct tgsi_full_instruction * inst = emit_data->inst;
3948 const struct tgsi_shader_info *info = &ctx->shader->selector->info;
3949 unsigned target = inst->Memory.Texture;
3950 char intrinsic_name[64];
3951 bool writeonly_memory = false;
3952
3953 if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
3954 store_emit_memory(ctx, emit_data);
3955 return;
3956 }
3957
3958 if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
3959 emit_waitcnt(ctx, VM_CNT);
3960
3961 writeonly_memory = is_oneway_access_only(inst, info,
3962 info->shader_buffers_load |
3963 info->shader_buffers_atomic,
3964 info->images_load |
3965 info->images_atomic);
3966
3967 if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
3968 store_emit_buffer(ctx, emit_data, writeonly_memory);
3969 return;
3970 }
3971
3972 if (target == TGSI_TEXTURE_BUFFER) {
3973 emit_data->output[emit_data->chan] = lp_build_intrinsic(
3974 builder, "llvm.amdgcn.buffer.store.format.v4f32",
3975 emit_data->dst_type, emit_data->args,
3976 emit_data->arg_count,
3977 get_store_intr_attribs(writeonly_memory));
3978 } else {
3979 ac_get_image_intr_name("llvm.amdgcn.image.store",
3980 LLVMTypeOf(emit_data->args[0]), /* vdata */
3981 LLVMTypeOf(emit_data->args[1]), /* coords */
3982 LLVMTypeOf(emit_data->args[2]), /* rsrc */
3983 intrinsic_name, sizeof(intrinsic_name));
3984
3985 emit_data->output[emit_data->chan] =
3986 lp_build_intrinsic(
3987 builder, intrinsic_name, emit_data->dst_type,
3988 emit_data->args, emit_data->arg_count,
3989 get_store_intr_attribs(writeonly_memory));
3990 }
3991 }
3992
3993 static void atomic_fetch_args(
3994 struct lp_build_tgsi_context * bld_base,
3995 struct lp_build_emit_data * emit_data)
3996 {
3997 struct si_shader_context *ctx = si_shader_context(bld_base);
3998 struct gallivm_state *gallivm = &ctx->gallivm;
3999 LLVMBuilderRef builder = gallivm->builder;
4000 const struct tgsi_full_instruction * inst = emit_data->inst;
4001 LLVMValueRef data1, data2;
4002 LLVMValueRef rsrc;
4003 LLVMValueRef tmp;
4004
4005 emit_data->dst_type = ctx->f32;
4006
4007 tmp = lp_build_emit_fetch(bld_base, inst, 2, 0);
4008 data1 = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
4009
4010 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
4011 tmp = lp_build_emit_fetch(bld_base, inst, 3, 0);
4012 data2 = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
4013 }
4014
4015 /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
4016 * of arguments, which is reversed relative to TGSI (and GLSL)
4017 */
4018 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
4019 emit_data->args[emit_data->arg_count++] = data2;
4020 emit_data->args[emit_data->arg_count++] = data1;
4021
4022 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
4023 LLVMValueRef offset;
4024
4025 rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
4026
4027 tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
4028 offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
4029
4030 buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
4031 offset, true, false);
4032 } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
4033 unsigned target = inst->Memory.Texture;
4034 LLVMValueRef coords;
4035
4036 image_fetch_rsrc(bld_base, &inst->Src[0], true, target, &rsrc);
4037 coords = image_fetch_coords(bld_base, inst, 1, rsrc);
4038
4039 if (target == TGSI_TEXTURE_BUFFER) {
4040 buffer_append_args(ctx, emit_data, rsrc, coords,
4041 ctx->i32_0, true, false);
4042 } else {
4043 emit_data->args[emit_data->arg_count++] = coords;
4044 emit_data->args[emit_data->arg_count++] = rsrc;
4045
4046 image_append_args(ctx, emit_data, target, true, false);
4047 }
4048 }
4049 }
4050
4051 static void atomic_emit_memory(struct si_shader_context *ctx,
4052 struct lp_build_emit_data *emit_data) {
4053 struct gallivm_state *gallivm = &ctx->gallivm;
4054 LLVMBuilderRef builder = gallivm->builder;
4055 const struct tgsi_full_instruction * inst = emit_data->inst;
4056 LLVMValueRef ptr, result, arg;
4057
4058 ptr = get_memory_ptr(ctx, inst, ctx->i32, 1);
4059
4060 arg = lp_build_emit_fetch(&ctx->bld_base, inst, 2, 0);
4061 arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
4062
4063 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
4064 LLVMValueRef new_data;
4065 new_data = lp_build_emit_fetch(&ctx->bld_base,
4066 inst, 3, 0);
4067
4068 new_data = LLVMBuildBitCast(builder, new_data, ctx->i32, "");
4069
4070 #if HAVE_LLVM >= 0x309
4071 result = LLVMBuildAtomicCmpXchg(builder, ptr, arg, new_data,
4072 LLVMAtomicOrderingSequentiallyConsistent,
4073 LLVMAtomicOrderingSequentiallyConsistent,
4074 false);
4075 #endif
4076
4077 result = LLVMBuildExtractValue(builder, result, 0, "");
4078 } else {
4079 LLVMAtomicRMWBinOp op;
4080
4081 switch(inst->Instruction.Opcode) {
4082 case TGSI_OPCODE_ATOMUADD:
4083 op = LLVMAtomicRMWBinOpAdd;
4084 break;
4085 case TGSI_OPCODE_ATOMXCHG:
4086 op = LLVMAtomicRMWBinOpXchg;
4087 break;
4088 case TGSI_OPCODE_ATOMAND:
4089 op = LLVMAtomicRMWBinOpAnd;
4090 break;
4091 case TGSI_OPCODE_ATOMOR:
4092 op = LLVMAtomicRMWBinOpOr;
4093 break;
4094 case TGSI_OPCODE_ATOMXOR:
4095 op = LLVMAtomicRMWBinOpXor;
4096 break;
4097 case TGSI_OPCODE_ATOMUMIN:
4098 op = LLVMAtomicRMWBinOpUMin;
4099 break;
4100 case TGSI_OPCODE_ATOMUMAX:
4101 op = LLVMAtomicRMWBinOpUMax;
4102 break;
4103 case TGSI_OPCODE_ATOMIMIN:
4104 op = LLVMAtomicRMWBinOpMin;
4105 break;
4106 case TGSI_OPCODE_ATOMIMAX:
4107 op = LLVMAtomicRMWBinOpMax;
4108 break;
4109 default:
4110 unreachable("unknown atomic opcode");
4111 }
4112
4113 result = LLVMBuildAtomicRMW(builder, op, ptr, arg,
4114 LLVMAtomicOrderingSequentiallyConsistent,
4115 false);
4116 }
4117 emit_data->output[emit_data->chan] = LLVMBuildBitCast(builder, result, emit_data->dst_type, "");
4118 }
4119
4120 static void atomic_emit(
4121 const struct lp_build_tgsi_action *action,
4122 struct lp_build_tgsi_context *bld_base,
4123 struct lp_build_emit_data *emit_data)
4124 {
4125 struct si_shader_context *ctx = si_shader_context(bld_base);
4126 struct gallivm_state *gallivm = &ctx->gallivm;
4127 LLVMBuilderRef builder = gallivm->builder;
4128 const struct tgsi_full_instruction * inst = emit_data->inst;
4129 char intrinsic_name[40];
4130 LLVMValueRef tmp;
4131
4132 if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
4133 atomic_emit_memory(ctx, emit_data);
4134 return;
4135 }
4136
4137 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
4138 inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
4139 snprintf(intrinsic_name, sizeof(intrinsic_name),
4140 "llvm.amdgcn.buffer.atomic.%s", action->intr_name);
4141 } else {
4142 LLVMValueRef coords;
4143 char coords_type[8];
4144
4145 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
4146 coords = emit_data->args[2];
4147 else
4148 coords = emit_data->args[1];
4149
4150 ac_build_type_name_for_intr(LLVMTypeOf(coords), coords_type, sizeof(coords_type));
4151 snprintf(intrinsic_name, sizeof(intrinsic_name),
4152 "llvm.amdgcn.image.atomic.%s.%s",
4153 action->intr_name, coords_type);
4154 }
4155
4156 tmp = lp_build_intrinsic(
4157 builder, intrinsic_name, ctx->i32,
4158 emit_data->args, emit_data->arg_count, 0);
4159 emit_data->output[emit_data->chan] =
4160 LLVMBuildBitCast(builder, tmp, ctx->f32, "");
4161 }
4162
4163 static void set_tex_fetch_args(struct si_shader_context *ctx,
4164 struct lp_build_emit_data *emit_data,
4165 unsigned target,
4166 LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
4167 LLVMValueRef *param, unsigned count,
4168 unsigned dmask)
4169 {
4170 struct gallivm_state *gallivm = &ctx->gallivm;
4171 struct ac_image_args args = {};
4172
4173 /* Pad to power of two vector */
4174 while (count < util_next_power_of_two(count))
4175 param[count++] = LLVMGetUndef(ctx->i32);
4176
4177 if (count > 1)
4178 args.addr = lp_build_gather_values(gallivm, param, count);
4179 else
4180 args.addr = param[0];
4181
4182 args.resource = res_ptr;
4183 args.sampler = samp_ptr;
4184 args.dmask = dmask;
4185 args.unorm = target == TGSI_TEXTURE_RECT ||
4186 target == TGSI_TEXTURE_SHADOWRECT;
4187 args.da = tgsi_is_array_sampler(target);
4188
4189 /* Ugly, but we seem to have no other choice right now. */
4190 STATIC_ASSERT(sizeof(args) <= sizeof(emit_data->args));
4191 memcpy(emit_data->args, &args, sizeof(args));
4192 }
4193
4194 static LLVMValueRef fix_resinfo(struct si_shader_context *ctx,
4195 unsigned target, LLVMValueRef out)
4196 {
4197 LLVMBuilderRef builder = ctx->gallivm.builder;
4198
4199 /* 1D textures are allocated and used as 2D on GFX9. */
4200 if (ctx->screen->b.chip_class >= GFX9 &&
4201 (target == TGSI_TEXTURE_1D_ARRAY ||
4202 target == TGSI_TEXTURE_SHADOW1D_ARRAY)) {
4203 LLVMValueRef layers =
4204 LLVMBuildExtractElement(builder, out,
4205 LLVMConstInt(ctx->i32, 2, 0), "");
4206 out = LLVMBuildInsertElement(builder, out, layers,
4207 ctx->i32_1, "");
4208 }
4209
4210 /* Divide the number of layers by 6 to get the number of cubes. */
4211 if (target == TGSI_TEXTURE_CUBE_ARRAY ||
4212 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
4213 LLVMValueRef imm2 = LLVMConstInt(ctx->i32, 2, 0);
4214
4215 LLVMValueRef z = LLVMBuildExtractElement(builder, out, imm2, "");
4216 z = LLVMBuildSDiv(builder, z, LLVMConstInt(ctx->i32, 6, 0), "");
4217
4218 out = LLVMBuildInsertElement(builder, out, z, imm2, "");
4219 }
4220 return out;
4221 }
4222
4223 static void resq_fetch_args(
4224 struct lp_build_tgsi_context * bld_base,
4225 struct lp_build_emit_data * emit_data)
4226 {
4227 struct si_shader_context *ctx = si_shader_context(bld_base);
4228 const struct tgsi_full_instruction *inst = emit_data->inst;
4229 const struct tgsi_full_src_register *reg = &inst->Src[0];
4230
4231 emit_data->dst_type = ctx->v4i32;
4232
4233 if (reg->Register.File == TGSI_FILE_BUFFER) {
4234 emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg);
4235 emit_data->arg_count = 1;
4236 } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
4237 image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
4238 &emit_data->args[0]);
4239 emit_data->arg_count = 1;
4240 } else {
4241 LLVMValueRef res_ptr;
4242 unsigned image_target;
4243
4244 if (inst->Memory.Texture == TGSI_TEXTURE_3D)
4245 image_target = TGSI_TEXTURE_2D_ARRAY;
4246 else
4247 image_target = inst->Memory.Texture;
4248
4249 image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
4250 &res_ptr);
4251 set_tex_fetch_args(ctx, emit_data, image_target,
4252 res_ptr, NULL, &ctx->i32_0, 1,
4253 0xf);
4254 }
4255 }
4256
4257 static void resq_emit(
4258 const struct lp_build_tgsi_action *action,
4259 struct lp_build_tgsi_context *bld_base,
4260 struct lp_build_emit_data *emit_data)
4261 {
4262 struct si_shader_context *ctx = si_shader_context(bld_base);
4263 struct gallivm_state *gallivm = &ctx->gallivm;
4264 LLVMBuilderRef builder = gallivm->builder;
4265 const struct tgsi_full_instruction *inst = emit_data->inst;
4266 LLVMValueRef out;
4267
4268 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
4269 out = LLVMBuildExtractElement(builder, emit_data->args[0],
4270 LLVMConstInt(ctx->i32, 2, 0), "");
4271 } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
4272 out = get_buffer_size(bld_base, emit_data->args[0]);
4273 } else {
4274 struct ac_image_args args;
4275
4276 memcpy(&args, emit_data->args, sizeof(args)); /* ugly */
4277 args.opcode = ac_image_get_resinfo;
4278 out = ac_build_image_opcode(&ctx->ac, &args);
4279
4280 out = fix_resinfo(ctx, inst->Memory.Texture, out);
4281 }
4282
4283 emit_data->output[emit_data->chan] = out;
4284 }
4285
4286 static const struct lp_build_tgsi_action tex_action;
4287
4288 enum desc_type {
4289 DESC_IMAGE,
4290 DESC_BUFFER,
4291 DESC_FMASK,
4292 DESC_SAMPLER,
4293 };
4294
4295 /**
4296 * Load an image view, fmask view. or sampler state descriptor.
4297 */
4298 static LLVMValueRef load_sampler_desc(struct si_shader_context *ctx,
4299 LLVMValueRef list, LLVMValueRef index,
4300 enum desc_type type)
4301 {
4302 struct gallivm_state *gallivm = &ctx->gallivm;
4303 LLVMBuilderRef builder = gallivm->builder;
4304
4305 switch (type) {
4306 case DESC_IMAGE:
4307 /* The image is at [0:7]. */
4308 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
4309 break;
4310 case DESC_BUFFER:
4311 /* The buffer is in [4:7]. */
4312 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4313 index = LLVMBuildAdd(builder, index, ctx->i32_1, "");
4314 list = LLVMBuildPointerCast(builder, list,
4315 const_array(ctx->v4i32, 0), "");
4316 break;
4317 case DESC_FMASK:
4318 /* The FMASK is at [8:15]. */
4319 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
4320 index = LLVMBuildAdd(builder, index, ctx->i32_1, "");
4321 break;
4322 case DESC_SAMPLER:
4323 /* The sampler state is at [12:15]. */
4324 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4325 index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 3, 0), "");
4326 list = LLVMBuildPointerCast(builder, list,
4327 const_array(ctx->v4i32, 0), "");
4328 break;
4329 }
4330
4331 return ac_build_indexed_load_const(&ctx->ac, list, index);
4332 }
4333
4334 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4335 *
4336 * SI-CI:
4337 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4338 * filtering manually. The driver sets img7 to a mask clearing
4339 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4340 * s_and_b32 samp0, samp0, img7
4341 *
4342 * VI:
4343 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4344 */
4345 static LLVMValueRef sici_fix_sampler_aniso(struct si_shader_context *ctx,
4346 LLVMValueRef res, LLVMValueRef samp)
4347 {
4348 LLVMBuilderRef builder = ctx->gallivm.builder;
4349 LLVMValueRef img7, samp0;
4350
4351 if (ctx->screen->b.chip_class >= VI)
4352 return samp;
4353
4354 img7 = LLVMBuildExtractElement(builder, res,
4355 LLVMConstInt(ctx->i32, 7, 0), "");
4356 samp0 = LLVMBuildExtractElement(builder, samp,
4357 ctx->i32_0, "");
4358 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
4359 return LLVMBuildInsertElement(builder, samp, samp0,
4360 ctx->i32_0, "");
4361 }
4362
4363 static void tex_fetch_ptrs(
4364 struct lp_build_tgsi_context *bld_base,
4365 struct lp_build_emit_data *emit_data,
4366 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr, LLVMValueRef *fmask_ptr)
4367 {
4368 struct si_shader_context *ctx = si_shader_context(bld_base);
4369 LLVMValueRef list = LLVMGetParam(ctx->main_fn, ctx->param_samplers);
4370 const struct tgsi_full_instruction *inst = emit_data->inst;
4371 const struct tgsi_full_src_register *reg;
4372 unsigned target = inst->Texture.Texture;
4373 unsigned sampler_src;
4374 LLVMValueRef index;
4375
4376 sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
4377 reg = &emit_data->inst->Src[sampler_src];
4378
4379 if (reg->Register.Indirect) {
4380 index = get_bounded_indirect_index(ctx,
4381 &reg->Indirect,
4382 reg->Register.Index,
4383 SI_NUM_SAMPLERS);
4384 } else {
4385 index = LLVMConstInt(ctx->i32, reg->Register.Index, 0);
4386 }
4387
4388 if (target == TGSI_TEXTURE_BUFFER)
4389 *res_ptr = load_sampler_desc(ctx, list, index, DESC_BUFFER);
4390 else
4391 *res_ptr = load_sampler_desc(ctx, list, index, DESC_IMAGE);
4392
4393 if (samp_ptr)
4394 *samp_ptr = NULL;
4395 if (fmask_ptr)
4396 *fmask_ptr = NULL;
4397
4398 if (target == TGSI_TEXTURE_2D_MSAA ||
4399 target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
4400 if (fmask_ptr)
4401 *fmask_ptr = load_sampler_desc(ctx, list, index,
4402 DESC_FMASK);
4403 } else if (target != TGSI_TEXTURE_BUFFER) {
4404 if (samp_ptr) {
4405 *samp_ptr = load_sampler_desc(ctx, list, index,
4406 DESC_SAMPLER);
4407 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
4408 }
4409 }
4410 }
4411
4412 static void txq_fetch_args(
4413 struct lp_build_tgsi_context *bld_base,
4414 struct lp_build_emit_data *emit_data)
4415 {
4416 struct si_shader_context *ctx = si_shader_context(bld_base);
4417 const struct tgsi_full_instruction *inst = emit_data->inst;
4418 unsigned target = inst->Texture.Texture;
4419 LLVMValueRef res_ptr;
4420 LLVMValueRef address;
4421
4422 tex_fetch_ptrs(bld_base, emit_data, &res_ptr, NULL, NULL);
4423
4424 if (target == TGSI_TEXTURE_BUFFER) {
4425 /* Read the size from the buffer descriptor directly. */
4426 emit_data->args[0] = get_buffer_size(bld_base, res_ptr);
4427 return;
4428 }
4429
4430 /* Textures - set the mip level. */
4431 address = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
4432
4433 set_tex_fetch_args(ctx, emit_data, target, res_ptr,
4434 NULL, &address, 1, 0xf);
4435 }
4436
4437 static void txq_emit(const struct lp_build_tgsi_action *action,
4438 struct lp_build_tgsi_context *bld_base,
4439 struct lp_build_emit_data *emit_data)
4440 {
4441 struct si_shader_context *ctx = si_shader_context(bld_base);
4442 struct ac_image_args args;
4443 unsigned target = emit_data->inst->Texture.Texture;
4444
4445 if (target == TGSI_TEXTURE_BUFFER) {
4446 /* Just return the buffer size. */
4447 emit_data->output[emit_data->chan] = emit_data->args[0];
4448 return;
4449 }
4450
4451 memcpy(&args, emit_data->args, sizeof(args)); /* ugly */
4452
4453 args.opcode = ac_image_get_resinfo;
4454 LLVMValueRef result = ac_build_image_opcode(&ctx->ac, &args);
4455
4456 emit_data->output[emit_data->chan] = fix_resinfo(ctx, target, result);
4457 }
4458
4459 static void tex_fetch_args(
4460 struct lp_build_tgsi_context *bld_base,
4461 struct lp_build_emit_data *emit_data)
4462 {
4463 struct si_shader_context *ctx = si_shader_context(bld_base);
4464 struct gallivm_state *gallivm = &ctx->gallivm;
4465 const struct tgsi_full_instruction *inst = emit_data->inst;
4466 unsigned opcode = inst->Instruction.Opcode;
4467 unsigned target = inst->Texture.Texture;
4468 LLVMValueRef coords[5], derivs[6];
4469 LLVMValueRef address[16];
4470 unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
4471 int ref_pos = tgsi_util_get_shadow_ref_src_index(target);
4472 unsigned count = 0;
4473 unsigned chan;
4474 unsigned num_deriv_channels = 0;
4475 bool has_offset = inst->Texture.NumOffsets > 0;
4476 LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
4477 unsigned dmask = 0xf;
4478
4479 tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
4480
4481 if (target == TGSI_TEXTURE_BUFFER) {
4482 emit_data->dst_type = ctx->v4f32;
4483 emit_data->args[0] = LLVMBuildBitCast(gallivm->builder, res_ptr,
4484 ctx->v16i8, "");
4485 emit_data->args[1] = ctx->i32_0;
4486 emit_data->args[2] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
4487 emit_data->arg_count = 3;
4488 return;
4489 }
4490
4491 /* Fetch and project texture coordinates */
4492 coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
4493 for (chan = 0; chan < 3; chan++ ) {
4494 coords[chan] = lp_build_emit_fetch(bld_base,
4495 emit_data->inst, 0,
4496 chan);
4497 if (opcode == TGSI_OPCODE_TXP)
4498 coords[chan] = lp_build_emit_llvm_binary(bld_base,
4499 TGSI_OPCODE_DIV,
4500 coords[chan],
4501 coords[3]);
4502 }
4503
4504 if (opcode == TGSI_OPCODE_TXP)
4505 coords[3] = bld_base->base.one;
4506
4507 /* Pack offsets. */
4508 if (has_offset &&
4509 opcode != TGSI_OPCODE_TXF &&
4510 opcode != TGSI_OPCODE_TXF_LZ) {
4511 /* The offsets are six-bit signed integers packed like this:
4512 * X=[5:0], Y=[13:8], and Z=[21:16].
4513 */
4514 LLVMValueRef offset[3], pack;
4515
4516 assert(inst->Texture.NumOffsets == 1);
4517
4518 for (chan = 0; chan < 3; chan++) {
4519 offset[chan] = lp_build_emit_fetch_texoffset(bld_base,
4520 emit_data->inst, 0, chan);
4521 offset[chan] = LLVMBuildAnd(gallivm->builder, offset[chan],
4522 LLVMConstInt(ctx->i32, 0x3f, 0), "");
4523 if (chan)
4524 offset[chan] = LLVMBuildShl(gallivm->builder, offset[chan],
4525 LLVMConstInt(ctx->i32, chan*8, 0), "");
4526 }
4527
4528 pack = LLVMBuildOr(gallivm->builder, offset[0], offset[1], "");
4529 pack = LLVMBuildOr(gallivm->builder, pack, offset[2], "");
4530 address[count++] = pack;
4531 }
4532
4533 /* Pack LOD bias value */
4534 if (opcode == TGSI_OPCODE_TXB)
4535 address[count++] = coords[3];
4536 if (opcode == TGSI_OPCODE_TXB2)
4537 address[count++] = lp_build_emit_fetch(bld_base, inst, 1, TGSI_CHAN_X);
4538
4539 /* Pack depth comparison value */
4540 if (tgsi_is_shadow_target(target) && opcode != TGSI_OPCODE_LODQ) {
4541 LLVMValueRef z;
4542
4543 if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
4544 z = lp_build_emit_fetch(bld_base, inst, 1, TGSI_CHAN_X);
4545 } else {
4546 assert(ref_pos >= 0);
4547 z = coords[ref_pos];
4548 }
4549
4550 /* TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4551 * so the depth comparison value isn't clamped for Z16 and
4552 * Z24 anymore. Do it manually here.
4553 *
4554 * It's unnecessary if the original texture format was
4555 * Z32_FLOAT, but we don't know that here.
