projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: shrink each vertex stream to the actually required size
[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 "radeon/radeon_elf_util.h"
37 #include "util/u_memory.h"
38 #include "util/u_string.h"
39 #include "tgsi/tgsi_build.h"
40 #include "tgsi/tgsi_util.h"
41 #include "tgsi/tgsi_dump.h"
42
43 #include "ac_llvm_util.h"
44 #include "si_shader_internal.h"
45 #include "si_pipe.h"
46 #include "sid.h"
47
48
49 static const char *scratch_rsrc_dword0_symbol =
50 "SCRATCH_RSRC_DWORD0";
51
52 static const char *scratch_rsrc_dword1_symbol =
53 "SCRATCH_RSRC_DWORD1";
54
55 struct si_shader_output_values
56 {
57 LLVMValueRef values[4];
58 unsigned semantic_name;
59 unsigned semantic_index;
60 ubyte vertex_stream[4];
61 };
62
63 static void si_init_shader_ctx(struct si_shader_context *ctx,
64 struct si_screen *sscreen,
65 struct si_shader *shader,
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 shader, struct si_shader_key *key,
73 FILE *f);
74
75 static void si_build_vs_prolog_function(struct si_shader_context *ctx,
76 union si_shader_part_key *key);
77 static void si_build_vs_epilog_function(struct si_shader_context *ctx,
78 union si_shader_part_key *key);
79 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
80 union si_shader_part_key *key);
81 static void si_build_ps_prolog_function(struct si_shader_context *ctx,
82 union si_shader_part_key *key);
83 static void si_build_ps_epilog_function(struct si_shader_context *ctx,
84 union si_shader_part_key *key);
85
86 /* Ideally pass the sample mask input to the PS epilog as v13, which
87 * is its usual location, so that the shader doesn't have to add v_mov.
88 */
89 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 13
90
91 /* The VS location of the PrimitiveID input is the same in the epilog,
92 * so that the main shader part doesn't have to move it.
93 */
94 #define VS_EPILOG_PRIMID_LOC 2
95
96 enum {
97 CONST_ADDR_SPACE = 2,
98 LOCAL_ADDR_SPACE = 3,
99 };
100
101 #define SENDMSG_GS 2
102 #define SENDMSG_GS_DONE 3
103
104 #define SENDMSG_GS_OP_NOP (0 << 4)
105 #define SENDMSG_GS_OP_CUT (1 << 4)
106 #define SENDMSG_GS_OP_EMIT (2 << 4)
107 #define SENDMSG_GS_OP_EMIT_CUT (3 << 4)
108
109 /**
110 * Returns a unique index for a semantic name and index. The index must be
111 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
112 * calculated.
113 */
114 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index)
115 {
116 switch (semantic_name) {
117 case TGSI_SEMANTIC_POSITION:
118 return 0;
119 case TGSI_SEMANTIC_PSIZE:
120 return 1;
121 case TGSI_SEMANTIC_CLIPDIST:
122 assert(index <= 1);
123 return 2 + index;
124 case TGSI_SEMANTIC_GENERIC:
125 if (index <= 63-4)
126 return 4 + index;
127
128 assert(!"invalid generic index");
129 return 0;
130
131 /* patch indices are completely separate and thus start from 0 */
132 case TGSI_SEMANTIC_TESSOUTER:
133 return 0;
134 case TGSI_SEMANTIC_TESSINNER:
135 return 1;
136 case TGSI_SEMANTIC_PATCH:
137 return 2 + index;
138
139 default:
140 assert(!"invalid semantic name");
141 return 0;
142 }
143 }
144
145 unsigned si_shader_io_get_unique_index2(unsigned name, unsigned index)
146 {
147 switch (name) {
148 case TGSI_SEMANTIC_FOG:
149 return 0;
150 case TGSI_SEMANTIC_LAYER:
151 return 1;
152 case TGSI_SEMANTIC_VIEWPORT_INDEX:
153 return 2;
154 case TGSI_SEMANTIC_PRIMID:
155 return 3;
156 case TGSI_SEMANTIC_COLOR: /* these alias */
157 case TGSI_SEMANTIC_BCOLOR:
158 return 4 + index;
159 case TGSI_SEMANTIC_TEXCOORD:
160 return 6 + index;
161 default:
162 assert(!"invalid semantic name");
163 return 0;
164 }
165 }
166
167 /**
168 * Get the value of a shader input parameter and extract a bitfield.
169 */
170 static LLVMValueRef unpack_param(struct si_shader_context *ctx,
171 unsigned param, unsigned rshift,
172 unsigned bitwidth)
173 {
174 struct gallivm_state *gallivm = &ctx->gallivm;
175 LLVMValueRef value = LLVMGetParam(ctx->main_fn,
176 param);
177
178 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
179 value = bitcast(&ctx->soa.bld_base,
180 TGSI_TYPE_UNSIGNED, value);
181
182 if (rshift)
183 value = LLVMBuildLShr(gallivm->builder, value,
184 lp_build_const_int32(gallivm, rshift), "");
185
186 if (rshift + bitwidth < 32) {
187 unsigned mask = (1 << bitwidth) - 1;
188 value = LLVMBuildAnd(gallivm->builder, value,
189 lp_build_const_int32(gallivm, mask), "");
190 }
191
192 return value;
193 }
194
195 static LLVMValueRef get_rel_patch_id(struct si_shader_context *ctx)
196 {
197 switch (ctx->type) {
198 case PIPE_SHADER_TESS_CTRL:
199 return unpack_param(ctx, SI_PARAM_REL_IDS, 0, 8);
200
201 case PIPE_SHADER_TESS_EVAL:
202 return LLVMGetParam(ctx->main_fn,
203 ctx->param_tes_rel_patch_id);
204
205 default:
206 assert(0);
207 return NULL;
208 }
209 }
210
211 /* Tessellation shaders pass outputs to the next shader using LDS.
212 *
213 * LS outputs = TCS inputs
214 * TCS outputs = TES inputs
215 *
216 * The LDS layout is:
217 * - TCS inputs for patch 0
218 * - TCS inputs for patch 1
219 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
220 * - ...
221 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
222 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
223 * - TCS outputs for patch 1
224 * - Per-patch TCS outputs for patch 1
225 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
226 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
227 * - ...
228 *
229 * All three shaders VS(LS), TCS, TES share the same LDS space.
230 */
231
232 static LLVMValueRef
233 get_tcs_in_patch_stride(struct si_shader_context *ctx)
234 {
235 if (ctx->type == PIPE_SHADER_VERTEX)
236 return unpack_param(ctx, SI_PARAM_LS_OUT_LAYOUT, 0, 13);
237 else if (ctx->type == PIPE_SHADER_TESS_CTRL)
238 return unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 0, 13);
239 else {
240 assert(0);
241 return NULL;
242 }
243 }
244
245 static LLVMValueRef
246 get_tcs_out_patch_stride(struct si_shader_context *ctx)
247 {
248 return unpack_param(ctx, SI_PARAM_TCS_OUT_LAYOUT, 0, 13);
249 }
250
251 static LLVMValueRef
252 get_tcs_out_patch0_offset(struct si_shader_context *ctx)
253 {
254 return lp_build_mul_imm(&ctx->soa.bld_base.uint_bld,
255 unpack_param(ctx,
256 SI_PARAM_TCS_OUT_OFFSETS,
257 0, 16),
258 4);
259 }
260
261 static LLVMValueRef
262 get_tcs_out_patch0_patch_data_offset(struct si_shader_context *ctx)
263 {
264 return lp_build_mul_imm(&ctx->soa.bld_base.uint_bld,
265 unpack_param(ctx,
266 SI_PARAM_TCS_OUT_OFFSETS,
267 16, 16),
268 4);
269 }
270
271 static LLVMValueRef
272 get_tcs_in_current_patch_offset(struct si_shader_context *ctx)
273 {
274 struct gallivm_state *gallivm = &ctx->gallivm;
275 LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
276 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
277
278 return LLVMBuildMul(gallivm->builder, patch_stride, rel_patch_id, "");
279 }
280
281 static LLVMValueRef
282 get_tcs_out_current_patch_offset(struct si_shader_context *ctx)
283 {
284 struct gallivm_state *gallivm = &ctx->gallivm;
285 LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
286 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
287 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
288
289 return LLVMBuildAdd(gallivm->builder, patch0_offset,
290 LLVMBuildMul(gallivm->builder, patch_stride,
291 rel_patch_id, ""),
292 "");
293 }
294
295 static LLVMValueRef
296 get_tcs_out_current_patch_data_offset(struct si_shader_context *ctx)
297 {
298 struct gallivm_state *gallivm = &ctx->gallivm;
299 LLVMValueRef patch0_patch_data_offset =
300 get_tcs_out_patch0_patch_data_offset(ctx);
301 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
302 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
303
304 return LLVMBuildAdd(gallivm->builder, patch0_patch_data_offset,
305 LLVMBuildMul(gallivm->builder, patch_stride,
306 rel_patch_id, ""),
307 "");
308 }
309
310 static LLVMValueRef build_gep0(struct si_shader_context *ctx,
311 LLVMValueRef base_ptr, LLVMValueRef index)
312 {
313 LLVMValueRef indices[2] = {
314 LLVMConstInt(ctx->i32, 0, 0),
315 index,
316 };
317 return LLVMBuildGEP(ctx->gallivm.builder, base_ptr,
318 indices, 2, "");
319 }
320
321 static void build_indexed_store(struct si_shader_context *ctx,
322 LLVMValueRef base_ptr, LLVMValueRef index,
323 LLVMValueRef value)
324 {
325 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
326 struct gallivm_state *gallivm = bld_base->base.gallivm;
327
328 LLVMBuildStore(gallivm->builder, value,
329 build_gep0(ctx, base_ptr, index));
330 }
331
332 /**
333 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
334 * It's equivalent to doing a load from &base_ptr[index].
335 *
336 * \param base_ptr Where the array starts.
337 * \param index The element index into the array.
338 * \param uniform Whether the base_ptr and index can be assumed to be
339 * dynamically uniform
340 */
341 static LLVMValueRef build_indexed_load(struct si_shader_context *ctx,
342 LLVMValueRef base_ptr, LLVMValueRef index,
343 bool uniform)
344 {
345 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
346 struct gallivm_state *gallivm = bld_base->base.gallivm;
347 LLVMValueRef pointer;
348
349 pointer = build_gep0(ctx, base_ptr, index);
350 if (uniform)
351 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
352 return LLVMBuildLoad(gallivm->builder, pointer, "");
353 }
354
355 /**
356 * Do a load from &base_ptr[index], but also add a flag that it's loading
357 * a constant from a dynamically uniform index.
358 */
359 static LLVMValueRef build_indexed_load_const(
360 struct si_shader_context *ctx,
361 LLVMValueRef base_ptr, LLVMValueRef index)
362 {
363 LLVMValueRef result = build_indexed_load(ctx, base_ptr, index, true);
364 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
365 return result;
366 }
367
368 static LLVMValueRef get_instance_index_for_fetch(
369 struct si_shader_context *radeon_bld,
370 unsigned param_start_instance, unsigned divisor)
371 {
372 struct si_shader_context *ctx =
373 si_shader_context(&radeon_bld->soa.bld_base);
374 struct gallivm_state *gallivm = radeon_bld->soa.bld_base.base.gallivm;
375
376 LLVMValueRef result = LLVMGetParam(radeon_bld->main_fn,
377 ctx->param_instance_id);
378
379 /* The division must be done before START_INSTANCE is added. */
380 if (divisor > 1)
381 result = LLVMBuildUDiv(gallivm->builder, result,
382 lp_build_const_int32(gallivm, divisor), "");
383
384 return LLVMBuildAdd(gallivm->builder, result,
385 LLVMGetParam(radeon_bld->main_fn, param_start_instance), "");
386 }
387
388 static void declare_input_vs(
389 struct si_shader_context *ctx,
390 unsigned input_index,
391 const struct tgsi_full_declaration *decl,
392 LLVMValueRef out[4])
393 {
394 struct lp_build_context *base = &ctx->soa.bld_base.base;
395 struct gallivm_state *gallivm = base->gallivm;
396
397 unsigned chan;
398 unsigned fix_fetch;
399
400 LLVMValueRef t_list_ptr;
401 LLVMValueRef t_offset;
402 LLVMValueRef t_list;
403 LLVMValueRef attribute_offset;
404 LLVMValueRef buffer_index;
405 LLVMValueRef args[3];
406 LLVMValueRef input;
407
408 /* Load the T list */
409 t_list_ptr = LLVMGetParam(ctx->main_fn, SI_PARAM_VERTEX_BUFFERS);
410
411 t_offset = lp_build_const_int32(gallivm, input_index);
412
413 t_list = build_indexed_load_const(ctx, t_list_ptr, t_offset);
414
415 /* Build the attribute offset */
416 attribute_offset = lp_build_const_int32(gallivm, 0);
417
418 buffer_index = LLVMGetParam(ctx->main_fn,
419 ctx->param_vertex_index0 +
420 input_index);
421
422 args[0] = t_list;
423 args[1] = attribute_offset;
424 args[2] = buffer_index;
425 input = lp_build_intrinsic(gallivm->builder,
426 "llvm.SI.vs.load.input", ctx->v4f32, args, 3,
427 LP_FUNC_ATTR_READNONE);
428
429 /* Break up the vec4 into individual components */
430 for (chan = 0; chan < 4; chan++) {
431 LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
432 out[chan] = LLVMBuildExtractElement(gallivm->builder,
433 input, llvm_chan, "");
434 }
435
436 fix_fetch = (ctx->shader->key.mono.vs.fix_fetch >> (2 * input_index)) & 3;
437 if (fix_fetch) {
438 /* The hardware returns an unsigned value; convert it to a
439 * signed one.
440 */
441 LLVMValueRef tmp = out[3];
442 LLVMValueRef c30 = LLVMConstInt(ctx->i32, 30, 0);
443
444 /* First, recover the sign-extended signed integer value. */
445 if (fix_fetch == SI_FIX_FETCH_A2_SSCALED)
446 tmp = LLVMBuildFPToUI(gallivm->builder, tmp, ctx->i32, "");
447 else
448 tmp = LLVMBuildBitCast(gallivm->builder, tmp, ctx->i32, "");
449
450 /* For the integer-like cases, do a natural sign extension.
451 *
452 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
453 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
454 * exponent.
455 */
456 tmp = LLVMBuildShl(gallivm->builder, tmp,
457 fix_fetch == SI_FIX_FETCH_A2_SNORM ?
458 LLVMConstInt(ctx->i32, 7, 0) : c30, "");
459 tmp = LLVMBuildAShr(gallivm->builder, tmp, c30, "");
460
461 /* Convert back to the right type. */
462 if (fix_fetch == SI_FIX_FETCH_A2_SNORM) {
463 LLVMValueRef clamp;
464 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
465 tmp = LLVMBuildSIToFP(gallivm->builder, tmp, ctx->f32, "");
466 clamp = LLVMBuildFCmp(gallivm->builder, LLVMRealULT, tmp, neg_one, "");
467 tmp = LLVMBuildSelect(gallivm->builder, clamp, neg_one, tmp, "");
468 } else if (fix_fetch == SI_FIX_FETCH_A2_SSCALED) {
469 tmp = LLVMBuildSIToFP(gallivm->builder, tmp, ctx->f32, "");
470 }
471
472 out[3] = tmp;
473 }
474 }
475
476 static LLVMValueRef get_primitive_id(struct lp_build_tgsi_context *bld_base,
477 unsigned swizzle)
478 {
479 struct si_shader_context *ctx = si_shader_context(bld_base);
480
481 if (swizzle > 0)
482 return bld_base->uint_bld.zero;
483
484 switch (ctx->type) {
485 case PIPE_SHADER_VERTEX:
486 return LLVMGetParam(ctx->main_fn,
487 ctx->param_vs_prim_id);
488 case PIPE_SHADER_TESS_CTRL:
489 return LLVMGetParam(ctx->main_fn,
490 SI_PARAM_PATCH_ID);
491 case PIPE_SHADER_TESS_EVAL:
492 return LLVMGetParam(ctx->main_fn,
493 ctx->param_tes_patch_id);
494 case PIPE_SHADER_GEOMETRY:
495 return LLVMGetParam(ctx->main_fn,
496 SI_PARAM_PRIMITIVE_ID);
497 default:
498 assert(0);
499 return bld_base->uint_bld.zero;
500 }
501 }
502
503 /**
504 * Return the value of tgsi_ind_register for indexing.
505 * This is the indirect index with the constant offset added to it.
506 */
507 static LLVMValueRef get_indirect_index(struct si_shader_context *ctx,
508 const struct tgsi_ind_register *ind,
509 int rel_index)
510 {
511 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
512 LLVMValueRef result;
513
514 result = ctx->soa.addr[ind->Index][ind->Swizzle];
515 result = LLVMBuildLoad(gallivm->builder, result, "");
516 result = LLVMBuildAdd(gallivm->builder, result,
517 lp_build_const_int32(gallivm, rel_index), "");
518 return result;
519 }
520
521 /**
522 * Like get_indirect_index, but restricts the return value to a (possibly
523 * undefined) value inside [0..num).
524 */
525 static LLVMValueRef get_bounded_indirect_index(struct si_shader_context *ctx,
526 const struct tgsi_ind_register *ind,
527 int rel_index, unsigned num)
528 {
529 LLVMValueRef result = get_indirect_index(ctx, ind, rel_index);
530
531 /* LLVM 3.8: If indirect resource indexing is used:
532 * - SI & CIK hang
533 * - VI crashes
534 */
535 if (HAVE_LLVM <= 0x0308)
536 return LLVMGetUndef(ctx->i32);
537
538 return si_llvm_bound_index(ctx, result, num);
539 }
540
541
542 /**
543 * Calculate a dword address given an input or output register and a stride.
544 */
545 static LLVMValueRef get_dw_address(struct si_shader_context *ctx,
546 const struct tgsi_full_dst_register *dst,
547 const struct tgsi_full_src_register *src,
548 LLVMValueRef vertex_dw_stride,
549 LLVMValueRef base_addr)
550 {
551 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
552 struct tgsi_shader_info *info = &ctx->shader->selector->info;
553 ubyte *name, *index, *array_first;
554 int first, param;
555 struct tgsi_full_dst_register reg;
556
557 /* Set the register description. The address computation is the same
558 * for sources and destinations. */
559 if (src) {
560 reg.Register.File = src->Register.File;
561 reg.Register.Index = src->Register.Index;
562 reg.Register.Indirect = src->Register.Indirect;
563 reg.Register.Dimension = src->Register.Dimension;
564 reg.Indirect = src->Indirect;
565 reg.Dimension = src->Dimension;
566 reg.DimIndirect = src->DimIndirect;
567 } else
568 reg = *dst;
569
570 /* If the register is 2-dimensional (e.g. an array of vertices
571 * in a primitive), calculate the base address of the vertex. */
572 if (reg.Register.Dimension) {
573 LLVMValueRef index;
574
575 if (reg.Dimension.Indirect)
576 index = get_indirect_index(ctx, &reg.DimIndirect,
577 reg.Dimension.Index);
578 else
579 index = lp_build_const_int32(gallivm, reg.Dimension.Index);
580
581 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
582 LLVMBuildMul(gallivm->builder, index,
583 vertex_dw_stride, ""), "");
584 }
585
586 /* Get information about the register. */
587 if (reg.Register.File == TGSI_FILE_INPUT) {
588 name = info->input_semantic_name;
589 index = info->input_semantic_index;
590 array_first = info->input_array_first;
591 } else if (reg.Register.File == TGSI_FILE_OUTPUT) {
592 name = info->output_semantic_name;
593 index = info->output_semantic_index;
594 array_first = info->output_array_first;
595 } else {
596 assert(0);
597 return NULL;
598 }
599
600 if (reg.Register.Indirect) {
601 /* Add the relative address of the element. */
602 LLVMValueRef ind_index;
603
604 if (reg.Indirect.ArrayID)
605 first = array_first[reg.Indirect.ArrayID];
606 else
607 first = reg.Register.Index;
608
609 ind_index = get_indirect_index(ctx, &reg.Indirect,
610 reg.Register.Index - first);
611
612 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
613 LLVMBuildMul(gallivm->builder, ind_index,
614 lp_build_const_int32(gallivm, 4), ""), "");
615
616 param = si_shader_io_get_unique_index(name[first], index[first]);
617 } else {
618 param = si_shader_io_get_unique_index(name[reg.Register.Index],
619 index[reg.Register.Index]);
620 }
621
622 /* Add the base address of the element. */
623 return LLVMBuildAdd(gallivm->builder, base_addr,
624 lp_build_const_int32(gallivm, param * 4), "");
625 }
626
627 /* The offchip buffer layout for TCS->TES is
628 *
629 * - attribute 0 of patch 0 vertex 0
630 * - attribute 0 of patch 0 vertex 1
631 * - attribute 0 of patch 0 vertex 2
632 * ...
633 * - attribute 0 of patch 1 vertex 0
634 * - attribute 0 of patch 1 vertex 1
635 * ...
636 * - attribute 1 of patch 0 vertex 0
637 * - attribute 1 of patch 0 vertex 1
638 * ...
639 * - per patch attribute 0 of patch 0
640 * - per patch attribute 0 of patch 1
641 * ...
642 *
643 * Note that every attribute has 4 components.
644 */
645 static LLVMValueRef get_tcs_tes_buffer_address(struct si_shader_context *ctx,
646 LLVMValueRef vertex_index,
647 LLVMValueRef param_index)
648 {
649 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
650 LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
651 LLVMValueRef param_stride, constant16;
652
653 vertices_per_patch = unpack_param(ctx, SI_PARAM_TCS_OFFCHIP_LAYOUT, 9, 6);
654 num_patches = unpack_param(ctx, SI_PARAM_TCS_OFFCHIP_LAYOUT, 0, 9);
655 total_vertices = LLVMBuildMul(gallivm->builder, vertices_per_patch,
656 num_patches, "");
657
658 constant16 = lp_build_const_int32(gallivm, 16);
659 if (vertex_index) {
660 base_addr = LLVMBuildMul(gallivm->builder, get_rel_patch_id(ctx),
661 vertices_per_patch, "");
662
663 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
664 vertex_index, "");
665
666 param_stride = total_vertices;
667 } else {
668 base_addr = get_rel_patch_id(ctx);
669 param_stride = num_patches;
670 }
671
672 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
673 LLVMBuildMul(gallivm->builder, param_index,
674 param_stride, ""), "");
675
676 base_addr = LLVMBuildMul(gallivm->builder, base_addr, constant16, "");
677
678 if (!vertex_index) {
679 LLVMValueRef patch_data_offset =
680 unpack_param(ctx, SI_PARAM_TCS_OFFCHIP_LAYOUT, 16, 16);
681
682 base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
683 patch_data_offset, "");
684 }
685 return base_addr;
686 }
687
688 static LLVMValueRef get_tcs_tes_buffer_address_from_reg(
689 struct si_shader_context *ctx,
690 const struct tgsi_full_dst_register *dst,
691 const struct tgsi_full_src_register *src)
692 {
693 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
694 struct tgsi_shader_info *info = &ctx->shader->selector->info;
695 ubyte *name, *index, *array_first;
696 struct tgsi_full_src_register reg;
697 LLVMValueRef vertex_index = NULL;
698 LLVMValueRef param_index = NULL;
699 unsigned param_index_base, param_base;
700
701 reg = src ? *src : tgsi_full_src_register_from_dst(dst);
702
703 if (reg.Register.Dimension) {
704
705 if (reg.Dimension.Indirect)
706 vertex_index = get_indirect_index(ctx, &reg.DimIndirect,
707 reg.Dimension.Index);
708 else
709 vertex_index = lp_build_const_int32(gallivm,
710 reg.Dimension.Index);
711 }
712
713 /* Get information about the register. */
714 if (reg.Register.File == TGSI_FILE_INPUT) {
715 name = info->input_semantic_name;
716 index = info->input_semantic_index;
717 array_first = info->input_array_first;
718 } else if (reg.Register.File == TGSI_FILE_OUTPUT) {
719 name = info->output_semantic_name;
720 index = info->output_semantic_index;
721 array_first = info->output_array_first;
722 } else {
723 assert(0);
724 return NULL;
725 }
726
727 if (reg.Register.Indirect) {
728 if (reg.Indirect.ArrayID)
729 param_base = array_first[reg.Indirect.ArrayID];
730 else
731 param_base = reg.Register.Index;
732
733 param_index = get_indirect_index(ctx, &reg.Indirect,
734 reg.Register.Index - param_base);
735
736 } else {
737 param_base = reg.Register.Index;
738 param_index = lp_build_const_int32(gallivm, 0);
739 }
740
741 param_index_base = si_shader_io_get_unique_index(name[param_base],
742 index[param_base]);
743
744 param_index = LLVMBuildAdd(gallivm->builder, param_index,
745 lp_build_const_int32(gallivm, param_index_base),
746 "");
747
748 return get_tcs_tes_buffer_address(ctx, vertex_index, param_index);
749 }
750
751 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
752 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
753 * or v4i32 (num_channels=3,4). */
754 static void build_tbuffer_store(struct si_shader_context *ctx,
755 LLVMValueRef rsrc,
756 LLVMValueRef vdata,
757 unsigned num_channels,
758 LLVMValueRef vaddr,
759 LLVMValueRef soffset,
760 unsigned inst_offset,
761 unsigned dfmt,
762 unsigned nfmt,
763 unsigned offen,
764 unsigned idxen,
765 unsigned glc,
766 unsigned slc,
767 unsigned tfe)
768 {
769 struct gallivm_state *gallivm = &ctx->gallivm;
770 LLVMValueRef args[] = {
771 rsrc,
772 vdata,
773 LLVMConstInt(ctx->i32, num_channels, 0),
774 vaddr,
775 soffset,
776 LLVMConstInt(ctx->i32, inst_offset, 0),
777 LLVMConstInt(ctx->i32, dfmt, 0),
778 LLVMConstInt(ctx->i32, nfmt, 0),
779 LLVMConstInt(ctx->i32, offen, 0),
780 LLVMConstInt(ctx->i32, idxen, 0),
781 LLVMConstInt(ctx->i32, glc, 0),
782 LLVMConstInt(ctx->i32, slc, 0),
783 LLVMConstInt(ctx->i32, tfe, 0)
784 };
785
786 /* The instruction offset field has 12 bits */
787 assert(offen || inst_offset < (1 << 12));
788
789 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
790 unsigned func = CLAMP(num_channels, 1, 3) - 1;
791 const char *types[] = {"i32", "v2i32", "v4i32"};
792 char name[256];
793 snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]);
794
795 lp_build_intrinsic(gallivm->builder, name, ctx->voidt,
796 args, ARRAY_SIZE(args), 0);
797 }
798
799 static void build_tbuffer_store_dwords(struct si_shader_context *ctx,
800 LLVMValueRef rsrc,
801 LLVMValueRef vdata,
802 unsigned num_channels,
803 LLVMValueRef vaddr,
804 LLVMValueRef soffset,
805 unsigned inst_offset)
806 {
807 static unsigned dfmt[] = {
808 V_008F0C_BUF_DATA_FORMAT_32,
809 V_008F0C_BUF_DATA_FORMAT_32_32,
810 V_008F0C_BUF_DATA_FORMAT_32_32_32,
811 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
812 };
813 assert(num_channels >= 1 && num_channels <= 4);
814
815 build_tbuffer_store(ctx, rsrc, vdata, num_channels, vaddr, soffset,
816 inst_offset, dfmt[num_channels-1],
817 V_008F0C_BUF_NUM_FORMAT_UINT, 1, 0, 1, 1, 0);
818 }
819
820 static LLVMValueRef build_buffer_load(struct si_shader_context *ctx,
821 LLVMValueRef rsrc,
822 int num_channels,
823 LLVMValueRef vindex,
824 LLVMValueRef voffset,
825 LLVMValueRef soffset,
826 unsigned inst_offset,
827 unsigned glc,
828 unsigned slc)
829 {
830 struct gallivm_state *gallivm = &ctx->gallivm;
831 unsigned func = CLAMP(num_channels, 1, 3) - 1;
832
833 if (HAVE_LLVM >= 0x309) {
834 LLVMValueRef args[] = {
835 LLVMBuildBitCast(gallivm->builder, rsrc, ctx->v4i32, ""),
836 vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
837 LLVMConstInt(ctx->i32, inst_offset, 0),
838 LLVMConstInt(ctx->i1, glc, 0),
839 LLVMConstInt(ctx->i1, slc, 0)
840 };
841
842 LLVMTypeRef types[] = {ctx->f32, LLVMVectorType(ctx->f32, 2),
843 ctx->v4f32};
844 const char *type_names[] = {"f32", "v2f32", "v4f32"};
845 char name[256];
846
847 if (voffset) {
848 args[2] = LLVMBuildAdd(gallivm->builder, args[2], voffset,
849 "");
850 }
851
852 if (soffset) {
853 args[2] = LLVMBuildAdd(gallivm->builder, args[2], soffset,
854 "");
855 }
856
857 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
858 type_names[func]);
859
860 return lp_build_intrinsic(gallivm->builder, name, types[func], args,
861 ARRAY_SIZE(args), LP_FUNC_ATTR_READONLY);
862 } else {
863 LLVMValueRef args[] = {
864 LLVMBuildBitCast(gallivm->builder, rsrc, ctx->v16i8, ""),
865 voffset ? voffset : vindex,
866 soffset,
867 LLVMConstInt(ctx->i32, inst_offset, 0),
868 LLVMConstInt(ctx->i32, voffset ? 1 : 0, 0), // offen
869 LLVMConstInt(ctx->i32, vindex ? 1 : 0, 0), //idxen
870 LLVMConstInt(ctx->i32, glc, 0),
871 LLVMConstInt(ctx->i32, slc, 0),
872 LLVMConstInt(ctx->i32, 0, 0), // TFE
873 };
874
875 LLVMTypeRef types[] = {ctx->i32, LLVMVectorType(ctx->i32, 2),
876 ctx->v4i32};
877 const char *type_names[] = {"i32", "v2i32", "v4i32"};
878 const char *arg_type = "i32";
879 char name[256];
880
881 if (voffset && vindex) {
882 LLVMValueRef vaddr[] = {vindex, voffset};
883
884 arg_type = "v2i32";
885 args[1] = lp_build_gather_values(gallivm, vaddr, 2);
886 }
887
888 snprintf(name, sizeof(name), "llvm.SI.buffer.load.dword.%s.%s",
889 type_names[func], arg_type);
890
891 return lp_build_intrinsic(gallivm->builder, name, types[func], args,
892 ARRAY_SIZE(args), LP_FUNC_ATTR_READONLY);
893 }
894 }
895
896 static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base,
897 enum tgsi_opcode_type type, unsigned swizzle,
898 LLVMValueRef buffer, LLVMValueRef offset,
899 LLVMValueRef base)
900 {
901 struct si_shader_context *ctx = si_shader_context(bld_base);
902 struct gallivm_state *gallivm = bld_base->base.gallivm;
903 LLVMValueRef value, value2;
904 LLVMTypeRef llvm_type = tgsi2llvmtype(bld_base, type);
905 LLVMTypeRef vec_type = LLVMVectorType(llvm_type, 4);
906
907 if (swizzle == ~0) {
908 value = build_buffer_load(ctx, buffer, 4, NULL, base, offset,
909 0, 1, 0);
910
911 return LLVMBuildBitCast(gallivm->builder, value, vec_type, "");
912 }
913
914 if (!tgsi_type_is_64bit(type)) {
915 value = build_buffer_load(ctx, buffer, 4, NULL, base, offset,
916 0, 1, 0);
917
918 value = LLVMBuildBitCast(gallivm->builder, value, vec_type, "");
919 return LLVMBuildExtractElement(gallivm->builder, value,
920 lp_build_const_int32(gallivm, swizzle), "");
921 }
922
923 value = build_buffer_load(ctx, buffer, 1, NULL, base, offset,
924 swizzle * 4, 1, 0);
925
926 value2 = build_buffer_load(ctx, buffer, 1, NULL, base, offset,
927 swizzle * 4 + 4, 1, 0);
928
929 return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
930 }
931
932 /**
933 * Load from LDS.
934 *
935 * \param type output value type
936 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
937 * \param dw_addr address in dwords
938 */
939 static LLVMValueRef lds_load(struct lp_build_tgsi_context *bld_base,
940 enum tgsi_opcode_type type, unsigned swizzle,
941 LLVMValueRef dw_addr)
942 {
943 struct si_shader_context *ctx = si_shader_context(bld_base);
944 struct gallivm_state *gallivm = bld_base->base.gallivm;
945 LLVMValueRef value;
946
947 if (swizzle == ~0) {
948 LLVMValueRef values[TGSI_NUM_CHANNELS];
949
950 for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; chan++)
951 values[chan] = lds_load(bld_base, type, chan, dw_addr);
952
953 return lp_build_gather_values(bld_base->base.gallivm, values,
954 TGSI_NUM_CHANNELS);
955 }
956
957 dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
958 lp_build_const_int32(gallivm, swizzle));
959
960 value = build_indexed_load(ctx, ctx->lds, dw_addr, false);
961 if (tgsi_type_is_64bit(type)) {
962 LLVMValueRef value2;
963 dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
964 lp_build_const_int32(gallivm, 1));
965 value2 = build_indexed_load(ctx, ctx->lds, dw_addr, false);
966 return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
967 }
968
969 return LLVMBuildBitCast(gallivm->builder, value,
970 tgsi2llvmtype(bld_base, type), "");
971 }
972
973 /**
974 * Store to LDS.
975 *
976 * \param swizzle offset (typically 0..3)
977 * \param dw_addr address in dwords
978 * \param value value to store
979 */
980 static void lds_store(struct lp_build_tgsi_context *bld_base,
981 unsigned swizzle, LLVMValueRef dw_addr,
982 LLVMValueRef value)
983 {
984 struct si_shader_context *ctx = si_shader_context(bld_base);
985 struct gallivm_state *gallivm = bld_base->base.gallivm;
986
987 dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
988 lp_build_const_int32(gallivm, swizzle));
989
990 value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
991 build_indexed_store(ctx, ctx->lds,
992 dw_addr, value);
993 }
994
995 static LLVMValueRef fetch_input_tcs(
996 struct lp_build_tgsi_context *bld_base,
997 const struct tgsi_full_src_register *reg,
998 enum tgsi_opcode_type type, unsigned swizzle)
999 {
1000 struct si_shader_context *ctx = si_shader_context(bld_base);
1001 LLVMValueRef dw_addr, stride;
1002
1003 stride = unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 13, 8);
1004 dw_addr = get_tcs_in_current_patch_offset(ctx);
1005 dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
1006
1007 return lds_load(bld_base, type, swizzle, dw_addr);
1008 }
1009
1010 static LLVMValueRef fetch_output_tcs(
1011 struct lp_build_tgsi_context *bld_base,
1012 const struct tgsi_full_src_register *reg,
1013 enum tgsi_opcode_type type, unsigned swizzle)
1014 {
1015 struct si_shader_context *ctx = si_shader_context(bld_base);
1016 LLVMValueRef dw_addr, stride;
1017
1018 if (reg->Register.Dimension) {
1019 stride = unpack_param(ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8);
1020 dw_addr = get_tcs_out_current_patch_offset(ctx);
1021 dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
1022 } else {
1023 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
1024 dw_addr = get_dw_address(ctx, NULL, reg, NULL, dw_addr);
1025 }
1026
1027 return lds_load(bld_base, type, swizzle, dw_addr);
1028 }
1029
1030 static LLVMValueRef fetch_input_tes(
1031 struct lp_build_tgsi_context *bld_base,
1032 const struct tgsi_full_src_register *reg,
1033 enum tgsi_opcode_type type, unsigned swizzle)
1034 {
1035 struct si_shader_context *ctx = si_shader_context(bld_base);
1036 struct gallivm_state *gallivm = bld_base->base.gallivm;
1037 LLVMValueRef rw_buffers, buffer, base, addr;
1038
1039 rw_buffers = LLVMGetParam(ctx->main_fn,
1040 SI_PARAM_RW_BUFFERS);
1041 buffer = build_indexed_load_const(ctx, rw_buffers,
1042 lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
1043
1044 base = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
1045 addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg);
1046
1047 return buffer_load(bld_base, type, swizzle, buffer, base, addr);
1048 }
1049
1050 static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
1051 const struct tgsi_full_instruction *inst,
1052 const struct tgsi_opcode_info *info,
1053 LLVMValueRef dst[4])
1054 {
1055 struct si_shader_context *ctx = si_shader_context(bld_base);
1056 struct gallivm_state *gallivm = bld_base->base.gallivm;
1057 const struct tgsi_full_dst_register *reg = &inst->Dst[0];
1058 unsigned chan_index;
1059 LLVMValueRef dw_addr, stride;
1060 LLVMValueRef rw_buffers, buffer, base, buf_addr;
1061 LLVMValueRef values[4];
1062
1063 /* Only handle per-patch and per-vertex outputs here.
1064 * Vectors will be lowered to scalars and this function will be called again.
1065 */
1066 if (reg->Register.File != TGSI_FILE_OUTPUT ||
1067 (dst[0] && LLVMGetTypeKind(LLVMTypeOf(dst[0])) == LLVMVectorTypeKind)) {
1068 si_llvm_emit_store(bld_base, inst, info, dst);
1069 return;
1070 }
1071
1072 if (reg->Register.Dimension) {
1073 stride = unpack_param(ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8);
1074 dw_addr = get_tcs_out_current_patch_offset(ctx);
1075 dw_addr = get_dw_address(ctx, reg, NULL, stride, dw_addr);
1076 } else {
1077 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
1078 dw_addr = get_dw_address(ctx, reg, NULL, NULL, dw_addr);
1079 }
1080
1081 rw_buffers = LLVMGetParam(ctx->main_fn,
1082 SI_PARAM_RW_BUFFERS);
1083 buffer = build_indexed_load_const(ctx, rw_buffers,
1084 lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
1085
1086 base = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
1087 buf_addr = get_tcs_tes_buffer_address_from_reg(ctx, reg, NULL);
1088
1089
1090 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan_index) {
1091 LLVMValueRef value = dst[chan_index];
1092
1093 if (inst->Instruction.Saturate)
1094 value = si_llvm_saturate(bld_base, value);
1095
1096 lds_store(bld_base, chan_index, dw_addr, value);
1097
1098 value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
1099 values[chan_index] = value;
1100
1101 if (inst->Dst[0].Register.WriteMask != 0xF) {
1102 build_tbuffer_store_dwords(ctx, buffer, value, 1,
1103 buf_addr, base,
1104 4 * chan_index);
1105 }
1106 }
1107
1108 if (inst->Dst[0].Register.WriteMask == 0xF) {
1109 LLVMValueRef value = lp_build_gather_values(bld_base->base.gallivm,
1110 values, 4);
1111 build_tbuffer_store_dwords(ctx, buffer, value, 4, buf_addr,
1112 base, 0);
1113 }
1114 }
1115
1116 static LLVMValueRef fetch_input_gs(
1117 struct lp_build_tgsi_context *bld_base,
1118 const struct tgsi_full_src_register *reg,
1119 enum tgsi_opcode_type type,
1120 unsigned swizzle)
1121 {
1122 struct lp_build_context *base = &bld_base->base;
1123 struct si_shader_context *ctx = si_shader_context(bld_base);
1124 struct si_shader *shader = ctx->shader;
1125 struct lp_build_context *uint = &ctx->soa.bld_base.uint_bld;
1126 struct gallivm_state *gallivm = base->gallivm;
1127 LLVMValueRef vtx_offset;
1128 LLVMValueRef args[9];
1129 unsigned vtx_offset_param;
1130 struct tgsi_shader_info *info = &shader->selector->info;
1131 unsigned semantic_name = info->input_semantic_name[reg->Register.Index];
1132 unsigned semantic_index = info->input_semantic_index[reg->Register.Index];
1133 unsigned param;
1134 LLVMValueRef value;
1135
1136 if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID)
1137 return get_primitive_id(bld_base, swizzle);
1138
1139 if (!reg->Register.Dimension)
1140 return NULL;
1141
1142 if (swizzle == ~0) {
1143 LLVMValueRef values[TGSI_NUM_CHANNELS];
1144 unsigned chan;
1145 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
1146 values[chan] = fetch_input_gs(bld_base, reg, type, chan);
1147 }
1148 return lp_build_gather_values(bld_base->base.gallivm, values,
1149 TGSI_NUM_CHANNELS);
1150 }
1151
1152 /* Get the vertex offset parameter */
1153 vtx_offset_param = reg->Dimension.Index;
1154 if (vtx_offset_param < 2) {
1155 vtx_offset_param += SI_PARAM_VTX0_OFFSET;
1156 } else {
1157 assert(vtx_offset_param < 6);
1158 vtx_offset_param += SI_PARAM_VTX2_OFFSET - 2;
1159 }
1160 vtx_offset = lp_build_mul_imm(uint,
1161 LLVMGetParam(ctx->main_fn,
1162 vtx_offset_param),
1163 4);
1164
1165 param = si_shader_io_get_unique_index(semantic_name, semantic_index);
1166 args[0] = ctx->esgs_ring;
1167 args[1] = vtx_offset;
1168 args[2] = lp_build_const_int32(gallivm, (param * 4 + swizzle) * 256);
1169 args[3] = uint->zero;
1170 args[4] = uint->one; /* OFFEN */
1171 args[5] = uint->zero; /* IDXEN */
1172 args[6] = uint->one; /* GLC */
1173 args[7] = uint->zero; /* SLC */
1174 args[8] = uint->zero; /* TFE */
1175
1176 value = lp_build_intrinsic(gallivm->builder,
1177 "llvm.SI.buffer.load.dword.i32.i32",
1178 ctx->i32, args, 9,
1179 LP_FUNC_ATTR_READONLY);
1180 if (tgsi_type_is_64bit(type)) {
1181 LLVMValueRef value2;
1182 args[2] = lp_build_const_int32(gallivm, (param * 4 + swizzle + 1) * 256);
1183 value2 = lp_build_intrinsic(gallivm->builder,
1184 "llvm.SI.buffer.load.dword.i32.i32",
1185 ctx->i32, args, 9,
1186 LP_FUNC_ATTR_READONLY);
1187 return si_llvm_emit_fetch_64bit(bld_base, type,
1188 value, value2);
1189 }
1190 return LLVMBuildBitCast(gallivm->builder,
1191 value,
1192 tgsi2llvmtype(bld_base, type), "");
1193 }
1194
1195 static int lookup_interp_param_index(unsigned interpolate, unsigned location)
1196 {
1197 switch (interpolate) {
1198 case TGSI_INTERPOLATE_CONSTANT:
1199 return 0;
1200
1201 case TGSI_INTERPOLATE_LINEAR:
1202 if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
1203 return SI_PARAM_LINEAR_SAMPLE;
1204 else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
1205 return SI_PARAM_LINEAR_CENTROID;
1206 else
1207 return SI_PARAM_LINEAR_CENTER;
1208 break;
1209 case TGSI_INTERPOLATE_COLOR:
1210 case TGSI_INTERPOLATE_PERSPECTIVE:
1211 if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
1212 return SI_PARAM_PERSP_SAMPLE;
1213 else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
1214 return SI_PARAM_PERSP_CENTROID;
1215 else
1216 return SI_PARAM_PERSP_CENTER;
1217 break;
1218 default:
1219 fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
1220 return -1;
1221 }
1222 }
1223
1224 static LLVMValueRef build_fs_interp(
1225 struct lp_build_tgsi_context *bld_base,
1226 LLVMValueRef llvm_chan,
1227 LLVMValueRef attr_number,
1228 LLVMValueRef params,
1229 LLVMValueRef i,
1230 LLVMValueRef j) {
1231
1232 struct si_shader_context *ctx = si_shader_context(bld_base);
1233 struct gallivm_state *gallivm = bld_base->base.gallivm;
1234 LLVMValueRef args[5];
1235 LLVMValueRef p1;
1236 if (HAVE_LLVM < 0x0400) {
1237 LLVMValueRef ij[2];
1238 ij[0] = LLVMBuildBitCast(gallivm->builder, i, ctx->i32, "");
1239 ij[1] = LLVMBuildBitCast(gallivm->builder, j, ctx->i32, "");
1240
1241 args[0] = llvm_chan;
1242 args[1] = attr_number;
1243 args[2] = params;
1244 args[3] = lp_build_gather_values(gallivm, ij, 2);
1245 return lp_build_intrinsic(gallivm->builder, "llvm.SI.fs.interp",
1246 ctx->f32, args, 4,
1247 LP_FUNC_ATTR_READNONE);
1248 }
1249
1250 args[0] = i;
1251 args[1] = llvm_chan;
1252 args[2] = attr_number;
1253 args[3] = params;
1254
1255 p1 = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.interp.p1",
1256 ctx->f32, args, 4, LP_FUNC_ATTR_READNONE);
1257
1258 args[0] = p1;
1259 args[1] = j;
1260 args[2] = llvm_chan;
1261 args[3] = attr_number;
1262 args[4] = params;
1263
1264 return lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.interp.p2",
1265 ctx->f32, args, 5, LP_FUNC_ATTR_READNONE);
1266 }
1267
1268 static LLVMValueRef build_fs_interp_mov(
1269 struct lp_build_tgsi_context *bld_base,
1270 LLVMValueRef parameter,
1271 LLVMValueRef llvm_chan,
1272 LLVMValueRef attr_number,
1273 LLVMValueRef params) {
1274
1275 struct si_shader_context *ctx = si_shader_context(bld_base);
1276 struct gallivm_state *gallivm = bld_base->base.gallivm;
1277 LLVMValueRef args[4];
1278 if (HAVE_LLVM < 0x0400) {
1279 args[0] = llvm_chan;
1280 args[1] = attr_number;
1281 args[2] = params;
1282
1283 return lp_build_intrinsic(gallivm->builder,
1284 "llvm.SI.fs.constant",
1285 ctx->f32, args, 3,
1286 LP_FUNC_ATTR_READNONE);
1287 }
1288
1289 args[0] = parameter;
1290 args[1] = llvm_chan;
1291 args[2] = attr_number;
1292 args[3] = params;
1293
1294 return lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.interp.mov",
1295 ctx->f32, args, 4, LP_FUNC_ATTR_READNONE);
1296 }
1297
1298 /**
1299 * Interpolate a fragment shader input.