4556 */
4557 if (ctx->screen->b.chip_class == VI)
4558 z = ac_build_clamp(&ctx->ac, z);
4559
4560 address[count++] = z;
4561 }
4562
4563 /* Pack user derivatives */
4564 if (opcode == TGSI_OPCODE_TXD) {
4565 int param, num_src_deriv_channels, num_dst_deriv_channels;
4566
4567 switch (target) {
4568 case TGSI_TEXTURE_3D:
4569 num_src_deriv_channels = 3;
4570 num_dst_deriv_channels = 3;
4571 num_deriv_channels = 3;
4572 break;
4573 case TGSI_TEXTURE_2D:
4574 case TGSI_TEXTURE_SHADOW2D:
4575 case TGSI_TEXTURE_RECT:
4576 case TGSI_TEXTURE_SHADOWRECT:
4577 case TGSI_TEXTURE_2D_ARRAY:
4578 case TGSI_TEXTURE_SHADOW2D_ARRAY:
4579 num_src_deriv_channels = 2;
4580 num_dst_deriv_channels = 2;
4581 num_deriv_channels = 2;
4582 break;
4583 case TGSI_TEXTURE_CUBE:
4584 case TGSI_TEXTURE_SHADOWCUBE:
4585 case TGSI_TEXTURE_CUBE_ARRAY:
4586 case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
4587 /* Cube derivatives will be converted to 2D. */
4588 num_src_deriv_channels = 3;
4589 num_dst_deriv_channels = 3;
4590 num_deriv_channels = 2;
4591 break;
4592 case TGSI_TEXTURE_1D:
4593 case TGSI_TEXTURE_SHADOW1D:
4594 case TGSI_TEXTURE_1D_ARRAY:
4595 case TGSI_TEXTURE_SHADOW1D_ARRAY:
4596 num_src_deriv_channels = 1;
4597
4598 /* 1D textures are allocated and used as 2D on GFX9. */
4599 if (ctx->screen->b.chip_class >= GFX9) {
4600 num_dst_deriv_channels = 2;
4601 num_deriv_channels = 2;
4602 } else {
4603 num_dst_deriv_channels = 1;
4604 num_deriv_channels = 1;
4605 }
4606 break;
4607 default:
4608 unreachable("invalid target");
4609 }
4610
4611 for (param = 0; param < 2; param++) {
4612 for (chan = 0; chan < num_src_deriv_channels; chan++)
4613 derivs[param * num_dst_deriv_channels + chan] =
4614 lp_build_emit_fetch(bld_base, inst, param+1, chan);
4615
4616 /* Fill in the rest with zeros. */
4617 for (chan = num_src_deriv_channels;
4618 chan < num_dst_deriv_channels; chan++)
4619 derivs[param * num_dst_deriv_channels + chan] =
4620 bld_base->base.zero;
4621 }
4622 }
4623
4624 if (target == TGSI_TEXTURE_CUBE ||
4625 target == TGSI_TEXTURE_CUBE_ARRAY ||
4626 target == TGSI_TEXTURE_SHADOWCUBE ||
4627 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
4628 ac_prepare_cube_coords(&ctx->ac,
4629 opcode == TGSI_OPCODE_TXD,
4630 target == TGSI_TEXTURE_CUBE_ARRAY ||
4631 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY,
4632 coords, derivs);
4633
4634 if (opcode == TGSI_OPCODE_TXD)
4635 for (int i = 0; i < num_deriv_channels * 2; i++)
4636 address[count++] = derivs[i];
4637
4638 /* Pack texture coordinates */
4639 address[count++] = coords[0];
4640 if (num_coords > 1)
4641 address[count++] = coords[1];
4642 if (num_coords > 2)
4643 address[count++] = coords[2];
4644
4645 /* 1D textures are allocated and used as 2D on GFX9. */
4646 if (ctx->screen->b.chip_class >= GFX9) {
4647 LLVMValueRef filler;
4648
4649 /* Use 0.5, so that we don't sample the border color. */
4650 if (opcode == TGSI_OPCODE_TXF)
4651 filler = ctx->i32_0;
4652 else
4653 filler = LLVMConstReal(ctx->f32, 0.5);
4654
4655 if (target == TGSI_TEXTURE_1D ||
4656 target == TGSI_TEXTURE_SHADOW1D) {
4657 address[count++] = filler;
4658 } else if (target == TGSI_TEXTURE_1D_ARRAY ||
4659 target == TGSI_TEXTURE_SHADOW1D_ARRAY) {
4660 address[count] = address[count - 1];
4661 address[count - 1] = filler;
4662 count++;
4663 }
4664 }
4665
4666 /* Pack LOD or sample index */
4667 if (opcode == TGSI_OPCODE_TXL || opcode == TGSI_OPCODE_TXF)
4668 address[count++] = coords[3];
4669 else if (opcode == TGSI_OPCODE_TXL2)
4670 address[count++] = lp_build_emit_fetch(bld_base, inst, 1, TGSI_CHAN_X);
4671
4672 if (count > 16) {
4673 assert(!"Cannot handle more than 16 texture address parameters");
4674 count = 16;
4675 }
4676
4677 for (chan = 0; chan < count; chan++ ) {
4678 address[chan] = LLVMBuildBitCast(gallivm->builder,
4679 address[chan], ctx->i32, "");
4680 }
4681
4682 /* Adjust the sample index according to FMASK.
4683 *
4684 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
4685 * which is the identity mapping. Each nibble says which physical sample
4686 * should be fetched to get that sample.
4687 *
4688 * For example, 0x11111100 means there are only 2 samples stored and
4689 * the second sample covers 3/4 of the pixel. When reading samples 0
4690 * and 1, return physical sample 0 (determined by the first two 0s
4691 * in FMASK), otherwise return physical sample 1.
4692 *
4693 * The sample index should be adjusted as follows:
4694 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
4695 */
4696 if (target == TGSI_TEXTURE_2D_MSAA ||
4697 target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
4698 struct lp_build_emit_data txf_emit_data = *emit_data;
4699 LLVMValueRef txf_address[4];
4700 /* We only need .xy for non-arrays, and .xyz for arrays. */
4701 unsigned txf_count = target == TGSI_TEXTURE_2D_MSAA ? 2 : 3;
4702 struct tgsi_full_instruction inst = {};
4703
4704 memcpy(txf_address, address, sizeof(txf_address));
4705
4706 /* Read FMASK using TXF_LZ. */
4707 inst.Instruction.Opcode = TGSI_OPCODE_TXF_LZ;
4708 inst.Texture.Texture = target;
4709 txf_emit_data.inst = &inst;
4710 txf_emit_data.chan = 0;
4711 set_tex_fetch_args(ctx, &txf_emit_data,
4712 target, fmask_ptr, NULL,
4713 txf_address, txf_count, 0xf);
4714 build_tex_intrinsic(&tex_action, bld_base, &txf_emit_data);
4715
4716 /* Initialize some constants. */
4717 LLVMValueRef four = LLVMConstInt(ctx->i32, 4, 0);
4718 LLVMValueRef F = LLVMConstInt(ctx->i32, 0xF, 0);
4719
4720 /* Apply the formula. */
4721 LLVMValueRef fmask =
4722 LLVMBuildExtractElement(gallivm->builder,
4723 txf_emit_data.output[0],
4724 ctx->i32_0, "");
4725
4726 unsigned sample_chan = txf_count; /* the sample index is last */
4727
4728 LLVMValueRef sample_index4 =
4729 LLVMBuildMul(gallivm->builder, address[sample_chan], four, "");
4730
4731 LLVMValueRef shifted_fmask =
4732 LLVMBuildLShr(gallivm->builder, fmask, sample_index4, "");
4733
4734 LLVMValueRef final_sample =
4735 LLVMBuildAnd(gallivm->builder, shifted_fmask, F, "");
4736
4737 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
4738 * resource descriptor is 0 (invalid),
4739 */
4740 LLVMValueRef fmask_desc =
4741 LLVMBuildBitCast(gallivm->builder, fmask_ptr,
4742 ctx->v8i32, "");
4743
4744 LLVMValueRef fmask_word1 =
4745 LLVMBuildExtractElement(gallivm->builder, fmask_desc,
4746 ctx->i32_1, "");
4747
4748 LLVMValueRef word1_is_nonzero =
4749 LLVMBuildICmp(gallivm->builder, LLVMIntNE,
4750 fmask_word1, ctx->i32_0, "");
4751
4752 /* Replace the MSAA sample index. */
4753 address[sample_chan] =
4754 LLVMBuildSelect(gallivm->builder, word1_is_nonzero,
4755 final_sample, address[sample_chan], "");
4756 }
4757
4758 if (opcode == TGSI_OPCODE_TXF ||
4759 opcode == TGSI_OPCODE_TXF_LZ) {
4760 /* add tex offsets */
4761 if (inst->Texture.NumOffsets) {
4762 struct lp_build_context *uint_bld = &bld_base->uint_bld;
4763 const struct tgsi_texture_offset *off = inst->TexOffsets;
4764
4765 assert(inst->Texture.NumOffsets == 1);
4766
4767 switch (target) {
4768 case TGSI_TEXTURE_3D:
4769 address[2] = lp_build_add(uint_bld, address[2],
4770 ctx->imms[off->Index * TGSI_NUM_CHANNELS + off->SwizzleZ]);
4771 /* fall through */
4772 case TGSI_TEXTURE_2D:
4773 case TGSI_TEXTURE_SHADOW2D:
4774 case TGSI_TEXTURE_RECT:
4775 case TGSI_TEXTURE_SHADOWRECT:
4776 case TGSI_TEXTURE_2D_ARRAY:
4777 case TGSI_TEXTURE_SHADOW2D_ARRAY:
4778 address[1] =
4779 lp_build_add(uint_bld, address[1],
4780 ctx->imms[off->Index * TGSI_NUM_CHANNELS + off->SwizzleY]);
4781 /* fall through */
4782 case TGSI_TEXTURE_1D:
4783 case TGSI_TEXTURE_SHADOW1D:
4784 case TGSI_TEXTURE_1D_ARRAY:
4785 case TGSI_TEXTURE_SHADOW1D_ARRAY:
4786 address[0] =
4787 lp_build_add(uint_bld, address[0],
4788 ctx->imms[off->Index * TGSI_NUM_CHANNELS + off->SwizzleX]);
4789 break;
4790 /* texture offsets do not apply to other texture targets */
4791 }
4792 }
4793 }
4794
4795 if (opcode == TGSI_OPCODE_TG4) {
4796 unsigned gather_comp = 0;
4797
4798 /* DMASK was repurposed for GATHER4. 4 components are always
4799 * returned and DMASK works like a swizzle - it selects
4800 * the component to fetch. The only valid DMASK values are
4801 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4802 * (red,red,red,red) etc.) The ISA document doesn't mention
4803 * this.
4804 */
4805
4806 /* Get the component index from src1.x for Gather4. */
4807 if (!tgsi_is_shadow_target(target)) {
4808 LLVMValueRef comp_imm;
4809 struct tgsi_src_register src1 = inst->Src[1].Register;
4810
4811 assert(src1.File == TGSI_FILE_IMMEDIATE);
4812
4813 comp_imm = ctx->imms[src1.Index * TGSI_NUM_CHANNELS + src1.SwizzleX];
4814 gather_comp = LLVMConstIntGetZExtValue(comp_imm);
4815 gather_comp = CLAMP(gather_comp, 0, 3);
4816 }
4817
4818 dmask = 1 << gather_comp;
4819 }
4820
4821 set_tex_fetch_args(ctx, emit_data, target, res_ptr,
4822 samp_ptr, address, count, dmask);
4823 }
4824
4825 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
4826 * incorrectly forces nearest filtering if the texture format is integer.
4827 * The only effect it has on Gather4, which always returns 4 texels for
4828 * bilinear filtering, is that the final coordinates are off by 0.5 of
4829 * the texel size.
4830 *
4831 * The workaround is to subtract 0.5 from the unnormalized coordinates,
4832 * or (0.5 / size) from the normalized coordinates.
4833 */
4834 static void si_lower_gather4_integer(struct si_shader_context *ctx,
4835 struct ac_image_args *args,
4836 unsigned target)
4837 {
4838 LLVMBuilderRef builder = ctx->gallivm.builder;
4839 LLVMValueRef coord = args->addr;
4840 LLVMValueRef half_texel[2];
4841 /* Texture coordinates start after:
4842 * {offset, bias, z-compare, derivatives}
4843 * Only the offset and z-compare can occur here.
4844 */
4845 unsigned coord_vgpr_index = (int)args->offset + (int)args->compare;
4846 int c;
4847
4848 if (target == TGSI_TEXTURE_RECT ||
4849 target == TGSI_TEXTURE_SHADOWRECT) {
4850 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
4851 } else {
4852 struct tgsi_full_instruction txq_inst = {};
4853 struct lp_build_emit_data txq_emit_data = {};
4854
4855 /* Query the texture size. */
4856 txq_inst.Texture.Texture = target;
4857 txq_emit_data.inst = &txq_inst;
4858 txq_emit_data.dst_type = ctx->v4i32;
4859 set_tex_fetch_args(ctx, &txq_emit_data, target,
4860 args->resource, NULL, &ctx->i32_0,
4861 1, 0xf);
4862 txq_emit(NULL, &ctx->bld_base, &txq_emit_data);
4863
4864 /* Compute -0.5 / size. */
4865 for (c = 0; c < 2; c++) {
4866 half_texel[c] =
4867 LLVMBuildExtractElement(builder, txq_emit_data.output[0],
4868 LLVMConstInt(ctx->i32, c, 0), "");
4869 half_texel[c] = LLVMBuildUIToFP(builder, half_texel[c], ctx->f32, "");
4870 half_texel[c] =
4871 lp_build_emit_llvm_unary(&ctx->bld_base,
4872 TGSI_OPCODE_RCP, half_texel[c]);
4873 half_texel[c] = LLVMBuildFMul(builder, half_texel[c],
4874 LLVMConstReal(ctx->f32, -0.5), "");
4875 }
4876 }
4877
4878 for (c = 0; c < 2; c++) {
4879 LLVMValueRef tmp;
4880 LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index + c, 0);
4881
4882 tmp = LLVMBuildExtractElement(builder, coord, index, "");
4883 tmp = LLVMBuildBitCast(builder, tmp, ctx->f32, "");
4884 tmp = LLVMBuildFAdd(builder, tmp, half_texel[c], "");
4885 tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
4886 coord = LLVMBuildInsertElement(builder, coord, tmp, index, "");
4887 }
4888
4889 args->addr = coord;
4890 }
4891
4892 static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
4893 struct lp_build_tgsi_context *bld_base,
4894 struct lp_build_emit_data *emit_data)
4895 {
4896 struct si_shader_context *ctx = si_shader_context(bld_base);
4897 const struct tgsi_full_instruction *inst = emit_data->inst;
4898 struct ac_image_args args;
4899 unsigned opcode = inst->Instruction.Opcode;
4900 unsigned target = inst->Texture.Texture;
4901
4902 if (target == TGSI_TEXTURE_BUFFER) {
4903 emit_data->output[emit_data->chan] =
4904 ac_build_buffer_load_format(&ctx->ac,
4905 emit_data->args[0],
4906 emit_data->args[2],
4907 emit_data->args[1],
4908 true);
4909 return;
4910 }
4911
4912 memcpy(&args, emit_data->args, sizeof(args)); /* ugly */
4913
4914 args.opcode = ac_image_sample;
4915 args.compare = tgsi_is_shadow_target(target);
4916 args.offset = inst->Texture.NumOffsets > 0;
4917
4918 switch (opcode) {
4919 case TGSI_OPCODE_TXF:
4920 case TGSI_OPCODE_TXF_LZ:
4921 args.opcode = opcode == TGSI_OPCODE_TXF_LZ ||
4922 target == TGSI_TEXTURE_2D_MSAA ||
4923 target == TGSI_TEXTURE_2D_ARRAY_MSAA ?
4924 ac_image_load : ac_image_load_mip;
4925 args.compare = false;
4926 args.offset = false;
4927 break;
4928 case TGSI_OPCODE_LODQ:
4929 args.opcode = ac_image_get_lod;
4930 args.compare = false;
4931 args.offset = false;
4932 break;
4933 case TGSI_OPCODE_TEX:
4934 case TGSI_OPCODE_TEX2:
4935 case TGSI_OPCODE_TXP:
4936 if (ctx->type != PIPE_SHADER_FRAGMENT)
4937 args.level_zero = true;
4938 break;
4939 case TGSI_OPCODE_TEX_LZ:
4940 args.level_zero = true;
4941 break;
4942 case TGSI_OPCODE_TXB:
4943 case TGSI_OPCODE_TXB2:
4944 assert(ctx->type == PIPE_SHADER_FRAGMENT);
4945 args.bias = true;
4946 break;
4947 case TGSI_OPCODE_TXL:
4948 case TGSI_OPCODE_TXL2:
4949 args.lod = true;
4950 break;
4951 case TGSI_OPCODE_TXD:
4952 args.deriv = true;
4953 break;
4954 case TGSI_OPCODE_TG4:
4955 args.opcode = ac_image_gather4;
4956 args.level_zero = true;
4957 break;
4958 default:
4959 assert(0);
4960 return;
4961 }
4962
4963 /* The hardware needs special lowering for Gather4 with integer formats. */
4964 if (ctx->screen->b.chip_class <= VI &&
4965 opcode == TGSI_OPCODE_TG4) {
4966 struct tgsi_shader_info *info = &ctx->shader->selector->info;
4967 /* This will also work with non-constant indexing because of how
4968 * glsl_to_tgsi works and we intent to preserve that behavior.
4969 */
4970 const unsigned src_idx = 2;
4971 unsigned sampler = inst->Src[src_idx].Register.Index;
4972
4973 assert(inst->Src[src_idx].Register.File == TGSI_FILE_SAMPLER);
4974
4975 if (info->sampler_type[sampler] == TGSI_RETURN_TYPE_SINT ||
4976 info->sampler_type[sampler] == TGSI_RETURN_TYPE_UINT)
4977 si_lower_gather4_integer(ctx, &args, target);
4978 }
4979
4980 emit_data->output[emit_data->chan] =
4981 ac_build_image_opcode(&ctx->ac, &args);
4982 }
4983
4984 static void si_llvm_emit_txqs(
4985 const struct lp_build_tgsi_action *action,
4986 struct lp_build_tgsi_context *bld_base,
4987 struct lp_build_emit_data *emit_data)
4988 {
4989 struct si_shader_context *ctx = si_shader_context(bld_base);
4990 struct gallivm_state *gallivm = &ctx->gallivm;
4991 LLVMBuilderRef builder = gallivm->builder;
4992 LLVMValueRef res, samples;
4993 LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
4994
4995 tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
4996
4997
4998 /* Read the samples from the descriptor directly. */
4999 res = LLVMBuildBitCast(builder, res_ptr, ctx->v8i32, "");
5000 samples = LLVMBuildExtractElement(
5001 builder, res,
5002 LLVMConstInt(ctx->i32, 3, 0), "");
5003 samples = LLVMBuildLShr(builder, samples,
5004 LLVMConstInt(ctx->i32, 16, 0), "");
5005 samples = LLVMBuildAnd(builder, samples,
5006 LLVMConstInt(ctx->i32, 0xf, 0), "");
5007 samples = LLVMBuildShl(builder, ctx->i32_1,
5008 samples, "");
5009
5010 emit_data->output[emit_data->chan] = samples;
5011 }
5012
5013 static void si_llvm_emit_ddxy(
5014 const struct lp_build_tgsi_action *action,
5015 struct lp_build_tgsi_context *bld_base,
5016 struct lp_build_emit_data *emit_data)
5017 {
5018 struct si_shader_context *ctx = si_shader_context(bld_base);
5019 struct gallivm_state *gallivm = &ctx->gallivm;
5020 unsigned opcode = emit_data->info->opcode;
5021 LLVMValueRef val;
5022 int idx;
5023 unsigned mask;
5024
5025 if (opcode == TGSI_OPCODE_DDX_FINE)
5026 mask = AC_TID_MASK_LEFT;
5027 else if (opcode == TGSI_OPCODE_DDY_FINE)
5028 mask = AC_TID_MASK_TOP;
5029 else
5030 mask = AC_TID_MASK_TOP_LEFT;
5031
5032 /* for DDX we want to next X pixel, DDY next Y pixel. */
5033 idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;
5034
5035 val = LLVMBuildBitCast(gallivm->builder, emit_data->args[0], ctx->i32, "");
5036 val = ac_build_ddxy(&ctx->ac, ctx->screen->has_ds_bpermute,
5037 mask, idx, ctx->lds, val);
5038 emit_data->output[emit_data->chan] = val;
5039 }
5040
5041 /*
5042 * this takes an I,J coordinate pair,
5043 * and works out the X and Y derivatives.
5044 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
5045 */
5046 static LLVMValueRef si_llvm_emit_ddxy_interp(
5047 struct lp_build_tgsi_context *bld_base,
5048 LLVMValueRef interp_ij)
5049 {
5050 struct si_shader_context *ctx = si_shader_context(bld_base);
5051 struct gallivm_state *gallivm = &ctx->gallivm;
5052 LLVMValueRef result[4], a;
5053 unsigned i;
5054
5055 for (i = 0; i < 2; i++) {
5056 a = LLVMBuildExtractElement(gallivm->builder, interp_ij,
5057 LLVMConstInt(ctx->i32, i, 0), "");
5058 result[i] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_DDX, a);
5059 result[2+i] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_DDY, a);
5060 }
5061
5062 return lp_build_gather_values(gallivm, result, 4);
5063 }
5064
5065 static void interp_fetch_args(
5066 struct lp_build_tgsi_context *bld_base,
5067 struct lp_build_emit_data *emit_data)
5068 {
5069 struct si_shader_context *ctx = si_shader_context(bld_base);
5070 struct gallivm_state *gallivm = &ctx->gallivm;
5071 const struct tgsi_full_instruction *inst = emit_data->inst;
5072
5073 if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) {
5074 /* offset is in second src, first two channels */
5075 emit_data->args[0] = lp_build_emit_fetch(bld_base,
5076 emit_data->inst, 1,
5077 TGSI_CHAN_X);
5078 emit_data->args[1] = lp_build_emit_fetch(bld_base,
5079 emit_data->inst, 1,
5080 TGSI_CHAN_Y);
5081 emit_data->arg_count = 2;
5082 } else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
5083 LLVMValueRef sample_position;
5084 LLVMValueRef sample_id;
5085 LLVMValueRef halfval = LLVMConstReal(ctx->f32, 0.5f);
5086
5087 /* fetch sample ID, then fetch its sample position,
5088 * and place into first two channels.
5089 */
5090 sample_id = lp_build_emit_fetch(bld_base,
5091 emit_data->inst, 1, TGSI_CHAN_X);
5092 sample_id = LLVMBuildBitCast(gallivm->builder, sample_id,
5093 ctx->i32, "");
5094 sample_position = load_sample_position(ctx, sample_id);
5095
5096 emit_data->args[0] = LLVMBuildExtractElement(gallivm->builder,
5097 sample_position,
5098 ctx->i32_0, "");
5099
5100 emit_data->args[0] = LLVMBuildFSub(gallivm->builder, emit_data->args[0], halfval, "");
5101 emit_data->args[1] = LLVMBuildExtractElement(gallivm->builder,
5102 sample_position,
5103 ctx->i32_1, "");
5104 emit_data->args[1] = LLVMBuildFSub(gallivm->builder, emit_data->args[1], halfval, "");
5105 emit_data->arg_count = 2;
5106 }
5107 }
5108
5109 static void build_interp_intrinsic(const struct lp_build_tgsi_action *action,
5110 struct lp_build_tgsi_context *bld_base,
5111 struct lp_build_emit_data *emit_data)
5112 {
5113 struct si_shader_context *ctx = si_shader_context(bld_base);
5114 struct si_shader *shader = ctx->shader;
5115 struct gallivm_state *gallivm = &ctx->gallivm;
5116 LLVMValueRef interp_param;
5117 const struct tgsi_full_instruction *inst = emit_data->inst;
5118 int input_index = inst->Src[0].Register.Index;
5119 int chan;
5120 int i;
5121 LLVMValueRef attr_number;
5122 LLVMValueRef params = LLVMGetParam(ctx->main_fn, SI_PARAM_PRIM_MASK);
5123 int interp_param_idx;
5124 unsigned interp = shader->selector->info.input_interpolate[input_index];
5125 unsigned location;
5126
5127 assert(inst->Src[0].Register.File == TGSI_FILE_INPUT);
5128
5129 if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
5130 inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE)
5131 location = TGSI_INTERPOLATE_LOC_CENTER;
5132 else
5133 location = TGSI_INTERPOLATE_LOC_CENTROID;
5134
5135 interp_param_idx = lookup_interp_param_index(interp, location);
5136 if (interp_param_idx == -1)
5137 return;
5138 else if (interp_param_idx)
5139 interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
5140 else
5141 interp_param = NULL;
5142
5143 attr_number = LLVMConstInt(ctx->i32, input_index, 0);
5144
5145 if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
5146 inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
5147 LLVMValueRef ij_out[2];
5148 LLVMValueRef ddxy_out = si_llvm_emit_ddxy_interp(bld_base, interp_param);
5149
5150 /*
5151 * take the I then J parameters, and the DDX/Y for it, and
5152 * calculate the IJ inputs for the interpolator.