1300 *
1301 * @param ctx context
1302 * @param input_index index of the input in hardware
1303 * @param semantic_name TGSI_SEMANTIC_*
1304 * @param semantic_index semantic index
1305 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1306 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1307 * @param interp_param interpolation weights (i,j)
1308 * @param prim_mask SI_PARAM_PRIM_MASK
1309 * @param face SI_PARAM_FRONT_FACE
1310 * @param result the return value (4 components)
1311 */
1312 static void interp_fs_input(struct si_shader_context *ctx,
1313 unsigned input_index,
1314 unsigned semantic_name,
1315 unsigned semantic_index,
1316 unsigned num_interp_inputs,
1317 unsigned colors_read_mask,
1318 LLVMValueRef interp_param,
1319 LLVMValueRef prim_mask,
1320 LLVMValueRef face,
1321 LLVMValueRef result[4])
1322 {
1323 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
1324 struct lp_build_context *base = &bld_base->base;
1325 struct lp_build_context *uint = &bld_base->uint_bld;
1326 struct gallivm_state *gallivm = base->gallivm;
1327 LLVMValueRef attr_number;
1328 LLVMValueRef i, j;
1329
1330 unsigned chan;
1331
1332 /* fs.constant returns the param from the middle vertex, so it's not
1333 * really useful for flat shading. It's meant to be used for custom
1334 * interpolation (but the intrinsic can't fetch from the other two
1335 * vertices).
1336 *
1337 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1338 * to do the right thing. The only reason we use fs.constant is that
1339 * fs.interp cannot be used on integers, because they can be equal
1340 * to NaN.
1341 *
1342 * When interp is false we will use fs.constant or for newer llvm,
1343 * amdgcn.interp.mov.
1344 */
1345 bool interp = interp_param != NULL;
1346
1347 attr_number = lp_build_const_int32(gallivm, input_index);
1348
1349 if (interp) {
1350 interp_param = LLVMBuildBitCast(gallivm->builder, interp_param,
1351 LLVMVectorType(ctx->f32, 2), "");
1352
1353 i = LLVMBuildExtractElement(gallivm->builder, interp_param,
1354 uint->zero, "");
1355 j = LLVMBuildExtractElement(gallivm->builder, interp_param,
1356 uint->one, "");
1357 }
1358
1359 if (semantic_name == TGSI_SEMANTIC_COLOR &&
1360 ctx->shader->key.part.ps.prolog.color_two_side) {
1361 LLVMValueRef is_face_positive;
1362 LLVMValueRef back_attr_number;
1363
1364 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1365 * otherwise it's at offset "num_inputs".
1366 */
1367 unsigned back_attr_offset = num_interp_inputs;
1368 if (semantic_index == 1 && colors_read_mask & 0xf)
1369 back_attr_offset += 1;
1370
1371 back_attr_number = lp_build_const_int32(gallivm, back_attr_offset);
1372
1373 is_face_positive = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
1374 face, uint->zero, "");
1375
1376 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
1377 LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
1378 LLVMValueRef front, back;
1379
1380 if (interp) {
1381 front = build_fs_interp(bld_base, llvm_chan,
1382 attr_number, prim_mask,
1383 i, j);
1384 back = build_fs_interp(bld_base, llvm_chan,
1385 back_attr_number, prim_mask,
1386 i, j);
1387 } else {
1388 front = build_fs_interp_mov(bld_base,
1389 lp_build_const_int32(gallivm, 2), /* P0 */
1390 llvm_chan, attr_number, prim_mask);
1391 back = build_fs_interp_mov(bld_base,
1392 lp_build_const_int32(gallivm, 2), /* P0 */
1393 llvm_chan, back_attr_number, prim_mask);
1394 }
1395
1396 result[chan] = LLVMBuildSelect(gallivm->builder,
1397 is_face_positive,
1398 front,
1399 back,
1400 "");
1401 }
1402 } else if (semantic_name == TGSI_SEMANTIC_FOG) {
1403 if (interp) {
1404 result[0] = build_fs_interp(bld_base, uint->zero,
1405 attr_number, prim_mask, i, j);
1406 } else {
1407 result[0] = build_fs_interp_mov(bld_base, uint->zero,
1408 lp_build_const_int32(gallivm, 2), /* P0 */
1409 attr_number, prim_mask);
1410 }
1411 result[1] =
1412 result[2] = lp_build_const_float(gallivm, 0.0f);
1413 result[3] = lp_build_const_float(gallivm, 1.0f);
1414 } else {
1415 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
1416 LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
1417
1418 if (interp) {
1419 result[chan] = build_fs_interp(bld_base,
1420 llvm_chan, attr_number, prim_mask, i, j);
1421 } else {
1422 result[chan] = build_fs_interp_mov(bld_base,
1423 lp_build_const_int32(gallivm, 2), /* P0 */
1424 llvm_chan, attr_number, prim_mask);
1425 }
1426 }
1427 }
1428 }
1429
1430 static void declare_input_fs(
1431 struct si_shader_context *radeon_bld,
1432 unsigned input_index,
1433 const struct tgsi_full_declaration *decl,
1434 LLVMValueRef out[4])
1435 {
1436 struct lp_build_context *base = &radeon_bld->soa.bld_base.base;
1437 struct si_shader_context *ctx =
1438 si_shader_context(&radeon_bld->soa.bld_base);
1439 struct si_shader *shader = ctx->shader;
1440 LLVMValueRef main_fn = radeon_bld->main_fn;
1441 LLVMValueRef interp_param = NULL;
1442 int interp_param_idx;
1443
1444 /* Get colors from input VGPRs (set by the prolog). */
1445 if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR) {
1446 unsigned i = decl->Semantic.Index;
1447 unsigned colors_read = shader->selector->info.colors_read;
1448 unsigned mask = colors_read >> (i * 4);
1449 unsigned offset = SI_PARAM_POS_FIXED_PT + 1 +
1450 (i ? util_bitcount(colors_read & 0xf) : 0);
1451
1452 out[0] = mask & 0x1 ? LLVMGetParam(main_fn, offset++) : base->undef;
1453 out[1] = mask & 0x2 ? LLVMGetParam(main_fn, offset++) : base->undef;
1454 out[2] = mask & 0x4 ? LLVMGetParam(main_fn, offset++) : base->undef;
1455 out[3] = mask & 0x8 ? LLVMGetParam(main_fn, offset++) : base->undef;
1456 return;
1457 }
1458
1459 interp_param_idx = lookup_interp_param_index(decl->Interp.Interpolate,
1460 decl->Interp.Location);
1461 if (interp_param_idx == -1)
1462 return;
1463 else if (interp_param_idx) {
1464 interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
1465 }
1466
1467 if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
1468 decl->Interp.Interpolate == TGSI_INTERPOLATE_COLOR &&
1469 ctx->shader->key.part.ps.prolog.flatshade_colors)
1470 interp_param = NULL; /* load the constant color */
1471
1472 interp_fs_input(ctx, input_index, decl->Semantic.Name,
1473 decl->Semantic.Index, shader->selector->info.num_inputs,
1474 shader->selector->info.colors_read, interp_param,
1475 LLVMGetParam(main_fn, SI_PARAM_PRIM_MASK),
1476 LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE),
1477 &out[0]);
1478 }
1479
1480 static LLVMValueRef get_sample_id(struct si_shader_context *radeon_bld)
1481 {
1482 return unpack_param(si_shader_context(&radeon_bld->soa.bld_base),
1483 SI_PARAM_ANCILLARY, 8, 4);
1484 }
1485
1486 /**
1487 * Set range metadata on an instruction. This can only be used on load and
1488 * call instructions. If you know an instruction can only produce the values
1489 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1490 * \p lo is the minimum value inclusive.
1491 * \p hi is the maximum value exclusive.
1492 */
1493 static void set_range_metadata(struct si_shader_context *ctx,
1494 LLVMValueRef value, unsigned lo, unsigned hi)
1495 {
1496 LLVMValueRef range_md, md_args[2];
1497 LLVMTypeRef type = LLVMTypeOf(value);
1498 LLVMContextRef context = LLVMGetTypeContext(type);
1499
1500 md_args[0] = LLVMConstInt(type, lo, false);
1501 md_args[1] = LLVMConstInt(type, hi, false);
1502 range_md = LLVMMDNodeInContext(context, md_args, 2);
1503 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1504 }
1505
1506 static LLVMValueRef get_thread_id(struct si_shader_context *ctx)
1507 {
1508 struct gallivm_state *gallivm = &ctx->gallivm;
1509 LLVMValueRef tid;
1510
1511 if (HAVE_LLVM < 0x0308) {
1512 tid = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid",
1513 ctx->i32, NULL, 0, LP_FUNC_ATTR_READNONE);
1514 } else {
1515 LLVMValueRef tid_args[2];
1516 tid_args[0] = lp_build_const_int32(gallivm, 0xffffffff);
1517 tid_args[1] = lp_build_const_int32(gallivm, 0);
1518 tid_args[1] = lp_build_intrinsic(gallivm->builder,
1519 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1520 tid_args, 2, LP_FUNC_ATTR_READNONE);
1521
1522 tid = lp_build_intrinsic(gallivm->builder,
1523 "llvm.amdgcn.mbcnt.hi", ctx->i32,
1524 tid_args, 2, LP_FUNC_ATTR_READNONE);
1525 }
1526 set_range_metadata(ctx, tid, 0, 64);
1527 return tid;
1528 }
1529
1530 /**
1531 * Load a dword from a constant buffer.
1532 */
1533 static LLVMValueRef buffer_load_const(struct si_shader_context *ctx,
1534 LLVMValueRef resource,
1535 LLVMValueRef offset)
1536 {
1537 LLVMBuilderRef builder = ctx->gallivm.builder;
1538 LLVMValueRef args[2] = {resource, offset};
1539
1540 return lp_build_intrinsic(builder, "llvm.SI.load.const", ctx->f32, args, 2,
1541 LP_FUNC_ATTR_READNONE);
1542 }
1543
1544 static LLVMValueRef load_sample_position(struct si_shader_context *radeon_bld, LLVMValueRef sample_id)
1545 {
1546 struct si_shader_context *ctx =
1547 si_shader_context(&radeon_bld->soa.bld_base);
1548 struct lp_build_context *uint_bld = &radeon_bld->soa.bld_base.uint_bld;
1549 struct gallivm_state *gallivm = &radeon_bld->gallivm;
1550 LLVMBuilderRef builder = gallivm->builder;
1551 LLVMValueRef desc = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
1552 LLVMValueRef buf_index = lp_build_const_int32(gallivm, SI_PS_CONST_SAMPLE_POSITIONS);
1553 LLVMValueRef resource = build_indexed_load_const(ctx, desc, buf_index);
1554
1555 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1556 LLVMValueRef offset0 = lp_build_mul_imm(uint_bld, sample_id, 8);
1557 LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, lp_build_const_int32(gallivm, 4), "");
1558
1559 LLVMValueRef pos[4] = {
1560 buffer_load_const(ctx, resource, offset0),
1561 buffer_load_const(ctx, resource, offset1),
1562 lp_build_const_float(gallivm, 0),
1563 lp_build_const_float(gallivm, 0)
1564 };
1565
1566 return lp_build_gather_values(gallivm, pos, 4);
1567 }
1568
1569 static void declare_system_value(
1570 struct si_shader_context *radeon_bld,
1571 unsigned index,
1572 const struct tgsi_full_declaration *decl)
1573 {
1574 struct si_shader_context *ctx =
1575 si_shader_context(&radeon_bld->soa.bld_base);
1576 struct lp_build_context *bld = &radeon_bld->soa.bld_base.base;
1577 struct gallivm_state *gallivm = &radeon_bld->gallivm;
1578 LLVMValueRef value = 0;
1579
1580 switch (decl->Semantic.Name) {
1581 case TGSI_SEMANTIC_INSTANCEID:
1582 value = LLVMGetParam(radeon_bld->main_fn,
1583 ctx->param_instance_id);
1584 break;
1585
1586 case TGSI_SEMANTIC_VERTEXID:
1587 value = LLVMBuildAdd(gallivm->builder,
1588 LLVMGetParam(radeon_bld->main_fn,
1589 ctx->param_vertex_id),
1590 LLVMGetParam(radeon_bld->main_fn,
1591 SI_PARAM_BASE_VERTEX), "");
1592 break;
1593
1594 case TGSI_SEMANTIC_VERTEXID_NOBASE:
1595 value = LLVMGetParam(radeon_bld->main_fn,
1596 ctx->param_vertex_id);
1597 break;
1598
1599 case TGSI_SEMANTIC_BASEVERTEX:
1600 value = LLVMGetParam(radeon_bld->main_fn,
1601 SI_PARAM_BASE_VERTEX);
1602 break;
1603
1604 case TGSI_SEMANTIC_BASEINSTANCE:
1605 value = LLVMGetParam(radeon_bld->main_fn,
1606 SI_PARAM_START_INSTANCE);
1607 break;
1608
1609 case TGSI_SEMANTIC_DRAWID:
1610 value = LLVMGetParam(radeon_bld->main_fn,
1611 SI_PARAM_DRAWID);
1612 break;
1613
1614 case TGSI_SEMANTIC_INVOCATIONID:
1615 if (ctx->type == PIPE_SHADER_TESS_CTRL)
1616 value = unpack_param(ctx, SI_PARAM_REL_IDS, 8, 5);
1617 else if (ctx->type == PIPE_SHADER_GEOMETRY)
1618 value = LLVMGetParam(radeon_bld->main_fn,
1619 SI_PARAM_GS_INSTANCE_ID);
1620 else
1621 assert(!"INVOCATIONID not implemented");
1622 break;
1623
1624 case TGSI_SEMANTIC_POSITION:
1625 {
1626 LLVMValueRef pos[4] = {
1627 LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_X_FLOAT),
1628 LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Y_FLOAT),
1629 LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Z_FLOAT),
1630 lp_build_emit_llvm_unary(&radeon_bld->soa.bld_base, TGSI_OPCODE_RCP,
1631 LLVMGetParam(radeon_bld->main_fn,
1632 SI_PARAM_POS_W_FLOAT)),
1633 };
1634 value = lp_build_gather_values(gallivm, pos, 4);
1635 break;
1636 }
1637
1638 case TGSI_SEMANTIC_FACE:
1639 value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_FRONT_FACE);
1640 break;
1641
1642 case TGSI_SEMANTIC_SAMPLEID:
1643 value = get_sample_id(radeon_bld);
1644 break;
1645
1646 case TGSI_SEMANTIC_SAMPLEPOS: {
1647 LLVMValueRef pos[4] = {
1648 LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_X_FLOAT),
1649 LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Y_FLOAT),
1650 lp_build_const_float(gallivm, 0),
1651 lp_build_const_float(gallivm, 0)
1652 };
1653 pos[0] = lp_build_emit_llvm_unary(&radeon_bld->soa.bld_base,
1654 TGSI_OPCODE_FRC, pos[0]);
1655 pos[1] = lp_build_emit_llvm_unary(&radeon_bld->soa.bld_base,
1656 TGSI_OPCODE_FRC, pos[1]);
1657 value = lp_build_gather_values(gallivm, pos, 4);
1658 break;
1659 }
1660
1661 case TGSI_SEMANTIC_SAMPLEMASK:
1662 /* This can only occur with the OpenGL Core profile, which
1663 * doesn't support smoothing.
1664 */
1665 value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_SAMPLE_COVERAGE);
1666 break;
1667
1668 case TGSI_SEMANTIC_TESSCOORD:
1669 {
1670 LLVMValueRef coord[4] = {
1671 LLVMGetParam(radeon_bld->main_fn, ctx->param_tes_u),
1672 LLVMGetParam(radeon_bld->main_fn, ctx->param_tes_v),
1673 bld->zero,
1674 bld->zero
1675 };
1676
1677 /* For triangles, the vector should be (u, v, 1-u-v). */
1678 if (ctx->shader->selector->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] ==
1679 PIPE_PRIM_TRIANGLES)
1680 coord[2] = lp_build_sub(bld, bld->one,
1681 lp_build_add(bld, coord[0], coord[1]));
1682
1683 value = lp_build_gather_values(gallivm, coord, 4);
1684 break;
1685 }
1686
1687 case TGSI_SEMANTIC_VERTICESIN:
1688 if (ctx->type == PIPE_SHADER_TESS_CTRL)
1689 value = unpack_param(ctx, SI_PARAM_TCS_OUT_LAYOUT, 26, 6);
1690 else if (ctx->type == PIPE_SHADER_TESS_EVAL)
1691 value = unpack_param(ctx, SI_PARAM_TCS_OFFCHIP_LAYOUT, 9, 7);
1692 else
1693 assert(!"invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1694 break;
1695
1696 case TGSI_SEMANTIC_TESSINNER:
1697 case TGSI_SEMANTIC_TESSOUTER:
1698 {
1699 LLVMValueRef rw_buffers, buffer, base, addr;
1700 int param = si_shader_io_get_unique_index(decl->Semantic.Name, 0);
1701
1702 rw_buffers = LLVMGetParam(ctx->main_fn,
1703 SI_PARAM_RW_BUFFERS);
1704 buffer = build_indexed_load_const(ctx, rw_buffers,
1705 lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
1706
1707 base = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
1708 addr = get_tcs_tes_buffer_address(ctx, NULL,
1709 lp_build_const_int32(gallivm, param));
1710
1711 value = buffer_load(&radeon_bld->soa.bld_base, TGSI_TYPE_FLOAT,
1712 ~0, buffer, base, addr);
1713
1714 break;
1715 }
1716
1717 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI:
1718 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI:
1719 {
1720 LLVMValueRef buf, slot, val[4];
1721 int i, offset;
1722
1723 slot = lp_build_const_int32(gallivm, SI_HS_CONST_DEFAULT_TESS_LEVELS);
1724 buf = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
1725 buf = build_indexed_load_const(ctx, buf, slot);
1726 offset = decl->Semantic.Name == TGSI_SEMANTIC_DEFAULT_TESSINNER_SI ? 4 : 0;
1727
1728 for (i = 0; i < 4; i++)
1729 val[i] = buffer_load_const(ctx, buf,
1730 lp_build_const_int32(gallivm, (offset + i) * 4));
1731 value = lp_build_gather_values(gallivm, val, 4);
1732 break;
1733 }
1734
1735 case TGSI_SEMANTIC_PRIMID:
1736 value = get_primitive_id(&radeon_bld->soa.bld_base, 0);
1737 break;
1738
1739 case TGSI_SEMANTIC_GRID_SIZE:
1740 value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_GRID_SIZE);
1741 break;
1742
1743 case TGSI_SEMANTIC_BLOCK_SIZE:
1744 {
1745 LLVMValueRef values[3];
1746 unsigned i;
1747 unsigned *properties = ctx->shader->selector->info.properties;
1748
1749 if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
1750 unsigned sizes[3] = {
1751 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
1752 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
1753 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
1754 };
1755
1756 for (i = 0; i < 3; ++i)
1757 values[i] = lp_build_const_int32(gallivm, sizes[i]);
1758
1759 value = lp_build_gather_values(gallivm, values, 3);
1760 } else {
1761 value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_BLOCK_SIZE);
1762 }
1763 break;
1764 }
1765
1766 case TGSI_SEMANTIC_BLOCK_ID:
1767 value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_BLOCK_ID);
1768 break;
1769
1770 case TGSI_SEMANTIC_THREAD_ID:
1771 value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_THREAD_ID);
1772 break;
1773
1774 #if HAVE_LLVM >= 0x0309
1775 case TGSI_SEMANTIC_HELPER_INVOCATION:
1776 value = lp_build_intrinsic(gallivm->builder,
1777 "llvm.amdgcn.ps.live",
1778 ctx->i1, NULL, 0,
1779 LP_FUNC_ATTR_READNONE);
1780 value = LLVMBuildNot(gallivm->builder, value, "");
1781 value = LLVMBuildSExt(gallivm->builder, value, ctx->i32, "");
1782 break;
1783 #endif
1784
1785 default:
1786 assert(!"unknown system value");
1787 return;
1788 }
1789
1790 radeon_bld->system_values[index] = value;
1791 }
1792
1793 static void declare_compute_memory(struct si_shader_context *radeon_bld,
1794 const struct tgsi_full_declaration *decl)
1795 {
1796 struct si_shader_context *ctx =
1797 si_shader_context(&radeon_bld->soa.bld_base);
1798 struct si_shader_selector *sel = ctx->shader->selector;
1799 struct gallivm_state *gallivm = &radeon_bld->gallivm;
1800
1801 LLVMTypeRef i8p = LLVMPointerType(ctx->i8, LOCAL_ADDR_SPACE);
1802 LLVMValueRef var;
1803
1804 assert(decl->Declaration.MemType == TGSI_MEMORY_TYPE_SHARED);
1805 assert(decl->Range.First == decl->Range.Last);
1806 assert(!ctx->shared_memory);
1807
1808 var = LLVMAddGlobalInAddressSpace(gallivm->module,
1809 LLVMArrayType(ctx->i8, sel->local_size),
1810 "compute_lds",
1811 LOCAL_ADDR_SPACE);
1812 LLVMSetAlignment(var, 4);
1813
1814 ctx->shared_memory = LLVMBuildBitCast(gallivm->builder, var, i8p, "");
1815 }
1816
1817 static LLVMValueRef load_const_buffer_desc(struct si_shader_context *ctx, int i)
1818 {
1819 LLVMValueRef list_ptr = LLVMGetParam(ctx->main_fn,
1820 SI_PARAM_CONST_BUFFERS);
1821
1822 return build_indexed_load_const(ctx, list_ptr,
1823 LLVMConstInt(ctx->i32, i, 0));
1824 }
1825
1826 static LLVMValueRef fetch_constant(
1827 struct lp_build_tgsi_context *bld_base,
1828 const struct tgsi_full_src_register *reg,
1829 enum tgsi_opcode_type type,
1830 unsigned swizzle)
1831 {
1832 struct si_shader_context *ctx = si_shader_context(bld_base);
1833 struct lp_build_context *base = &bld_base->base;
1834 const struct tgsi_ind_register *ireg = &reg->Indirect;
1835 unsigned buf, idx;
1836
1837 LLVMValueRef addr, bufp;
1838 LLVMValueRef result;
1839
1840 if (swizzle == LP_CHAN_ALL) {
1841 unsigned chan;
1842 LLVMValueRef values[4];
1843 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
1844 values[chan] = fetch_constant(bld_base, reg, type, chan);
1845
1846 return lp_build_gather_values(bld_base->base.gallivm, values, 4);
1847 }
1848
1849 buf = reg->Register.Dimension ? reg->Dimension.Index : 0;
1850 idx = reg->Register.Index * 4 + swizzle;
1851
1852 if (reg->Register.Dimension && reg->Dimension.Indirect) {
1853 LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, SI_PARAM_CONST_BUFFERS);
1854 LLVMValueRef index;
1855 index = get_bounded_indirect_index(ctx, &reg->DimIndirect,
1856 reg->Dimension.Index,
1857 SI_NUM_CONST_BUFFERS);
1858 bufp = build_indexed_load_const(ctx, ptr, index);
1859 } else
1860 bufp = load_const_buffer_desc(ctx, buf);
1861
1862 if (reg->Register.Indirect) {
1863 addr = ctx->soa.addr[ireg->Index][ireg->Swizzle];
1864 addr = LLVMBuildLoad(base->gallivm->builder, addr, "load addr reg");
1865 addr = lp_build_mul_imm(&bld_base->uint_bld, addr, 16);
1866 addr = lp_build_add(&bld_base->uint_bld, addr,
1867 lp_build_const_int32(base->gallivm, idx * 4));
1868 } else {
1869 addr = LLVMConstInt(ctx->i32, idx * 4, 0);
1870 }
1871
1872 result = buffer_load_const(ctx, bufp, addr);
1873
1874 if (!tgsi_type_is_64bit(type))
1875 result = bitcast(bld_base, type, result);
1876 else {
1877 LLVMValueRef addr2, result2;
1878
1879 addr2 = lp_build_add(&bld_base->uint_bld, addr,
1880 LLVMConstInt(ctx->i32, 4, 0));
1881 result2 = buffer_load_const(ctx, bufp, addr2);
1882
1883 result = si_llvm_emit_fetch_64bit(bld_base, type,
1884 result, result2);
1885 }
1886 return result;
1887 }
1888
1889 /* Upper 16 bits must be zero. */
1890 static LLVMValueRef si_llvm_pack_two_int16(struct gallivm_state *gallivm,
1891 LLVMValueRef val[2])
1892 {
1893 return LLVMBuildOr(gallivm->builder, val[0],
1894 LLVMBuildShl(gallivm->builder, val[1],
1895 lp_build_const_int32(gallivm, 16),
1896 ""), "");
1897 }
1898
1899 /* Upper 16 bits are ignored and will be dropped. */
1900 static LLVMValueRef si_llvm_pack_two_int32_as_int16(struct gallivm_state *gallivm,
1901 LLVMValueRef val[2])
1902 {
1903 LLVMValueRef v[2] = {
1904 LLVMBuildAnd(gallivm->builder, val[0],
1905 lp_build_const_int32(gallivm, 0xffff), ""),
1906 val[1],
1907 };
1908 return si_llvm_pack_two_int16(gallivm, v);
1909 }
1910
1911 /* Initialize arguments for the shader export intrinsic */
1912 static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
1913 LLVMValueRef *values,
1914 unsigned target,
1915 LLVMValueRef *args)
1916 {
1917 struct si_shader_context *ctx = si_shader_context(bld_base);
1918 struct lp_build_context *uint =
1919 &ctx->soa.bld_base.uint_bld;
1920 struct lp_build_context *base = &bld_base->base;
1921 struct gallivm_state *gallivm = base->gallivm;
1922 LLVMBuilderRef builder = base->gallivm->builder;
1923 LLVMValueRef val[4];
1924 unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
1925 unsigned chan;
1926 bool is_int8;
1927
1928 /* Default is 0xf. Adjusted below depending on the format. */
1929 args[0] = lp_build_const_int32(base->gallivm, 0xf); /* writemask */
1930
1931 /* Specify whether the EXEC mask represents the valid mask */
1932 args[1] = uint->zero;
1933
1934 /* Specify whether this is the last export */
1935 args[2] = uint->zero;
1936
1937 /* Specify the target we are exporting */
1938 args[3] = lp_build_const_int32(base->gallivm, target);
1939
1940 if (ctx->type == PIPE_SHADER_FRAGMENT) {
1941 const struct si_shader_key *key = &ctx->shader->key;
1942 unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
1943 int cbuf = target - V_008DFC_SQ_EXP_MRT;
1944
1945 assert(cbuf >= 0 && cbuf < 8);
1946 spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
1947 is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
1948 }
1949
1950 args[4] = uint->zero; /* COMPR flag */
1951 args[5] = base->undef;
1952 args[6] = base->undef;
1953 args[7] = base->undef;
1954 args[8] = base->undef;
1955
1956 switch (spi_shader_col_format) {
1957 case V_028714_SPI_SHADER_ZERO:
1958 args[0] = uint->zero; /* writemask */
1959 args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_NULL);
1960 break;
1961
1962 case V_028714_SPI_SHADER_32_R:
1963 args[0] = uint->one; /* writemask */
1964 args[5] = values[0];
1965 break;
1966
1967 case V_028714_SPI_SHADER_32_GR:
1968 args[0] = lp_build_const_int32(base->gallivm, 0x3); /* writemask */
1969 args[5] = values[0];
1970 args[6] = values[1];
1971 break;
1972
1973 case V_028714_SPI_SHADER_32_AR:
1974 args[0] = lp_build_const_int32(base->gallivm, 0x9); /* writemask */
1975 args[5] = values[0];
1976 args[8] = values[3];
1977 break;
1978
1979 case V_028714_SPI_SHADER_FP16_ABGR:
1980 args[4] = uint->one; /* COMPR flag */
1981
1982 for (chan = 0; chan < 2; chan++) {
1983 LLVMValueRef pack_args[2] = {
1984 values[2 * chan],
1985 values[2 * chan + 1]
1986 };
1987 LLVMValueRef packed;
1988
1989 packed = lp_build_intrinsic(base->gallivm->builder,
1990 "llvm.SI.packf16",
1991 ctx->i32, pack_args, 2,
1992 LP_FUNC_ATTR_READNONE);
1993 args[chan + 5] =
1994 LLVMBuildBitCast(base->gallivm->builder,
1995 packed, ctx->f32, "");
1996 }
1997 break;
1998
1999 case V_028714_SPI_SHADER_UNORM16_ABGR:
2000 for (chan = 0; chan < 4; chan++) {
2001 val[chan] = si_llvm_saturate(bld_base, values[chan]);
2002 val[chan] = LLVMBuildFMul(builder, val[chan],
2003 lp_build_const_float(gallivm, 65535), "");
2004 val[chan] = LLVMBuildFAdd(builder, val[chan],
2005 lp_build_const_float(gallivm, 0.5), "");
2006 val[chan] = LLVMBuildFPToUI(builder, val[chan],
2007 ctx->i32, "");
2008 }
2009
2010 args[4] = uint->one; /* COMPR flag */
2011 args[5] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2012 si_llvm_pack_two_int16(gallivm, val));
2013 args[6] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2014 si_llvm_pack_two_int16(gallivm, val+2));
2015 break;
2016
2017 case V_028714_SPI_SHADER_SNORM16_ABGR:
2018 for (chan = 0; chan < 4; chan++) {
2019 /* Clamp between [-1, 1]. */
2020 val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MIN,
2021 values[chan],
2022 lp_build_const_float(gallivm, 1));
2023 val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MAX,
2024 val[chan],
2025 lp_build_const_float(gallivm, -1));
2026 /* Convert to a signed integer in [-32767, 32767]. */
2027 val[chan] = LLVMBuildFMul(builder, val[chan],
2028 lp_build_const_float(gallivm, 32767), "");
2029 /* If positive, add 0.5, else add -0.5. */
2030 val[chan] = LLVMBuildFAdd(builder, val[chan],
2031 LLVMBuildSelect(builder,
2032 LLVMBuildFCmp(builder, LLVMRealOGE,
2033 val[chan], base->zero, ""),
2034 lp_build_const_float(gallivm, 0.5),
2035 lp_build_const_float(gallivm, -0.5), ""), "");
2036 val[chan] = LLVMBuildFPToSI(builder, val[chan], ctx->i32, "");
2037 }
2038
2039 args[4] = uint->one; /* COMPR flag */
2040 args[5] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2041 si_llvm_pack_two_int32_as_int16(gallivm, val));
2042 args[6] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2043 si_llvm_pack_two_int32_as_int16(gallivm, val+2));
2044 break;
2045
2046 case V_028714_SPI_SHADER_UINT16_ABGR: {
2047 LLVMValueRef max = lp_build_const_int32(gallivm, is_int8 ?
2048 255 : 65535);
2049 /* Clamp. */
2050 for (chan = 0; chan < 4; chan++) {
2051 val[chan] = bitcast(bld_base, TGSI_TYPE_UNSIGNED, values[chan]);
2052 val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_UMIN,
2053 val[chan], max);
2054 }
2055
2056 args[4] = uint->one; /* COMPR flag */
2057 args[5] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2058 si_llvm_pack_two_int16(gallivm, val));
2059 args[6] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2060 si_llvm_pack_two_int16(gallivm, val+2));
2061 break;
2062 }
2063
2064 case V_028714_SPI_SHADER_SINT16_ABGR: {
2065 LLVMValueRef max = lp_build_const_int32(gallivm, is_int8 ?
2066 127 : 32767);
2067 LLVMValueRef min = lp_build_const_int32(gallivm, is_int8 ?
2068 -128 : -32768);
2069 /* Clamp. */
2070 for (chan = 0; chan < 4; chan++) {
2071 val[chan] = bitcast(bld_base, TGSI_TYPE_UNSIGNED, values[chan]);
2072 val[chan] = lp_build_emit_llvm_binary(bld_base,
2073 TGSI_OPCODE_IMIN,
2074 val[chan], max);
2075 val[chan] = lp_build_emit_llvm_binary(bld_base,
2076 TGSI_OPCODE_IMAX,
2077 val[chan], min);
2078 }
2079
2080 args[4] = uint->one; /* COMPR flag */
2081 args[5] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2082 si_llvm_pack_two_int32_as_int16(gallivm, val));
2083 args[6] = bitcast(bld_base, TGSI_TYPE_FLOAT,
2084 si_llvm_pack_two_int32_as_int16(gallivm, val+2));
2085 break;
2086 }
2087
2088 case V_028714_SPI_SHADER_32_ABGR:
2089 memcpy(&args[5], values, sizeof(values[0]) * 4);
2090 break;
2091 }
2092 }
2093
2094 static void si_alpha_test(struct lp_build_tgsi_context *bld_base,
2095 LLVMValueRef alpha)
2096 {
2097 struct si_shader_context *ctx = si_shader_context(bld_base);
2098 struct gallivm_state *gallivm = bld_base->base.gallivm;
2099
2100 if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
2101 LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
2102 SI_PARAM_ALPHA_REF);
2103
2104 LLVMValueRef alpha_pass =
2105 lp_build_cmp(&bld_base->base,
2106 ctx->shader->key.part.ps.epilog.alpha_func,
2107 alpha, alpha_ref);
2108 LLVMValueRef arg =
2109 lp_build_select(&bld_base->base,
2110 alpha_pass,
2111 lp_build_const_float(gallivm, 1.0f),
2112 lp_build_const_float(gallivm, -1.0f));
2113
2114 lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
2115 ctx->voidt, &arg, 1, 0);
2116 } else {
2117 lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kilp",
2118 ctx->voidt, NULL, 0, 0);
2119 }
2120 }
2121
2122 static LLVMValueRef si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context *bld_base,
2123 LLVMValueRef alpha,
2124 unsigned samplemask_param)
2125 {
2126 struct si_shader_context *ctx = si_shader_context(bld_base);
2127 struct gallivm_state *gallivm = bld_base->base.gallivm;
2128 LLVMValueRef coverage;
2129
2130 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2131 coverage = LLVMGetParam(ctx->main_fn,
2132 samplemask_param);
2133 coverage = bitcast(bld_base, TGSI_TYPE_SIGNED, coverage);
2134
2135 coverage = lp_build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
2136 ctx->i32,
2137 &coverage, 1, LP_FUNC_ATTR_READNONE);
2138
2139 coverage = LLVMBuildUIToFP(gallivm->builder, coverage,
2140 ctx->f32, "");
2141
2142 coverage = LLVMBuildFMul(gallivm->builder, coverage,
2143 lp_build_const_float(gallivm,
2144 1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");
2145
2146 return LLVMBuildFMul(gallivm->builder, alpha, coverage, "");
2147 }
2148
2149 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context *bld_base,
2150 LLVMValueRef (*pos)[9], LLVMValueRef *out_elts)
2151 {
2152 struct si_shader_context *ctx = si_shader_context(bld_base);
2153 struct lp_build_context *base = &bld_base->base;
2154 struct lp_build_context *uint = &ctx->soa.bld_base.uint_bld;
2155 unsigned reg_index;
2156 unsigned chan;
2157 unsigned const_chan;
2158 LLVMValueRef base_elt;
2159 LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
2160 LLVMValueRef constbuf_index = lp_build_const_int32(base->gallivm,
2161 SI_VS_CONST_CLIP_PLANES);
2162 LLVMValueRef const_resource = build_indexed_load_const(ctx, ptr, constbuf_index);
2163
2164 for (reg_index = 0; reg_index < 2; reg_index ++) {
2165 LLVMValueRef *args = pos[2 + reg_index];
2166
2167 args[5] =
2168 args[6] =
2169 args[7] =
2170 args[8] = lp_build_const_float(base->gallivm, 0.0f);
2171
2172 /* Compute dot products of position and user clip plane vectors */
2173 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
2174 for (const_chan = 0; const_chan < TGSI_NUM_CHANNELS; const_chan++) {
2175 args[1] = lp_build_const_int32(base->gallivm,
2176 ((reg_index * 4 + chan) * 4 +
2177 const_chan) * 4);
2178 base_elt = buffer_load_const(ctx, const_resource,
2179 args[1]);
2180 args[5 + chan] =
2181 lp_build_add(base, args[5 + chan],
2182 lp_build_mul(base, base_elt,
2183 out_elts[const_chan]));
2184 }
2185 }
2186
2187 args[0] = lp_build_const_int32(base->gallivm, 0xf);
2188 args[1] = uint->zero;
2189 args[2] = uint->zero;
2190 args[3] = lp_build_const_int32(base->gallivm,
2191 V_008DFC_SQ_EXP_POS + 2 + reg_index);
2192 args[4] = uint->zero;
2193 }
2194 }
2195
2196 static void si_dump_streamout(struct pipe_stream_output_info *so)
2197 {
2198 unsigned i;
2199
2200 if (so->num_outputs)
2201 fprintf(stderr, "STREAMOUT\n");
2202
2203 for (i = 0; i < so->num_outputs; i++) {
2204 unsigned mask = ((1 << so->output[i].num_components) - 1) <<
2205 so->output[i].start_component;
2206 fprintf(stderr, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2207 i, so->output[i].output_buffer,
2208 so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1,
2209 so->output[i].register_index,
2210 mask & 1 ? "x" : "",
2211 mask & 2 ? "y" : "",
2212 mask & 4 ? "z" : "",
2213 mask & 8 ? "w" : "");
2214 }
2215 }
2216
2217 static void emit_streamout_output(struct si_shader_context *ctx,
2218 LLVMValueRef const *so_buffers,
2219 LLVMValueRef const *so_write_offsets,
2220 struct pipe_stream_output *stream_out,
2221 struct si_shader_output_values *shader_out)
2222 {
2223 struct gallivm_state *gallivm = &ctx->gallivm;
2224 LLVMBuilderRef builder = gallivm->builder;
2225 unsigned buf_idx = stream_out->output_buffer;
2226 unsigned start = stream_out->start_component;
2227 unsigned num_comps = stream_out->num_components;
2228 LLVMValueRef out[4];
2229
2230 assert(num_comps && num_comps <= 4);
2231 if (!num_comps || num_comps > 4)
2232 return;
2233
2234 /* Load the output as int. */
2235 for (int j = 0; j < num_comps; j++) {
2236 assert(stream_out->stream == shader_out->vertex_stream[start + j]);
2237
2238 out[j] = LLVMBuildBitCast(builder,
2239 shader_out->values[start + j],
2240 ctx->i32, "");
2241 }
2242
2243 /* Pack the output. */
2244 LLVMValueRef vdata = NULL;
2245
2246 switch (num_comps) {
2247 case 1: /* as i32 */
2248 vdata = out[0];
2249 break;
2250 case 2: /* as v2i32 */
2251 case 3: /* as v4i32 (aligned to 4) */
2252 case 4: /* as v4i32 */
2253 vdata = LLVMGetUndef(LLVMVectorType(ctx->i32, util_next_power_of_two(num_comps)));
2254 for (int j = 0; j < num_comps; j++) {
2255 vdata = LLVMBuildInsertElement(builder, vdata, out[j],
2256 LLVMConstInt(ctx->i32, j, 0), "");
2257 }
2258 break;
2259 }
2260
2261 build_tbuffer_store_dwords(ctx, so_buffers[buf_idx],
2262 vdata, num_comps,
2263 so_write_offsets[buf_idx],
2264 LLVMConstInt(ctx->i32, 0, 0),
2265 stream_out->dst_offset * 4);
2266 }
2267
2268 /**
2269 * Write streamout data to buffers for vertex stream @p stream (different
2270 * vertex streams can occur for GS copy shaders).