5153 * temp1 = ddx * offset/sample.x + I;
5154 * interp_param.I = ddy * offset/sample.y + temp1;
5155 * temp1 = ddx * offset/sample.x + J;
5156 * interp_param.J = ddy * offset/sample.y + temp1;
5157 */
5158 for (i = 0; i < 2; i++) {
5159 LLVMValueRef ix_ll = LLVMConstInt(ctx->i32, i, 0);
5160 LLVMValueRef iy_ll = LLVMConstInt(ctx->i32, i + 2, 0);
5161 LLVMValueRef ddx_el = LLVMBuildExtractElement(gallivm->builder,
5162 ddxy_out, ix_ll, "");
5163 LLVMValueRef ddy_el = LLVMBuildExtractElement(gallivm->builder,
5164 ddxy_out, iy_ll, "");
5165 LLVMValueRef interp_el = LLVMBuildExtractElement(gallivm->builder,
5166 interp_param, ix_ll, "");
5167 LLVMValueRef temp1, temp2;
5168
5169 interp_el = LLVMBuildBitCast(gallivm->builder, interp_el,
5170 ctx->f32, "");
5171
5172 temp1 = LLVMBuildFMul(gallivm->builder, ddx_el, emit_data->args[0], "");
5173
5174 temp1 = LLVMBuildFAdd(gallivm->builder, temp1, interp_el, "");
5175
5176 temp2 = LLVMBuildFMul(gallivm->builder, ddy_el, emit_data->args[1], "");
5177
5178 ij_out[i] = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
5179 }
5180 interp_param = lp_build_gather_values(gallivm, ij_out, 2);
5181 }
5182
5183 for (chan = 0; chan < 4; chan++) {
5184 LLVMValueRef llvm_chan;
5185 unsigned schan;
5186
5187 schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);
5188 llvm_chan = LLVMConstInt(ctx->i32, schan, 0);
5189
5190 if (interp_param) {
5191 interp_param = LLVMBuildBitCast(gallivm->builder,
5192 interp_param, LLVMVectorType(ctx->f32, 2), "");
5193 LLVMValueRef i = LLVMBuildExtractElement(
5194 gallivm->builder, interp_param, ctx->i32_0, "");
5195 LLVMValueRef j = LLVMBuildExtractElement(
5196 gallivm->builder, interp_param, ctx->i32_1, "");
5197 emit_data->output[chan] = ac_build_fs_interp(&ctx->ac,
5198 llvm_chan, attr_number, params,
5199 i, j);
5200 } else {
5201 emit_data->output[chan] = ac_build_fs_interp_mov(&ctx->ac,
5202 LLVMConstInt(ctx->i32, 2, 0), /* P0 */
5203 llvm_chan, attr_number, params);
5204 }
5205 }
5206 }
5207
5208 static LLVMValueRef si_emit_ballot(struct si_shader_context *ctx,
5209 LLVMValueRef value)
5210 {
5211 struct gallivm_state *gallivm = &ctx->gallivm;
5212 LLVMValueRef args[3] = {
5213 value,
5214 ctx->i32_0,
5215 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
5216 };
5217
5218 /* We currently have no other way to prevent LLVM from lifting the icmp
5219 * calls to a dominating basic block.
5220 */
5221 emit_optimization_barrier(ctx, &args[0]);
5222
5223 if (LLVMTypeOf(args[0]) != ctx->i32)
5224 args[0] = LLVMBuildBitCast(gallivm->builder, args[0], ctx->i32, "");
5225
5226 return lp_build_intrinsic(gallivm->builder,
5227 "llvm.amdgcn.icmp.i32",
5228 ctx->i64, args, 3,
5229 LP_FUNC_ATTR_NOUNWIND |
5230 LP_FUNC_ATTR_READNONE |
5231 LP_FUNC_ATTR_CONVERGENT);
5232 }
5233
5234 static void vote_all_emit(
5235 const struct lp_build_tgsi_action *action,
5236 struct lp_build_tgsi_context *bld_base,
5237 struct lp_build_emit_data *emit_data)
5238 {
5239 struct si_shader_context *ctx = si_shader_context(bld_base);
5240 struct gallivm_state *gallivm = &ctx->gallivm;
5241 LLVMValueRef active_set, vote_set;
5242 LLVMValueRef tmp;
5243
5244 active_set = si_emit_ballot(ctx, ctx->i32_1);
5245 vote_set = si_emit_ballot(ctx, emit_data->args[0]);
5246
5247 tmp = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, vote_set, active_set, "");
5248 emit_data->output[emit_data->chan] =
5249 LLVMBuildSExt(gallivm->builder, tmp, ctx->i32, "");
5250 }
5251
5252 static void vote_any_emit(
5253 const struct lp_build_tgsi_action *action,
5254 struct lp_build_tgsi_context *bld_base,
5255 struct lp_build_emit_data *emit_data)
5256 {
5257 struct si_shader_context *ctx = si_shader_context(bld_base);
5258 struct gallivm_state *gallivm = &ctx->gallivm;
5259 LLVMValueRef vote_set;
5260 LLVMValueRef tmp;
5261
5262 vote_set = si_emit_ballot(ctx, emit_data->args[0]);
5263
5264 tmp = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
5265 vote_set, LLVMConstInt(ctx->i64, 0, 0), "");
5266 emit_data->output[emit_data->chan] =
5267 LLVMBuildSExt(gallivm->builder, tmp, ctx->i32, "");
5268 }
5269
5270 static void vote_eq_emit(
5271 const struct lp_build_tgsi_action *action,
5272 struct lp_build_tgsi_context *bld_base,
5273 struct lp_build_emit_data *emit_data)
5274 {
5275 struct si_shader_context *ctx = si_shader_context(bld_base);
5276 struct gallivm_state *gallivm = &ctx->gallivm;
5277 LLVMValueRef active_set, vote_set;
5278 LLVMValueRef all, none, tmp;
5279
5280 active_set = si_emit_ballot(ctx, ctx->i32_1);
5281 vote_set = si_emit_ballot(ctx, emit_data->args[0]);
5282
5283 all = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, vote_set, active_set, "");
5284 none = LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
5285 vote_set, LLVMConstInt(ctx->i64, 0, 0), "");
5286 tmp = LLVMBuildOr(gallivm->builder, all, none, "");
5287 emit_data->output[emit_data->chan] =
5288 LLVMBuildSExt(gallivm->builder, tmp, ctx->i32, "");
5289 }
5290
5291 static void ballot_emit(
5292 const struct lp_build_tgsi_action *action,
5293 struct lp_build_tgsi_context *bld_base,
5294 struct lp_build_emit_data *emit_data)
5295 {
5296 struct si_shader_context *ctx = si_shader_context(bld_base);
5297 LLVMBuilderRef builder = ctx->gallivm.builder;
5298 LLVMValueRef tmp;
5299
5300 tmp = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
5301 tmp = si_emit_ballot(ctx, tmp);
5302 tmp = LLVMBuildBitCast(builder, tmp, ctx->v2i32, "");
5303
5304 emit_data->output[0] = LLVMBuildExtractElement(builder, tmp, ctx->i32_0, "");
5305 emit_data->output[1] = LLVMBuildExtractElement(builder, tmp, ctx->i32_1, "");
5306 }
5307
5308 static void read_invoc_fetch_args(
5309 struct lp_build_tgsi_context *bld_base,
5310 struct lp_build_emit_data *emit_data)
5311 {
5312 emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
5313 0, emit_data->src_chan);
5314
5315 /* Always read the source invocation (= lane) from the X channel. */
5316 emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
5317 1, TGSI_CHAN_X);
5318 emit_data->arg_count = 2;
5319 }
5320
5321 static void read_lane_emit(
5322 const struct lp_build_tgsi_action *action,
5323 struct lp_build_tgsi_context *bld_base,
5324 struct lp_build_emit_data *emit_data)
5325 {
5326 struct si_shader_context *ctx = si_shader_context(bld_base);
5327 LLVMBuilderRef builder = ctx->gallivm.builder;
5328
5329 /* We currently have no other way to prevent LLVM from lifting the icmp
5330 * calls to a dominating basic block.
5331 */
5332 emit_optimization_barrier(ctx, &emit_data->args[0]);
5333
5334 for (unsigned i = 0; i < emit_data->arg_count; ++i) {
5335 emit_data->args[i] = LLVMBuildBitCast(builder, emit_data->args[i],
5336 ctx->i32, "");
5337 }
5338
5339 emit_data->output[emit_data->chan] =
5340 ac_build_intrinsic(&ctx->ac, action->intr_name,
5341 ctx->i32, emit_data->args, emit_data->arg_count,
5342 AC_FUNC_ATTR_READNONE |
5343 AC_FUNC_ATTR_CONVERGENT);
5344 }
5345
5346 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base,
5347 struct lp_build_emit_data *emit_data)
5348 {
5349 struct si_shader_context *ctx = si_shader_context(bld_base);
5350 struct tgsi_src_register src0 = emit_data->inst->Src[0].Register;
5351 LLVMValueRef imm;
5352 unsigned stream;
5353
5354 assert(src0.File == TGSI_FILE_IMMEDIATE);
5355
5356 imm = ctx->imms[src0.Index * TGSI_NUM_CHANNELS + src0.SwizzleX];
5357 stream = LLVMConstIntGetZExtValue(imm) & 0x3;
5358 return stream;
5359 }
5360
5361 /* Emit one vertex from the geometry shader */
5362 static void si_llvm_emit_vertex(
5363 const struct lp_build_tgsi_action *action,
5364 struct lp_build_tgsi_context *bld_base,
5365 struct lp_build_emit_data *emit_data)
5366 {
5367 struct si_shader_context *ctx = si_shader_context(bld_base);
5368 struct lp_build_context *uint = &bld_base->uint_bld;
5369 struct si_shader *shader = ctx->shader;
5370 struct tgsi_shader_info *info = &shader->selector->info;
5371 struct gallivm_state *gallivm = &ctx->gallivm;
5372 struct lp_build_if_state if_state;
5373 LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
5374 ctx->param_gs2vs_offset);
5375 LLVMValueRef gs_next_vertex;
5376 LLVMValueRef can_emit, kill;
5377 unsigned chan, offset;
5378 int i;
5379 unsigned stream;
5380
5381 stream = si_llvm_get_stream(bld_base, emit_data);
5382
5383 /* Write vertex attribute values to GSVS ring */
5384 gs_next_vertex = LLVMBuildLoad(gallivm->builder,
5385 ctx->gs_next_vertex[stream],
5386 "");
5387
5388 /* If this thread has already emitted the declared maximum number of
5389 * vertices, skip the write: excessive vertex emissions are not
5390 * supposed to have any effect.
5391 *
5392 * If the shader has no writes to memory, kill it instead. This skips
5393 * further memory loads and may allow LLVM to skip to the end
5394 * altogether.
5395 */
5396 can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULT, gs_next_vertex,
5397 LLVMConstInt(ctx->i32,
5398 shader->selector->gs_max_out_vertices, 0), "");
5399
5400 bool use_kill = !info->writes_memory;
5401 if (use_kill) {
5402 kill = lp_build_select(&bld_base->base, can_emit,
5403 LLVMConstReal(ctx->f32, 1.0f),
5404 LLVMConstReal(ctx->f32, -1.0f));
5405
5406 ac_build_kill(&ctx->ac, kill);
5407 } else {
5408 lp_build_if(&if_state, gallivm, can_emit);
5409 }
5410
5411 offset = 0;
5412 for (i = 0; i < info->num_outputs; i++) {
5413 LLVMValueRef *out_ptr = ctx->outputs[i];
5414
5415 for (chan = 0; chan < 4; chan++) {
5416 if (!(info->output_usagemask[i] & (1 << chan)) ||
5417 ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
5418 continue;
5419
5420 LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
5421 LLVMValueRef voffset =
5422 LLVMConstInt(ctx->i32, offset *
5423 shader->selector->gs_max_out_vertices, 0);
5424 offset++;
5425
5426 voffset = lp_build_add(uint, voffset, gs_next_vertex);
5427 voffset = lp_build_mul_imm(uint, voffset, 4);
5428
5429 out_val = LLVMBuildBitCast(gallivm->builder, out_val, ctx->i32, "");
5430
5431 ac_build_buffer_store_dword(&ctx->ac,
5432 ctx->gsvs_ring[stream],
5433 out_val, 1,
5434 voffset, soffset, 0,
5435 1, 1, true, true);
5436 }
5437 }
5438
5439 gs_next_vertex = lp_build_add(uint, gs_next_vertex,
5440 ctx->i32_1);
5441
5442 LLVMBuildStore(gallivm->builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
5443
5444 /* Signal vertex emission */
5445 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
5446 LLVMGetParam(ctx->main_fn, ctx->param_gs_wave_id));
5447 if (!use_kill)
5448 lp_build_endif(&if_state);
5449 }
5450
5451 /* Cut one primitive from the geometry shader */
5452 static void si_llvm_emit_primitive(
5453 const struct lp_build_tgsi_action *action,
5454 struct lp_build_tgsi_context *bld_base,
5455 struct lp_build_emit_data *emit_data)
5456 {
5457 struct si_shader_context *ctx = si_shader_context(bld_base);
5458 unsigned stream;
5459
5460 /* Signal primitive cut */
5461 stream = si_llvm_get_stream(bld_base, emit_data);
5462 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
5463 LLVMGetParam(ctx->main_fn, ctx->param_gs_wave_id));
5464 }
5465
5466 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
5467 struct lp_build_tgsi_context *bld_base,
5468 struct lp_build_emit_data *emit_data)
5469 {
5470 struct si_shader_context *ctx = si_shader_context(bld_base);
5471 struct gallivm_state *gallivm = &ctx->gallivm;
5472
5473 /* SI only (thanks to a hw bug workaround):
5474 * The real barrier instruction isn’t needed, because an entire patch
5475 * always fits into a single wave.
5476 */
5477 if (HAVE_LLVM >= 0x0309 &&
5478 ctx->screen->b.chip_class == SI &&
5479 ctx->type == PIPE_SHADER_TESS_CTRL) {
5480 emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
5481 return;
5482 }
5483
5484 lp_build_intrinsic(gallivm->builder,
5485 HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
5486 : "llvm.AMDGPU.barrier.local",
5487 ctx->voidt, NULL, 0, LP_FUNC_ATTR_CONVERGENT);
5488 }
5489
5490 static const struct lp_build_tgsi_action tex_action = {
5491 .fetch_args = tex_fetch_args,
5492 .emit = build_tex_intrinsic,
5493 };
5494
5495 static const struct lp_build_tgsi_action interp_action = {
5496 .fetch_args = interp_fetch_args,
5497 .emit = build_interp_intrinsic,
5498 };
5499
5500 static void si_create_function(struct si_shader_context *ctx,
5501 const char *name,
5502 LLVMTypeRef *returns, unsigned num_returns,
5503 LLVMTypeRef *params, unsigned num_params,
5504 int last_sgpr)
5505 {
5506 int i;
5507
5508 si_llvm_create_func(ctx, name, returns, num_returns,
5509 params, num_params);
5510 si_llvm_shader_type(ctx->main_fn, ctx->type);
5511 ctx->return_value = LLVMGetUndef(ctx->return_type);
5512
5513 for (i = 0; i <= last_sgpr; ++i) {
5514 LLVMValueRef P = LLVMGetParam(ctx->main_fn, i);
5515
5516 /* The combination of:
5517 * - ByVal
5518 * - dereferenceable
5519 * - invariant.load
5520 * allows the optimization passes to move loads and reduces
5521 * SGPR spilling significantly.
5522 */
5523 if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
5524 lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_BYVAL);
5525 lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_NOALIAS);
5526 ac_add_attr_dereferenceable(P, UINT64_MAX);
5527 } else
5528 lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_INREG);
5529 }
5530
5531 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5532 "no-signed-zeros-fp-math",
5533 "true");
5534
5535 if (ctx->screen->b.debug_flags & DBG_UNSAFE_MATH) {
5536 /* These were copied from some LLVM test. */
5537 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5538 "less-precise-fpmad",
5539 "true");
5540 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5541 "no-infs-fp-math",
5542 "true");
5543 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5544 "no-nans-fp-math",
5545 "true");
5546 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5547 "unsafe-fp-math",
5548 "true");
5549 }
5550 }
5551
5552 static void declare_streamout_params(struct si_shader_context *ctx,
5553 struct pipe_stream_output_info *so,
5554 LLVMTypeRef *params, LLVMTypeRef i32,
5555 unsigned *num_params)
5556 {
5557 int i;
5558
5559 /* Streamout SGPRs. */
5560 if (so->num_outputs) {
5561 if (ctx->type != PIPE_SHADER_TESS_EVAL)
5562 params[ctx->param_streamout_config = (*num_params)++] = i32;
5563 else
5564 ctx->param_streamout_config = *num_params - 1;
5565
5566 params[ctx->param_streamout_write_index = (*num_params)++] = i32;
5567 }
5568 /* A streamout buffer offset is loaded if the stride is non-zero. */
5569 for (i = 0; i < 4; i++) {
5570 if (!so->stride[i])
5571 continue;
5572
5573 params[ctx->param_streamout_offset[i] = (*num_params)++] = i32;
5574 }
5575 }
5576
5577 static unsigned llvm_get_type_size(LLVMTypeRef type)
5578 {
5579 LLVMTypeKind kind = LLVMGetTypeKind(type);
5580
5581 switch (kind) {
5582 case LLVMIntegerTypeKind:
5583 return LLVMGetIntTypeWidth(type) / 8;
5584 case LLVMFloatTypeKind:
5585 return 4;
5586 case LLVMPointerTypeKind:
5587 return 8;
5588 case LLVMVectorTypeKind:
5589 return LLVMGetVectorSize(type) *
5590 llvm_get_type_size(LLVMGetElementType(type));
5591 case LLVMArrayTypeKind:
5592 return LLVMGetArrayLength(type) *
5593 llvm_get_type_size(LLVMGetElementType(type));
5594 default:
5595 assert(0);
5596 return 0;
5597 }
5598 }
5599
5600 static void declare_tess_lds(struct si_shader_context *ctx)
5601 {
5602 struct gallivm_state *gallivm = &ctx->gallivm;
5603
5604 unsigned lds_size = ctx->screen->b.chip_class >= CIK ? 65536 : 32768;
5605 ctx->lds = LLVMBuildIntToPtr(gallivm->builder, ctx->i32_0,
5606 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), LOCAL_ADDR_SPACE),
5607 "tess_lds");
5608 }
5609
5610 static unsigned si_get_max_workgroup_size(struct si_shader *shader)
5611 {
5612 const unsigned *properties = shader->selector->info.properties;
5613 unsigned max_work_group_size =
5614 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
5615 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
5616 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
5617
5618 if (!max_work_group_size) {
5619 /* This is a variable group size compute shader,
5620 * compile it for the maximum possible group size.
5621 */
5622 max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
5623 }
5624 return max_work_group_size;
5625 }
5626
5627 static void create_function(struct si_shader_context *ctx)
5628 {
5629 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
5630 struct gallivm_state *gallivm = &ctx->gallivm;
5631 struct si_shader *shader = ctx->shader;
5632 LLVMTypeRef params[SI_NUM_PARAMS + SI_MAX_ATTRIBS], v3i32;
5633 LLVMTypeRef returns[16+32*4];
5634 unsigned i, last_sgpr, num_params = 0, num_return_sgprs;
5635 unsigned num_returns = 0;
5636 unsigned num_prolog_vgprs = 0;
5637
5638 v3i32 = LLVMVectorType(ctx->i32, 3);
5639
5640 params[ctx->param_rw_buffers = num_params++] =
5641 const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
5642 params[ctx->param_const_buffers = num_params++] =
5643 const_array(ctx->v16i8, SI_NUM_CONST_BUFFERS);
5644 params[ctx->param_samplers = num_params++] =
5645 const_array(ctx->v8i32, SI_NUM_SAMPLERS);
5646 params[ctx->param_images = num_params++] =
5647 const_array(ctx->v8i32, SI_NUM_IMAGES);
5648 params[ctx->param_shader_buffers = num_params++] =
5649 const_array(ctx->v4i32, SI_NUM_SHADER_BUFFERS);
5650
5651 switch (ctx->type) {
5652 case PIPE_SHADER_VERTEX:
5653 params[ctx->param_vertex_buffers = num_params++] =
5654 const_array(ctx->v16i8, SI_NUM_VERTEX_BUFFERS);
5655 params[ctx->param_base_vertex = num_params++] = ctx->i32;
5656 params[ctx->param_start_instance = num_params++] = ctx->i32;
5657 params[ctx->param_draw_id = num_params++] = ctx->i32;
5658 params[ctx->param_vs_state_bits = num_params++] = ctx->i32;
5659
5660 if (shader->key.as_es) {
5661 params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
5662 } else if (shader->key.as_ls) {
5663 /* no extra parameters */
5664 } else {
5665 if (shader->is_gs_copy_shader)
5666 num_params = ctx->param_rw_buffers + 1;
5667
5668 /* The locations of the other parameters are assigned dynamically. */
5669 declare_streamout_params(ctx, &shader->selector->so,
5670 params, ctx->i32, &num_params);
5671 }
5672
5673 last_sgpr = num_params-1;
5674
5675 /* VGPRs */
5676 params[ctx->param_vertex_id = num_params++] = ctx->i32;
5677 params[ctx->param_rel_auto_id = num_params++] = ctx->i32;
5678 params[ctx->param_vs_prim_id = num_params++] = ctx->i32;
5679 params[ctx->param_instance_id = num_params++] = ctx->i32;
5680
5681 if (!shader->is_gs_copy_shader) {
5682 /* Vertex load indices. */
5683 ctx->param_vertex_index0 = num_params;
5684
5685 for (i = 0; i < shader->selector->info.num_inputs; i++)
5686 params[num_params++] = ctx->i32;
5687
5688 num_prolog_vgprs += shader->selector->info.num_inputs;
5689
5690 /* PrimitiveID output. */
5691 if (!shader->key.as_es && !shader->key.as_ls)
5692 for (i = 0; i <= VS_EPILOG_PRIMID_LOC; i++)
5693 returns[num_returns++] = ctx->f32;
5694 }
5695 break;
5696
5697 case PIPE_SHADER_TESS_CTRL:
5698 params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
5699 params[ctx->param_tcs_out_lds_offsets = num_params++] = ctx->i32;
5700 params[ctx->param_tcs_out_lds_layout = num_params++] = ctx->i32;
5701 params[ctx->param_vs_state_bits = num_params++] = ctx->i32;
5702 params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
5703 params[ctx->param_tcs_factor_offset = num_params++] = ctx->i32;
5704 last_sgpr = num_params - 1;
5705
5706 /* VGPRs */
5707 params[ctx->param_tcs_patch_id = num_params++] = ctx->i32;
5708 params[ctx->param_tcs_rel_ids = num_params++] = ctx->i32;
5709
5710 /* param_tcs_offchip_offset and param_tcs_factor_offset are
5711 * placed after the user SGPRs.
5712 */
5713 for (i = 0; i < GFX6_TCS_NUM_USER_SGPR + 2; i++)
5714 returns[num_returns++] = ctx->i32; /* SGPRs */
5715
5716 for (i = 0; i < 3; i++)
5717 returns[num_returns++] = ctx->f32; /* VGPRs */
5718 break;
5719
5720 case PIPE_SHADER_TESS_EVAL:
5721 params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
5722
5723 if (shader->key.as_es) {
5724 params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
5725 params[num_params++] = ctx->i32;
5726 params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
5727 } else {
5728 params[num_params++] = ctx->i32;
5729 declare_streamout_params(ctx, &shader->selector->so,
5730 params, ctx->i32, &num_params);
5731 params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
5732 }
5733 last_sgpr = num_params - 1;
5734
5735 /* VGPRs */
5736 params[ctx->param_tes_u = num_params++] = ctx->f32;
5737 params[ctx->param_tes_v = num_params++] = ctx->f32;
5738 params[ctx->param_tes_rel_patch_id = num_params++] = ctx->i32;
5739 params[ctx->param_tes_patch_id = num_params++] = ctx->i32;
5740
5741 /* PrimitiveID output. */
5742 if (!shader->key.as_es)
5743 for (i = 0; i <= VS_EPILOG_PRIMID_LOC; i++)
5744 returns[num_returns++] = ctx->f32;
5745 break;
5746
5747 case PIPE_SHADER_GEOMETRY:
5748 params[ctx->param_gs2vs_offset = num_params++] = ctx->i32;
5749 params[ctx->param_gs_wave_id = num_params++] = ctx->i32;
5750 last_sgpr = num_params - 1;
5751
5752 /* VGPRs */
5753 params[ctx->param_gs_vtx0_offset = num_params++] = ctx->i32;
5754 params[ctx->param_gs_vtx1_offset = num_params++] = ctx->i32;
5755 params[ctx->param_gs_prim_id = num_params++] = ctx->i32;
5756 params[ctx->param_gs_vtx2_offset = num_params++] = ctx->i32;
5757 params[ctx->param_gs_vtx3_offset = num_params++] = ctx->i32;
5758 params[ctx->param_gs_vtx4_offset = num_params++] = ctx->i32;
5759 params[ctx->param_gs_vtx5_offset = num_params++] = ctx->i32;
5760 params[ctx->param_gs_instance_id = num_params++] = ctx->i32;
5761 break;
5762
5763 case PIPE_SHADER_FRAGMENT:
5764 params[SI_PARAM_ALPHA_REF] = ctx->f32;
5765 params[SI_PARAM_PRIM_MASK] = ctx->i32;
5766 last_sgpr = SI_PARAM_PRIM_MASK;
5767 params[SI_PARAM_PERSP_SAMPLE] = ctx->v2i32;
5768 params[SI_PARAM_PERSP_CENTER] = ctx->v2i32;
5769 params[SI_PARAM_PERSP_CENTROID] = ctx->v2i32;
5770 params[SI_PARAM_PERSP_PULL_MODEL] = v3i32;
5771 params[SI_PARAM_LINEAR_SAMPLE] = ctx->v2i32;
5772 params[SI_PARAM_LINEAR_CENTER] = ctx->v2i32;
5773 params[SI_PARAM_LINEAR_CENTROID] = ctx->v2i32;
5774 params[SI_PARAM_LINE_STIPPLE_TEX] = ctx->f32;
5775 params[SI_PARAM_POS_X_FLOAT] = ctx->f32;
5776 params[SI_PARAM_POS_Y_FLOAT] = ctx->f32;
5777 params[SI_PARAM_POS_Z_FLOAT] = ctx->f32;
5778 params[SI_PARAM_POS_W_FLOAT] = ctx->f32;
5779 params[SI_PARAM_FRONT_FACE] = ctx->i32;
5780 shader->info.face_vgpr_index = 20;
5781 params[SI_PARAM_ANCILLARY] = ctx->i32;
5782 params[SI_PARAM_SAMPLE_COVERAGE] = ctx->f32;
5783 params[SI_PARAM_POS_FIXED_PT] = ctx->i32;
5784 num_params = SI_PARAM_POS_FIXED_PT+1;
5785
5786 /* Color inputs from the prolog. */
5787 if (shader->selector->info.colors_read) {
5788 unsigned num_color_elements =
5789 util_bitcount(shader->selector->info.colors_read);
5790
5791 assert(num_params + num_color_elements <= ARRAY_SIZE(params));
5792 for (i = 0; i < num_color_elements; i++)
5793 params[num_params++] = ctx->f32;
5794
5795 num_prolog_vgprs += num_color_elements;
5796 }
5797
5798 /* Outputs for the epilog. */
5799 num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
5800 num_returns =
5801 num_return_sgprs +
5802 util_bitcount(shader->selector->info.colors_written) * 4 +
5803 shader->selector->info.writes_z +
5804 shader->selector->info.writes_stencil +
5805 shader->selector->info.writes_samplemask +
5806 1 /* SampleMaskIn */;
5807
5808 num_returns = MAX2(num_returns,
5809 num_return_sgprs +
5810 PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
5811
5812 for (i = 0; i < num_return_sgprs; i++)
5813 returns[i] = ctx->i32;
5814 for (; i < num_returns; i++)
5815 returns[i] = ctx->f32;
5816 break;
5817
5818 case PIPE_SHADER_COMPUTE:
5819 params[SI_PARAM_GRID_SIZE] = v3i32;
5820 params[SI_PARAM_BLOCK_SIZE] = v3i32;
5821 params[SI_PARAM_BLOCK_ID] = v3i32;
5822 last_sgpr = SI_PARAM_BLOCK_ID;
5823
5824 params[SI_PARAM_THREAD_ID] = v3i32;
5825 num_params = SI_PARAM_THREAD_ID + 1;
5826 break;
5827 default:
5828 assert(0 && "unimplemented shader");
5829 return;
5830 }
5831
5832 assert(num_params <= ARRAY_SIZE(params));
5833
5834 si_create_function(ctx, "main", returns, num_returns, params,
5835 num_params, last_sgpr);
5836
5837 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5838 if (ctx->type == PIPE_SHADER_FRAGMENT &&
5839 ctx->separate_prolog) {
5840 si_llvm_add_attribute(ctx->main_fn,
5841 "InitialPSInputAddr",
5842 S_0286D0_PERSP_SAMPLE_ENA(1) |
5843 S_0286D0_PERSP_CENTER_ENA(1) |
5844 S_0286D0_PERSP_CENTROID_ENA(1) |
5845 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5846 S_0286D0_LINEAR_CENTER_ENA(1) |
5847 S_0286D0_LINEAR_CENTROID_ENA(1) |
5848 S_0286D0_FRONT_FACE_ENA(1) |
5849 S_0286D0_POS_FIXED_PT_ENA(1));
5850 } else if (ctx->type == PIPE_SHADER_COMPUTE) {
5851 si_llvm_add_attribute(ctx->main_fn,
5852 "amdgpu-max-work-group-size",
5853 si_get_max_workgroup_size(shader));
5854 }
5855
5856 shader->info.num_input_sgprs = 0;
5857 shader->info.num_input_vgprs = 0;
5858
5859 for (i = 0; i <= last_sgpr; ++i)
5860 shader->info.num_input_sgprs += llvm_get_type_size(params[i]) / 4;
5861
5862 for (; i < num_params; ++i)
5863 shader->info.num_input_vgprs += llvm_get_type_size(params[i]) / 4;
5864
5865 assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
5866 shader->info.num_input_vgprs -= num_prolog_vgprs;
5867
5868 if (!ctx->screen->has_ds_bpermute &&
5869 bld_base->info &&
5870 (bld_base->info->opcode_count[TGSI_OPCODE_DDX] > 0 ||
5871 bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0 ||
5872 bld_base->info->opcode_count[TGSI_OPCODE_DDX_FINE] > 0 ||
5873 bld_base->info->opcode_count[TGSI_OPCODE_DDY_FINE] > 0 ||
5874 bld_base->info->opcode_count[TGSI_OPCODE_INTERP_OFFSET] > 0 ||
5875 bld_base->info->opcode_count[TGSI_OPCODE_INTERP_SAMPLE] > 0))
5876 ctx->lds =
5877 LLVMAddGlobalInAddressSpace(gallivm->module,
5878 LLVMArrayType(ctx->i32, 64),
5879 "ddxy_lds",
5880 LOCAL_ADDR_SPACE);
5881
5882 if ((ctx->type == PIPE_SHADER_VERTEX && shader->key.as_ls) ||
5883 ctx->type == PIPE_SHADER_TESS_CTRL)
5884 declare_tess_lds(ctx);
5885 }
5886
5887 /**
5888 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5889 * for later use.
5890 */
5891 static void preload_ring_buffers(struct si_shader_context *ctx)
5892 {
5893 struct gallivm_state *gallivm = &ctx->gallivm;
5894 LLVMBuilderRef builder = gallivm->builder;
5895
5896 LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
5897 ctx->param_rw_buffers);
5898
5899 if ((ctx->type == PIPE_SHADER_VERTEX &&
5900 ctx->shader->key.as_es) ||
5901 (ctx->type == PIPE_SHADER_TESS_EVAL &&
5902 ctx->shader->key.as_es) ||
5903 ctx->type == PIPE_SHADER_GEOMETRY) {
5904 unsigned ring =
5905 ctx->type == PIPE_SHADER_GEOMETRY ? SI_GS_RING_ESGS
5906 : SI_ES_RING_ESGS;
5907 LLVMValueRef offset = LLVMConstInt(ctx->i32, ring, 0);
5908
5909 ctx->esgs_ring =
5910 ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
5911 }
5912
5913 if (ctx->shader->is_gs_copy_shader) {
5914 LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
5915
5916 ctx->gsvs_ring[0] =
5917 ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
5918 } else if (ctx->type == PIPE_SHADER_GEOMETRY) {
5919 const struct si_shader_selector *sel = ctx->shader->selector;
5920 LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
5921 LLVMValueRef base_ring;
5922
5923 base_ring = ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
5924
5925 /* The conceptual layout of the GSVS ring is
5926 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5927 * but the real memory layout is swizzled across
5928 * threads:
5929 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5930 * t16v0c0 ..
5931 * Override the buffer descriptor accordingly.
5932 */
5933 LLVMTypeRef v2i64 = LLVMVectorType(ctx->i64, 2);
5934 uint64_t stream_offset = 0;
5935
5936 for (unsigned stream = 0; stream < 4; ++stream) {
5937 unsigned num_components;
5938 unsigned stride;
5939 unsigned num_records;
5940 LLVMValueRef ring, tmp;
5941
5942 num_components = sel->info.num_stream_output_components[stream];
5943 if (!num_components)
5944 continue;
5945
5946 stride = 4 * num_components * sel->gs_max_out_vertices;
5947
5948 /* Limit on the stride field for <= CIK. */
5949 assert(stride < (1 << 14));
5950
5951 num_records = 64;
5952
5953 ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
5954 tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_0, "");
5955 tmp = LLVMBuildAdd(builder, tmp,
5956 LLVMConstInt(ctx->i64,
5957 stream_offset, 0), "");
5958 stream_offset += stride * 64;
5959
5960 ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_0, "");
5961 ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
5962 tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_1, "");
5963 tmp = LLVMBuildOr(builder, tmp,
5964 LLVMConstInt(ctx->i32,
5965 S_008F04_STRIDE(stride) |
5966 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5967 ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_1, "");
5968 ring = LLVMBuildInsertElement(builder, ring,
5969 LLVMConstInt(ctx->i32, num_records, 0),
5970 LLVMConstInt(ctx->i32, 2, 0), "");
5971 ring = LLVMBuildInsertElement(builder, ring,
5972 LLVMConstInt(ctx->i32,
5973 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
5974 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
5975 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
5976 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
5977 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
5978 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
5979 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
5980 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5981 S_008F0C_ADD_TID_ENABLE(1),
5982 0),
5983 LLVMConstInt(ctx->i32, 3, 0), "");
5984 ring = LLVMBuildBitCast(builder, ring, ctx->v16i8, "");
5985
5986 ctx->gsvs_ring[stream] = ring;
5987 }
5988 }
5989 }
5990
5991 static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
5992 LLVMValueRef param_rw_buffers,
5993 unsigned param_pos_fixed_pt)
5994 {
5995 struct gallivm_state *gallivm = &ctx->gallivm;
5996 LLVMBuilderRef builder = gallivm->builder;
5997 LLVMValueRef slot, desc, offset, row, bit, address[2];
5998
5999 /* Use the fixed-point gl_FragCoord input.
6000 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
6001 * per coordinate to get the repeating effect.
6002 */
6003 address[0] = unpack_param(ctx, param_pos_fixed_pt, 0, 5);
6004 address[1] = unpack_param(ctx, param_pos_fixed_pt, 16, 5);
6005
6006 /* Load the buffer descriptor. */
6007 slot = LLVMConstInt(ctx->i32, SI_PS_CONST_POLY_STIPPLE, 0);
6008 desc = ac_build_indexed_load_const(&ctx->ac, param_rw_buffers, slot);
6009
6010 /* The stipple pattern is 32x32, each row has 32 bits. */
6011 offset = LLVMBuildMul(builder, address[1],
6012 LLVMConstInt(ctx->i32, 4, 0), "");
6013 row = buffer_load_const(ctx, desc, offset);
6014 row = LLVMBuildBitCast(builder, row, ctx->i32, "");
6015 bit = LLVMBuildLShr(builder, row, address[0], "");
6016 bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
6017
6018 /* The intrinsic kills the thread if arg < 0. */
6019 bit = LLVMBuildSelect(builder, bit, LLVMConstReal(ctx->f32, 0),
6020 LLVMConstReal(ctx->f32, -1), "");
6021 ac_build_kill(&ctx->ac, bit);
6022 }
6023
6024 void si_shader_binary_read_config(struct ac_shader_binary *binary,
6025 struct si_shader_config *conf,
6026 unsigned symbol_offset)
6027 {
6028 unsigned i;
6029 const unsigned char *config =
6030 ac_shader_binary_config_start(binary, symbol_offset);
6031 bool really_needs_scratch = false;
6032
6033 /* LLVM adds SGPR spills to the scratch size.
6034 * Find out if we really need the scratch buffer.
6035 */
6036 for (i = 0; i < binary->reloc_count; i++) {
6037 const struct ac_shader_reloc *reloc = &binary->relocs[i];
6038
6039 if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name) ||
6040 !strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
6041 really_needs_scratch = true;
6042 break;
6043 }
6044 }
6045
6046 /* XXX: We may be able to emit some of these values directly rather than
6047 * extracting fields to be emitted later.
6048 */
6049
6050 for (i = 0; i < binary->config_size_per_symbol; i+= 8) {
6051 unsigned reg = util_le32_to_cpu(*(uint32_t*)(config + i));
6052 unsigned value = util_le32_to_cpu(*(uint32_t*)(config + i + 4));
6053 switch (reg) {
6054 case R_00B028_SPI_SHADER_PGM_RSRC1_PS:
6055 case R_00B128_SPI_SHADER_PGM_RSRC1_VS:
6056 case R_00B228_SPI_SHADER_PGM_RSRC1_GS:
6057 case R_00B848_COMPUTE_PGM_RSRC1:
6058 conf->num_sgprs = MAX2(conf->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
6059 conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
6060 conf->float_mode = G_00B028_FLOAT_MODE(value);
6061 conf->rsrc1 = value;
6062 break;
6063 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
6064 conf->lds_size = MAX2(conf->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
6065 break;
6066 case R_00B84C_COMPUTE_PGM_RSRC2:
6067 conf->lds_size = MAX2(conf->lds_size, G_00B84C_LDS_SIZE(value));
6068 conf->rsrc2 = value;
6069 break;
6070 case R_0286CC_SPI_PS_INPUT_ENA:
6071 conf->spi_ps_input_ena = value;
6072 break;
6073 case R_0286D0_SPI_PS_INPUT_ADDR:
6074 conf->spi_ps_input_addr = value;
6075 break;
6076 case R_0286E8_SPI_TMPRING_SIZE:
6077 case R_00B860_COMPUTE_TMPRING_SIZE:
6078 /* WAVESIZE is in units of 256 dwords. */
6079 if (really_needs_scratch)
6080 conf->scratch_bytes_per_wave =
6081 G_00B860_WAVESIZE(value) * 256 * 4;
6082 break;
6083 case 0x4: /* SPILLED_SGPRS */
6084 conf->spilled_sgprs = value;
6085 break;
6086 case 0x8: /* SPILLED_VGPRS */
6087 conf->spilled_vgprs = value;
6088 break;
6089 default:
6090 {
6091 static bool printed;
6092
6093 if (!printed) {
6094 fprintf(stderr, "Warning: LLVM emitted unknown "
6095 "config register: 0x%x\n", reg);
6096 printed = true;
6097 }
6098 }
6099 break;
6100 }
6101 }
6102
6103 if (!conf->spi_ps_input_addr)
6104 conf->spi_ps_input_addr = conf->spi_ps_input_ena;
6105 }
6106
6107 void si_shader_apply_scratch_relocs(struct si_context *sctx,
6108 struct si_shader *shader,
6109 struct si_shader_config *config,
6110 uint64_t scratch_va)
6111 {
6112 unsigned i;
6113 uint32_t scratch_rsrc_dword0 = scratch_va;
6114 uint32_t scratch_rsrc_dword1 =
6115 S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
6116
6117 /* Enable scratch coalescing if LLVM sets ELEMENT_SIZE & INDEX_STRIDE
6118 * correctly.
6119 */
6120 if (HAVE_LLVM >= 0x0309)
6121 scratch_rsrc_dword1 |= S_008F04_SWIZZLE_ENABLE(1);
6122 else
6123 scratch_rsrc_dword1 |=
6124 S_008F04_STRIDE(config->scratch_bytes_per_wave / 64);
6125
6126 for (i = 0 ; i < shader->binary.reloc_count; i++) {
6127 const struct ac_shader_reloc *reloc =
6128 &shader->binary.relocs[i];
6129 if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name)) {
6130 util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset,
6131 &scratch_rsrc_dword0, 4);
6132 } else if (!strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
6133 util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset,
6134 &scratch_rsrc_dword1, 4);
6135 }
6136 }
6137 }
6138
6139 static unsigned si_get_shader_binary_size(struct si_shader *shader)
6140 {
6141 unsigned size = shader->binary.code_size;
6142
6143 if (shader->prolog)
6144 size += shader->prolog->binary.code_size;
6145 if (shader->previous_stage)
6146 size += shader->previous_stage->binary.code_size;
6147 if (shader->epilog)
6148 size += shader->epilog->binary.code_size;
6149 return size;
6150 }
6151
6152 int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader)
6153 {
6154 const struct ac_shader_binary *prolog =
6155 shader->prolog ? &shader->prolog->binary : NULL;
6156 const struct ac_shader_binary *previous_stage =
6157 shader->previous_stage ? &shader->previous_stage->binary : NULL;
6158 const struct ac_shader_binary *epilog =
6159 shader->epilog ? &shader->epilog->binary : NULL;
6160 const struct ac_shader_binary *mainb = &shader->binary;
6161 unsigned bo_size = si_get_shader_binary_size(shader) +
6162 (!epilog ? mainb->rodata_size : 0);
6163 unsigned char *ptr;
6164
6165 assert(!prolog || !prolog->rodata_size);
6166 assert(!previous_stage || !previous_stage->rodata_size);
6167 assert((!prolog && !previous_stage && !epilog) || !mainb->rodata_size);
6168 assert(!epilog || !epilog->rodata_size);
6169
6170 /* GFX9 can fetch at most 128 bytes past the end of the shader.
6171 * Prevent VM faults.
6172 */
6173 if (sscreen->b.chip_class >= GFX9)
6174 bo_size += 128;
6175
6176 r600_resource_reference(&shader->bo, NULL);
6177 shader->bo = (struct r600_resource*)
6178 pipe_buffer_create(&sscreen->b.b, 0,
6179 PIPE_USAGE_IMMUTABLE,
6180 align(bo_size, SI_CPDMA_ALIGNMENT));
6181 if (!shader->bo)
6182 return -ENOMEM;
6183
6184 /* Upload. */
6185 ptr = sscreen->b.ws->buffer_map(shader->bo->buf, NULL,
6186 PIPE_TRANSFER_READ_WRITE |
6187 PIPE_TRANSFER_UNSYNCHRONIZED);
6188
6189 if (prolog) {
6190 util_memcpy_cpu_to_le32(ptr, prolog->code, prolog->code_size);
6191 ptr += prolog->code_size;
6192 }
6193 if (previous_stage) {
6194 util_memcpy_cpu_to_le32(ptr, previous_stage->code,
6195 previous_stage->code_size);
6196 ptr += previous_stage->code_size;
6197 }
6198
6199 util_memcpy_cpu_to_le32(ptr, mainb->code, mainb->code_size);
6200 ptr += mainb->code_size;
6201
6202 if (epilog)
6203 util_memcpy_cpu_to_le32(ptr, epilog->code, epilog->code_size);
6204 else if (mainb->rodata_size > 0)
6205 util_memcpy_cpu_to_le32(ptr, mainb->rodata, mainb->rodata_size);
6206
6207 sscreen->b.ws->buffer_unmap(shader->bo->buf);
6208 return 0;
6209 }
6210
6211 static void si_shader_dump_disassembly(const struct ac_shader_binary *binary,
6212 struct pipe_debug_callback *debug,
6213 const char *name, FILE *file)
6214 {
6215 char *line, *p;
6216 unsigned i, count;
6217
6218 if (binary->disasm_string) {
6219 fprintf(file, "Shader %s disassembly:\n", name);
6220 fprintf(file, "%s", binary->disasm_string);
6221
6222 if (debug && debug->debug_message) {
6223 /* Very long debug messages are cut off, so send the
6224 * disassembly one line at a time. This causes more
6225 * overhead, but on the plus side it simplifies
6226 * parsing of resulting logs.
6227 */
6228 pipe_debug_message(debug, SHADER_INFO,
6229 "Shader Disassembly Begin");
6230
6231 line = binary->disasm_string;
6232 while (*line) {
6233 p = util_strchrnul(line, '\n');
6234 count = p - line;
6235
6236 if (count) {
6237 pipe_debug_message(debug, SHADER_INFO,
6238 "%.*s", count, line);
6239 }
6240
6241 if (!*p)
6242 break;
6243 line = p + 1;
6244 }
6245
6246 pipe_debug_message(debug, SHADER_INFO,
6247 "Shader Disassembly End");
6248 }
6249 } else {
6250 fprintf(file, "Shader %s binary:\n", name);
6251 for (i = 0; i < binary->code_size; i += 4) {
6252 fprintf(file, "@0x%x: %02x%02x%02x%02x\n", i,
6253 binary->code[i + 3], binary->code[i + 2],
6254 binary->code[i + 1], binary->code[i]);
6255 }
6256 }
6257 }
6258
6259 static void si_shader_dump_stats(struct si_screen *sscreen,
6260 struct si_shader *shader,
6261 struct pipe_debug_callback *debug,
6262 unsigned processor,
6263 FILE *file,
6264 bool check_debug_option)
6265 {
6266 struct si_shader_config *conf = &shader->config;
6267 unsigned num_inputs = shader->selector ? shader->selector->info.num_inputs : 0;
6268 unsigned code_size = si_get_shader_binary_size(shader);
6269 unsigned lds_increment = sscreen->b.chip_class >= CIK ? 512 : 256;
6270 unsigned lds_per_wave = 0;
6271 unsigned max_simd_waves = 10;
6272
6273 /* Compute LDS usage for PS. */
6274 switch (processor) {
6275 case PIPE_SHADER_FRAGMENT:
6276 /* The minimum usage per wave is (num_inputs * 48). The maximum
6277 * usage is (num_inputs * 48 * 16).
6278 * We can get anything in between and it varies between waves.
6279 *
6280 * The 48 bytes per input for a single primitive is equal to
6281 * 4 bytes/component * 4 components/input * 3 points.
6282 *
6283 * Other stages don't know the size at compile time or don't
6284 * allocate LDS per wave, but instead they do it per thread group.