2271 */
2272 static void si_llvm_emit_streamout(struct si_shader_context *ctx,
2273 struct si_shader_output_values *outputs,
2274 unsigned noutput, unsigned stream)
2275 {
2276 struct si_shader_selector *sel = ctx->shader->selector;
2277 struct pipe_stream_output_info *so = &sel->so;
2278 struct gallivm_state *gallivm = &ctx->gallivm;
2279 LLVMBuilderRef builder = gallivm->builder;
2280 int i;
2281 struct lp_build_if_state if_ctx;
2282
2283 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2284 LLVMValueRef so_vtx_count =
2285 unpack_param(ctx, ctx->param_streamout_config, 16, 7);
2286
2287 LLVMValueRef tid = get_thread_id(ctx);
2288
2289 /* can_emit = tid < so_vtx_count; */
2290 LLVMValueRef can_emit =
2291 LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
2292
2293 /* Emit the streamout code conditionally. This actually avoids
2294 * out-of-bounds buffer access. The hw tells us via the SGPR
2295 * (so_vtx_count) which threads are allowed to emit streamout data. */
2296 lp_build_if(&if_ctx, gallivm, can_emit);
2297 {
2298 /* The buffer offset is computed as follows:
2299 * ByteOffset = streamout_offset[buffer_id]*4 +
2300 * (streamout_write_index + thread_id)*stride[buffer_id] +
2301 * attrib_offset
2302 */
2303
2304 LLVMValueRef so_write_index =
2305 LLVMGetParam(ctx->main_fn,
2306 ctx->param_streamout_write_index);
2307
2308 /* Compute (streamout_write_index + thread_id). */
2309 so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
2310
2311 /* Load the descriptor and compute the write offset for each
2312 * enabled buffer. */
2313 LLVMValueRef so_write_offset[4] = {};
2314 LLVMValueRef so_buffers[4];
2315 LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
2316 SI_PARAM_RW_BUFFERS);
2317
2318 for (i = 0; i < 4; i++) {
2319 if (!so->stride[i])
2320 continue;
2321
2322 LLVMValueRef offset = lp_build_const_int32(gallivm,
2323 SI_VS_STREAMOUT_BUF0 + i);
2324
2325 so_buffers[i] = build_indexed_load_const(ctx, buf_ptr, offset);
2326
2327 LLVMValueRef so_offset = LLVMGetParam(ctx->main_fn,
2328 ctx->param_streamout_offset[i]);
2329 so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
2330
2331 so_write_offset[i] = LLVMBuildMul(builder, so_write_index,
2332 LLVMConstInt(ctx->i32, so->stride[i]*4, 0), "");
2333 so_write_offset[i] = LLVMBuildAdd(builder, so_write_offset[i], so_offset, "");
2334 }
2335
2336 /* Write streamout data. */
2337 for (i = 0; i < so->num_outputs; i++) {
2338 unsigned reg = so->output[i].register_index;
2339
2340 if (reg >= noutput)
2341 continue;
2342
2343 if (stream != so->output[i].stream)
2344 continue;
2345
2346 emit_streamout_output(ctx, so_buffers, so_write_offset,
2347 &so->output[i], &outputs[reg]);
2348 }
2349 }
2350 lp_build_endif(&if_ctx);
2351 }
2352
2353
2354 /* Generate export instructions for hardware VS shader stage */
2355 static void si_llvm_export_vs(struct lp_build_tgsi_context *bld_base,
2356 struct si_shader_output_values *outputs,
2357 unsigned noutput)
2358 {
2359 struct si_shader_context *ctx = si_shader_context(bld_base);
2360 struct si_shader *shader = ctx->shader;
2361 struct lp_build_context *base = &bld_base->base;
2362 struct lp_build_context *uint =
2363 &ctx->soa.bld_base.uint_bld;
2364 LLVMValueRef args[9];
2365 LLVMValueRef pos_args[4][9] = { { 0 } };
2366 LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL, viewport_index_value = NULL;
2367 unsigned semantic_name, semantic_index;
2368 unsigned target;
2369 unsigned param_count = 0;
2370 unsigned pos_idx;
2371 int i;
2372
2373 for (i = 0; i < noutput; i++) {
2374 semantic_name = outputs[i].semantic_name;
2375 semantic_index = outputs[i].semantic_index;
2376 bool export_param = true;
2377
2378 switch (semantic_name) {
2379 case TGSI_SEMANTIC_POSITION: /* ignore these */
2380 case TGSI_SEMANTIC_PSIZE:
2381 case TGSI_SEMANTIC_CLIPVERTEX:
2382 case TGSI_SEMANTIC_EDGEFLAG:
2383 break;
2384 case TGSI_SEMANTIC_GENERIC:
2385 case TGSI_SEMANTIC_CLIPDIST:
2386 if (shader->key.opt.hw_vs.kill_outputs &
2387 (1ull << si_shader_io_get_unique_index(semantic_name, semantic_index)))
2388 export_param = false;
2389 break;
2390 default:
2391 if (shader->key.opt.hw_vs.kill_outputs2 &
2392 (1u << si_shader_io_get_unique_index2(semantic_name, semantic_index)))
2393 export_param = false;
2394 break;
2395 }
2396
2397 if (outputs[i].vertex_stream[0] != 0 &&
2398 outputs[i].vertex_stream[1] != 0 &&
2399 outputs[i].vertex_stream[2] != 0 &&
2400 outputs[i].vertex_stream[3] != 0)
2401 export_param = false;
2402
2403 handle_semantic:
2404 /* Select the correct target */
2405 switch(semantic_name) {
2406 case TGSI_SEMANTIC_PSIZE:
2407 psize_value = outputs[i].values[0];
2408 continue;
2409 case TGSI_SEMANTIC_EDGEFLAG:
2410 edgeflag_value = outputs[i].values[0];
2411 continue;
2412 case TGSI_SEMANTIC_LAYER:
2413 layer_value = outputs[i].values[0];
2414 semantic_name = TGSI_SEMANTIC_GENERIC;
2415 goto handle_semantic;
2416 case TGSI_SEMANTIC_VIEWPORT_INDEX:
2417 viewport_index_value = outputs[i].values[0];
2418 semantic_name = TGSI_SEMANTIC_GENERIC;
2419 goto handle_semantic;
2420 case TGSI_SEMANTIC_POSITION:
2421 target = V_008DFC_SQ_EXP_POS;
2422 break;
2423 case TGSI_SEMANTIC_COLOR:
2424 case TGSI_SEMANTIC_BCOLOR:
2425 if (!export_param)
2426 continue;
2427 target = V_008DFC_SQ_EXP_PARAM + param_count;
2428 assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
2429 shader->info.vs_output_param_offset[i] = param_count;
2430 param_count++;
2431 break;
2432 case TGSI_SEMANTIC_CLIPDIST:
2433 if (shader->key.opt.hw_vs.clip_disable) {
2434 semantic_name = TGSI_SEMANTIC_GENERIC;
2435 goto handle_semantic;
2436 }
2437 target = V_008DFC_SQ_EXP_POS + 2 + semantic_index;
2438 break;
2439 case TGSI_SEMANTIC_CLIPVERTEX:
2440 if (shader->key.opt.hw_vs.clip_disable)
2441 continue;
2442 si_llvm_emit_clipvertex(bld_base, pos_args, outputs[i].values);
2443 continue;
2444 case TGSI_SEMANTIC_PRIMID:
2445 case TGSI_SEMANTIC_FOG:
2446 case TGSI_SEMANTIC_TEXCOORD:
2447 case TGSI_SEMANTIC_GENERIC:
2448 if (!export_param)
2449 continue;
2450 target = V_008DFC_SQ_EXP_PARAM + param_count;
2451 assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
2452 shader->info.vs_output_param_offset[i] = param_count;
2453 param_count++;
2454 break;
2455 default:
2456 target = 0;
2457 fprintf(stderr,
2458 "Warning: SI unhandled vs output type:%d\n",
2459 semantic_name);
2460 }
2461
2462 si_llvm_init_export_args(bld_base, outputs[i].values, target, args);
2463
2464 if (target >= V_008DFC_SQ_EXP_POS &&
2465 target <= (V_008DFC_SQ_EXP_POS + 3)) {
2466 memcpy(pos_args[target - V_008DFC_SQ_EXP_POS],
2467 args, sizeof(args));
2468 } else {
2469 lp_build_intrinsic(base->gallivm->builder,
2470 "llvm.SI.export", ctx->voidt,
2471 args, 9, 0);
2472 }
2473
2474 if (semantic_name == TGSI_SEMANTIC_CLIPDIST) {
2475 semantic_name = TGSI_SEMANTIC_GENERIC;
2476 goto handle_semantic;
2477 }
2478 }
2479
2480 shader->info.nr_param_exports = param_count;
2481
2482 /* We need to add the position output manually if it's missing. */
2483 if (!pos_args[0][0]) {
2484 pos_args[0][0] = lp_build_const_int32(base->gallivm, 0xf); /* writemask */
2485 pos_args[0][1] = uint->zero; /* EXEC mask */
2486 pos_args[0][2] = uint->zero; /* last export? */
2487 pos_args[0][3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_POS);
2488 pos_args[0][4] = uint->zero; /* COMPR flag */
2489 pos_args[0][5] = base->zero; /* X */
2490 pos_args[0][6] = base->zero; /* Y */
2491 pos_args[0][7] = base->zero; /* Z */
2492 pos_args[0][8] = base->one; /* W */
2493 }
2494
2495 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2496 if (shader->selector->info.writes_psize ||
2497 shader->selector->info.writes_edgeflag ||
2498 shader->selector->info.writes_viewport_index ||
2499 shader->selector->info.writes_layer) {
2500 pos_args[1][0] = lp_build_const_int32(base->gallivm, /* writemask */
2501 shader->selector->info.writes_psize |
2502 (shader->selector->info.writes_edgeflag << 1) |
2503 (shader->selector->info.writes_layer << 2) |
2504 (shader->selector->info.writes_viewport_index << 3));
2505 pos_args[1][1] = uint->zero; /* EXEC mask */
2506 pos_args[1][2] = uint->zero; /* last export? */
2507 pos_args[1][3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_POS + 1);
2508 pos_args[1][4] = uint->zero; /* COMPR flag */
2509 pos_args[1][5] = base->zero; /* X */
2510 pos_args[1][6] = base->zero; /* Y */
2511 pos_args[1][7] = base->zero; /* Z */
2512 pos_args[1][8] = base->zero; /* W */
2513
2514 if (shader->selector->info.writes_psize)
2515 pos_args[1][5] = psize_value;
2516
2517 if (shader->selector->info.writes_edgeflag) {
2518 /* The output is a float, but the hw expects an integer
2519 * with the first bit containing the edge flag. */
2520 edgeflag_value = LLVMBuildFPToUI(base->gallivm->builder,
2521 edgeflag_value,
2522 ctx->i32, "");
2523 edgeflag_value = lp_build_min(&bld_base->int_bld,
2524 edgeflag_value,
2525 bld_base->int_bld.one);
2526
2527 /* The LLVM intrinsic expects a float. */
2528 pos_args[1][6] = LLVMBuildBitCast(base->gallivm->builder,
2529 edgeflag_value,
2530 ctx->f32, "");
2531 }
2532
2533 if (shader->selector->info.writes_layer)
2534 pos_args[1][7] = layer_value;
2535
2536 if (shader->selector->info.writes_viewport_index)
2537 pos_args[1][8] = viewport_index_value;
2538 }
2539
2540 for (i = 0; i < 4; i++)
2541 if (pos_args[i][0])
2542 shader->info.nr_pos_exports++;
2543
2544 pos_idx = 0;
2545 for (i = 0; i < 4; i++) {
2546 if (!pos_args[i][0])
2547 continue;
2548
2549 /* Specify the target we are exporting */
2550 pos_args[i][3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_POS + pos_idx++);
2551
2552 if (pos_idx == shader->info.nr_pos_exports)
2553 /* Specify that this is the last export */
2554 pos_args[i][2] = uint->one;
2555
2556 lp_build_intrinsic(base->gallivm->builder, "llvm.SI.export",
2557 ctx->voidt, pos_args[i], 9, 0);
2558 }
2559 }
2560
2561 /**
2562 * Forward all outputs from the vertex shader to the TES. This is only used
2563 * for the fixed function TCS.
2564 */
2565 static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
2566 {
2567 struct si_shader_context *ctx = si_shader_context(bld_base);
2568 struct gallivm_state *gallivm = bld_base->base.gallivm;
2569 LLVMValueRef invocation_id, rw_buffers, buffer, buffer_offset;
2570 LLVMValueRef lds_vertex_stride, lds_vertex_offset, lds_base;
2571 uint64_t inputs;
2572
2573 invocation_id = unpack_param(ctx, SI_PARAM_REL_IDS, 8, 5);
2574
2575 rw_buffers = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
2576 buffer = build_indexed_load_const(ctx, rw_buffers,
2577 lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
2578
2579 buffer_offset = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
2580
2581 lds_vertex_stride = unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 13, 8);
2582 lds_vertex_offset = LLVMBuildMul(gallivm->builder, invocation_id,
2583 lds_vertex_stride, "");
2584 lds_base = get_tcs_in_current_patch_offset(ctx);
2585 lds_base = LLVMBuildAdd(gallivm->builder, lds_base, lds_vertex_offset, "");
2586
2587 inputs = ctx->shader->key.mono.tcs.inputs_to_copy;
2588 while (inputs) {
2589 unsigned i = u_bit_scan64(&inputs);
2590
2591 LLVMValueRef lds_ptr = LLVMBuildAdd(gallivm->builder, lds_base,
2592 lp_build_const_int32(gallivm, 4 * i),
2593 "");
2594
2595 LLVMValueRef buffer_addr = get_tcs_tes_buffer_address(ctx,
2596 invocation_id,
2597 lp_build_const_int32(gallivm, i));
2598
2599 LLVMValueRef value = lds_load(bld_base, TGSI_TYPE_SIGNED, ~0,
2600 lds_ptr);
2601
2602 build_tbuffer_store_dwords(ctx, buffer, value, 4, buffer_addr,
2603 buffer_offset, 0);
2604 }
2605 }
2606
2607 static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
2608 LLVMValueRef rel_patch_id,
2609 LLVMValueRef invocation_id,
2610 LLVMValueRef tcs_out_current_patch_data_offset)
2611 {
2612 struct si_shader_context *ctx = si_shader_context(bld_base);
2613 struct gallivm_state *gallivm = bld_base->base.gallivm;
2614 struct si_shader *shader = ctx->shader;
2615 unsigned tess_inner_index, tess_outer_index;
2616 LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
2617 LLVMValueRef out[6], vec0, vec1, rw_buffers, tf_base;
2618 unsigned stride, outer_comps, inner_comps, i;
2619 struct lp_build_if_state if_ctx, inner_if_ctx;
2620
2621 si_llvm_emit_barrier(NULL, bld_base, NULL);
2622
2623 /* Do this only for invocation 0, because the tess levels are per-patch,
2624 * not per-vertex.
2625 *
2626 * This can't jump, because invocation 0 executes this. It should
2627 * at least mask out the loads and stores for other invocations.
2628 */
2629 lp_build_if(&if_ctx, gallivm,
2630 LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
2631 invocation_id, bld_base->uint_bld.zero, ""));
2632
2633 /* Determine the layout of one tess factor element in the buffer. */
2634 switch (shader->key.part.tcs.epilog.prim_mode) {
2635 case PIPE_PRIM_LINES:
2636 stride = 2; /* 2 dwords, 1 vec2 store */
2637 outer_comps = 2;
2638 inner_comps = 0;
2639 break;
2640 case PIPE_PRIM_TRIANGLES:
2641 stride = 4; /* 4 dwords, 1 vec4 store */
2642 outer_comps = 3;
2643 inner_comps = 1;
2644 break;
2645 case PIPE_PRIM_QUADS:
2646 stride = 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2647 outer_comps = 4;
2648 inner_comps = 2;
2649 break;
2650 default:
2651 assert(0);
2652 return;
2653 }
2654
2655 /* Load tess_inner and tess_outer from LDS.
2656 * Any invocation can write them, so we can't get them from a temporary.
2657 */
2658 tess_inner_index = si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER, 0);
2659 tess_outer_index = si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER, 0);
2660
2661 lds_base = tcs_out_current_patch_data_offset;
2662 lds_inner = LLVMBuildAdd(gallivm->builder, lds_base,
2663 lp_build_const_int32(gallivm,
2664 tess_inner_index * 4), "");
2665 lds_outer = LLVMBuildAdd(gallivm->builder, lds_base,
2666 lp_build_const_int32(gallivm,
2667 tess_outer_index * 4), "");
2668
2669 if (shader->key.part.tcs.epilog.prim_mode == PIPE_PRIM_LINES) {
2670 /* For isolines, the hardware expects tess factors in the
2671 * reverse order from what GLSL / TGSI specify.
2672 */
2673 out[0] = lds_load(bld_base, TGSI_TYPE_SIGNED, 1, lds_outer);
2674 out[1] = lds_load(bld_base, TGSI_TYPE_SIGNED, 0, lds_outer);
2675 } else {
2676 for (i = 0; i < outer_comps; i++)
2677 out[i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_outer);
2678 for (i = 0; i < inner_comps; i++)
2679 out[outer_comps+i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_inner);
2680 }
2681
2682 /* Convert the outputs to vectors for stores. */
2683 vec0 = lp_build_gather_values(gallivm, out, MIN2(stride, 4));
2684 vec1 = NULL;
2685
2686 if (stride > 4)
2687 vec1 = lp_build_gather_values(gallivm, out+4, stride - 4);
2688
2689 /* Get the buffer. */
2690 rw_buffers = LLVMGetParam(ctx->main_fn,
2691 SI_PARAM_RW_BUFFERS);
2692 buffer = build_indexed_load_const(ctx, rw_buffers,
2693 lp_build_const_int32(gallivm, SI_HS_RING_TESS_FACTOR));
2694
2695 /* Get the offset. */
2696 tf_base = LLVMGetParam(ctx->main_fn,
2697 SI_PARAM_TESS_FACTOR_OFFSET);
2698 byteoffset = LLVMBuildMul(gallivm->builder, rel_patch_id,
2699 lp_build_const_int32(gallivm, 4 * stride), "");
2700
2701 lp_build_if(&inner_if_ctx, gallivm,
2702 LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
2703 rel_patch_id, bld_base->uint_bld.zero, ""));
2704
2705 /* Store the dynamic HS control word. */
2706 build_tbuffer_store_dwords(ctx, buffer,
2707 lp_build_const_int32(gallivm, 0x80000000),
2708 1, lp_build_const_int32(gallivm, 0), tf_base, 0);
2709
2710 lp_build_endif(&inner_if_ctx);
2711
2712 /* Store the tessellation factors. */
2713 build_tbuffer_store_dwords(ctx, buffer, vec0,
2714 MIN2(stride, 4), byteoffset, tf_base, 4);
2715 if (vec1)
2716 build_tbuffer_store_dwords(ctx, buffer, vec1,
2717 stride - 4, byteoffset, tf_base, 20);
2718 lp_build_endif(&if_ctx);
2719 }
2720
2721 /* This only writes the tessellation factor levels. */
2722 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context *bld_base)
2723 {
2724 struct si_shader_context *ctx = si_shader_context(bld_base);
2725 LLVMValueRef rel_patch_id, invocation_id, tf_lds_offset;
2726
2727 si_copy_tcs_inputs(bld_base);
2728
2729 rel_patch_id = get_rel_patch_id(ctx);
2730 invocation_id = unpack_param(ctx, SI_PARAM_REL_IDS, 8, 5);
2731 tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
2732
2733 /* Return epilog parameters from this function. */
2734 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2735 LLVMValueRef ret = ctx->return_value;
2736 LLVMValueRef rw_buffers, rw0, rw1, tf_soffset;
2737 unsigned vgpr;
2738
2739 /* RW_BUFFERS pointer */
2740 rw_buffers = LLVMGetParam(ctx->main_fn,
2741 SI_PARAM_RW_BUFFERS);
2742 rw_buffers = LLVMBuildPtrToInt(builder, rw_buffers, ctx->i64, "");
2743 rw_buffers = LLVMBuildBitCast(builder, rw_buffers, ctx->v2i32, "");
2744 rw0 = LLVMBuildExtractElement(builder, rw_buffers,
2745 bld_base->uint_bld.zero, "");
2746 rw1 = LLVMBuildExtractElement(builder, rw_buffers,
2747 bld_base->uint_bld.one, "");
2748 ret = LLVMBuildInsertValue(builder, ret, rw0, 0, "");
2749 ret = LLVMBuildInsertValue(builder, ret, rw1, 1, "");
2750
2751 /* Tess factor buffer soffset is after user SGPRs. */
2752 tf_soffset = LLVMGetParam(ctx->main_fn,
2753 SI_PARAM_TESS_FACTOR_OFFSET);
2754 ret = LLVMBuildInsertValue(builder, ret, tf_soffset,
2755 SI_TCS_NUM_USER_SGPR + 1, "");
2756
2757 /* VGPRs */
2758 rel_patch_id = bitcast(bld_base, TGSI_TYPE_FLOAT, rel_patch_id);
2759 invocation_id = bitcast(bld_base, TGSI_TYPE_FLOAT, invocation_id);
2760 tf_lds_offset = bitcast(bld_base, TGSI_TYPE_FLOAT, tf_lds_offset);
2761
2762 vgpr = SI_TCS_NUM_USER_SGPR + 2;
2763 ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
2764 ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");
2765 ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
2766 ctx->return_value = ret;
2767 }
2768
2769 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context *bld_base)
2770 {
2771 struct si_shader_context *ctx = si_shader_context(bld_base);
2772 struct si_shader *shader = ctx->shader;
2773 struct tgsi_shader_info *info = &shader->selector->info;
2774 struct gallivm_state *gallivm = bld_base->base.gallivm;
2775 unsigned i, chan;
2776 LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
2777 ctx->param_rel_auto_id);
2778 LLVMValueRef vertex_dw_stride =
2779 unpack_param(ctx, SI_PARAM_LS_OUT_LAYOUT, 13, 8);
2780 LLVMValueRef base_dw_addr = LLVMBuildMul(gallivm->builder, vertex_id,
2781 vertex_dw_stride, "");
2782
2783 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2784 * its inputs from it. */
2785 for (i = 0; i < info->num_outputs; i++) {
2786 LLVMValueRef *out_ptr = ctx->soa.outputs[i];
2787 unsigned name = info->output_semantic_name[i];
2788 unsigned index = info->output_semantic_index[i];
2789 int param = si_shader_io_get_unique_index(name, index);
2790 LLVMValueRef dw_addr = LLVMBuildAdd(gallivm->builder, base_dw_addr,
2791 lp_build_const_int32(gallivm, param * 4), "");
2792
2793 for (chan = 0; chan < 4; chan++) {
2794 lds_store(bld_base, chan, dw_addr,
2795 LLVMBuildLoad(gallivm->builder, out_ptr[chan], ""));
2796 }
2797 }
2798 }
2799
2800 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context *bld_base)
2801 {
2802 struct si_shader_context *ctx = si_shader_context(bld_base);
2803 struct gallivm_state *gallivm = bld_base->base.gallivm;
2804 struct si_shader *es = ctx->shader;
2805 struct tgsi_shader_info *info = &es->selector->info;
2806 LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
2807 ctx->param_es2gs_offset);
2808 unsigned chan;
2809 int i;
2810
2811 for (i = 0; i < info->num_outputs; i++) {
2812 LLVMValueRef *out_ptr =
2813 ctx->soa.outputs[i];
2814 int param_index;
2815
2816 if (info->output_semantic_name[i] == TGSI_SEMANTIC_VIEWPORT_INDEX ||
2817 info->output_semantic_name[i] == TGSI_SEMANTIC_LAYER)
2818 continue;
2819
2820 param_index = si_shader_io_get_unique_index(info->output_semantic_name[i],
2821 info->output_semantic_index[i]);
2822
2823 for (chan = 0; chan < 4; chan++) {
2824 LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
2825 out_val = LLVMBuildBitCast(gallivm->builder, out_val, ctx->i32, "");
2826
2827 build_tbuffer_store(ctx,
2828 ctx->esgs_ring,
2829 out_val, 1,
2830 LLVMGetUndef(ctx->i32), soffset,
2831 (4 * param_index + chan) * 4,
2832 V_008F0C_BUF_DATA_FORMAT_32,
2833 V_008F0C_BUF_NUM_FORMAT_UINT,
2834 0, 0, 1, 1, 0);
2835 }
2836 }
2837 }
2838
2839 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context *bld_base)
2840 {
2841 struct si_shader_context *ctx = si_shader_context(bld_base);
2842 struct gallivm_state *gallivm = bld_base->base.gallivm;
2843 LLVMValueRef args[2];
2844
2845 args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_NOP | SENDMSG_GS_DONE);
2846 args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
2847 lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
2848 ctx->voidt, args, 2, 0);
2849 }
2850
2851 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context *bld_base)
2852 {
2853 struct si_shader_context *ctx = si_shader_context(bld_base);
2854 struct gallivm_state *gallivm = bld_base->base.gallivm;
2855 struct tgsi_shader_info *info = &ctx->shader->selector->info;
2856 struct si_shader_output_values *outputs = NULL;
2857 int i,j;
2858
2859 assert(!ctx->shader->is_gs_copy_shader);
2860
2861 outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
2862
2863 /* Vertex color clamping.
2864 *
2865 * This uses a state constant loaded in a user data SGPR and
2866 * an IF statement is added that clamps all colors if the constant
2867 * is true.