6285 */
6286 lds_per_wave = conf->lds_size * lds_increment +
6287 align(num_inputs * 48, lds_increment);
6288 break;
6289 case PIPE_SHADER_COMPUTE:
6290 if (shader->selector) {
6291 unsigned max_workgroup_size =
6292 si_get_max_workgroup_size(shader);
6293 lds_per_wave = (conf->lds_size * lds_increment) /
6294 DIV_ROUND_UP(max_workgroup_size, 64);
6295 }
6296 break;
6297 }
6298
6299 /* Compute the per-SIMD wave counts. */
6300 if (conf->num_sgprs) {
6301 if (sscreen->b.chip_class >= VI)
6302 max_simd_waves = MIN2(max_simd_waves, 800 / conf->num_sgprs);
6303 else
6304 max_simd_waves = MIN2(max_simd_waves, 512 / conf->num_sgprs);
6305 }
6306
6307 if (conf->num_vgprs)
6308 max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
6309
6310 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
6311 * 16KB makes some SIMDs unoccupied). */
6312 if (lds_per_wave)
6313 max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
6314
6315 if (!check_debug_option ||
6316 r600_can_dump_shader(&sscreen->b, processor)) {
6317 if (processor == PIPE_SHADER_FRAGMENT) {
6318 fprintf(file, "*** SHADER CONFIG ***\n"
6319 "SPI_PS_INPUT_ADDR = 0x%04x\n"
6320 "SPI_PS_INPUT_ENA = 0x%04x\n",
6321 conf->spi_ps_input_addr, conf->spi_ps_input_ena);
6322 }
6323
6324 fprintf(file, "*** SHADER STATS ***\n"
6325 "SGPRS: %d\n"
6326 "VGPRS: %d\n"
6327 "Spilled SGPRs: %d\n"
6328 "Spilled VGPRs: %d\n"
6329 "Private memory VGPRs: %d\n"
6330 "Code Size: %d bytes\n"
6331 "LDS: %d blocks\n"
6332 "Scratch: %d bytes per wave\n"
6333 "Max Waves: %d\n"
6334 "********************\n\n\n",
6335 conf->num_sgprs, conf->num_vgprs,
6336 conf->spilled_sgprs, conf->spilled_vgprs,
6337 conf->private_mem_vgprs, code_size,
6338 conf->lds_size, conf->scratch_bytes_per_wave,
6339 max_simd_waves);
6340 }
6341
6342 pipe_debug_message(debug, SHADER_INFO,
6343 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
6344 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
6345 "Spilled VGPRs: %d PrivMem VGPRs: %d",
6346 conf->num_sgprs, conf->num_vgprs, code_size,
6347 conf->lds_size, conf->scratch_bytes_per_wave,
6348 max_simd_waves, conf->spilled_sgprs,
6349 conf->spilled_vgprs, conf->private_mem_vgprs);
6350 }
6351
6352 const char *si_get_shader_name(struct si_shader *shader, unsigned processor)
6353 {
6354 switch (processor) {
6355 case PIPE_SHADER_VERTEX:
6356 if (shader->key.as_es)
6357 return "Vertex Shader as ES";
6358 else if (shader->key.as_ls)
6359 return "Vertex Shader as LS";
6360 else
6361 return "Vertex Shader as VS";
6362 case PIPE_SHADER_TESS_CTRL:
6363 return "Tessellation Control Shader";
6364 case PIPE_SHADER_TESS_EVAL:
6365 if (shader->key.as_es)
6366 return "Tessellation Evaluation Shader as ES";
6367 else
6368 return "Tessellation Evaluation Shader as VS";
6369 case PIPE_SHADER_GEOMETRY:
6370 if (shader->is_gs_copy_shader)
6371 return "GS Copy Shader as VS";
6372 else
6373 return "Geometry Shader";
6374 case PIPE_SHADER_FRAGMENT:
6375 return "Pixel Shader";
6376 case PIPE_SHADER_COMPUTE:
6377 return "Compute Shader";
6378 default:
6379 return "Unknown Shader";
6380 }
6381 }
6382
6383 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
6384 struct pipe_debug_callback *debug, unsigned processor,
6385 FILE *file, bool check_debug_option)
6386 {
6387 if (!check_debug_option ||
6388 r600_can_dump_shader(&sscreen->b, processor))
6389 si_dump_shader_key(processor, shader, file);
6390
6391 if (!check_debug_option && shader->binary.llvm_ir_string) {
6392 fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
6393 si_get_shader_name(shader, processor));
6394 fprintf(file, "%s\n", shader->binary.llvm_ir_string);
6395 }
6396
6397 if (!check_debug_option ||
6398 (r600_can_dump_shader(&sscreen->b, processor) &&
6399 !(sscreen->b.debug_flags & DBG_NO_ASM))) {
6400 fprintf(file, "\n%s:\n", si_get_shader_name(shader, processor));
6401
6402 if (shader->prolog)
6403 si_shader_dump_disassembly(&shader->prolog->binary,
6404 debug, "prolog", file);
6405 if (shader->previous_stage)
6406 si_shader_dump_disassembly(&shader->previous_stage->binary,
6407 debug, "previous stage", file);
6408
6409 si_shader_dump_disassembly(&shader->binary, debug, "main", file);
6410
6411 if (shader->epilog)
6412 si_shader_dump_disassembly(&shader->epilog->binary,
6413 debug, "epilog", file);
6414 fprintf(file, "\n");
6415 }
6416
6417 si_shader_dump_stats(sscreen, shader, debug, processor, file,
6418 check_debug_option);
6419 }
6420
6421 int si_compile_llvm(struct si_screen *sscreen,
6422 struct ac_shader_binary *binary,
6423 struct si_shader_config *conf,
6424 LLVMTargetMachineRef tm,
6425 LLVMModuleRef mod,
6426 struct pipe_debug_callback *debug,
6427 unsigned processor,
6428 const char *name)
6429 {
6430 int r = 0;
6431 unsigned count = p_atomic_inc_return(&sscreen->b.num_compilations);
6432
6433 if (r600_can_dump_shader(&sscreen->b, processor)) {
6434 fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
6435
6436 if (!(sscreen->b.debug_flags & (DBG_NO_IR | DBG_PREOPT_IR))) {
6437 fprintf(stderr, "%s LLVM IR:\n\n", name);
6438 ac_dump_module(mod);
6439 fprintf(stderr, "\n");
6440 }
6441 }
6442
6443 if (sscreen->record_llvm_ir) {
6444 char *ir = LLVMPrintModuleToString(mod);
6445 binary->llvm_ir_string = strdup(ir);
6446 LLVMDisposeMessage(ir);
6447 }
6448
6449 if (!si_replace_shader(count, binary)) {
6450 r = si_llvm_compile(mod, binary, tm, debug);
6451 if (r)
6452 return r;
6453 }
6454
6455 si_shader_binary_read_config(binary, conf, 0);
6456
6457 /* Enable 64-bit and 16-bit denormals, because there is no performance
6458 * cost.
6459 *
6460 * If denormals are enabled, all floating-point output modifiers are
6461 * ignored.
6462 *
6463 * Don't enable denormals for 32-bit floats, because:
6464 * - Floating-point output modifiers would be ignored by the hw.
6465 * - Some opcodes don't support denormals, such as v_mad_f32. We would
6466 * have to stop using those.
6467 * - SI & CI would be very slow.
6468 */
6469 conf->float_mode |= V_00B028_FP_64_DENORMS;
6470
6471 FREE(binary->config);
6472 FREE(binary->global_symbol_offsets);
6473 binary->config = NULL;
6474 binary->global_symbol_offsets = NULL;
6475
6476 /* Some shaders can't have rodata because their binaries can be
6477 * concatenated.
6478 */
6479 if (binary->rodata_size &&
6480 (processor == PIPE_SHADER_VERTEX ||
6481 processor == PIPE_SHADER_TESS_CTRL ||
6482 processor == PIPE_SHADER_TESS_EVAL ||
6483 processor == PIPE_SHADER_FRAGMENT)) {
6484 fprintf(stderr, "radeonsi: The shader can't have rodata.");
6485 return -EINVAL;
6486 }
6487
6488 return r;
6489 }
6490
6491 static void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
6492 {
6493 if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
6494 LLVMBuildRetVoid(ctx->gallivm.builder);
6495 else
6496 LLVMBuildRet(ctx->gallivm.builder, ret);
6497 }
6498
6499 /* Generate code for the hardware VS shader stage to go with a geometry shader */
6500 struct si_shader *
6501 si_generate_gs_copy_shader(struct si_screen *sscreen,
6502 LLVMTargetMachineRef tm,
6503 struct si_shader_selector *gs_selector,
6504 struct pipe_debug_callback *debug)
6505 {
6506 struct si_shader_context ctx;
6507 struct si_shader *shader;
6508 struct gallivm_state *gallivm = &ctx.gallivm;
6509 LLVMBuilderRef builder;
6510 struct lp_build_tgsi_context *bld_base = &ctx.bld_base;
6511 struct lp_build_context *uint = &bld_base->uint_bld;
6512 struct si_shader_output_values *outputs;
6513 struct tgsi_shader_info *gsinfo = &gs_selector->info;
6514 int i, r;
6515
6516 outputs = MALLOC(gsinfo->num_outputs * sizeof(outputs[0]));
6517
6518 if (!outputs)
6519 return NULL;
6520
6521 shader = CALLOC_STRUCT(si_shader);
6522 if (!shader) {
6523 FREE(outputs);
6524 return NULL;
6525 }
6526
6527
6528 shader->selector = gs_selector;
6529 shader->is_gs_copy_shader = true;
6530
6531 si_init_shader_ctx(&ctx, sscreen, tm);
6532 ctx.shader = shader;
6533 ctx.type = PIPE_SHADER_VERTEX;
6534
6535 builder = gallivm->builder;
6536
6537 create_function(&ctx);
6538 preload_ring_buffers(&ctx);
6539
6540 LLVMValueRef voffset =
6541 lp_build_mul_imm(uint, LLVMGetParam(ctx.main_fn,
6542 ctx.param_vertex_id), 4);
6543
6544 /* Fetch the vertex stream ID.*/
6545 LLVMValueRef stream_id;
6546
6547 if (gs_selector->so.num_outputs)
6548 stream_id = unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
6549 else
6550 stream_id = ctx.i32_0;
6551
6552 /* Fill in output information. */
6553 for (i = 0; i < gsinfo->num_outputs; ++i) {
6554 outputs[i].semantic_name = gsinfo->output_semantic_name[i];
6555 outputs[i].semantic_index = gsinfo->output_semantic_index[i];
6556
6557 for (int chan = 0; chan < 4; chan++) {
6558 outputs[i].vertex_stream[chan] =
6559 (gsinfo->output_streams[i] >> (2 * chan)) & 3;
6560 }
6561 }
6562
6563 LLVMBasicBlockRef end_bb;
6564 LLVMValueRef switch_inst;
6565
6566 end_bb = LLVMAppendBasicBlockInContext(gallivm->context, ctx.main_fn, "end");
6567 switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);
6568
6569 for (int stream = 0; stream < 4; stream++) {
6570 LLVMBasicBlockRef bb;
6571 unsigned offset;
6572
6573 if (!gsinfo->num_stream_output_components[stream])
6574 continue;
6575
6576 if (stream > 0 && !gs_selector->so.num_outputs)
6577 continue;
6578
6579 bb = LLVMInsertBasicBlockInContext(gallivm->context, end_bb, "out");
6580 LLVMAddCase(switch_inst, LLVMConstInt(ctx.i32, stream, 0), bb);
6581 LLVMPositionBuilderAtEnd(builder, bb);
6582
6583 /* Fetch vertex data from GSVS ring */
6584 offset = 0;
6585 for (i = 0; i < gsinfo->num_outputs; ++i) {
6586 for (unsigned chan = 0; chan < 4; chan++) {
6587 if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
6588 outputs[i].vertex_stream[chan] != stream) {
6589 outputs[i].values[chan] = ctx.bld_base.base.undef;
6590 continue;
6591 }
6592
6593 LLVMValueRef soffset = LLVMConstInt(ctx.i32,
6594 offset * gs_selector->gs_max_out_vertices * 16 * 4, 0);
6595 offset++;
6596
6597 outputs[i].values[chan] =
6598 ac_build_buffer_load(&ctx.ac,
6599 ctx.gsvs_ring[0], 1,
6600 ctx.i32_0, voffset,
6601 soffset, 0, 1, 1, true);
6602 }
6603 }
6604
6605 /* Streamout and exports. */
6606 if (gs_selector->so.num_outputs) {
6607 si_llvm_emit_streamout(&ctx, outputs,
6608 gsinfo->num_outputs,
6609 stream);
6610 }
6611
6612 if (stream == 0)
6613 si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
6614
6615 LLVMBuildBr(builder, end_bb);
6616 }
6617
6618 LLVMPositionBuilderAtEnd(builder, end_bb);
6619
6620 LLVMBuildRetVoid(gallivm->builder);
6621
6622 /* Dump LLVM IR before any optimization passes */
6623 if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
6624 r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
6625 ac_dump_module(ctx.gallivm.module);
6626
6627 si_llvm_finalize_module(&ctx,
6628 r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_GEOMETRY));
6629
6630 r = si_compile_llvm(sscreen, &ctx.shader->binary,
6631 &ctx.shader->config, ctx.tm,
6632 ctx.gallivm.module,
6633 debug, PIPE_SHADER_GEOMETRY,
6634 "GS Copy Shader");
6635 if (!r) {
6636 if (r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
6637 fprintf(stderr, "GS Copy Shader:\n");
6638 si_shader_dump(sscreen, ctx.shader, debug,
6639 PIPE_SHADER_GEOMETRY, stderr, true);
6640 r = si_shader_binary_upload(sscreen, ctx.shader);
6641 }
6642
6643 si_llvm_dispose(&ctx);
6644
6645 FREE(outputs);
6646
6647 if (r != 0) {
6648 FREE(shader);
6649 shader = NULL;
6650 }
6651 return shader;
6652 }
6653
6654 static void si_dump_shader_key_vs(struct si_shader_key *key,
6655 struct si_vs_prolog_bits *prolog,
6656 const char *prefix, FILE *f)
6657 {
6658 fprintf(f, " %s.instance_divisors = {", prefix);
6659 for (int i = 0; i < ARRAY_SIZE(prolog->instance_divisors); i++) {
6660 fprintf(f, !i ? "%u" : ", %u",
6661 prolog->instance_divisors[i]);
6662 }
6663 fprintf(f, "}\n");
6664
6665 fprintf(f, " mono.vs.fix_fetch = {");
6666 for (int i = 0; i < SI_MAX_ATTRIBS; i++)
6667 fprintf(f, !i ? "%u" : ", %u", key->mono.vs_fix_fetch[i]);
6668 fprintf(f, "}\n");
6669 }
6670
6671 static void si_dump_shader_key(unsigned processor, struct si_shader *shader,
6672 FILE *f)
6673 {
6674 struct si_shader_key *key = &shader->key;
6675
6676 fprintf(f, "SHADER KEY\n");
6677
6678 switch (processor) {
6679 case PIPE_SHADER_VERTEX:
6680 si_dump_shader_key_vs(key, &key->part.vs.prolog,
6681 "part.vs.prolog", f);
6682 fprintf(f, " as_es = %u\n", key->as_es);
6683 fprintf(f, " as_ls = %u\n", key->as_ls);
6684 fprintf(f, " part.vs.epilog.export_prim_id = %u\n",
6685 key->part.vs.epilog.export_prim_id);
6686 break;
6687
6688 case PIPE_SHADER_TESS_CTRL:
6689 if (shader->selector->screen->b.chip_class >= GFX9) {
6690 si_dump_shader_key_vs(key, &key->part.tcs.ls_prolog,
6691 "part.tcs.ls_prolog", f);
6692 }
6693 fprintf(f, " part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
6694 fprintf(f, " mono.ff_tcs_inputs_to_copy = 0x%"PRIx64"\n", key->mono.ff_tcs_inputs_to_copy);
6695 break;
6696
6697 case PIPE_SHADER_TESS_EVAL:
6698 fprintf(f, " part.tes.epilog.export_prim_id = %u\n", key->part.tes.epilog.export_prim_id);
6699 fprintf(f, " as_es = %u\n", key->as_es);
6700 break;
6701
6702 case PIPE_SHADER_GEOMETRY:
6703 fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
6704 break;
6705
6706 case PIPE_SHADER_COMPUTE:
6707 break;
6708
6709 case PIPE_SHADER_FRAGMENT:
6710 fprintf(f, " part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
6711 fprintf(f, " part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
6712 fprintf(f, " part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
6713 fprintf(f, " part.ps.prolog.force_persp_sample_interp = %u\n", key->part.ps.prolog.force_persp_sample_interp);
6714 fprintf(f, " part.ps.prolog.force_linear_sample_interp = %u\n", key->part.ps.prolog.force_linear_sample_interp);
6715 fprintf(f, " part.ps.prolog.force_persp_center_interp = %u\n", key->part.ps.prolog.force_persp_center_interp);
6716 fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
6717 fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
6718 fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
6719 fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
6720 fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
6721 fprintf(f, " part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
6722 fprintf(f, " part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
6723 fprintf(f, " part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
6724 fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
6725 fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
6726 fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
6727 break;
6728
6729 default:
6730 assert(0);
6731 }
6732
6733 if ((processor == PIPE_SHADER_GEOMETRY ||
6734 processor == PIPE_SHADER_TESS_EVAL ||
6735 processor == PIPE_SHADER_VERTEX) &&
6736 !key->as_es && !key->as_ls) {
6737 fprintf(f, " opt.hw_vs.kill_outputs = 0x%"PRIx64"\n", key->opt.hw_vs.kill_outputs);
6738 fprintf(f, " opt.hw_vs.kill_outputs2 = 0x%x\n", key->opt.hw_vs.kill_outputs2);
6739 fprintf(f, " opt.hw_vs.clip_disable = %u\n", key->opt.hw_vs.clip_disable);
6740 }
6741 }
6742
6743 static void si_init_shader_ctx(struct si_shader_context *ctx,
6744 struct si_screen *sscreen,
6745 LLVMTargetMachineRef tm)
6746 {
6747 struct lp_build_tgsi_context *bld_base;
6748 struct lp_build_tgsi_action tmpl = {};
6749
6750 si_llvm_context_init(ctx, sscreen, tm);
6751
6752 bld_base = &ctx->bld_base;
6753 bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
6754
6755 bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID] = interp_action;
6756 bld_base->op_actions[TGSI_OPCODE_INTERP_SAMPLE] = interp_action;
6757 bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET] = interp_action;
6758
6759 bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
6760 bld_base->op_actions[TGSI_OPCODE_TEX_LZ] = tex_action;
6761 bld_base->op_actions[TGSI_OPCODE_TEX2] = tex_action;
6762 bld_base->op_actions[TGSI_OPCODE_TXB] = tex_action;
6763 bld_base->op_actions[TGSI_OPCODE_TXB2] = tex_action;
6764 bld_base->op_actions[TGSI_OPCODE_TXD] = tex_action;
6765 bld_base->op_actions[TGSI_OPCODE_TXF] = tex_action;
6766 bld_base->op_actions[TGSI_OPCODE_TXF_LZ] = tex_action;
6767 bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action;
6768 bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action;
6769 bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
6770 bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = txq_fetch_args;
6771 bld_base->op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
6772 bld_base->op_actions[TGSI_OPCODE_TG4] = tex_action;
6773 bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action;
6774 bld_base->op_actions[TGSI_OPCODE_TXQS].emit = si_llvm_emit_txqs;
6775
6776 bld_base->op_actions[TGSI_OPCODE_LOAD].fetch_args = load_fetch_args;
6777 bld_base->op_actions[TGSI_OPCODE_LOAD].emit = load_emit;
6778 bld_base->op_actions[TGSI_OPCODE_STORE].fetch_args = store_fetch_args;
6779 bld_base->op_actions[TGSI_OPCODE_STORE].emit = store_emit;
6780 bld_base->op_actions[TGSI_OPCODE_RESQ].fetch_args = resq_fetch_args;
6781 bld_base->op_actions[TGSI_OPCODE_RESQ].emit = resq_emit;
6782
6783 tmpl.fetch_args = atomic_fetch_args;
6784 tmpl.emit = atomic_emit;
6785 bld_base->op_actions[TGSI_OPCODE_ATOMUADD] = tmpl;
6786 bld_base->op_actions[TGSI_OPCODE_ATOMUADD].intr_name = "add";
6787 bld_base->op_actions[TGSI_OPCODE_ATOMXCHG] = tmpl;
6788 bld_base->op_actions[TGSI_OPCODE_ATOMXCHG].intr_name = "swap";
6789 bld_base->op_actions[TGSI_OPCODE_ATOMCAS] = tmpl;
6790 bld_base->op_actions[TGSI_OPCODE_ATOMCAS].intr_name = "cmpswap";
6791 bld_base->op_actions[TGSI_OPCODE_ATOMAND] = tmpl;
6792 bld_base->op_actions[TGSI_OPCODE_ATOMAND].intr_name = "and";
6793 bld_base->op_actions[TGSI_OPCODE_ATOMOR] = tmpl;
6794 bld_base->op_actions[TGSI_OPCODE_ATOMOR].intr_name = "or";
6795 bld_base->op_actions[TGSI_OPCODE_ATOMXOR] = tmpl;
6796 bld_base->op_actions[TGSI_OPCODE_ATOMXOR].intr_name = "xor";
6797 bld_base->op_actions[TGSI_OPCODE_ATOMUMIN] = tmpl;
6798 bld_base->op_actions[TGSI_OPCODE_ATOMUMIN].intr_name = "umin";
6799 bld_base->op_actions[TGSI_OPCODE_ATOMUMAX] = tmpl;
6800 bld_base->op_actions[TGSI_OPCODE_ATOMUMAX].intr_name = "umax";
6801 bld_base->op_actions[TGSI_OPCODE_ATOMIMIN] = tmpl;
6802 bld_base->op_actions[TGSI_OPCODE_ATOMIMIN].intr_name = "smin";
6803 bld_base->op_actions[TGSI_OPCODE_ATOMIMAX] = tmpl;
6804 bld_base->op_actions[TGSI_OPCODE_ATOMIMAX].intr_name = "smax";
6805
6806 bld_base->op_actions[TGSI_OPCODE_MEMBAR].emit = membar_emit;
6807
6808 bld_base->op_actions[TGSI_OPCODE_CLOCK].emit = clock_emit;
6809
6810 bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
6811 bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
6812 bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
6813 bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;
6814
6815 bld_base->op_actions[TGSI_OPCODE_VOTE_ALL].emit = vote_all_emit;
6816 bld_base->op_actions[TGSI_OPCODE_VOTE_ANY].emit = vote_any_emit;
6817 bld_base->op_actions[TGSI_OPCODE_VOTE_EQ].emit = vote_eq_emit;
6818 bld_base->op_actions[TGSI_OPCODE_BALLOT].emit = ballot_emit;
6819 bld_base->op_actions[TGSI_OPCODE_READ_FIRST].intr_name = "llvm.amdgcn.readfirstlane";
6820 bld_base->op_actions[TGSI_OPCODE_READ_FIRST].emit = read_lane_emit;
6821 bld_base->op_actions[TGSI_OPCODE_READ_INVOC].intr_name = "llvm.amdgcn.readlane";
6822 bld_base->op_actions[TGSI_OPCODE_READ_INVOC].fetch_args = read_invoc_fetch_args;
6823 bld_base->op_actions[TGSI_OPCODE_READ_INVOC].emit = read_lane_emit;
6824
6825 bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
6826 bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
6827 bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
6828 }
6829
6830 static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx)
6831 {
6832 struct si_shader *shader = ctx->shader;
6833 struct tgsi_shader_info *info = &shader->selector->info;
6834
6835 if (ctx->type == PIPE_SHADER_FRAGMENT ||
6836 ctx->type == PIPE_SHADER_COMPUTE ||
6837 shader->key.as_es ||
6838 shader->key.as_ls)
6839 return;
6840
6841 ac_eliminate_const_vs_outputs(&ctx->ac,
6842 ctx->main_fn,
6843 shader->info.vs_output_param_offset,
6844 info->num_outputs,
6845 &shader->info.nr_param_exports);
6846 }
6847
6848 static void si_count_scratch_private_memory(struct si_shader_context *ctx)
6849 {
6850 ctx->shader->config.private_mem_vgprs = 0;
6851
6852 /* Process all LLVM instructions. */
6853 LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(ctx->main_fn);
6854 while (bb) {
6855 LLVMValueRef next = LLVMGetFirstInstruction(bb);
6856
6857 while (next) {
6858 LLVMValueRef inst = next;
6859 next = LLVMGetNextInstruction(next);
6860
6861 if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
6862 continue;
6863
6864 LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
6865 /* No idea why LLVM aligns allocas to 4 elements. */
6866 unsigned alignment = LLVMGetAlignment(inst);
6867 unsigned dw_size = align(llvm_get_type_size(type) / 4, alignment);
6868 ctx->shader->config.private_mem_vgprs += dw_size;
6869 }
6870 bb = LLVMGetNextBasicBlock(bb);
6871 }
6872 }
6873
6874 static bool si_compile_tgsi_main(struct si_shader_context *ctx,
6875 struct si_shader *shader)
6876 {
6877 struct si_shader_selector *sel = shader->selector;
6878 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
6879
6880 switch (ctx->type) {
6881 case PIPE_SHADER_VERTEX:
6882 ctx->load_input = declare_input_vs;
6883 if (shader->key.as_ls)
6884 bld_base->emit_epilogue = si_llvm_emit_ls_epilogue;
6885 else if (shader->key.as_es)
6886 bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
6887 else
6888 bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
6889 break;
6890 case PIPE_SHADER_TESS_CTRL:
6891 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tcs;
6892 bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = fetch_output_tcs;
6893 bld_base->emit_store = store_output_tcs;
6894 bld_base->emit_epilogue = si_llvm_emit_tcs_epilogue;
6895 break;
6896 case PIPE_SHADER_TESS_EVAL:
6897 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tes;
6898 if (shader->key.as_es)
6899 bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
6900 else
6901 bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
6902 break;
6903 case PIPE_SHADER_GEOMETRY:
6904 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_gs;
6905 bld_base->emit_epilogue = si_llvm_emit_gs_epilogue;
6906 break;
6907 case PIPE_SHADER_FRAGMENT:
6908 ctx->load_input = declare_input_fs;
6909 bld_base->emit_epilogue = si_llvm_return_fs_outputs;
6910 break;
6911 case PIPE_SHADER_COMPUTE:
6912 ctx->declare_memory_region = declare_compute_memory;
6913 break;
6914 default:
6915 assert(!"Unsupported shader type");
6916 return false;
6917 }
6918
6919 create_function(ctx);
6920 preload_ring_buffers(ctx);
6921
6922 if (ctx->type == PIPE_SHADER_GEOMETRY) {
6923 int i;
6924 for (i = 0; i < 4; i++) {
6925 ctx->gs_next_vertex[i] =
6926 lp_build_alloca(&ctx->gallivm,
6927 ctx->i32, "");
6928 }
6929 }
6930
6931 if (!lp_build_tgsi_llvm(bld_base, sel->tokens)) {
6932 fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n");
6933 return false;
6934 }
6935
6936 si_llvm_build_ret(ctx, ctx->return_value);
6937 return true;
6938 }
6939
6940 /**
6941 * Compute the VS prolog key, which contains all the information needed to
6942 * build the VS prolog function, and set shader->info bits where needed.