2868 */
2869 if (ctx->type == PIPE_SHADER_VERTEX) {
2870 struct lp_build_if_state if_ctx;
2871 LLVMValueRef cond = NULL;
2872 LLVMValueRef addr, val;
2873
2874 for (i = 0; i < info->num_outputs; i++) {
2875 if (info->output_semantic_name[i] != TGSI_SEMANTIC_COLOR &&
2876 info->output_semantic_name[i] != TGSI_SEMANTIC_BCOLOR)
2877 continue;
2878
2879 /* We've found a color. */
2880 if (!cond) {
2881 /* The state is in the first bit of the user SGPR. */
2882 cond = LLVMGetParam(ctx->main_fn,
2883 SI_PARAM_VS_STATE_BITS);
2884 cond = LLVMBuildTrunc(gallivm->builder, cond,
2885 ctx->i1, "");
2886 lp_build_if(&if_ctx, gallivm, cond);
2887 }
2888
2889 for (j = 0; j < 4; j++) {
2890 addr = ctx->soa.outputs[i][j];
2891 val = LLVMBuildLoad(gallivm->builder, addr, "");
2892 val = si_llvm_saturate(bld_base, val);
2893 LLVMBuildStore(gallivm->builder, val, addr);
2894 }
2895 }
2896
2897 if (cond)
2898 lp_build_endif(&if_ctx);
2899 }
2900
2901 for (i = 0; i < info->num_outputs; i++) {
2902 outputs[i].semantic_name = info->output_semantic_name[i];
2903 outputs[i].semantic_index = info->output_semantic_index[i];
2904
2905 for (j = 0; j < 4; j++) {
2906 outputs[i].values[j] =
2907 LLVMBuildLoad(gallivm->builder,
2908 ctx->soa.outputs[i][j],
2909 "");
2910 outputs[i].vertex_stream[j] =
2911 (info->output_streams[i] >> (2 * j)) & 3;
2912 }
2913
2914 }
2915
2916 /* Return the primitive ID from the LLVM function. */
2917 ctx->return_value =
2918 LLVMBuildInsertValue(gallivm->builder,
2919 ctx->return_value,
2920 bitcast(bld_base, TGSI_TYPE_FLOAT,
2921 get_primitive_id(bld_base, 0)),
2922 VS_EPILOG_PRIMID_LOC, "");
2923
2924 if (ctx->shader->selector->so.num_outputs)
2925 si_llvm_emit_streamout(ctx, outputs, i, 0);
2926 si_llvm_export_vs(bld_base, outputs, i);
2927 FREE(outputs);
2928 }
2929
2930 struct si_ps_exports {
2931 unsigned num;
2932 LLVMValueRef args[10][9];
2933 };
2934
2935 unsigned si_get_spi_shader_z_format(bool writes_z, bool writes_stencil,
2936 bool writes_samplemask)
2937 {
2938 if (writes_z) {
2939 /* Z needs 32 bits. */
2940 if (writes_samplemask)
2941 return V_028710_SPI_SHADER_32_ABGR;
2942 else if (writes_stencil)
2943 return V_028710_SPI_SHADER_32_GR;
2944 else
2945 return V_028710_SPI_SHADER_32_R;
2946 } else if (writes_stencil || writes_samplemask) {
2947 /* Both stencil and sample mask need only 16 bits. */
2948 return V_028710_SPI_SHADER_UINT16_ABGR;
2949 } else {
2950 return V_028710_SPI_SHADER_ZERO;
2951 }
2952 }
2953
2954 static void si_export_mrt_z(struct lp_build_tgsi_context *bld_base,
2955 LLVMValueRef depth, LLVMValueRef stencil,
2956 LLVMValueRef samplemask, struct si_ps_exports *exp)
2957 {
2958 struct si_shader_context *ctx = si_shader_context(bld_base);
2959 struct lp_build_context *base = &bld_base->base;
2960 struct lp_build_context *uint = &bld_base->uint_bld;
2961 LLVMValueRef args[9];
2962 unsigned mask = 0;
2963 unsigned format = si_get_spi_shader_z_format(depth != NULL,
2964 stencil != NULL,
2965 samplemask != NULL);
2966
2967 assert(depth || stencil || samplemask);
2968
2969 args[1] = uint->one; /* whether the EXEC mask is valid */
2970 args[2] = uint->one; /* DONE bit */
2971
2972 /* Specify the target we are exporting */
2973 args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRTZ);
2974
2975 args[4] = uint->zero; /* COMP flag */
2976 args[5] = base->undef; /* R, depth */
2977 args[6] = base->undef; /* G, stencil test value[0:7], stencil op value[8:15] */
2978 args[7] = base->undef; /* B, sample mask */
2979 args[8] = base->undef; /* A, alpha to mask */
2980
2981 if (format == V_028710_SPI_SHADER_UINT16_ABGR) {
2982 assert(!depth);
2983 args[4] = uint->one; /* COMPR flag */
2984
2985 if (stencil) {
2986 /* Stencil should be in X[23:16]. */
2987 stencil = bitcast(bld_base, TGSI_TYPE_UNSIGNED, stencil);
2988 stencil = LLVMBuildShl(base->gallivm->builder, stencil,
2989 LLVMConstInt(ctx->i32, 16, 0), "");
2990 args[5] = bitcast(bld_base, TGSI_TYPE_FLOAT, stencil);
2991 mask |= 0x3;
2992 }
2993 if (samplemask) {
2994 /* SampleMask should be in Y[15:0]. */
2995 args[6] = samplemask;
2996 mask |= 0xc;
2997 }
2998 } else {
2999 if (depth) {
3000 args[5] = depth;
3001 mask |= 0x1;
3002 }
3003 if (stencil) {
3004 args[6] = stencil;
3005 mask |= 0x2;
3006 }
3007 if (samplemask) {
3008 args[7] = samplemask;
3009 mask |= 0x4;
3010 }
3011 }
3012
3013 /* SI (except OLAND and HAINAN) has a bug that it only looks
3014 * at the X writemask component. */
3015 if (ctx->screen->b.chip_class == SI &&
3016 ctx->screen->b.family != CHIP_OLAND &&
3017 ctx->screen->b.family != CHIP_HAINAN)
3018 mask |= 0x1;
3019
3020 /* Specify which components to enable */
3021 args[0] = lp_build_const_int32(base->gallivm, mask);
3022
3023 memcpy(exp->args[exp->num++], args, sizeof(args));
3024 }
3025
3026 static void si_export_mrt_color(struct lp_build_tgsi_context *bld_base,
3027 LLVMValueRef *color, unsigned index,
3028 unsigned samplemask_param,
3029 bool is_last, struct si_ps_exports *exp)
3030 {
3031 struct si_shader_context *ctx = si_shader_context(bld_base);
3032 struct lp_build_context *base = &bld_base->base;
3033 int i;
3034
3035 /* Clamp color */
3036 if (ctx->shader->key.part.ps.epilog.clamp_color)
3037 for (i = 0; i < 4; i++)
3038 color[i] = si_llvm_saturate(bld_base, color[i]);
3039
3040 /* Alpha to one */
3041 if (ctx->shader->key.part.ps.epilog.alpha_to_one)
3042 color[3] = base->one;
3043
3044 /* Alpha test */
3045 if (index == 0 &&
3046 ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
3047 si_alpha_test(bld_base, color[3]);
3048
3049 /* Line & polygon smoothing */
3050 if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
3051 color[3] = si_scale_alpha_by_sample_mask(bld_base, color[3],
3052 samplemask_param);
3053
3054 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3055 if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
3056 LLVMValueRef args[8][9];
3057 int c, last = -1;
3058
3059 /* Get the export arguments, also find out what the last one is. */
3060 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
3061 si_llvm_init_export_args(bld_base, color,
3062 V_008DFC_SQ_EXP_MRT + c, args[c]);
3063 if (args[c][0] != bld_base->uint_bld.zero)
3064 last = c;
3065 }
3066
3067 /* Emit all exports. */
3068 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
3069 if (is_last && last == c) {
3070 args[c][1] = bld_base->uint_bld.one; /* whether the EXEC mask is valid */
3071 args[c][2] = bld_base->uint_bld.one; /* DONE bit */
3072 } else if (args[c][0] == bld_base->uint_bld.zero)
3073 continue; /* unnecessary NULL export */
3074
3075 memcpy(exp->args[exp->num++], args[c], sizeof(args[c]));
3076 }
3077 } else {
3078 LLVMValueRef args[9];
3079
3080 /* Export */
3081 si_llvm_init_export_args(bld_base, color, V_008DFC_SQ_EXP_MRT + index,
3082 args);
3083 if (is_last) {
3084 args[1] = bld_base->uint_bld.one; /* whether the EXEC mask is valid */
3085 args[2] = bld_base->uint_bld.one; /* DONE bit */
3086 } else if (args[0] == bld_base->uint_bld.zero)
3087 return; /* unnecessary NULL export */
3088
3089 memcpy(exp->args[exp->num++], args, sizeof(args));
3090 }
3091 }
3092
3093 static void si_emit_ps_exports(struct si_shader_context *ctx,
3094 struct si_ps_exports *exp)
3095 {
3096 for (unsigned i = 0; i < exp->num; i++)
3097 lp_build_intrinsic(ctx->gallivm.builder,
3098 "llvm.SI.export", ctx->voidt,
3099 exp->args[i], 9, 0);
3100 }
3101
3102 static void si_export_null(struct lp_build_tgsi_context *bld_base)
3103 {
3104 struct si_shader_context *ctx = si_shader_context(bld_base);
3105 struct lp_build_context *base = &bld_base->base;
3106 struct lp_build_context *uint = &bld_base->uint_bld;
3107 LLVMValueRef args[9];
3108
3109 args[0] = lp_build_const_int32(base->gallivm, 0x0); /* enabled channels */
3110 args[1] = uint->one; /* whether the EXEC mask is valid */
3111 args[2] = uint->one; /* DONE bit */
3112 args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_NULL);
3113 args[4] = uint->zero; /* COMPR flag (0 = 32-bit export) */
3114 args[5] = base->undef; /* R */
3115 args[6] = base->undef; /* G */
3116 args[7] = base->undef; /* B */
3117 args[8] = base->undef; /* A */
3118
3119 lp_build_intrinsic(base->gallivm->builder, "llvm.SI.export",
3120 ctx->voidt, args, 9, 0);
3121 }
3122
3123 /**
3124 * Return PS outputs in this order:
3125 *
3126 * v[0:3] = color0.xyzw
3127 * v[4:7] = color1.xyzw
3128 * ...
3129 * vN+0 = Depth
3130 * vN+1 = Stencil
3131 * vN+2 = SampleMask
3132 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3133 *
3134 * The alpha-ref SGPR is returned via its original location.
3135 */
3136 static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context *bld_base)
3137 {
3138 struct si_shader_context *ctx = si_shader_context(bld_base);
3139 struct si_shader *shader = ctx->shader;
3140 struct lp_build_context *base = &bld_base->base;
3141 struct tgsi_shader_info *info = &shader->selector->info;
3142 LLVMBuilderRef builder = base->gallivm->builder;
3143 unsigned i, j, first_vgpr, vgpr;
3144
3145 LLVMValueRef color[8][4] = {};
3146 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
3147 LLVMValueRef ret;
3148
3149 /* Read the output values. */
3150 for (i = 0; i < info->num_outputs; i++) {
3151 unsigned semantic_name = info->output_semantic_name[i];
3152 unsigned semantic_index = info->output_semantic_index[i];
3153
3154 switch (semantic_name) {
3155 case TGSI_SEMANTIC_COLOR:
3156 assert(semantic_index < 8);
3157 for (j = 0; j < 4; j++) {
3158 LLVMValueRef ptr = ctx->soa.outputs[i][j];
3159 LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
3160 color[semantic_index][j] = result;
3161 }
3162 break;
3163 case TGSI_SEMANTIC_POSITION:
3164 depth = LLVMBuildLoad(builder,
3165 ctx->soa.outputs[i][2], "");
3166 break;
3167 case TGSI_SEMANTIC_STENCIL:
3168 stencil = LLVMBuildLoad(builder,
3169 ctx->soa.outputs[i][1], "");
3170 break;
3171 case TGSI_SEMANTIC_SAMPLEMASK:
3172 samplemask = LLVMBuildLoad(builder,
3173 ctx->soa.outputs[i][0], "");
3174 break;
3175 default:
3176 fprintf(stderr, "Warning: SI unhandled fs output type:%d\n",
3177 semantic_name);
3178 }
3179 }
3180
3181 /* Fill the return structure. */
3182 ret = ctx->return_value;
3183
3184 /* Set SGPRs. */
3185 ret = LLVMBuildInsertValue(builder, ret,
3186 bitcast(bld_base, TGSI_TYPE_SIGNED,
3187 LLVMGetParam(ctx->main_fn,
3188 SI_PARAM_ALPHA_REF)),
3189 SI_SGPR_ALPHA_REF, "");
3190
3191 /* Set VGPRs */
3192 first_vgpr = vgpr = SI_SGPR_ALPHA_REF + 1;
3193 for (i = 0; i < ARRAY_SIZE(color); i++) {
3194 if (!color[i][0])
3195 continue;
3196
3197 for (j = 0; j < 4; j++)
3198 ret = LLVMBuildInsertValue(builder, ret, color[i][j], vgpr++, "");
3199 }
3200 if (depth)
3201 ret = LLVMBuildInsertValue(builder, ret, depth, vgpr++, "");
3202 if (stencil)
3203 ret = LLVMBuildInsertValue(builder, ret, stencil, vgpr++, "");
3204 if (samplemask)
3205 ret = LLVMBuildInsertValue(builder, ret, samplemask, vgpr++, "");
3206
3207 /* Add the input sample mask for smoothing at the end. */
3208 if (vgpr < first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC)
3209 vgpr = first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC;
3210 ret = LLVMBuildInsertValue(builder, ret,
3211 LLVMGetParam(ctx->main_fn,
3212 SI_PARAM_SAMPLE_COVERAGE), vgpr++, "");
3213
3214 ctx->return_value = ret;
3215 }
3216
3217 /**
3218 * Given a v8i32 resource descriptor for a buffer, extract the size of the
3219 * buffer in number of elements and return it as an i32.
3220 */
3221 static LLVMValueRef get_buffer_size(
3222 struct lp_build_tgsi_context *bld_base,
3223 LLVMValueRef descriptor)
3224 {
3225 struct si_shader_context *ctx = si_shader_context(bld_base);
3226 struct gallivm_state *gallivm = bld_base->base.gallivm;
3227 LLVMBuilderRef builder = gallivm->builder;
3228 LLVMValueRef size =
3229 LLVMBuildExtractElement(builder, descriptor,
3230 lp_build_const_int32(gallivm, 2), "");
3231
3232 if (ctx->screen->b.chip_class >= VI) {
3233 /* On VI, the descriptor contains the size in bytes,
3234 * but TXQ must return the size in elements.
3235 * The stride is always non-zero for resources using TXQ.
3236 */
3237 LLVMValueRef stride =
3238 LLVMBuildExtractElement(builder, descriptor,
3239 lp_build_const_int32(gallivm, 1), "");
3240 stride = LLVMBuildLShr(builder, stride,
3241 lp_build_const_int32(gallivm, 16), "");
3242 stride = LLVMBuildAnd(builder, stride,
3243 lp_build_const_int32(gallivm, 0x3FFF), "");
3244
3245 size = LLVMBuildUDiv(builder, size, stride, "");
3246 }
3247
3248 return size;
3249 }
3250
3251 /**
3252 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
3253 * intrinsic names).
3254 */
3255 static void build_type_name_for_intr(
3256 LLVMTypeRef type,
3257 char *buf, unsigned bufsize)
3258 {
3259 LLVMTypeRef elem_type = type;
3260
3261 assert(bufsize >= 8);
3262
3263 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
3264 int ret = snprintf(buf, bufsize, "v%u",
3265 LLVMGetVectorSize(type));
3266 if (ret < 0) {
3267 char *type_name = LLVMPrintTypeToString(type);
3268 fprintf(stderr, "Error building type name for: %s\n",
3269 type_name);
3270 return;
3271 }
3272 elem_type = LLVMGetElementType(type);
3273 buf += ret;
3274 bufsize -= ret;
3275 }
3276 switch (LLVMGetTypeKind(elem_type)) {
3277 default: break;
3278 case LLVMIntegerTypeKind:
3279 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
3280 break;
3281 case LLVMFloatTypeKind:
3282 snprintf(buf, bufsize, "f32");
3283 break;
3284 case LLVMDoubleTypeKind:
3285 snprintf(buf, bufsize, "f64");
3286 break;
3287 }
3288 }
3289
3290 static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
3291 struct lp_build_tgsi_context *bld_base,
3292 struct lp_build_emit_data *emit_data);
3293
3294 /* Prevent optimizations (at least of memory accesses) across the current
3295 * point in the program by emitting empty inline assembly that is marked as
3296 * having side effects.
3297 */
3298 #if 0 /* unused currently */
3299 static void emit_optimization_barrier(struct si_shader_context *ctx)
3300 {
3301 LLVMBuilderRef builder = ctx->gallivm.builder;
3302 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
3303 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, "", "", true, false);
3304 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
3305 }
3306 #endif
3307
3308 /* Combine these with & instead of |. */
3309 #define NOOP_WAITCNT 0xf7f
3310 #define LGKM_CNT 0x07f
3311 #define VM_CNT 0xf70
3312
3313 static void emit_waitcnt(struct si_shader_context *ctx, unsigned simm16)
3314 {
3315 struct gallivm_state *gallivm = &ctx->gallivm;
3316 LLVMBuilderRef builder = gallivm->builder;
3317 LLVMValueRef args[1] = {
3318 lp_build_const_int32(gallivm, simm16)
3319 };
3320 lp_build_intrinsic(builder, "llvm.amdgcn.s.waitcnt",
3321 ctx->voidt, args, 1, 0);
3322 }
3323
3324 static void membar_emit(
3325 const struct lp_build_tgsi_action *action,
3326 struct lp_build_tgsi_context *bld_base,
3327 struct lp_build_emit_data *emit_data)
3328 {
3329 struct si_shader_context *ctx = si_shader_context(bld_base);
3330 LLVMValueRef src0 = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0);
3331 unsigned flags = LLVMConstIntGetZExtValue(src0);
3332 unsigned waitcnt = NOOP_WAITCNT;
3333
3334 if (flags & TGSI_MEMBAR_THREAD_GROUP)
3335 waitcnt &= VM_CNT & LGKM_CNT;
3336
3337 if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER |
3338 TGSI_MEMBAR_SHADER_BUFFER |
3339 TGSI_MEMBAR_SHADER_IMAGE))
3340 waitcnt &= VM_CNT;
3341
3342 if (flags & TGSI_MEMBAR_SHARED)
3343 waitcnt &= LGKM_CNT;
3344
3345 if (waitcnt != NOOP_WAITCNT)
3346 emit_waitcnt(ctx, waitcnt);
3347 }
3348
3349 static LLVMValueRef
3350 shader_buffer_fetch_rsrc(struct si_shader_context *ctx,
3351 const struct tgsi_full_src_register *reg)
3352 {
3353 LLVMValueRef index;
3354 LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
3355 SI_PARAM_SHADER_BUFFERS);
3356
3357 if (!reg->Register.Indirect)
3358 index = LLVMConstInt(ctx->i32, reg->Register.Index, 0);
3359 else
3360 index = get_bounded_indirect_index(ctx, &reg->Indirect,
3361 reg->Register.Index,
3362 SI_NUM_SHADER_BUFFERS);
3363
3364 return build_indexed_load_const(ctx, rsrc_ptr, index);
3365 }
3366
3367 static bool tgsi_is_array_sampler(unsigned target)
3368 {
3369 return target == TGSI_TEXTURE_1D_ARRAY ||
3370 target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
3371 target == TGSI_TEXTURE_2D_ARRAY ||
3372 target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
3373 target == TGSI_TEXTURE_CUBE_ARRAY ||
3374 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY ||
3375 target == TGSI_TEXTURE_2D_ARRAY_MSAA;
3376 }
3377
3378 static bool tgsi_is_array_image(unsigned target)
3379 {
3380 return target == TGSI_TEXTURE_3D ||
3381 target == TGSI_TEXTURE_CUBE ||
3382 target == TGSI_TEXTURE_1D_ARRAY ||
3383 target == TGSI_TEXTURE_2D_ARRAY ||
3384 target == TGSI_TEXTURE_CUBE_ARRAY ||
3385 target == TGSI_TEXTURE_2D_ARRAY_MSAA;
3386 }
3387
3388 /**
3389 * Given a 256-bit resource descriptor, force the DCC enable bit to off.
3390 *
3391 * At least on Tonga, executing image stores on images with DCC enabled and
3392 * non-trivial can eventually lead to lockups. This can occur when an
3393 * application binds an image as read-only but then uses a shader that writes
3394 * to it. The OpenGL spec allows almost arbitrarily bad behavior (including
3395 * program termination) in this case, but it doesn't cost much to be a bit
3396 * nicer: disabling DCC in the shader still leads to undefined results but
3397 * avoids the lockup.
3398 */
3399 static LLVMValueRef force_dcc_off(struct si_shader_context *ctx,
3400 LLVMValueRef rsrc)
3401 {
3402 if (ctx->screen->b.chip_class <= CIK) {
3403 return rsrc;
3404 } else {
3405 LLVMBuilderRef builder = ctx->gallivm.builder;
3406 LLVMValueRef i32_6 = LLVMConstInt(ctx->i32, 6, 0);
3407 LLVMValueRef i32_C = LLVMConstInt(ctx->i32, C_008F28_COMPRESSION_EN, 0);
3408 LLVMValueRef tmp;
3409
3410 tmp = LLVMBuildExtractElement(builder, rsrc, i32_6, "");
3411 tmp = LLVMBuildAnd(builder, tmp, i32_C, "");
3412 return LLVMBuildInsertElement(builder, rsrc, tmp, i32_6, "");
3413 }
3414 }
3415
3416 static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
3417 {
3418 return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
3419 CONST_ADDR_SPACE);
3420 }
3421
3422 /**
3423 * Load the resource descriptor for \p image.
3424 */
3425 static void
3426 image_fetch_rsrc(
3427 struct lp_build_tgsi_context *bld_base,
3428 const struct tgsi_full_src_register *image,
3429 bool is_store, unsigned target,
3430 LLVMValueRef *rsrc)
3431 {
3432 struct si_shader_context *ctx = si_shader_context(bld_base);
3433 LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
3434 SI_PARAM_IMAGES);
3435 LLVMValueRef index, tmp;
3436 bool dcc_off = target != TGSI_TEXTURE_BUFFER && is_store;
3437
3438 assert(image->Register.File == TGSI_FILE_IMAGE);
3439
3440 if (!image->Register.Indirect) {
3441 const struct tgsi_shader_info *info = bld_base->info;
3442
3443 index = LLVMConstInt(ctx->i32, image->Register.Index, 0);
3444
3445 if (info->images_writemask & (1 << image->Register.Index) &&
3446 target != TGSI_TEXTURE_BUFFER)
3447 dcc_off = true;
3448 } else {
3449 /* From the GL_ARB_shader_image_load_store extension spec:
3450 *
3451 * If a shader performs an image load, store, or atomic
3452 * operation using an image variable declared as an array,
3453 * and if the index used to select an individual element is
3454 * negative or greater than or equal to the size of the
3455 * array, the results of the operation are undefined but may
3456 * not lead to termination.
3457 */
3458 index = get_bounded_indirect_index(ctx, &image->Indirect,
3459 image->Register.Index,
3460 SI_NUM_IMAGES);
3461 }
3462
3463 if (target == TGSI_TEXTURE_BUFFER) {
3464 LLVMBuilderRef builder = ctx->gallivm.builder;
3465
3466 rsrc_ptr = LLVMBuildPointerCast(builder, rsrc_ptr,
3467 const_array(ctx->v4i32, 0), "");
3468 index = LLVMBuildMul(builder, index,
3469 LLVMConstInt(ctx->i32, 2, 0), "");
3470 index = LLVMBuildAdd(builder, index,
3471 LLVMConstInt(ctx->i32, 1, 0), "");
3472 *rsrc = build_indexed_load_const(ctx, rsrc_ptr, index);
3473 return;
3474 }
3475
3476 tmp = build_indexed_load_const(ctx, rsrc_ptr, index);
3477 if (dcc_off)
3478 tmp = force_dcc_off(ctx, tmp);
3479 *rsrc = tmp;
3480 }
3481
3482 static LLVMValueRef image_fetch_coords(
3483 struct lp_build_tgsi_context *bld_base,
3484 const struct tgsi_full_instruction *inst,
3485 unsigned src)
3486 {
3487 struct gallivm_state *gallivm = bld_base->base.gallivm;
3488 LLVMBuilderRef builder = gallivm->builder;
3489 unsigned target = inst->Memory.Texture;
3490 unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
3491 LLVMValueRef coords[4];
3492 LLVMValueRef tmp;
3493 int chan;
3494
3495 for (chan = 0; chan < num_coords; ++chan) {
3496 tmp = lp_build_emit_fetch(bld_base, inst, src, chan);
3497 tmp = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
3498 coords[chan] = tmp;
3499 }
3500
3501 if (num_coords == 1)
3502 return coords[0];
3503
3504 if (num_coords == 3) {
3505 /* LLVM has difficulties lowering 3-element vectors. */
3506 coords[3] = bld_base->uint_bld.undef;
3507 num_coords = 4;
3508 }
3509
3510 return lp_build_gather_values(gallivm, coords, num_coords);
3511 }
3512
3513 /**
3514 * Append the extra mode bits that are used by image load and store.
3515 */
3516 static void image_append_args(
3517 struct si_shader_context *ctx,
3518 struct lp_build_emit_data * emit_data,
3519 unsigned target,
3520 bool atomic,
3521 bool force_glc)
3522 {
3523 const struct tgsi_full_instruction *inst = emit_data->inst;
3524 LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
3525 LLVMValueRef i1true = LLVMConstInt(ctx->i1, 1, 0);
3526 LLVMValueRef r128 = i1false;
3527 LLVMValueRef da = tgsi_is_array_image(target) ? i1true : i1false;
3528 LLVMValueRef glc =
3529 force_glc ||
3530 inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE) ?