6943 *
6944 * \param info Shader info of the vertex shader.
6945 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
6946 * \param prolog_key Key of the VS prolog
6947 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
6948 * \param key Output shader part key.
6949 */
6950 static void si_get_vs_prolog_key(const struct tgsi_shader_info *info,
6951 unsigned num_input_sgprs,
6952 const struct si_vs_prolog_bits *prolog_key,
6953 struct si_shader *shader_out,
6954 union si_shader_part_key *key)
6955 {
6956 memset(key, 0, sizeof(*key));
6957 key->vs_prolog.states = *prolog_key;
6958 key->vs_prolog.num_input_sgprs = num_input_sgprs;
6959 key->vs_prolog.last_input = MAX2(1, info->num_inputs) - 1;
6960
6961 /* Set the instanceID flag. */
6962 for (unsigned i = 0; i < info->num_inputs; i++)
6963 if (key->vs_prolog.states.instance_divisors[i])
6964 shader_out->info.uses_instanceid = true;
6965 }
6966
6967 /**
6968 * Compute the VS epilog key, which contains all the information needed to
6969 * build the VS epilog function, and set the PrimitiveID output offset.
6970 */
6971 static void si_get_vs_epilog_key(struct si_shader *shader,
6972 struct si_vs_epilog_bits *states,
6973 union si_shader_part_key *key)
6974 {
6975 memset(key, 0, sizeof(*key));
6976 key->vs_epilog.states = *states;
6977
6978 /* Set up the PrimitiveID output. */
6979 if (shader->key.part.vs.epilog.export_prim_id) {
6980 unsigned index = shader->selector->info.num_outputs;
6981 unsigned offset = shader->info.nr_param_exports++;
6982
6983 key->vs_epilog.prim_id_param_offset = offset;
6984 assert(index < ARRAY_SIZE(shader->info.vs_output_param_offset));
6985 shader->info.vs_output_param_offset[index] = offset;
6986 }
6987 }
6988
6989 /**
6990 * Compute the PS prolog key, which contains all the information needed to
6991 * build the PS prolog function, and set related bits in shader->config.
6992 */
6993 static void si_get_ps_prolog_key(struct si_shader *shader,
6994 union si_shader_part_key *key,
6995 bool separate_prolog)
6996 {
6997 struct tgsi_shader_info *info = &shader->selector->info;
6998
6999 memset(key, 0, sizeof(*key));
7000 key->ps_prolog.states = shader->key.part.ps.prolog;
7001 key->ps_prolog.colors_read = info->colors_read;
7002 key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
7003 key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
7004 key->ps_prolog.wqm = info->uses_derivatives &&
7005 (key->ps_prolog.colors_read ||
7006 key->ps_prolog.states.force_persp_sample_interp ||
7007 key->ps_prolog.states.force_linear_sample_interp ||
7008 key->ps_prolog.states.force_persp_center_interp ||
7009 key->ps_prolog.states.force_linear_center_interp ||
7010 key->ps_prolog.states.bc_optimize_for_persp ||
7011 key->ps_prolog.states.bc_optimize_for_linear);
7012
7013 if (info->colors_read) {
7014 unsigned *color = shader->selector->color_attr_index;
7015
7016 if (shader->key.part.ps.prolog.color_two_side) {
7017 /* BCOLORs are stored after the last input. */
7018 key->ps_prolog.num_interp_inputs = info->num_inputs;
7019 key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
7020 shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
7021 }
7022
7023 for (unsigned i = 0; i < 2; i++) {
7024 unsigned interp = info->input_interpolate[color[i]];
7025 unsigned location = info->input_interpolate_loc[color[i]];
7026
7027 if (!(info->colors_read & (0xf << i*4)))
7028 continue;
7029
7030 key->ps_prolog.color_attr_index[i] = color[i];
7031
7032 if (shader->key.part.ps.prolog.flatshade_colors &&
7033 interp == TGSI_INTERPOLATE_COLOR)
7034 interp = TGSI_INTERPOLATE_CONSTANT;
7035
7036 switch (interp) {
7037 case TGSI_INTERPOLATE_CONSTANT:
7038 key->ps_prolog.color_interp_vgpr_index[i] = -1;
7039 break;
7040 case TGSI_INTERPOLATE_PERSPECTIVE:
7041 case TGSI_INTERPOLATE_COLOR:
7042 /* Force the interpolation location for colors here. */
7043 if (shader->key.part.ps.prolog.force_persp_sample_interp)
7044 location = TGSI_INTERPOLATE_LOC_SAMPLE;
7045 if (shader->key.part.ps.prolog.force_persp_center_interp)
7046 location = TGSI_INTERPOLATE_LOC_CENTER;
7047
7048 switch (location) {
7049 case TGSI_INTERPOLATE_LOC_SAMPLE:
7050 key->ps_prolog.color_interp_vgpr_index[i] = 0;
7051 shader->config.spi_ps_input_ena |=
7052 S_0286CC_PERSP_SAMPLE_ENA(1);
7053 break;
7054 case TGSI_INTERPOLATE_LOC_CENTER:
7055 key->ps_prolog.color_interp_vgpr_index[i] = 2;
7056 shader->config.spi_ps_input_ena |=
7057 S_0286CC_PERSP_CENTER_ENA(1);
7058 break;
7059 case TGSI_INTERPOLATE_LOC_CENTROID:
7060 key->ps_prolog.color_interp_vgpr_index[i] = 4;
7061 shader->config.spi_ps_input_ena |=
7062 S_0286CC_PERSP_CENTROID_ENA(1);
7063 break;
7064 default:
7065 assert(0);
7066 }
7067 break;
7068 case TGSI_INTERPOLATE_LINEAR:
7069 /* Force the interpolation location for colors here. */
7070 if (shader->key.part.ps.prolog.force_linear_sample_interp)
7071 location = TGSI_INTERPOLATE_LOC_SAMPLE;
7072 if (shader->key.part.ps.prolog.force_linear_center_interp)
7073 location = TGSI_INTERPOLATE_LOC_CENTER;
7074
7075 /* The VGPR assignment for non-monolithic shaders
7076 * works because InitialPSInputAddr is set on the
7077 * main shader and PERSP_PULL_MODEL is never used.
7078 */
7079 switch (location) {
7080 case TGSI_INTERPOLATE_LOC_SAMPLE:
7081 key->ps_prolog.color_interp_vgpr_index[i] =
7082 separate_prolog ? 6 : 9;
7083 shader->config.spi_ps_input_ena |=
7084 S_0286CC_LINEAR_SAMPLE_ENA(1);
7085 break;
7086 case TGSI_INTERPOLATE_LOC_CENTER:
7087 key->ps_prolog.color_interp_vgpr_index[i] =
7088 separate_prolog ? 8 : 11;
7089 shader->config.spi_ps_input_ena |=
7090 S_0286CC_LINEAR_CENTER_ENA(1);
7091 break;
7092 case TGSI_INTERPOLATE_LOC_CENTROID:
7093 key->ps_prolog.color_interp_vgpr_index[i] =
7094 separate_prolog ? 10 : 13;
7095 shader->config.spi_ps_input_ena |=
7096 S_0286CC_LINEAR_CENTROID_ENA(1);
7097 break;
7098 default:
7099 assert(0);
7100 }
7101 break;
7102 default:
7103 assert(0);
7104 }
7105 }
7106 }
7107 }
7108
7109 /**
7110 * Check whether a PS prolog is required based on the key.
7111 */
7112 static bool si_need_ps_prolog(const union si_shader_part_key *key)
7113 {
7114 return key->ps_prolog.colors_read ||
7115 key->ps_prolog.states.force_persp_sample_interp ||
7116 key->ps_prolog.states.force_linear_sample_interp ||
7117 key->ps_prolog.states.force_persp_center_interp ||
7118 key->ps_prolog.states.force_linear_center_interp ||
7119 key->ps_prolog.states.bc_optimize_for_persp ||
7120 key->ps_prolog.states.bc_optimize_for_linear ||
7121 key->ps_prolog.states.poly_stipple;
7122 }
7123
7124 /**
7125 * Compute the PS epilog key, which contains all the information needed to
7126 * build the PS epilog function.
7127 */
7128 static void si_get_ps_epilog_key(struct si_shader *shader,
7129 union si_shader_part_key *key)
7130 {
7131 struct tgsi_shader_info *info = &shader->selector->info;
7132 memset(key, 0, sizeof(*key));
7133 key->ps_epilog.colors_written = info->colors_written;
7134 key->ps_epilog.writes_z = info->writes_z;
7135 key->ps_epilog.writes_stencil = info->writes_stencil;
7136 key->ps_epilog.writes_samplemask = info->writes_samplemask;
7137 key->ps_epilog.states = shader->key.part.ps.epilog;
7138 }
7139
7140 /**
7141 * Build the GS prolog function. Rotate the input vertices for triangle strips
7142 * with adjacency.
7143 */
7144 static void si_build_gs_prolog_function(struct si_shader_context *ctx,
7145 union si_shader_part_key *key)
7146 {
7147 const unsigned num_sgprs = SI_GS_NUM_USER_SGPR + 2;
7148 const unsigned num_vgprs = 8;
7149 struct gallivm_state *gallivm = &ctx->gallivm;
7150 LLVMBuilderRef builder = gallivm->builder;
7151 LLVMTypeRef params[32];
7152 LLVMTypeRef returns[32];
7153 LLVMValueRef func, ret;
7154
7155 for (unsigned i = 0; i < num_sgprs; ++i) {
7156 params[i] = ctx->i32;
7157 returns[i] = ctx->i32;
7158 }
7159
7160 for (unsigned i = 0; i < num_vgprs; ++i) {
7161 params[num_sgprs + i] = ctx->i32;
7162 returns[num_sgprs + i] = ctx->f32;
7163 }
7164
7165 /* Create the function. */
7166 si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs,
7167 params, num_sgprs + num_vgprs, num_sgprs - 1);
7168 func = ctx->main_fn;
7169
7170 /* Copy inputs to outputs. This should be no-op, as the registers match,
7171 * but it will prevent the compiler from overwriting them unintentionally.
7172 */
7173 ret = ctx->return_value;
7174 for (unsigned i = 0; i < num_sgprs; i++) {
7175 LLVMValueRef p = LLVMGetParam(func, i);
7176 ret = LLVMBuildInsertValue(builder, ret, p, i, "");
7177 }
7178 for (unsigned i = 0; i < num_vgprs; i++) {
7179 LLVMValueRef p = LLVMGetParam(func, num_sgprs + i);
7180 p = LLVMBuildBitCast(builder, p, ctx->f32, "");
7181 ret = LLVMBuildInsertValue(builder, ret, p, num_sgprs + i, "");
7182 }
7183
7184 if (key->gs_prolog.states.tri_strip_adj_fix) {
7185 /* Remap the input vertices for every other primitive. */
7186 const unsigned vtx_params[6] = {
7187 num_sgprs,
7188 num_sgprs + 1,
7189 num_sgprs + 3,
7190 num_sgprs + 4,
7191 num_sgprs + 5,
7192 num_sgprs + 6
7193 };
7194 LLVMValueRef prim_id, rotate;
7195
7196 prim_id = LLVMGetParam(func, num_sgprs + 2);
7197 rotate = LLVMBuildTrunc(builder, prim_id, ctx->i1, "");
7198
7199 for (unsigned i = 0; i < 6; ++i) {
7200 LLVMValueRef base, rotated, actual;
7201 base = LLVMGetParam(func, vtx_params[i]);
7202 rotated = LLVMGetParam(func, vtx_params[(i + 4) % 6]);
7203 actual = LLVMBuildSelect(builder, rotate, rotated, base, "");
7204 actual = LLVMBuildBitCast(builder, actual, ctx->f32, "");
7205 ret = LLVMBuildInsertValue(builder, ret, actual, vtx_params[i], "");
7206 }
7207 }
7208
7209 LLVMBuildRet(builder, ret);
7210 }
7211
7212 /**
7213 * Given a list of shader part functions, build a wrapper function that
7214 * runs them in sequence to form a monolithic shader.
7215 */
7216 static void si_build_wrapper_function(struct si_shader_context *ctx,
7217 LLVMValueRef *parts,
7218 unsigned num_parts,
7219 unsigned main_part)
7220 {
7221 struct gallivm_state *gallivm = &ctx->gallivm;
7222 LLVMBuilderRef builder = ctx->gallivm.builder;
7223 /* PS epilog has one arg per color component */
7224 LLVMTypeRef param_types[48];
7225 LLVMValueRef out[48];
7226 LLVMTypeRef function_type;
7227 unsigned num_params;
7228 unsigned num_out;
7229 MAYBE_UNUSED unsigned num_out_sgpr; /* used in debug checks */
7230 unsigned num_sgprs, num_vgprs;
7231 unsigned last_sgpr_param;
7232 unsigned gprs;
7233
7234 for (unsigned i = 0; i < num_parts; ++i) {
7235 lp_add_function_attr(parts[i], -1, LP_FUNC_ATTR_ALWAYSINLINE);
7236 LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
7237 }
7238
7239 /* The parameters of the wrapper function correspond to those of the
7240 * first part in terms of SGPRs and VGPRs, but we use the types of the
7241 * main part to get the right types. This is relevant for the
7242 * dereferenceable attribute on descriptor table pointers.
7243 */
7244 num_sgprs = 0;
7245 num_vgprs = 0;
7246
7247 function_type = LLVMGetElementType(LLVMTypeOf(parts[0]));
7248 num_params = LLVMCountParamTypes(function_type);
7249
7250 for (unsigned i = 0; i < num_params; ++i) {
7251 LLVMValueRef param = LLVMGetParam(parts[0], i);
7252
7253 if (ac_is_sgpr_param(param)) {
7254 assert(num_vgprs == 0);
7255 num_sgprs += llvm_get_type_size(LLVMTypeOf(param)) / 4;
7256 } else {
7257 num_vgprs += llvm_get_type_size(LLVMTypeOf(param)) / 4;
7258 }
7259 }
7260 assert(num_vgprs + num_sgprs <= ARRAY_SIZE(param_types));
7261
7262 num_params = 0;
7263 last_sgpr_param = 0;
7264 gprs = 0;
7265 while (gprs < num_sgprs + num_vgprs) {
7266 LLVMValueRef param = LLVMGetParam(parts[main_part], num_params);
7267 unsigned size;
7268
7269 param_types[num_params] = LLVMTypeOf(param);
7270 if (gprs < num_sgprs)
7271 last_sgpr_param = num_params;
7272 size = llvm_get_type_size(param_types[num_params]) / 4;
7273 num_params++;
7274
7275 assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
7276 assert(gprs + size <= num_sgprs + num_vgprs &&
7277 (gprs >= num_sgprs || gprs + size <= num_sgprs));
7278
7279 gprs += size;
7280 }
7281
7282 si_create_function(ctx, "wrapper", NULL, 0, param_types, num_params, last_sgpr_param);
7283
7284 /* Record the arguments of the function as if they were an output of
7285 * a previous part.
7286 */
7287 num_out = 0;
7288 num_out_sgpr = 0;
7289
7290 for (unsigned i = 0; i < num_params; ++i) {
7291 LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
7292 LLVMTypeRef param_type = LLVMTypeOf(param);
7293 LLVMTypeRef out_type = i <= last_sgpr_param ? ctx->i32 : ctx->f32;
7294 unsigned size = llvm_get_type_size(param_type) / 4;
7295
7296 if (size == 1) {
7297 if (param_type != out_type)
7298 param = LLVMBuildBitCast(builder, param, out_type, "");
7299 out[num_out++] = param;
7300 } else {
7301 LLVMTypeRef vector_type = LLVMVectorType(out_type, size);
7302
7303 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
7304 param = LLVMBuildPtrToInt(builder, param, ctx->i64, "");
7305 param_type = ctx->i64;
7306 }
7307
7308 if (param_type != vector_type)
7309 param = LLVMBuildBitCast(builder, param, vector_type, "");
7310
7311 for (unsigned j = 0; j < size; ++j)
7312 out[num_out++] = LLVMBuildExtractElement(
7313 builder, param, LLVMConstInt(ctx->i32, j, 0), "");
7314 }
7315
7316 if (i <= last_sgpr_param)
7317 num_out_sgpr = num_out;
7318 }
7319
7320 /* Now chain the parts. */
7321 for (unsigned part = 0; part < num_parts; ++part) {
7322 LLVMValueRef in[48];
7323 LLVMValueRef ret;
7324 LLVMTypeRef ret_type;
7325 unsigned out_idx = 0;
7326
7327 num_params = LLVMCountParams(parts[part]);
7328 assert(num_params <= ARRAY_SIZE(param_types));
7329
7330 /* Derive arguments for the next part from outputs of the
7331 * previous one.
7332 */
7333 for (unsigned param_idx = 0; param_idx < num_params; ++param_idx) {
7334 LLVMValueRef param;
7335 LLVMTypeRef param_type;
7336 bool is_sgpr;
7337 unsigned param_size;
7338 LLVMValueRef arg = NULL;
7339
7340 param = LLVMGetParam(parts[part], param_idx);
7341 param_type = LLVMTypeOf(param);
7342 param_size = llvm_get_type_size(param_type) / 4;
7343 is_sgpr = ac_is_sgpr_param(param);
7344
7345 if (is_sgpr) {
7346 #if HAVE_LLVM < 0x0400
7347 LLVMRemoveAttribute(param, LLVMByValAttribute);
7348 #else
7349 unsigned kind_id = LLVMGetEnumAttributeKindForName("byval", 5);
7350 LLVMRemoveEnumAttributeAtIndex(parts[part], param_idx + 1, kind_id);
7351 #endif
7352 lp_add_function_attr(parts[part], param_idx + 1, LP_FUNC_ATTR_INREG);
7353 }
7354
7355 assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));
7356 assert(is_sgpr || out_idx >= num_out_sgpr);
7357
7358 if (param_size == 1)
7359 arg = out[out_idx];
7360 else
7361 arg = lp_build_gather_values(gallivm, &out[out_idx], param_size);
7362
7363 if (LLVMTypeOf(arg) != param_type) {
7364 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
7365 arg = LLVMBuildBitCast(builder, arg, ctx->i64, "");
7366 arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
7367 } else {
7368 arg = LLVMBuildBitCast(builder, arg, param_type, "");
7369 }
7370 }
7371
7372 in[param_idx] = arg;
7373 out_idx += param_size;
7374 }
7375
7376 ret = LLVMBuildCall(builder, parts[part], in, num_params, "");
7377 ret_type = LLVMTypeOf(ret);
7378
7379 /* Extract the returned GPRs. */
7380 num_out = 0;
7381 num_out_sgpr = 0;
7382
7383 if (LLVMGetTypeKind(ret_type) != LLVMVoidTypeKind) {
7384 assert(LLVMGetTypeKind(ret_type) == LLVMStructTypeKind);
7385
7386 unsigned ret_size = LLVMCountStructElementTypes(ret_type);
7387
7388 for (unsigned i = 0; i < ret_size; ++i) {
7389 LLVMValueRef val =
7390 LLVMBuildExtractValue(builder, ret, i, "");
7391
7392 out[num_out++] = val;
7393
7394 if (LLVMTypeOf(val) == ctx->i32) {
7395 assert(num_out_sgpr + 1 == num_out);
7396 num_out_sgpr = num_out;
7397 }
7398 }
7399 }
7400 }
7401
7402 LLVMBuildRetVoid(builder);
7403 }
7404
7405 int si_compile_tgsi_shader(struct si_screen *sscreen,
7406 LLVMTargetMachineRef tm,
7407 struct si_shader *shader,
7408 bool is_monolithic,
7409 struct pipe_debug_callback *debug)
7410 {
7411 struct si_shader_selector *sel = shader->selector;
7412 struct si_shader_context ctx;
7413 int r = -1;
7414
7415 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
7416 * conversion fails. */
7417 if (r600_can_dump_shader(&sscreen->b, sel->info.processor) &&
7418 !(sscreen->b.debug_flags & DBG_NO_TGSI)) {
7419 tgsi_dump(sel->tokens, 0);
7420 si_dump_streamout(&sel->so);
7421 }
7422
7423 si_init_shader_ctx(&ctx, sscreen, tm);
7424 si_llvm_context_set_tgsi(&ctx, shader);
7425 ctx.separate_prolog = !is_monolithic;
7426
7427 memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
7428 sizeof(shader->info.vs_output_param_offset));
7429
7430 shader->info.uses_instanceid = sel->info.uses_instanceid;
7431
7432 ctx.load_system_value = declare_system_value;
7433
7434 if (!si_compile_tgsi_main(&ctx, shader)) {
7435 si_llvm_dispose(&ctx);
7436 return -1;
7437 }
7438
7439 if (is_monolithic && ctx.type == PIPE_SHADER_VERTEX) {
7440 LLVMValueRef parts[3];
7441 bool need_prolog;
7442 bool need_epilog;
7443
7444 need_prolog = sel->vs_needs_prolog;
7445 need_epilog = !shader->key.as_es && !shader->key.as_ls;
7446
7447 parts[need_prolog ? 1 : 0] = ctx.main_fn;
7448
7449 if (need_prolog) {
7450 union si_shader_part_key prolog_key;
7451 si_get_vs_prolog_key(&sel->info,
7452 shader->info.num_input_sgprs,
7453 &shader->key.part.vs.prolog,
7454 shader, &prolog_key);
7455 si_build_vs_prolog_function(&ctx, &prolog_key);
7456 parts[0] = ctx.main_fn;
7457 }
7458
7459 if (need_epilog) {
7460 union si_shader_part_key epilog_key;
7461 si_get_vs_epilog_key(shader, &shader->key.part.vs.epilog, &epilog_key);
7462 si_build_vs_epilog_function(&ctx, &epilog_key);
7463 parts[need_prolog ? 2 : 1] = ctx.main_fn;
7464 }
7465
7466 si_build_wrapper_function(&ctx, parts, 1 + need_prolog + need_epilog,
7467 need_prolog ? 1 : 0);
7468 } else if (is_monolithic && ctx.type == PIPE_SHADER_TESS_CTRL) {
7469 LLVMValueRef parts[2];
7470 union si_shader_part_key epilog_key;
7471
7472 parts[0] = ctx.main_fn;
7473
7474 memset(&epilog_key, 0, sizeof(epilog_key));
7475 epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
7476 si_build_tcs_epilog_function(&ctx, &epilog_key);
7477 parts[1] = ctx.main_fn;
7478
7479 si_build_wrapper_function(&ctx, parts, 2, 0);
7480 } else if (is_monolithic && ctx.type == PIPE_SHADER_TESS_EVAL &&
7481 !shader->key.as_es) {
7482 LLVMValueRef parts[2];
7483 union si_shader_part_key epilog_key;
7484
7485 parts[0] = ctx.main_fn;
7486
7487 si_get_vs_epilog_key(shader, &shader->key.part.tes.epilog, &epilog_key);
7488 si_build_vs_epilog_function(&ctx, &epilog_key);
7489 parts[1] = ctx.main_fn;
7490
7491 si_build_wrapper_function(&ctx, parts, 2, 0);
7492 } else if (is_monolithic && ctx.type == PIPE_SHADER_GEOMETRY) {
7493 LLVMValueRef parts[2];
7494 union si_shader_part_key prolog_key;
7495
7496 parts[1] = ctx.main_fn;
7497
7498 memset(&prolog_key, 0, sizeof(prolog_key));
7499 prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
7500 si_build_gs_prolog_function(&ctx, &prolog_key);
7501 parts[0] = ctx.main_fn;
7502
7503 si_build_wrapper_function(&ctx, parts, 2, 1);
7504 } else if (is_monolithic && ctx.type == PIPE_SHADER_FRAGMENT) {
7505 LLVMValueRef parts[3];
7506 union si_shader_part_key prolog_key;
7507 union si_shader_part_key epilog_key;
7508 bool need_prolog;
7509
7510 si_get_ps_prolog_key(shader, &prolog_key, false);
7511 need_prolog = si_need_ps_prolog(&prolog_key);
7512
7513 parts[need_prolog ? 1 : 0] = ctx.main_fn;
7514
7515 if (need_prolog) {
7516 si_build_ps_prolog_function(&ctx, &prolog_key);
7517 parts[0] = ctx.main_fn;
7518 }
7519
7520 si_get_ps_epilog_key(shader, &epilog_key);
7521 si_build_ps_epilog_function(&ctx, &epilog_key);
7522 parts[need_prolog ? 2 : 1] = ctx.main_fn;
7523
7524 si_build_wrapper_function(&ctx, parts, need_prolog ? 3 : 2, need_prolog ? 1 : 0);
7525 }
7526
7527 /* Dump LLVM IR before any optimization passes */
7528 if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
7529 r600_can_dump_shader(&sscreen->b, ctx.type))
7530 LLVMDumpModule(ctx.gallivm.module);
7531
7532 si_llvm_finalize_module(&ctx,
7533 r600_extra_shader_checks(&sscreen->b, ctx.type));
7534
7535 /* Post-optimization transformations and analysis. */
7536 si_eliminate_const_vs_outputs(&ctx);
7537
7538 if ((debug && debug->debug_message) ||
7539 r600_can_dump_shader(&sscreen->b, ctx.type))
7540 si_count_scratch_private_memory(&ctx);
7541
7542 /* Compile to bytecode. */
7543 r = si_compile_llvm(sscreen, &shader->binary, &shader->config, tm,
7544 ctx.gallivm.module, debug, ctx.type, "TGSI shader");
7545 si_llvm_dispose(&ctx);
7546 if (r) {
7547 fprintf(stderr, "LLVM failed to compile shader\n");
7548 return r;
7549 }
7550
7551 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
7552 * LLVM 3.9svn has this bug.
7553 */
7554 if (sel->type == PIPE_SHADER_COMPUTE) {
7555 unsigned wave_size = 64;
7556 unsigned max_vgprs = 256;
7557 unsigned max_sgprs = sscreen->b.chip_class >= VI ? 800 : 512;
7558 unsigned max_sgprs_per_wave = 128;
7559 unsigned max_block_threads = si_get_max_workgroup_size(shader);
7560 unsigned min_waves_per_cu = DIV_ROUND_UP(max_block_threads, wave_size);
7561 unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);
7562
7563 max_vgprs = max_vgprs / min_waves_per_simd;
7564 max_sgprs = MIN2(max_sgprs / min_waves_per_simd, max_sgprs_per_wave);
7565
7566 if (shader->config.num_sgprs > max_sgprs ||
7567 shader->config.num_vgprs > max_vgprs) {
7568 fprintf(stderr, "LLVM failed to compile a shader correctly: "
7569 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
7570 shader->config.num_sgprs, shader->config.num_vgprs,
7571 max_sgprs, max_vgprs);
7572
7573 /* Just terminate the process, because dependent
7574 * shaders can hang due to bad input data, but use
7575 * the env var to allow shader-db to work.
7576 */
7577 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
7578 abort();
7579 }
7580 }
7581
7582 /* Add the scratch offset to input SGPRs. */
7583 if (shader->config.scratch_bytes_per_wave)
7584 shader->info.num_input_sgprs += 1; /* scratch byte offset */
7585
7586 /* Calculate the number of fragment input VGPRs. */
7587 if (ctx.type == PIPE_SHADER_FRAGMENT) {
7588 shader->info.num_input_vgprs = 0;
7589 shader->info.face_vgpr_index = -1;
7590
7591 if (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_addr))
7592 shader->info.num_input_vgprs += 2;
7593 if (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr))
7594 shader->info.num_input_vgprs += 2;
7595 if (G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_addr))
7596 shader->info.num_input_vgprs += 2;
7597 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader->config.spi_ps_input_addr))
7598 shader->info.num_input_vgprs += 3;
7599 if (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_addr))
7600 shader->info.num_input_vgprs += 2;
7601 if (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr))
7602 shader->info.num_input_vgprs += 2;
7603 if (G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_addr))
7604 shader->info.num_input_vgprs += 2;
7605 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader->config.spi_ps_input_addr))
7606 shader->info.num_input_vgprs += 1;
7607 if (G_0286CC_POS_X_FLOAT_ENA(shader->config.spi_ps_input_addr))
7608 shader->info.num_input_vgprs += 1;
7609 if (G_0286CC_POS_Y_FLOAT_ENA(shader->config.spi_ps_input_addr))
7610 shader->info.num_input_vgprs += 1;
7611 if (G_0286CC_POS_Z_FLOAT_ENA(shader->config.spi_ps_input_addr))
7612 shader->info.num_input_vgprs += 1;
7613 if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_addr))
7614 shader->info.num_input_vgprs += 1;
7615 if (G_0286CC_FRONT_FACE_ENA(shader->config.spi_ps_input_addr)) {
7616 shader->info.face_vgpr_index = shader->info.num_input_vgprs;
7617 shader->info.num_input_vgprs += 1;
7618 }
7619 if (G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr))
7620 shader->info.num_input_vgprs += 1;
7621 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader->config.spi_ps_input_addr))
7622 shader->info.num_input_vgprs += 1;
7623 if (G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr))
7624 shader->info.num_input_vgprs += 1;
7625 }
7626
7627 return 0;
7628 }
7629
7630 /**
7631 * Create, compile and return a shader part (prolog or epilog).
7632 *
7633 * \param sscreen screen
7634 * \param list list of shader parts of the same category
7635 * \param type shader type
7636 * \param key shader part key
7637 * \param prolog whether the part being requested is a prolog
7638 * \param tm LLVM target machine
7639 * \param debug debug callback
7640 * \param build the callback responsible for building the main function
7641 * \return non-NULL on success
7642 */
7643 static struct si_shader_part *
7644 si_get_shader_part(struct si_screen *sscreen,
7645 struct si_shader_part **list,
7646 enum pipe_shader_type type,
7647 bool prolog,
7648 union si_shader_part_key *key,
7649 LLVMTargetMachineRef tm,
7650 struct pipe_debug_callback *debug,
7651 void (*build)(struct si_shader_context *,
7652 union si_shader_part_key *),
7653 const char *name)
7654 {
7655 struct si_shader_part *result;
7656
7657 mtx_lock(&sscreen->shader_parts_mutex);
7658
7659 /* Find existing. */
7660 for (result = *list; result; result = result->next) {
7661 if (memcmp(&result->key, key, sizeof(*key)) == 0) {
7662 mtx_unlock(&sscreen->shader_parts_mutex);
7663 return result;
7664 }
7665 }
7666
7667 /* Compile a new one. */
7668 result = CALLOC_STRUCT(si_shader_part);
7669 result->key = *key;
7670
7671 struct si_shader shader = {};
7672 struct si_shader_context ctx;
7673 struct gallivm_state *gallivm = &ctx.gallivm;
7674
7675 si_init_shader_ctx(&ctx, sscreen, tm);
7676 ctx.shader = &shader;
7677 ctx.type = type;
7678
7679 switch (type) {
7680 case PIPE_SHADER_VERTEX:
7681 break;
7682 case PIPE_SHADER_TESS_CTRL:
7683 assert(!prolog);
7684 shader.key.part.tcs.epilog = key->tcs_epilog.states;
7685 break;
7686 case PIPE_SHADER_GEOMETRY:
7687 assert(prolog);
7688 break;
7689 case PIPE_SHADER_FRAGMENT:
7690 if (prolog)
7691 shader.key.part.ps.prolog = key->ps_prolog.states;
7692 else
7693 shader.key.part.ps.epilog = key->ps_epilog.states;
7694 break;
7695 default:
7696 unreachable("bad shader part");
7697 }
7698
7699 build(&ctx, key);
7700
7701 /* Compile. */
7702 si_llvm_finalize_module(&ctx,
7703 r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
7704
7705 if (si_compile_llvm(sscreen, &result->binary, &result->config, tm,
7706 gallivm->module, debug, ctx.type, name)) {
7707 FREE(result);
7708 result = NULL;
7709 goto out;
7710 }
7711
7712 result->next = *list;
7713 *list = result;
7714
7715 out:
7716 si_llvm_dispose(&ctx);
7717 mtx_unlock(&sscreen->shader_parts_mutex);
7718 return result;
7719 }
7720
7721 /**
7722 * Build the vertex shader prolog function.
7723 *
7724 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7725 * All inputs are returned unmodified. The vertex load indices are
7726 * stored after them, which will be used by the API VS for fetching inputs.
7727 *
7728 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7729 * input_v0,
7730 * input_v1,
7731 * input_v2,
7732 * input_v3,
7733 * (VertexID + BaseVertex),
7734 * (InstanceID + StartInstance),
7735 * (InstanceID / 2 + StartInstance)
7736 */
7737 static void si_build_vs_prolog_function(struct si_shader_context *ctx,
7738 union si_shader_part_key *key)
7739 {
7740 struct gallivm_state *gallivm = &ctx->gallivm;
7741 LLVMTypeRef *params, *returns;
7742 LLVMValueRef ret, func;
7743 int last_sgpr, num_params, num_returns, i;
7744
7745 ctx->param_vertex_id = key->vs_prolog.num_input_sgprs;
7746 ctx->param_instance_id = key->vs_prolog.num_input_sgprs + 3;
7747
7748 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7749 params = alloca((key->vs_prolog.num_input_sgprs + 4) *
7750 sizeof(LLVMTypeRef));
7751 returns = alloca((key->vs_prolog.num_input_sgprs + 4 +
7752 key->vs_prolog.last_input + 1) *
7753 sizeof(LLVMTypeRef));
7754 num_params = 0;
7755 num_returns = 0;
7756
7757 /* Declare input and output SGPRs. */
7758 num_params = 0;
7759 for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
7760 params[num_params++] = ctx->i32;
7761 returns[num_returns++] = ctx->i32;
7762 }
7763 last_sgpr = num_params - 1;
7764
7765 /* 4 preloaded VGPRs (outputs must be floats) */
7766 for (i = 0; i < 4; i++) {
7767 params[num_params++] = ctx->i32;
7768 returns[num_returns++] = ctx->f32;
7769 }
7770
7771 /* Vertex load indices. */
7772 for (i = 0; i <= key->vs_prolog.last_input; i++)
7773 returns[num_returns++] = ctx->f32;
7774
7775 /* Create the function. */
7776 si_create_function(ctx, "vs_prolog", returns, num_returns, params,
7777 num_params, last_sgpr);
7778 func = ctx->main_fn;
7779
7780 /* Copy inputs to outputs. This should be no-op, as the registers match,
7781 * but it will prevent the compiler from overwriting them unintentionally.
7782 */
7783 ret = ctx->return_value;
7784 for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
7785 LLVMValueRef p = LLVMGetParam(func, i);
7786 ret = LLVMBuildInsertValue(gallivm->builder, ret, p, i, "");
7787 }
7788 for (i = num_params - 4; i < num_params; i++) {
7789 LLVMValueRef p = LLVMGetParam(func, i);
7790 p = LLVMBuildBitCast(gallivm->builder, p, ctx->f32, "");
7791 ret = LLVMBuildInsertValue(gallivm->builder, ret, p, i, "");
7792 }
7793
7794 /* Compute vertex load indices from instance divisors. */
7795 for (i = 0; i <= key->vs_prolog.last_input; i++) {
7796 unsigned divisor = key->vs_prolog.states.instance_divisors[i];
7797 LLVMValueRef index;
7798
7799 if (divisor) {
7800 /* InstanceID / Divisor + StartInstance */
7801 index = get_instance_index_for_fetch(ctx,
7802 SI_SGPR_START_INSTANCE,
7803 divisor);
7804 } else {
7805 /* VertexID + BaseVertex */
7806 index = LLVMBuildAdd(gallivm->builder,
7807 LLVMGetParam(func, ctx->param_vertex_id),
7808 LLVMGetParam(func, SI_SGPR_BASE_VERTEX), "");
7809 }
7810
7811 index = LLVMBuildBitCast(gallivm->builder, index, ctx->f32, "");
7812 ret = LLVMBuildInsertValue(gallivm->builder, ret, index,
7813 num_params++, "");
7814 }
7815
7816 si_llvm_build_ret(ctx, ret);
7817 }
7818
7819 /**
7820 * Build the vertex shader epilog function. This is also used by the tessellation
7821 * evaluation shader compiled as VS.
7822 *
7823 * The input is PrimitiveID.
7824 *
7825 * If PrimitiveID is required by the pixel shader, export it.
7826 * Otherwise, do nothing.
7827 */
7828 static void si_build_vs_epilog_function(struct si_shader_context *ctx,
7829 union si_shader_part_key *key)
7830 {
7831 struct gallivm_state *gallivm = &ctx->gallivm;
7832 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
7833 LLVMTypeRef params[5];
7834 int num_params, i;
7835
7836 /* Declare input VGPRs. */
7837 num_params = key->vs_epilog.states.export_prim_id ?
7838 (VS_EPILOG_PRIMID_LOC + 1) : 0;
7839 assert(num_params <= ARRAY_SIZE(params));
7840
7841 for (i = 0; i < num_params; i++)
7842 params[i] = ctx->f32;
7843
7844 /* Create the function. */
7845 si_create_function(ctx, "vs_epilog", NULL, 0, params, num_params, -1);
7846
7847 /* Emit exports. */
7848 if (key->vs_epilog.states.export_prim_id) {
7849 struct lp_build_context *base = &bld_base->base;
7850 struct ac_export_args args;
7851
7852 args.enabled_channels = 0x1; /* enabled channels */
7853 args.valid_mask = 0; /* whether the EXEC mask is valid */
7854 args.done = 0; /* DONE bit */
7855 args.target = V_008DFC_SQ_EXP_PARAM +
7856 key->vs_epilog.prim_id_param_offset;
7857 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
7858 args.out[0] = LLVMGetParam(ctx->main_fn,
7859 VS_EPILOG_PRIMID_LOC); /* X */
7860 args.out[1] = base->undef; /* Y */
7861 args.out[2] = base->undef; /* Z */
7862 args.out[3] = base->undef; /* W */
7863
7864 ac_build_export(&ctx->ac, &args);
7865 }
7866
7867 LLVMBuildRetVoid(gallivm->builder);
7868 }
7869
7870 static bool si_get_vs_prolog(struct si_screen *sscreen,
7871 LLVMTargetMachineRef tm,
7872 struct si_shader *shader,
7873 struct pipe_debug_callback *debug,
7874 struct si_shader *main_part,
7875 const struct si_vs_prolog_bits *key)
7876 {
7877 struct si_shader_selector *vs = main_part->selector;
7878
7879 /* The prolog is a no-op if there are no inputs. */
7880 if (!vs->vs_needs_prolog)
7881 return true;
7882
7883 /* Get the prolog. */
7884 union si_shader_part_key prolog_key;
7885 si_get_vs_prolog_key(&vs->info, main_part->info.num_input_sgprs,
7886 key, shader, &prolog_key);
7887
7888 shader->prolog =
7889 si_get_shader_part(sscreen, &sscreen->vs_prologs,
7890 PIPE_SHADER_VERTEX, true, &prolog_key, tm,
7891 debug, si_build_vs_prolog_function,
7892 "Vertex Shader Prolog");
7893 return shader->prolog != NULL;
7894 }
7895
7896 /**
7897 * Create & compile a vertex shader epilog. This a helper used by VS and TES.
7898 */
7899 static bool si_get_vs_epilog(struct si_screen *sscreen,
7900 LLVMTargetMachineRef tm,
7901 struct si_shader *shader,
7902 struct pipe_debug_callback *debug,
7903 struct si_vs_epilog_bits *states)
7904 {
7905 union si_shader_part_key epilog_key;
7906
7907 si_get_vs_epilog_key(shader, states, &epilog_key);
7908
7909 shader->epilog = si_get_shader_part(sscreen, &sscreen->vs_epilogs,
7910 PIPE_SHADER_VERTEX, true,
7911 &epilog_key, tm, debug,
7912 si_build_vs_epilog_function,
7913 "Vertex Shader Epilog");
7914 return shader->epilog != NULL;
7915 }
7916
7917 /**
7918 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7919 */
7920 static bool si_shader_select_vs_parts(struct si_screen *sscreen,
7921 LLVMTargetMachineRef tm,
7922 struct si_shader *shader,
7923 struct pipe_debug_callback *debug)
7924 {
7925 if (!si_get_vs_prolog(sscreen, tm, shader, debug, shader,
7926 &shader->key.part.vs.prolog))
7927 return false;
7928
7929 /* Get the epilog. */
7930 if (!shader->key.as_es && !shader->key.as_ls &&
7931 !si_get_vs_epilog(sscreen, tm, shader, debug,
7932 &shader->key.part.vs.epilog))
7933 return false;
7934
7935 return true;
7936 }
7937
7938 /**
7939 * Select and compile (or reuse) TES parts (epilog).
7940 */
7941 static bool si_shader_select_tes_parts(struct si_screen *sscreen,
7942 LLVMTargetMachineRef tm,
7943 struct si_shader *shader,
7944 struct pipe_debug_callback *debug)
7945 {
7946 if (shader->key.as_es)
7947 return true;
7948
7949 /* TES compiled as VS. */
7950 return si_get_vs_epilog(sscreen, tm, shader, debug,
7951 &shader->key.part.tes.epilog);
7952 }
7953
7954 /**
7955 * Compile the TCS epilog function. This writes tesselation factors to memory
7956 * based on the output primitive type of the tesselator (determined by TES).
7957 */
7958 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
7959 union si_shader_part_key *key)
7960 {
7961 struct gallivm_state *gallivm = &ctx->gallivm;
7962 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
7963 LLVMTypeRef params[16];
7964 LLVMValueRef func;
7965 int last_sgpr, num_params = 0;
7966
7967 /* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
7968 params[ctx->param_rw_buffers = num_params++] =
7969 const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
7970 params[ctx->param_const_buffers = num_params++] = ctx->i64;
7971 params[ctx->param_samplers = num_params++] = ctx->i64;
7972 params[ctx->param_images = num_params++] = ctx->i64;
7973 params[ctx->param_shader_buffers = num_params++] = ctx->i64;
7974 params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
7975 params[ctx->param_tcs_out_lds_offsets = num_params++] = ctx->i32;
7976 params[ctx->param_tcs_out_lds_layout = num_params++] = ctx->i32;
7977 params[ctx->param_vs_state_bits = num_params++] = ctx->i32;
7978 params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
7979 params[ctx->param_tcs_factor_offset = num_params++] = ctx->i32;
7980 last_sgpr = num_params - 1;
7981
7982 params[num_params++] = ctx->i32; /* patch index within the wave (REL_PATCH_ID) */
7983 params[num_params++] = ctx->i32; /* invocation ID within the patch */
7984 params[num_params++] = ctx->i32; /* LDS offset where tess factors should be loaded from */
7985
7986 /* Create the function. */
7987 si_create_function(ctx, "tcs_epilog", NULL, 0, params, num_params, last_sgpr);
7988 declare_tess_lds(ctx);
7989 func = ctx->main_fn;
7990
7991 si_write_tess_factors(bld_base,
7992 LLVMGetParam(func, last_sgpr + 1),
7993 LLVMGetParam(func, last_sgpr + 2),
7994 LLVMGetParam(func, last_sgpr + 3));
7995
7996 LLVMBuildRetVoid(gallivm->builder);
7997 }
7998
7999 /**
8000 * Select and compile (or reuse) TCS parts (epilog).
8001 */
8002 static bool si_shader_select_tcs_parts(struct si_screen *sscreen,
8003 LLVMTargetMachineRef tm,
8004 struct si_shader *shader,
8005 struct pipe_debug_callback *debug)
8006 {
8007 if (sscreen->b.chip_class >= GFX9) {
8008 struct si_shader *ls_main_part =
8009 shader->key.part.tcs.ls->main_shader_part_ls;
8010
8011 if (!si_get_vs_prolog(sscreen, tm, shader, debug, ls_main_part,
8012 &shader->key.part.tcs.ls_prolog))
8013 return false;
8014
8015 shader->previous_stage = ls_main_part;
8016 }
8017
8018 /* Get the epilog. */
8019 union si_shader_part_key epilog_key;
8020 memset(&epilog_key, 0, sizeof(epilog_key));
8021 epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
8022
8023 shader->epilog = si_get_shader_part(sscreen, &sscreen->tcs_epilogs,
8024 PIPE_SHADER_TESS_CTRL, false,
8025 &epilog_key, tm, debug,
8026 si_build_tcs_epilog_function,
8027 "Tessellation Control Shader Epilog");
8028 return shader->epilog != NULL;
8029 }
8030
8031 /**
8032 * Select and compile (or reuse) GS parts (prolog).
8033 */
8034 static bool si_shader_select_gs_parts(struct si_screen *sscreen,
8035 LLVMTargetMachineRef tm,
8036 struct si_shader *shader,
8037 struct pipe_debug_callback *debug)
8038 {
8039 union si_shader_part_key prolog_key;
8040
8041 if (!shader->key.part.gs.prolog.tri_strip_adj_fix)
8042 return true;
8043
8044 memset(&prolog_key, 0, sizeof(prolog_key));
8045 prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
8046
8047 shader->prolog = si_get_shader_part(sscreen, &sscreen->gs_prologs,
8048 PIPE_SHADER_GEOMETRY, true,
8049 &prolog_key, tm, debug,
8050 si_build_gs_prolog_function,
8051 "Geometry Shader Prolog");
8052 return shader->prolog != NULL;
8053 }
8054
8055 /**
8056 * Build the pixel shader prolog function. This handles:
8057 * - two-side color selection and interpolation
8058 * - overriding interpolation parameters for the API PS
8059 * - polygon stippling
8060 *
8061 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
8062 * overriden by other states. (e.g. per-sample interpolation)
8063 * Interpolated colors are stored after the preloaded VGPRs.