3531 i1true : i1false;
3532 LLVMValueRef slc = i1false;
3533 LLVMValueRef lwe = i1false;
3534
3535 if (atomic || (HAVE_LLVM <= 0x0309)) {
3536 emit_data->args[emit_data->arg_count++] = r128;
3537 emit_data->args[emit_data->arg_count++] = da;
3538 if (!atomic) {
3539 emit_data->args[emit_data->arg_count++] = glc;
3540 }
3541 emit_data->args[emit_data->arg_count++] = slc;
3542 return;
3543 }
3544
3545 /* HAVE_LLVM >= 0x0400 */
3546 emit_data->args[emit_data->arg_count++] = glc;
3547 emit_data->args[emit_data->arg_count++] = slc;
3548 emit_data->args[emit_data->arg_count++] = lwe;
3549 emit_data->args[emit_data->arg_count++] = da;
3550 }
3551
3552 /**
3553 * Append the resource and indexing arguments for buffer intrinsics.
3554 *
3555 * \param rsrc the v4i32 buffer resource
3556 * \param index index into the buffer (stride-based)
3557 * \param offset byte offset into the buffer
3558 */
3559 static void buffer_append_args(
3560 struct si_shader_context *ctx,
3561 struct lp_build_emit_data *emit_data,
3562 LLVMValueRef rsrc,
3563 LLVMValueRef index,
3564 LLVMValueRef offset,
3565 bool atomic,
3566 bool force_glc)
3567 {
3568 const struct tgsi_full_instruction *inst = emit_data->inst;
3569 LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
3570 LLVMValueRef i1true = LLVMConstInt(ctx->i1, 1, 0);
3571
3572 emit_data->args[emit_data->arg_count++] = rsrc;
3573 emit_data->args[emit_data->arg_count++] = index; /* vindex */
3574 emit_data->args[emit_data->arg_count++] = offset; /* voffset */
3575 if (!atomic) {
3576 emit_data->args[emit_data->arg_count++] =
3577 force_glc ||
3578 inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE) ?
3579 i1true : i1false; /* glc */
3580 }
3581 emit_data->args[emit_data->arg_count++] = i1false; /* slc */
3582 }
3583
3584 static void load_fetch_args(
3585 struct lp_build_tgsi_context * bld_base,
3586 struct lp_build_emit_data * emit_data)
3587 {
3588 struct si_shader_context *ctx = si_shader_context(bld_base);
3589 struct gallivm_state *gallivm = bld_base->base.gallivm;
3590 const struct tgsi_full_instruction * inst = emit_data->inst;
3591 unsigned target = inst->Memory.Texture;
3592 LLVMValueRef rsrc;
3593
3594 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
3595
3596 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
3597 LLVMBuilderRef builder = gallivm->builder;
3598 LLVMValueRef offset;
3599 LLVMValueRef tmp;
3600
3601 rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
3602
3603 tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
3604 offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
3605
3606 buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
3607 offset, false, false);
3608 } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
3609 LLVMValueRef coords;
3610
3611 image_fetch_rsrc(bld_base, &inst->Src[0], false, target, &rsrc);
3612 coords = image_fetch_coords(bld_base, inst, 1);
3613
3614 if (target == TGSI_TEXTURE_BUFFER) {
3615 buffer_append_args(ctx, emit_data, rsrc, coords,
3616 bld_base->uint_bld.zero, false, false);
3617 } else {
3618 emit_data->args[0] = coords;
3619 emit_data->args[1] = rsrc;
3620 emit_data->args[2] = lp_build_const_int32(gallivm, 15); /* dmask */
3621 emit_data->arg_count = 3;
3622
3623 image_append_args(ctx, emit_data, target, false, false);
3624 }
3625 }
3626 }
3627
3628 static void load_emit_buffer(struct si_shader_context *ctx,
3629 struct lp_build_emit_data *emit_data)
3630 {
3631 const struct tgsi_full_instruction *inst = emit_data->inst;
3632 struct gallivm_state *gallivm = &ctx->gallivm;
3633 LLVMBuilderRef builder = gallivm->builder;
3634 uint writemask = inst->Dst[0].Register.WriteMask;
3635 uint count = util_last_bit(writemask);
3636 const char *intrinsic_name;
3637 LLVMTypeRef dst_type;
3638
3639 switch (count) {
3640 case 1:
3641 intrinsic_name = "llvm.amdgcn.buffer.load.f32";
3642 dst_type = ctx->f32;
3643 break;
3644 case 2:
3645 intrinsic_name = "llvm.amdgcn.buffer.load.v2f32";
3646 dst_type = LLVMVectorType(ctx->f32, 2);
3647 break;
3648 default: // 3 & 4
3649 intrinsic_name = "llvm.amdgcn.buffer.load.v4f32";
3650 dst_type = ctx->v4f32;
3651 count = 4;
3652 }
3653
3654 emit_data->output[emit_data->chan] = lp_build_intrinsic(
3655 builder, intrinsic_name, dst_type,
3656 emit_data->args, emit_data->arg_count,
3657 LP_FUNC_ATTR_READONLY);
3658 }
3659
3660 static LLVMValueRef get_memory_ptr(struct si_shader_context *ctx,
3661 const struct tgsi_full_instruction *inst,
3662 LLVMTypeRef type, int arg)
3663 {
3664 struct gallivm_state *gallivm = &ctx->gallivm;
3665 LLVMBuilderRef builder = gallivm->builder;
3666 LLVMValueRef offset, ptr;
3667 int addr_space;
3668
3669 offset = lp_build_emit_fetch(&ctx->soa.bld_base, inst, arg, 0);
3670 offset = LLVMBuildBitCast(builder, offset, ctx->i32, "");
3671
3672 ptr = ctx->shared_memory;
3673 ptr = LLVMBuildGEP(builder, ptr, &offset, 1, "");
3674 addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3675 ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, addr_space), "");
3676
3677 return ptr;
3678 }
3679
3680 static void load_emit_memory(
3681 struct si_shader_context *ctx,
3682 struct lp_build_emit_data *emit_data)
3683 {
3684 const struct tgsi_full_instruction *inst = emit_data->inst;
3685 struct lp_build_context *base = &ctx->soa.bld_base.base;
3686 struct gallivm_state *gallivm = &ctx->gallivm;
3687 LLVMBuilderRef builder = gallivm->builder;
3688 unsigned writemask = inst->Dst[0].Register.WriteMask;
3689 LLVMValueRef channels[4], ptr, derived_ptr, index;
3690 int chan;
3691
3692 ptr = get_memory_ptr(ctx, inst, base->elem_type, 1);
3693
3694 for (chan = 0; chan < 4; ++chan) {
3695 if (!(writemask & (1 << chan))) {
3696 channels[chan] = LLVMGetUndef(base->elem_type);
3697 continue;
3698 }
3699
3700 index = lp_build_const_int32(gallivm, chan);
3701 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
3702 channels[chan] = LLVMBuildLoad(builder, derived_ptr, "");
3703 }
3704 emit_data->output[emit_data->chan] = lp_build_gather_values(gallivm, channels, 4);
3705 }
3706
3707 static void get_image_intr_name(const char *base_name,
3708 LLVMTypeRef data_type,
3709 LLVMTypeRef coords_type,
3710 LLVMTypeRef rsrc_type,
3711 char *out_name, unsigned out_len)
3712 {
3713 char coords_type_name[8];
3714
3715 build_type_name_for_intr(coords_type, coords_type_name,
3716 sizeof(coords_type_name));
3717
3718 if (HAVE_LLVM <= 0x0309) {
3719 snprintf(out_name, out_len, "%s.%s", base_name, coords_type_name);
3720 } else {
3721 char data_type_name[8];
3722 char rsrc_type_name[8];
3723
3724 build_type_name_for_intr(data_type, data_type_name,
3725 sizeof(data_type_name));
3726 build_type_name_for_intr(rsrc_type, rsrc_type_name,
3727 sizeof(rsrc_type_name));
3728 snprintf(out_name, out_len, "%s.%s.%s.%s", base_name,
3729 data_type_name, coords_type_name, rsrc_type_name);
3730 }
3731 }
3732
3733 static void load_emit(
3734 const struct lp_build_tgsi_action *action,
3735 struct lp_build_tgsi_context *bld_base,
3736 struct lp_build_emit_data *emit_data)
3737 {
3738 struct si_shader_context *ctx = si_shader_context(bld_base);
3739 struct gallivm_state *gallivm = bld_base->base.gallivm;
3740 LLVMBuilderRef builder = gallivm->builder;
3741 const struct tgsi_full_instruction * inst = emit_data->inst;
3742 char intrinsic_name[64];
3743
3744 if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
3745 load_emit_memory(ctx, emit_data);
3746 return;
3747 }
3748
3749 if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
3750 emit_waitcnt(ctx, VM_CNT);
3751
3752 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
3753 load_emit_buffer(ctx, emit_data);
3754 return;
3755 }
3756
3757 if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
3758 emit_data->output[emit_data->chan] =
3759 lp_build_intrinsic(
3760 builder, "llvm.amdgcn.buffer.load.format.v4f32", emit_data->dst_type,
3761 emit_data->args, emit_data->arg_count,
3762 LP_FUNC_ATTR_READONLY);
3763 } else {
3764 get_image_intr_name("llvm.amdgcn.image.load",
3765 emit_data->dst_type, /* vdata */
3766 LLVMTypeOf(emit_data->args[0]), /* coords */
3767 LLVMTypeOf(emit_data->args[1]), /* rsrc */
3768 intrinsic_name, sizeof(intrinsic_name));
3769
3770 emit_data->output[emit_data->chan] =
3771 lp_build_intrinsic(
3772 builder, intrinsic_name, emit_data->dst_type,
3773 emit_data->args, emit_data->arg_count,
3774 LP_FUNC_ATTR_READONLY);
3775 }
3776 }
3777
3778 static void store_fetch_args(
3779 struct lp_build_tgsi_context * bld_base,
3780 struct lp_build_emit_data * emit_data)
3781 {
3782 struct si_shader_context *ctx = si_shader_context(bld_base);
3783 struct gallivm_state *gallivm = bld_base->base.gallivm;
3784 LLVMBuilderRef builder = gallivm->builder;
3785 const struct tgsi_full_instruction * inst = emit_data->inst;
3786 struct tgsi_full_src_register memory;
3787 LLVMValueRef chans[4];
3788 LLVMValueRef data;
3789 LLVMValueRef rsrc;
3790 unsigned chan;
3791
3792 emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
3793
3794 for (chan = 0; chan < 4; ++chan) {
3795 chans[chan] = lp_build_emit_fetch(bld_base, inst, 1, chan);
3796 }
3797 data = lp_build_gather_values(gallivm, chans, 4);
3798
3799 emit_data->args[emit_data->arg_count++] = data;
3800
3801 memory = tgsi_full_src_register_from_dst(&inst->Dst[0]);
3802
3803 if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
3804 LLVMValueRef offset;
3805 LLVMValueRef tmp;
3806
3807 rsrc = shader_buffer_fetch_rsrc(ctx, &memory);
3808
3809 tmp = lp_build_emit_fetch(bld_base, inst, 0, 0);
3810 offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
3811
3812 buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
3813 offset, false, false);
3814 } else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE) {
3815 unsigned target = inst->Memory.Texture;
3816 LLVMValueRef coords;
3817
3818 /* 8bit/16bit TC L1 write corruption bug on SI.
3819 * All store opcodes not aligned to a dword are affected.
3820 *
3821 * The only way to get unaligned stores in radeonsi is through
3822 * shader images.
3823 */
3824 bool force_glc = ctx->screen->b.chip_class == SI;
3825
3826 coords = image_fetch_coords(bld_base, inst, 0);
3827
3828 if (target == TGSI_TEXTURE_BUFFER) {
3829 image_fetch_rsrc(bld_base, &memory, true, target, &rsrc);
3830 buffer_append_args(ctx, emit_data, rsrc, coords,
3831 bld_base->uint_bld.zero, false, force_glc);
3832 } else {
3833 emit_data->args[1] = coords;
3834 image_fetch_rsrc(bld_base, &memory, true, target,
3835 &emit_data->args[2]);
3836 emit_data->args[3] = lp_build_const_int32(gallivm, 15); /* dmask */
3837 emit_data->arg_count = 4;
3838
3839 image_append_args(ctx, emit_data, target, false, force_glc);
3840 }
3841 }
3842 }
3843
3844 static void store_emit_buffer(
3845 struct si_shader_context *ctx,
3846 struct lp_build_emit_data *emit_data)
3847 {
3848 const struct tgsi_full_instruction *inst = emit_data->inst;
3849 struct gallivm_state *gallivm = &ctx->gallivm;
3850 LLVMBuilderRef builder = gallivm->builder;
3851 struct lp_build_context *uint_bld = &ctx->soa.bld_base.uint_bld;
3852 LLVMValueRef base_data = emit_data->args[0];
3853 LLVMValueRef base_offset = emit_data->args[3];
3854 unsigned writemask = inst->Dst[0].Register.WriteMask;
3855
3856 while (writemask) {
3857 int start, count;
3858 const char *intrinsic_name;
3859 LLVMValueRef data;
3860 LLVMValueRef offset;
3861 LLVMValueRef tmp;
3862
3863 u_bit_scan_consecutive_range(&writemask, &start, &count);
3864
3865 /* Due to an LLVM limitation, split 3-element writes
3866 * into a 2-element and a 1-element write. */
3867 if (count == 3) {
3868 writemask |= 1 << (start + 2);
3869 count = 2;
3870 }
3871
3872 if (count == 4) {
3873 data = base_data;
3874 intrinsic_name = "llvm.amdgcn.buffer.store.v4f32";
3875 } else if (count == 2) {
3876 LLVMTypeRef v2f32 = LLVMVectorType(ctx->f32, 2);
3877
3878 tmp = LLVMBuildExtractElement(
3879 builder, base_data,
3880 lp_build_const_int32(gallivm, start), "");
3881 data = LLVMBuildInsertElement(
3882 builder, LLVMGetUndef(v2f32), tmp,
3883 uint_bld->zero, "");
3884
3885 tmp = LLVMBuildExtractElement(
3886 builder, base_data,
3887 lp_build_const_int32(gallivm, start + 1), "");
3888 data = LLVMBuildInsertElement(
3889 builder, data, tmp, uint_bld->one, "");
3890
3891 intrinsic_name = "llvm.amdgcn.buffer.store.v2f32";
3892 } else {
3893 assert(count == 1);
3894 data = LLVMBuildExtractElement(
3895 builder, base_data,
3896 lp_build_const_int32(gallivm, start), "");
3897 intrinsic_name = "llvm.amdgcn.buffer.store.f32";
3898 }
3899
3900 offset = base_offset;
3901 if (start != 0) {
3902 offset = LLVMBuildAdd(
3903 builder, offset,
3904 lp_build_const_int32(gallivm, start * 4), "");
3905 }
3906
3907 emit_data->args[0] = data;
3908 emit_data->args[3] = offset;
3909
3910 lp_build_intrinsic(
3911 builder, intrinsic_name, emit_data->dst_type,
3912 emit_data->args, emit_data->arg_count, 0);
3913 }
3914 }
3915
3916 static void store_emit_memory(
3917 struct si_shader_context *ctx,
3918 struct lp_build_emit_data *emit_data)
3919 {
3920 const struct tgsi_full_instruction *inst = emit_data->inst;
3921 struct gallivm_state *gallivm = &ctx->gallivm;
3922 struct lp_build_context *base = &ctx->soa.bld_base.base;
3923 LLVMBuilderRef builder = gallivm->builder;
3924 unsigned writemask = inst->Dst[0].Register.WriteMask;
3925 LLVMValueRef ptr, derived_ptr, data, index;
3926 int chan;
3927
3928 ptr = get_memory_ptr(ctx, inst, base->elem_type, 0);
3929
3930 for (chan = 0; chan < 4; ++chan) {
3931 if (!(writemask & (1 << chan))) {
3932 continue;
3933 }
3934 data = lp_build_emit_fetch(&ctx->soa.bld_base, inst, 1, chan);
3935 index = lp_build_const_int32(gallivm, chan);
3936 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
3937 LLVMBuildStore(builder, data, derived_ptr);
3938 }
3939 }
3940
3941 static void store_emit(
3942 const struct lp_build_tgsi_action *action,
3943 struct lp_build_tgsi_context *bld_base,
3944 struct lp_build_emit_data *emit_data)
3945 {
3946 struct si_shader_context *ctx = si_shader_context(bld_base);
3947 struct gallivm_state *gallivm = bld_base->base.gallivm;
3948 LLVMBuilderRef builder = gallivm->builder;
3949 const struct tgsi_full_instruction * inst = emit_data->inst;
3950 unsigned target = inst->Memory.Texture;
3951 char intrinsic_name[64];
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 if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
3962 store_emit_buffer(ctx, emit_data);
3963 return;
3964 }
3965
3966 if (target == TGSI_TEXTURE_BUFFER) {
3967 emit_data->output[emit_data->chan] = lp_build_intrinsic(
3968 builder, "llvm.amdgcn.buffer.store.format.v4f32",
3969 emit_data->dst_type, emit_data->args,
3970 emit_data->arg_count, 0);
3971 } else {
3972 get_image_intr_name("llvm.amdgcn.image.store",
3973 LLVMTypeOf(emit_data->args[0]), /* vdata */
3974 LLVMTypeOf(emit_data->args[1]), /* coords */
3975 LLVMTypeOf(emit_data->args[2]), /* rsrc */
3976 intrinsic_name, sizeof(intrinsic_name));
3977
3978 emit_data->output[emit_data->chan] =
3979 lp_build_intrinsic(
3980 builder, intrinsic_name, emit_data->dst_type,
3981 emit_data->args, emit_data->arg_count, 0);
3982 }
3983 }
3984
3985 static void atomic_fetch_args(
3986 struct lp_build_tgsi_context * bld_base,
3987 struct lp_build_emit_data * emit_data)
3988 {
3989 struct si_shader_context *ctx = si_shader_context(bld_base);
3990 struct gallivm_state *gallivm = bld_base->base.gallivm;
3991 LLVMBuilderRef builder = gallivm->builder;
3992 const struct tgsi_full_instruction * inst = emit_data->inst;
3993 LLVMValueRef data1, data2;
3994 LLVMValueRef rsrc;
3995 LLVMValueRef tmp;
3996
3997 emit_data->dst_type = bld_base->base.elem_type;
3998
3999 tmp = lp_build_emit_fetch(bld_base, inst, 2, 0);
4000 data1 = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
4001
4002 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
4003 tmp = lp_build_emit_fetch(bld_base, inst, 3, 0);
4004 data2 = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
4005 }
4006
4007 /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
4008 * of arguments, which is reversed relative to TGSI (and GLSL)
4009 */
4010 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
4011 emit_data->args[emit_data->arg_count++] = data2;
4012 emit_data->args[emit_data->arg_count++] = data1;
4013
4014 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
4015 LLVMValueRef offset;
4016
4017 rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
4018
4019 tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
4020 offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
4021
4022 buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
4023 offset, true, false);
4024 } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
4025 unsigned target = inst->Memory.Texture;
4026 LLVMValueRef coords;
4027
4028 image_fetch_rsrc(bld_base, &inst->Src[0], true, target, &rsrc);
4029 coords = image_fetch_coords(bld_base, inst, 1);
4030
4031 if (target == TGSI_TEXTURE_BUFFER) {
4032 buffer_append_args(ctx, emit_data, rsrc, coords,
4033 bld_base->uint_bld.zero, true, false);
4034 } else {
4035 emit_data->args[emit_data->arg_count++] = coords;
4036 emit_data->args[emit_data->arg_count++] = rsrc;
4037
4038 image_append_args(ctx, emit_data, target, true, false);
4039 }
4040 }
4041 }
4042
4043 static void atomic_emit_memory(struct si_shader_context *ctx,
4044 struct lp_build_emit_data *emit_data) {
4045 struct gallivm_state *gallivm = &ctx->gallivm;
4046 LLVMBuilderRef builder = gallivm->builder;
4047 const struct tgsi_full_instruction * inst = emit_data->inst;
4048 LLVMValueRef ptr, result, arg;
4049
4050 ptr = get_memory_ptr(ctx, inst, ctx->i32, 1);
4051
4052 arg = lp_build_emit_fetch(&ctx->soa.bld_base, inst, 2, 0);
4053 arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
4054
4055 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
4056 LLVMValueRef new_data;
4057 new_data = lp_build_emit_fetch(&ctx->soa.bld_base,
4058 inst, 3, 0);
4059
4060 new_data = LLVMBuildBitCast(builder, new_data, ctx->i32, "");
4061
4062 #if HAVE_LLVM >= 0x309
4063 result = LLVMBuildAtomicCmpXchg(builder, ptr, arg, new_data,
4064 LLVMAtomicOrderingSequentiallyConsistent,
4065 LLVMAtomicOrderingSequentiallyConsistent,
4066 false);
4067 #endif
4068
4069 result = LLVMBuildExtractValue(builder, result, 0, "");
4070 } else {
4071 LLVMAtomicRMWBinOp op;
4072
4073 switch(inst->Instruction.Opcode) {
4074 case TGSI_OPCODE_ATOMUADD:
4075 op = LLVMAtomicRMWBinOpAdd;
4076 break;
4077 case TGSI_OPCODE_ATOMXCHG:
4078 op = LLVMAtomicRMWBinOpXchg;
4079 break;
4080 case TGSI_OPCODE_ATOMAND:
4081 op = LLVMAtomicRMWBinOpAnd;
4082 break;
4083 case TGSI_OPCODE_ATOMOR:
4084 op = LLVMAtomicRMWBinOpOr;
4085 break;
4086 case TGSI_OPCODE_ATOMXOR:
4087 op = LLVMAtomicRMWBinOpXor;
4088 break;
4089 case TGSI_OPCODE_ATOMUMIN:
4090 op = LLVMAtomicRMWBinOpUMin;
4091 break;
4092 case TGSI_OPCODE_ATOMUMAX:
4093 op = LLVMAtomicRMWBinOpUMax;
4094 break;
4095 case TGSI_OPCODE_ATOMIMIN:
4096 op = LLVMAtomicRMWBinOpMin;
4097 break;
4098 case TGSI_OPCODE_ATOMIMAX:
4099 op = LLVMAtomicRMWBinOpMax;
4100 break;
4101 default:
4102 unreachable("unknown atomic opcode");
4103 }
4104
4105 result = LLVMBuildAtomicRMW(builder, op, ptr, arg,
4106 LLVMAtomicOrderingSequentiallyConsistent,
4107 false);
4108 }
4109 emit_data->output[emit_data->chan] = LLVMBuildBitCast(builder, result, emit_data->dst_type, "");
4110 }
4111
4112 static void atomic_emit(
4113 const struct lp_build_tgsi_action *action,
4114 struct lp_build_tgsi_context *bld_base,
4115 struct lp_build_emit_data *emit_data)
4116 {
4117 struct si_shader_context *ctx = si_shader_context(bld_base);
4118 struct gallivm_state *gallivm = bld_base->base.gallivm;
4119 LLVMBuilderRef builder = gallivm->builder;
4120 const struct tgsi_full_instruction * inst = emit_data->inst;
4121 char intrinsic_name[40];
4122 LLVMValueRef tmp;
4123
4124 if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
4125 atomic_emit_memory(ctx, emit_data);
4126 return;
4127 }
4128
4129 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
4130 inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
4131 snprintf(intrinsic_name, sizeof(intrinsic_name),
4132 "llvm.amdgcn.buffer.atomic.%s", action->intr_name);
4133 } else {
4134 LLVMValueRef coords;
4135 char coords_type[8];
4136
4137 if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
4138 coords = emit_data->args[2];
4139 else
4140 coords = emit_data->args[1];
4141
4142 build_type_name_for_intr(LLVMTypeOf(coords), coords_type, sizeof(coords_type));
4143 snprintf(intrinsic_name, sizeof(intrinsic_name),
4144 "llvm.amdgcn.image.atomic.%s.%s",
4145 action->intr_name, coords_type);
4146 }
4147
4148 tmp = lp_build_intrinsic(
4149 builder, intrinsic_name, bld_base->uint_bld.elem_type,
4150 emit_data->args, emit_data->arg_count, 0);
4151 emit_data->output[emit_data->chan] =
4152 LLVMBuildBitCast(builder, tmp, bld_base->base.elem_type, "");
4153 }
4154
4155 static void resq_fetch_args(
4156 struct lp_build_tgsi_context * bld_base,
4157 struct lp_build_emit_data * emit_data)
4158 {
4159 struct si_shader_context *ctx = si_shader_context(bld_base);
4160 struct gallivm_state *gallivm = bld_base->base.gallivm;
4161 const struct tgsi_full_instruction *inst = emit_data->inst;
4162 const struct tgsi_full_src_register *reg = &inst->Src[0];
4163
4164 emit_data->dst_type = ctx->v4i32;
4165
4166 if (reg->Register.File == TGSI_FILE_BUFFER) {
4167 emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg);
4168 emit_data->arg_count = 1;
4169 } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
4170 image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
4171 &emit_data->args[0]);
4172 emit_data->arg_count = 1;
4173 } else {
4174 emit_data->args[0] = bld_base->uint_bld.zero; /* mip level */
4175 image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
4176 &emit_data->args[1]);
4177 emit_data->args[2] = lp_build_const_int32(gallivm, 15); /* dmask */
4178 emit_data->args[3] = bld_base->uint_bld.zero; /* unorm */
4179 emit_data->args[4] = bld_base->uint_bld.zero; /* r128 */
4180 emit_data->args[5] = tgsi_is_array_image(inst->Memory.Texture) ?