8064 */
8065 static void si_build_ps_prolog_function(struct si_shader_context *ctx,
8066 union si_shader_part_key *key)
8067 {
8068 struct gallivm_state *gallivm = &ctx->gallivm;
8069 LLVMTypeRef *params;
8070 LLVMValueRef ret, func;
8071 int last_sgpr, num_params, num_returns, i, num_color_channels;
8072
8073 assert(si_need_ps_prolog(key));
8074
8075 /* Number of inputs + 8 color elements. */
8076 params = alloca((key->ps_prolog.num_input_sgprs +
8077 key->ps_prolog.num_input_vgprs + 8) *
8078 sizeof(LLVMTypeRef));
8079
8080 /* Declare inputs. */
8081 num_params = 0;
8082 for (i = 0; i < key->ps_prolog.num_input_sgprs; i++)
8083 params[num_params++] = ctx->i32;
8084 last_sgpr = num_params - 1;
8085
8086 for (i = 0; i < key->ps_prolog.num_input_vgprs; i++)
8087 params[num_params++] = ctx->f32;
8088
8089 /* Declare outputs (same as inputs + add colors if needed) */
8090 num_returns = num_params;
8091 num_color_channels = util_bitcount(key->ps_prolog.colors_read);
8092 for (i = 0; i < num_color_channels; i++)
8093 params[num_returns++] = ctx->f32;
8094
8095 /* Create the function. */
8096 si_create_function(ctx, "ps_prolog", params, num_returns, params,
8097 num_params, last_sgpr);
8098 func = ctx->main_fn;
8099
8100 /* Copy inputs to outputs. This should be no-op, as the registers match,
8101 * but it will prevent the compiler from overwriting them unintentionally.
8102 */
8103 ret = ctx->return_value;
8104 for (i = 0; i < num_params; i++) {
8105 LLVMValueRef p = LLVMGetParam(func, i);
8106 ret = LLVMBuildInsertValue(gallivm->builder, ret, p, i, "");
8107 }
8108
8109 /* Polygon stippling. */
8110 if (key->ps_prolog.states.poly_stipple) {
8111 /* POS_FIXED_PT is always last. */
8112 unsigned pos = key->ps_prolog.num_input_sgprs +
8113 key->ps_prolog.num_input_vgprs - 1;
8114 LLVMValueRef ptr[2], list;
8115
8116 /* Get the pointer to rw buffers. */
8117 ptr[0] = LLVMGetParam(func, SI_SGPR_RW_BUFFERS);
8118 ptr[1] = LLVMGetParam(func, SI_SGPR_RW_BUFFERS_HI);
8119 list = lp_build_gather_values(gallivm, ptr, 2);
8120 list = LLVMBuildBitCast(gallivm->builder, list, ctx->i64, "");
8121 list = LLVMBuildIntToPtr(gallivm->builder, list,
8122 const_array(ctx->v16i8, SI_NUM_RW_BUFFERS), "");
8123
8124 si_llvm_emit_polygon_stipple(ctx, list, pos);
8125 }
8126
8127 if (key->ps_prolog.states.bc_optimize_for_persp ||
8128 key->ps_prolog.states.bc_optimize_for_linear) {
8129 unsigned i, base = key->ps_prolog.num_input_sgprs;
8130 LLVMValueRef center[2], centroid[2], tmp, bc_optimize;
8131
8132 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
8133 * The hw doesn't compute CENTROID if the whole wave only
8134 * contains fully-covered quads.
8135 *
8136 * PRIM_MASK is after user SGPRs.
8137 */
8138 bc_optimize = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
8139 bc_optimize = LLVMBuildLShr(gallivm->builder, bc_optimize,
8140 LLVMConstInt(ctx->i32, 31, 0), "");
8141 bc_optimize = LLVMBuildTrunc(gallivm->builder, bc_optimize,
8142 ctx->i1, "");
8143
8144 if (key->ps_prolog.states.bc_optimize_for_persp) {
8145 /* Read PERSP_CENTER. */
8146 for (i = 0; i < 2; i++)
8147 center[i] = LLVMGetParam(func, base + 2 + i);
8148 /* Read PERSP_CENTROID. */
8149 for (i = 0; i < 2; i++)
8150 centroid[i] = LLVMGetParam(func, base + 4 + i);
8151 /* Select PERSP_CENTROID. */
8152 for (i = 0; i < 2; i++) {
8153 tmp = LLVMBuildSelect(gallivm->builder, bc_optimize,
8154 center[i], centroid[i], "");
8155 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8156 tmp, base + 4 + i, "");
8157 }
8158 }
8159 if (key->ps_prolog.states.bc_optimize_for_linear) {
8160 /* Read LINEAR_CENTER. */
8161 for (i = 0; i < 2; i++)
8162 center[i] = LLVMGetParam(func, base + 8 + i);
8163 /* Read LINEAR_CENTROID. */
8164 for (i = 0; i < 2; i++)
8165 centroid[i] = LLVMGetParam(func, base + 10 + i);
8166 /* Select LINEAR_CENTROID. */
8167 for (i = 0; i < 2; i++) {
8168 tmp = LLVMBuildSelect(gallivm->builder, bc_optimize,
8169 center[i], centroid[i], "");
8170 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8171 tmp, base + 10 + i, "");
8172 }
8173 }
8174 }
8175
8176 /* Force per-sample interpolation. */
8177 if (key->ps_prolog.states.force_persp_sample_interp) {
8178 unsigned i, base = key->ps_prolog.num_input_sgprs;
8179 LLVMValueRef persp_sample[2];
8180
8181 /* Read PERSP_SAMPLE. */
8182 for (i = 0; i < 2; i++)
8183 persp_sample[i] = LLVMGetParam(func, base + i);
8184 /* Overwrite PERSP_CENTER. */
8185 for (i = 0; i < 2; i++)
8186 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8187 persp_sample[i], base + 2 + i, "");
8188 /* Overwrite PERSP_CENTROID. */
8189 for (i = 0; i < 2; i++)
8190 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8191 persp_sample[i], base + 4 + i, "");
8192 }
8193 if (key->ps_prolog.states.force_linear_sample_interp) {
8194 unsigned i, base = key->ps_prolog.num_input_sgprs;
8195 LLVMValueRef linear_sample[2];
8196
8197 /* Read LINEAR_SAMPLE. */
8198 for (i = 0; i < 2; i++)
8199 linear_sample[i] = LLVMGetParam(func, base + 6 + i);
8200 /* Overwrite LINEAR_CENTER. */
8201 for (i = 0; i < 2; i++)
8202 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8203 linear_sample[i], base + 8 + i, "");
8204 /* Overwrite LINEAR_CENTROID. */
8205 for (i = 0; i < 2; i++)
8206 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8207 linear_sample[i], base + 10 + i, "");
8208 }
8209
8210 /* Force center interpolation. */
8211 if (key->ps_prolog.states.force_persp_center_interp) {
8212 unsigned i, base = key->ps_prolog.num_input_sgprs;
8213 LLVMValueRef persp_center[2];
8214
8215 /* Read PERSP_CENTER. */
8216 for (i = 0; i < 2; i++)
8217 persp_center[i] = LLVMGetParam(func, base + 2 + i);
8218 /* Overwrite PERSP_SAMPLE. */
8219 for (i = 0; i < 2; i++)
8220 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8221 persp_center[i], base + i, "");
8222 /* Overwrite PERSP_CENTROID. */
8223 for (i = 0; i < 2; i++)
8224 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8225 persp_center[i], base + 4 + i, "");
8226 }
8227 if (key->ps_prolog.states.force_linear_center_interp) {
8228 unsigned i, base = key->ps_prolog.num_input_sgprs;
8229 LLVMValueRef linear_center[2];
8230
8231 /* Read LINEAR_CENTER. */
8232 for (i = 0; i < 2; i++)
8233 linear_center[i] = LLVMGetParam(func, base + 8 + i);
8234 /* Overwrite LINEAR_SAMPLE. */
8235 for (i = 0; i < 2; i++)
8236 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8237 linear_center[i], base + 6 + i, "");
8238 /* Overwrite LINEAR_CENTROID. */
8239 for (i = 0; i < 2; i++)
8240 ret = LLVMBuildInsertValue(gallivm->builder, ret,
8241 linear_center[i], base + 10 + i, "");
8242 }
8243
8244 /* Interpolate colors. */
8245 for (i = 0; i < 2; i++) {
8246 unsigned writemask = (key->ps_prolog.colors_read >> (i * 4)) & 0xf;
8247 unsigned face_vgpr = key->ps_prolog.num_input_sgprs +
8248 key->ps_prolog.face_vgpr_index;
8249 LLVMValueRef interp[2], color[4];
8250 LLVMValueRef interp_ij = NULL, prim_mask = NULL, face = NULL;
8251
8252 if (!writemask)
8253 continue;
8254
8255 /* If the interpolation qualifier is not CONSTANT (-1). */
8256 if (key->ps_prolog.color_interp_vgpr_index[i] != -1) {
8257 unsigned interp_vgpr = key->ps_prolog.num_input_sgprs +
8258 key->ps_prolog.color_interp_vgpr_index[i];
8259
8260 /* Get the (i,j) updated by bc_optimize handling. */
8261 interp[0] = LLVMBuildExtractValue(gallivm->builder, ret,
8262 interp_vgpr, "");
8263 interp[1] = LLVMBuildExtractValue(gallivm->builder, ret,
8264 interp_vgpr + 1, "");
8265 interp_ij = lp_build_gather_values(gallivm, interp, 2);
8266 }
8267
8268 /* Use the absolute location of the input. */
8269 prim_mask = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
8270
8271 if (key->ps_prolog.states.color_two_side) {
8272 face = LLVMGetParam(func, face_vgpr);
8273 face = LLVMBuildBitCast(gallivm->builder, face, ctx->i32, "");
8274 }
8275
8276 interp_fs_input(ctx,
8277 key->ps_prolog.color_attr_index[i],
8278 TGSI_SEMANTIC_COLOR, i,
8279 key->ps_prolog.num_interp_inputs,
8280 key->ps_prolog.colors_read, interp_ij,
8281 prim_mask, face, color);
8282
8283 while (writemask) {
8284 unsigned chan = u_bit_scan(&writemask);
8285 ret = LLVMBuildInsertValue(gallivm->builder, ret, color[chan],
8286 num_params++, "");
8287 }
8288 }
8289
8290 /* Tell LLVM to insert WQM instruction sequence when needed. */
8291 if (key->ps_prolog.wqm) {
8292 LLVMAddTargetDependentFunctionAttr(func,
8293 "amdgpu-ps-wqm-outputs", "");
8294 }
8295
8296 si_llvm_build_ret(ctx, ret);
8297 }
8298
8299 /**
8300 * Build the pixel shader epilog function. This handles everything that must be
8301 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
8302 */
8303 static void si_build_ps_epilog_function(struct si_shader_context *ctx,
8304 union si_shader_part_key *key)
8305 {
8306 struct gallivm_state *gallivm = &ctx->gallivm;
8307 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
8308 LLVMTypeRef params[16+8*4+3];
8309 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
8310 int last_sgpr, num_params = 0, i;
8311 struct si_ps_exports exp = {};
8312
8313 /* Declare input SGPRs. */
8314 params[ctx->param_rw_buffers = num_params++] = ctx->i64;
8315 params[ctx->param_const_buffers = num_params++] = ctx->i64;
8316 params[ctx->param_samplers = num_params++] = ctx->i64;
8317 params[ctx->param_images = num_params++] = ctx->i64;
8318 params[ctx->param_shader_buffers = num_params++] = ctx->i64;
8319 assert(num_params == SI_PARAM_ALPHA_REF);
8320 params[SI_PARAM_ALPHA_REF] = ctx->f32;
8321 last_sgpr = SI_PARAM_ALPHA_REF;
8322
8323 /* Declare input VGPRs. */
8324 num_params = (last_sgpr + 1) +
8325 util_bitcount(key->ps_epilog.colors_written) * 4 +
8326 key->ps_epilog.writes_z +
8327 key->ps_epilog.writes_stencil +
8328 key->ps_epilog.writes_samplemask;
8329
8330 num_params = MAX2(num_params,
8331 last_sgpr + 1 + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
8332
8333 assert(num_params <= ARRAY_SIZE(params));
8334
8335 for (i = last_sgpr + 1; i < num_params; i++)
8336 params[i] = ctx->f32;
8337
8338 /* Create the function. */
8339 si_create_function(ctx, "ps_epilog", NULL, 0, params, num_params, last_sgpr);
8340 /* Disable elimination of unused inputs. */
8341 si_llvm_add_attribute(ctx->main_fn,
8342 "InitialPSInputAddr", 0xffffff);
8343
8344 /* Process colors. */
8345 unsigned vgpr = last_sgpr + 1;
8346 unsigned colors_written = key->ps_epilog.colors_written;
8347 int last_color_export = -1;
8348
8349 /* Find the last color export. */
8350 if (!key->ps_epilog.writes_z &&
8351 !key->ps_epilog.writes_stencil &&
8352 !key->ps_epilog.writes_samplemask) {
8353 unsigned spi_format = key->ps_epilog.states.spi_shader_col_format;
8354
8355 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
8356 if (colors_written == 0x1 && key->ps_epilog.states.last_cbuf > 0) {
8357 /* Just set this if any of the colorbuffers are enabled. */
8358 if (spi_format &
8359 ((1llu << (4 * (key->ps_epilog.states.last_cbuf + 1))) - 1))
8360 last_color_export = 0;
8361 } else {
8362 for (i = 0; i < 8; i++)
8363 if (colors_written & (1 << i) &&
8364 (spi_format >> (i * 4)) & 0xf)
8365 last_color_export = i;
8366 }
8367 }
8368
8369 while (colors_written) {
8370 LLVMValueRef color[4];
8371 int mrt = u_bit_scan(&colors_written);
8372
8373 for (i = 0; i < 4; i++)
8374 color[i] = LLVMGetParam(ctx->main_fn, vgpr++);
8375
8376 si_export_mrt_color(bld_base, color, mrt,
8377 num_params - 1,
8378 mrt == last_color_export, &exp);
8379 }
8380
8381 /* Process depth, stencil, samplemask. */
8382 if (key->ps_epilog.writes_z)
8383 depth = LLVMGetParam(ctx->main_fn, vgpr++);
8384 if (key->ps_epilog.writes_stencil)
8385 stencil = LLVMGetParam(ctx->main_fn, vgpr++);
8386 if (key->ps_epilog.writes_samplemask)
8387 samplemask = LLVMGetParam(ctx->main_fn, vgpr++);
8388
8389 if (depth || stencil || samplemask)
8390 si_export_mrt_z(bld_base, depth, stencil, samplemask, &exp);
8391 else if (last_color_export == -1)
8392 si_export_null(bld_base);
8393
8394 if (exp.num)
8395 si_emit_ps_exports(ctx, &exp);
8396
8397 /* Compile. */
8398 LLVMBuildRetVoid(gallivm->builder);
8399 }
8400
8401 /**
8402 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
8403 */
8404 static bool si_shader_select_ps_parts(struct si_screen *sscreen,
8405 LLVMTargetMachineRef tm,
8406 struct si_shader *shader,
8407 struct pipe_debug_callback *debug)
8408 {
8409 union si_shader_part_key prolog_key;
8410 union si_shader_part_key epilog_key;
8411
8412 /* Get the prolog. */
8413 si_get_ps_prolog_key(shader, &prolog_key, true);
8414
8415 /* The prolog is a no-op if these aren't set. */
8416 if (si_need_ps_prolog(&prolog_key)) {
8417 shader->prolog =
8418 si_get_shader_part(sscreen, &sscreen->ps_prologs,
8419 PIPE_SHADER_FRAGMENT, true,
8420 &prolog_key, tm, debug,
8421 si_build_ps_prolog_function,
8422 "Fragment Shader Prolog");
8423 if (!shader->prolog)
8424 return false;
8425 }
8426
8427 /* Get the epilog. */
8428 si_get_ps_epilog_key(shader, &epilog_key);
8429
8430 shader->epilog =
8431 si_get_shader_part(sscreen, &sscreen->ps_epilogs,
8432 PIPE_SHADER_FRAGMENT, false,
8433 &epilog_key, tm, debug,
8434 si_build_ps_epilog_function,
8435 "Fragment Shader Epilog");
8436 if (!shader->epilog)
8437 return false;
8438
8439 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
8440 if (shader->key.part.ps.prolog.poly_stipple) {
8441 shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
8442 assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
8443 }
8444
8445 /* Set up the enable bits for per-sample shading if needed. */
8446 if (shader->key.part.ps.prolog.force_persp_sample_interp &&
8447 (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
8448 G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8449 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
8450 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
8451 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
8452 }
8453 if (shader->key.part.ps.prolog.force_linear_sample_interp &&
8454 (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
8455 G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8456 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
8457 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
8458 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
8459 }
8460 if (shader->key.part.ps.prolog.force_persp_center_interp &&
8461 (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
8462 G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8463 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
8464 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
8465 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
8466 }
8467 if (shader->key.part.ps.prolog.force_linear_center_interp &&
8468 (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
8469 G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8470 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
8471 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
8472 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
8473 }
8474
8475 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
8476 if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_ena) &&
8477 !(shader->config.spi_ps_input_ena & 0xf)) {
8478 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
8479 assert(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr));
8480 }
8481
8482 /* At least one pair of interpolation weights must be enabled. */
8483 if (!(shader->config.spi_ps_input_ena & 0x7f)) {
8484 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
8485 assert(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr));
8486 }
8487
8488 /* The sample mask input is always enabled, because the API shader always
8489 * passes it through to the epilog. Disable it here if it's unused.
8490 */
8491 if (!shader->key.part.ps.epilog.poly_line_smoothing &&
8492 !shader->selector->info.reads_samplemask)
8493 shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;
8494
8495 return true;
8496 }
8497
8498 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
8499 unsigned *lds_size)
8500 {
8501 /* SPI barrier management bug:
8502 * Make sure we have at least 4k of LDS in use to avoid the bug.
8503 * It applies to workgroup sizes of more than one wavefront.
8504 */
8505 if (sscreen->b.family == CHIP_BONAIRE ||
8506 sscreen->b.family == CHIP_KABINI ||
8507 sscreen->b.family == CHIP_MULLINS)
8508 *lds_size = MAX2(*lds_size, 8);
8509 }
8510
8511 static void si_fix_resource_usage(struct si_screen *sscreen,
8512 struct si_shader *shader)
8513 {
8514 unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */
8515
8516 shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
8517
8518 if (shader->selector->type == PIPE_SHADER_COMPUTE &&
8519 si_get_max_workgroup_size(shader) > 64) {
8520 si_multiwave_lds_size_workaround(sscreen,
8521 &shader->config.lds_size);
8522 }
8523 }
8524
8525 int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
8526 struct si_shader *shader,
8527 struct pipe_debug_callback *debug)
8528 {
8529 struct si_shader_selector *sel = shader->selector;
8530 struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key);
8531 int r;
8532
8533 /* LS, ES, VS are compiled on demand if the main part hasn't been
8534 * compiled for that stage.
8535 *
8536 * Vertex shaders are compiled on demand when a vertex fetch
8537 * workaround must be applied.
8538 */
8539 if (shader->is_monolithic) {
8540 /* Monolithic shader (compiled as a whole, has many variants,
8541 * may take a long time to compile).
8542 */
8543 r = si_compile_tgsi_shader(sscreen, tm, shader, true, debug);
8544 if (r)
8545 return r;
8546 } else {
8547 /* The shader consists of 2-3 parts:
8548 *
8549 * - the middle part is the user shader, it has 1 variant only
8550 * and it was compiled during the creation of the shader
8551 * selector
8552 * - the prolog part is inserted at the beginning
8553 * - the epilog part is inserted at the end
8554 *
8555 * The prolog and epilog have many (but simple) variants.
8556 */
8557
8558 /* Copy the compiled TGSI shader data over. */
8559 shader->is_binary_shared = true;
8560 shader->binary = mainp->binary;
8561 shader->config = mainp->config;
8562 shader->info.num_input_sgprs = mainp->info.num_input_sgprs;
8563 shader->info.num_input_vgprs = mainp->info.num_input_vgprs;
8564 shader->info.face_vgpr_index = mainp->info.face_vgpr_index;
8565 memcpy(shader->info.vs_output_param_offset,
8566 mainp->info.vs_output_param_offset,
8567 sizeof(mainp->info.vs_output_param_offset));
8568 shader->info.uses_instanceid = mainp->info.uses_instanceid;
8569 shader->info.nr_pos_exports = mainp->info.nr_pos_exports;
8570 shader->info.nr_param_exports = mainp->info.nr_param_exports;
8571
8572 /* Select prologs and/or epilogs. */
8573 switch (sel->type) {
8574 case PIPE_SHADER_VERTEX:
8575 if (!si_shader_select_vs_parts(sscreen, tm, shader, debug))
8576 return -1;
8577 break;
8578 case PIPE_SHADER_TESS_CTRL:
8579 if (!si_shader_select_tcs_parts(sscreen, tm, shader, debug))
8580 return -1;
8581 break;
8582 case PIPE_SHADER_TESS_EVAL:
8583 if (!si_shader_select_tes_parts(sscreen, tm, shader, debug))
8584 return -1;
8585 break;
8586 case PIPE_SHADER_GEOMETRY:
8587 if (!si_shader_select_gs_parts(sscreen, tm, shader, debug))
8588 return -1;
8589 break;
8590 case PIPE_SHADER_FRAGMENT:
8591 if (!si_shader_select_ps_parts(sscreen, tm, shader, debug))
8592 return -1;
8593
8594 /* Make sure we have at least as many VGPRs as there
8595 * are allocated inputs.
8596 */
8597 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8598 shader->info.num_input_vgprs);
8599 break;
8600 }
8601
8602 /* Update SGPR and VGPR counts. */
8603 if (shader->prolog) {
8604 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
8605 shader->prolog->config.num_sgprs);
8606 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8607 shader->prolog->config.num_vgprs);
8608 }
8609 if (shader->previous_stage) {
8610 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
8611 shader->previous_stage->config.num_sgprs);
8612 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8613 shader->previous_stage->config.num_vgprs);
8614 shader->config.spilled_sgprs =
8615 MAX2(shader->config.spilled_sgprs,
8616 shader->previous_stage->config.spilled_sgprs);
8617 shader->config.spilled_vgprs =
8618 MAX2(shader->config.spilled_vgprs,
8619 shader->previous_stage->config.spilled_vgprs);
8620 shader->config.private_mem_vgprs =
8621 MAX2(shader->config.private_mem_vgprs,
8622 shader->previous_stage->config.private_mem_vgprs);
8623 shader->config.scratch_bytes_per_wave =
8624 MAX2(shader->config.scratch_bytes_per_wave,
8625 shader->previous_stage->config.scratch_bytes_per_wave);
8626 shader->info.uses_instanceid |=
8627 shader->previous_stage->info.uses_instanceid;
8628 }
8629 if (shader->epilog) {
8630 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
8631 shader->epilog->config.num_sgprs);
8632 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8633 shader->epilog->config.num_vgprs);
8634 }
8635 }
8636
8637 si_fix_resource_usage(sscreen, shader);
8638 si_shader_dump(sscreen, shader, debug, sel->info.processor,
8639 stderr, true);
8640
8641 /* Upload. */
8642 r = si_shader_binary_upload(sscreen, shader);
8643 if (r) {
8644 fprintf(stderr, "LLVM failed to upload shader\n");
8645 return r;
8646 }
8647
8648 return 0;
8649 }
8650
8651 void si_shader_destroy(struct si_shader *shader)
8652 {
8653 if (shader->scratch_bo)
8654 r600_resource_reference(&shader->scratch_bo, NULL);
8655
8656 r600_resource_reference(&shader->bo, NULL);
8657
8658 if (!shader->is_binary_shared)
8659 radeon_shader_binary_clean(&shader->binary);
8660
8661 free(shader->shader_log);
8662 }