4181 bld_base->uint_bld.one : bld_base->uint_bld.zero; /* da */
4182 emit_data->args[6] = bld_base->uint_bld.zero; /* glc */
4183 emit_data->args[7] = bld_base->uint_bld.zero; /* slc */
4184 emit_data->args[8] = bld_base->uint_bld.zero; /* tfe */
4185 emit_data->args[9] = bld_base->uint_bld.zero; /* lwe */
4186 emit_data->arg_count = 10;
4187 }
4188 }
4189
4190 static void resq_emit(
4191 const struct lp_build_tgsi_action *action,
4192 struct lp_build_tgsi_context *bld_base,
4193 struct lp_build_emit_data *emit_data)
4194 {
4195 struct gallivm_state *gallivm = bld_base->base.gallivm;
4196 LLVMBuilderRef builder = gallivm->builder;
4197 const struct tgsi_full_instruction *inst = emit_data->inst;
4198 LLVMValueRef out;
4199
4200 if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
4201 out = LLVMBuildExtractElement(builder, emit_data->args[0],
4202 lp_build_const_int32(gallivm, 2), "");
4203 } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
4204 out = get_buffer_size(bld_base, emit_data->args[0]);
4205 } else {
4206 out = lp_build_intrinsic(
4207 builder, "llvm.SI.getresinfo.i32", emit_data->dst_type,
4208 emit_data->args, emit_data->arg_count,
4209 LP_FUNC_ATTR_READNONE);
4210
4211 /* Divide the number of layers by 6 to get the number of cubes. */
4212 if (inst->Memory.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
4213 LLVMValueRef imm2 = lp_build_const_int32(gallivm, 2);
4214 LLVMValueRef imm6 = lp_build_const_int32(gallivm, 6);
4215
4216 LLVMValueRef z = LLVMBuildExtractElement(builder, out, imm2, "");
4217 z = LLVMBuildSDiv(builder, z, imm6, "");
4218 out = LLVMBuildInsertElement(builder, out, z, imm2, "");
4219 }
4220 }
4221
4222 emit_data->output[emit_data->chan] = out;
4223 }
4224
4225 static void set_tex_fetch_args(struct si_shader_context *ctx,
4226 struct lp_build_emit_data *emit_data,
4227 unsigned opcode, unsigned target,
4228 LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
4229 LLVMValueRef *param, unsigned count,
4230 unsigned dmask)
4231 {
4232 struct gallivm_state *gallivm = &ctx->gallivm;
4233 unsigned num_args;
4234 unsigned is_rect = target == TGSI_TEXTURE_RECT;
4235
4236 /* Pad to power of two vector */
4237 while (count < util_next_power_of_two(count))
4238 param[count++] = LLVMGetUndef(ctx->i32);
4239
4240 /* Texture coordinates. */
4241 if (count > 1)
4242 emit_data->args[0] = lp_build_gather_values(gallivm, param, count);
4243 else
4244 emit_data->args[0] = param[0];
4245
4246 /* Resource. */
4247 emit_data->args[1] = res_ptr;
4248 num_args = 2;
4249
4250 if (opcode == TGSI_OPCODE_TXF || opcode == TGSI_OPCODE_TXQ)
4251 emit_data->dst_type = ctx->v4i32;
4252 else {
4253 emit_data->dst_type = ctx->v4f32;
4254
4255 emit_data->args[num_args++] = samp_ptr;
4256 }
4257
4258 emit_data->args[num_args++] = lp_build_const_int32(gallivm, dmask);
4259 emit_data->args[num_args++] = lp_build_const_int32(gallivm, is_rect); /* unorm */
4260 emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* r128 */
4261 emit_data->args[num_args++] = lp_build_const_int32(gallivm,
4262 tgsi_is_array_sampler(target)); /* da */
4263 emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* glc */
4264 emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* slc */
4265 emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* tfe */
4266 emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* lwe */
4267
4268 emit_data->arg_count = num_args;
4269 }
4270
4271 static const struct lp_build_tgsi_action tex_action;
4272
4273 enum desc_type {
4274 DESC_IMAGE,
4275 DESC_BUFFER,
4276 DESC_FMASK,
4277 DESC_SAMPLER,
4278 };
4279
4280 /**
4281 * Load an image view, fmask view. or sampler state descriptor.
4282 */
4283 static LLVMValueRef load_sampler_desc_custom(struct si_shader_context *ctx,
4284 LLVMValueRef list, LLVMValueRef index,
4285 enum desc_type type)
4286 {
4287 struct gallivm_state *gallivm = &ctx->gallivm;
4288 LLVMBuilderRef builder = gallivm->builder;
4289
4290 switch (type) {
4291 case DESC_IMAGE:
4292 /* The image is at [0:7]. */
4293 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
4294 break;
4295 case DESC_BUFFER:
4296 /* The buffer is in [4:7]. */
4297 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4298 index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
4299 list = LLVMBuildPointerCast(builder, list,
4300 const_array(ctx->v4i32, 0), "");
4301 break;
4302 case DESC_FMASK:
4303 /* The FMASK is at [8:15]. */
4304 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
4305 index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
4306 break;
4307 case DESC_SAMPLER:
4308 /* The sampler state is at [12:15]. */
4309 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4310 index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 3, 0), "");
4311 list = LLVMBuildPointerCast(builder, list,
4312 const_array(ctx->v4i32, 0), "");
4313 break;
4314 }
4315
4316 return build_indexed_load_const(ctx, list, index);
4317 }
4318
4319 static LLVMValueRef load_sampler_desc(struct si_shader_context *ctx,
4320 LLVMValueRef index, enum desc_type type)
4321 {
4322 LLVMValueRef list = LLVMGetParam(ctx->main_fn,
4323 SI_PARAM_SAMPLERS);
4324
4325 return load_sampler_desc_custom(ctx, list, index, type);
4326 }
4327
4328 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4329 *
4330 * SI-CI:
4331 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4332 * filtering manually. The driver sets img7 to a mask clearing
4333 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4334 * s_and_b32 samp0, samp0, img7
4335 *
4336 * VI:
4337 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4338 */
4339 static LLVMValueRef sici_fix_sampler_aniso(struct si_shader_context *ctx,
4340 LLVMValueRef res, LLVMValueRef samp)
4341 {
4342 LLVMBuilderRef builder = ctx->gallivm.builder;
4343 LLVMValueRef img7, samp0;
4344
4345 if (ctx->screen->b.chip_class >= VI)
4346 return samp;
4347
4348 img7 = LLVMBuildExtractElement(builder, res,
4349 LLVMConstInt(ctx->i32, 7, 0), "");
4350 samp0 = LLVMBuildExtractElement(builder, samp,
4351 LLVMConstInt(ctx->i32, 0, 0), "");
4352 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
4353 return LLVMBuildInsertElement(builder, samp, samp0,
4354 LLVMConstInt(ctx->i32, 0, 0), "");
4355 }
4356
4357 static void tex_fetch_ptrs(
4358 struct lp_build_tgsi_context *bld_base,
4359 struct lp_build_emit_data *emit_data,
4360 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr, LLVMValueRef *fmask_ptr)
4361 {
4362 struct si_shader_context *ctx = si_shader_context(bld_base);
4363 const struct tgsi_full_instruction *inst = emit_data->inst;
4364 unsigned target = inst->Texture.Texture;
4365 unsigned sampler_src;
4366 unsigned sampler_index;
4367 LLVMValueRef index;
4368
4369 sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
4370 sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
4371
4372 if (emit_data->inst->Src[sampler_src].Register.Indirect) {
4373 const struct tgsi_full_src_register *reg = &emit_data->inst->Src[sampler_src];
4374
4375 index = get_bounded_indirect_index(ctx,
4376 &reg->Indirect,
4377 reg->Register.Index,
4378 SI_NUM_SAMPLERS);
4379 } else {
4380 index = LLVMConstInt(ctx->i32, sampler_index, 0);
4381 }
4382
4383 if (target == TGSI_TEXTURE_BUFFER)
4384 *res_ptr = load_sampler_desc(ctx, index, DESC_BUFFER);
4385 else
4386 *res_ptr = load_sampler_desc(ctx, index, DESC_IMAGE);
4387
4388 if (samp_ptr)
4389 *samp_ptr = NULL;
4390 if (fmask_ptr)
4391 *fmask_ptr = NULL;
4392
4393 if (target == TGSI_TEXTURE_2D_MSAA ||
4394 target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
4395 if (fmask_ptr)
4396 *fmask_ptr = load_sampler_desc(ctx, index, DESC_FMASK);
4397 } else if (target != TGSI_TEXTURE_BUFFER) {
4398 if (samp_ptr) {
4399 *samp_ptr = load_sampler_desc(ctx, index, DESC_SAMPLER);
4400 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
4401 }
4402 }
4403 }
4404
4405 static void txq_fetch_args(
4406 struct lp_build_tgsi_context *bld_base,
4407 struct lp_build_emit_data *emit_data)
4408 {
4409 struct si_shader_context *ctx = si_shader_context(bld_base);
4410 const struct tgsi_full_instruction *inst = emit_data->inst;
4411 unsigned target = inst->Texture.Texture;
4412 LLVMValueRef res_ptr;
4413 LLVMValueRef address;
4414
4415 tex_fetch_ptrs(bld_base, emit_data, &res_ptr, NULL, NULL);
4416
4417 if (target == TGSI_TEXTURE_BUFFER) {
4418 /* Read the size from the buffer descriptor directly. */
4419 emit_data->args[0] = get_buffer_size(bld_base, res_ptr);
4420 return;
4421 }
4422
4423 /* Textures - set the mip level. */
4424 address = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
4425
4426 set_tex_fetch_args(ctx, emit_data, TGSI_OPCODE_TXQ, target, res_ptr,
4427 NULL, &address, 1, 0xf);
4428 }
4429
4430 static void txq_emit(const struct lp_build_tgsi_action *action,
4431 struct lp_build_tgsi_context *bld_base,
4432 struct lp_build_emit_data *emit_data)
4433 {
4434 struct lp_build_context *base = &bld_base->base;
4435 unsigned target = emit_data->inst->Texture.Texture;
4436
4437 if (target == TGSI_TEXTURE_BUFFER) {
4438 /* Just return the buffer size. */
4439 emit_data->output[emit_data->chan] = emit_data->args[0];
4440 return;
4441 }
4442
4443 emit_data->output[emit_data->chan] = lp_build_intrinsic(
4444 base->gallivm->builder, "llvm.SI.getresinfo.i32",
4445 emit_data->dst_type, emit_data->args, emit_data->arg_count,
4446 LP_FUNC_ATTR_READNONE);
4447
4448 /* Divide the number of layers by 6 to get the number of cubes. */
4449 if (target == TGSI_TEXTURE_CUBE_ARRAY ||
4450 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
4451 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
4452 LLVMValueRef two = lp_build_const_int32(bld_base->base.gallivm, 2);
4453 LLVMValueRef six = lp_build_const_int32(bld_base->base.gallivm, 6);
4454
4455 LLVMValueRef v4 = emit_data->output[emit_data->chan];
4456 LLVMValueRef z = LLVMBuildExtractElement(builder, v4, two, "");
4457 z = LLVMBuildSDiv(builder, z, six, "");
4458
4459 emit_data->output[emit_data->chan] =
4460 LLVMBuildInsertElement(builder, v4, z, two, "");
4461 }
4462 }
4463
4464 static void tex_fetch_args(
4465 struct lp_build_tgsi_context *bld_base,
4466 struct lp_build_emit_data *emit_data)
4467 {
4468 struct si_shader_context *ctx = si_shader_context(bld_base);
4469 struct gallivm_state *gallivm = bld_base->base.gallivm;
4470 const struct tgsi_full_instruction *inst = emit_data->inst;
4471 unsigned opcode = inst->Instruction.Opcode;
4472 unsigned target = inst->Texture.Texture;
4473 LLVMValueRef coords[5], derivs[6];
4474 LLVMValueRef address[16];
4475 unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
4476 int ref_pos = tgsi_util_get_shadow_ref_src_index(target);
4477 unsigned count = 0;
4478 unsigned chan;
4479 unsigned num_deriv_channels = 0;
4480 bool has_offset = inst->Texture.NumOffsets > 0;
4481 LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
4482 unsigned dmask = 0xf;
4483
4484 tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
4485
4486 if (target == TGSI_TEXTURE_BUFFER) {
4487 emit_data->dst_type = ctx->v4f32;
4488 emit_data->args[0] = LLVMBuildBitCast(gallivm->builder, res_ptr,
4489 ctx->v16i8, "");
4490 emit_data->args[1] = bld_base->uint_bld.zero;
4491 emit_data->args[2] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
4492 emit_data->arg_count = 3;
4493 return;
4494 }
4495
4496 /* Fetch and project texture coordinates */
4497 coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
4498 for (chan = 0; chan < 3; chan++ ) {
4499 coords[chan] = lp_build_emit_fetch(bld_base,
4500 emit_data->inst, 0,
4501 chan);
4502 if (opcode == TGSI_OPCODE_TXP)
4503 coords[chan] = lp_build_emit_llvm_binary(bld_base,
4504 TGSI_OPCODE_DIV,
4505 coords[chan],
4506 coords[3]);
4507 }
4508
4509 if (opcode == TGSI_OPCODE_TXP)
4510 coords[3] = bld_base->base.one;
4511
4512 /* Pack offsets. */
4513 if (has_offset && opcode != TGSI_OPCODE_TXF) {
4514 /* The offsets are six-bit signed integers packed like this:
4515 * X=[5:0], Y=[13:8], and Z=[21:16].
4516 */
4517 LLVMValueRef offset[3], pack;
4518
4519 assert(inst->Texture.NumOffsets == 1);
4520
4521 for (chan = 0; chan < 3; chan++) {
4522 offset[chan] = lp_build_emit_fetch_texoffset(bld_base,
4523 emit_data->inst, 0, chan);
4524 offset[chan] = LLVMBuildAnd(gallivm->builder, offset[chan],
4525 lp_build_const_int32(gallivm, 0x3f), "");
4526 if (chan)
4527 offset[chan] = LLVMBuildShl(gallivm->builder, offset[chan],
4528 lp_build_const_int32(gallivm, chan*8), "");
4529 }
4530
4531 pack = LLVMBuildOr(gallivm->builder, offset[0], offset[1], "");
4532 pack = LLVMBuildOr(gallivm->builder, pack, offset[2], "");
4533 address[count++] = pack;
4534 }
4535
4536 /* Pack LOD bias value */
4537 if (opcode == TGSI_OPCODE_TXB)
4538 address[count++] = coords[3];
4539 if (opcode == TGSI_OPCODE_TXB2)
4540 address[count++] = lp_build_emit_fetch(bld_base, inst, 1, TGSI_CHAN_X);
4541
4542 /* Pack depth comparison value */
4543 if (tgsi_is_shadow_target(target) && opcode != TGSI_OPCODE_LODQ) {
4544 LLVMValueRef z;
4545
4546 if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
4547 z = lp_build_emit_fetch(bld_base, inst, 1, TGSI_CHAN_X);
4548 } else {
4549 assert(ref_pos >= 0);
4550 z = coords[ref_pos];
4551 }
4552
4553 /* TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4554 * so the depth comparison value isn't clamped for Z16 and
4555 * Z24 anymore. Do it manually here.
4556 *
4557 * It's unnecessary if the original texture format was
4558 * Z32_FLOAT, but we don't know that here.
4559 */
4560 if (ctx->screen->b.chip_class == VI)
4561 z = si_llvm_saturate(bld_base, z);
4562
4563 address[count++] = z;
4564 }
4565
4566 /* Pack user derivatives */
4567 if (opcode == TGSI_OPCODE_TXD) {
4568 int param, num_src_deriv_channels;
4569
4570 switch (target) {
4571 case TGSI_TEXTURE_3D:
4572 num_src_deriv_channels = 3;
4573 num_deriv_channels = 3;
4574 break;
4575 case TGSI_TEXTURE_2D:
4576 case TGSI_TEXTURE_SHADOW2D:
4577 case TGSI_TEXTURE_RECT:
4578 case TGSI_TEXTURE_SHADOWRECT:
4579 case TGSI_TEXTURE_2D_ARRAY:
4580 case TGSI_TEXTURE_SHADOW2D_ARRAY:
4581 num_src_deriv_channels = 2;
4582 num_deriv_channels = 2;
4583 break;
4584 case TGSI_TEXTURE_CUBE:
4585 case TGSI_TEXTURE_SHADOWCUBE:
4586 case TGSI_TEXTURE_CUBE_ARRAY:
4587 case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
4588 /* Cube derivatives will be converted to 2D. */
4589 num_src_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 num_deriv_channels = 1;
4598 break;
4599 default:
4600 unreachable("invalid target");
4601 }
4602
4603 for (param = 0; param < 2; param++)
4604 for (chan = 0; chan < num_src_deriv_channels; chan++)
4605 derivs[param * num_src_deriv_channels + chan] =
4606 lp_build_emit_fetch(bld_base, inst, param+1, chan);
4607 }
4608
4609 if (target == TGSI_TEXTURE_CUBE ||
4610 target == TGSI_TEXTURE_CUBE_ARRAY ||
4611 target == TGSI_TEXTURE_SHADOWCUBE ||
4612 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
4613 si_prepare_cube_coords(bld_base, emit_data, coords, derivs);
4614
4615 if (opcode == TGSI_OPCODE_TXD)
4616 for (int i = 0; i < num_deriv_channels * 2; i++)
4617 address[count++] = derivs[i];
4618
4619 /* Pack texture coordinates */
4620 address[count++] = coords[0];
4621 if (num_coords > 1)
4622 address[count++] = coords[1];
4623 if (num_coords > 2)
4624 address[count++] = coords[2];
4625
4626 /* Pack LOD or sample index */
4627 if (opcode == TGSI_OPCODE_TXL || opcode == TGSI_OPCODE_TXF)
4628 address[count++] = coords[3];
4629 else if (opcode == TGSI_OPCODE_TXL2)
4630 address[count++] = lp_build_emit_fetch(bld_base, inst, 1, TGSI_CHAN_X);
4631
4632 if (count > 16) {
4633 assert(!"Cannot handle more than 16 texture address parameters");
4634 count = 16;
4635 }
4636
4637 for (chan = 0; chan < count; chan++ ) {
4638 address[chan] = LLVMBuildBitCast(gallivm->builder,
4639 address[chan], ctx->i32, "");
4640 }
4641
4642 /* Adjust the sample index according to FMASK.
4643 *
4644 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
4645 * which is the identity mapping. Each nibble says which physical sample
4646 * should be fetched to get that sample.
4647 *
4648 * For example, 0x11111100 means there are only 2 samples stored and
4649 * the second sample covers 3/4 of the pixel. When reading samples 0
4650 * and 1, return physical sample 0 (determined by the first two 0s
4651 * in FMASK), otherwise return physical sample 1.
4652 *
4653 * The sample index should be adjusted as follows:
4654 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
4655 */
4656 if (target == TGSI_TEXTURE_2D_MSAA ||
4657 target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
4658 struct lp_build_context *uint_bld = &bld_base->uint_bld;
4659 struct lp_build_emit_data txf_emit_data = *emit_data;
4660 LLVMValueRef txf_address[4];
4661 unsigned txf_count = count;
4662 struct tgsi_full_instruction inst = {};
4663
4664 memcpy(txf_address, address, sizeof(txf_address));
4665
4666 if (target == TGSI_TEXTURE_2D_MSAA) {
4667 txf_address[2] = bld_base->uint_bld.zero;
4668 }
4669 txf_address[3] = bld_base->uint_bld.zero;
4670
4671 /* Read FMASK using TXF. */
4672 inst.Instruction.Opcode = TGSI_OPCODE_TXF;
4673 inst.Texture.Texture = target;
4674 txf_emit_data.inst = &inst;
4675 txf_emit_data.chan = 0;
4676 set_tex_fetch_args(ctx, &txf_emit_data, TGSI_OPCODE_TXF,
4677 target, fmask_ptr, NULL,
4678 txf_address, txf_count, 0xf);
4679 build_tex_intrinsic(&tex_action, bld_base, &txf_emit_data);
4680
4681 /* Initialize some constants. */
4682 LLVMValueRef four = LLVMConstInt(ctx->i32, 4, 0);
4683 LLVMValueRef F = LLVMConstInt(ctx->i32, 0xF, 0);
4684
4685 /* Apply the formula. */
4686 LLVMValueRef fmask =
4687 LLVMBuildExtractElement(gallivm->builder,
4688 txf_emit_data.output[0],
4689 uint_bld->zero, "");
4690
4691 unsigned sample_chan = target == TGSI_TEXTURE_2D_MSAA ? 2 : 3;
4692
4693 LLVMValueRef sample_index4 =
4694 LLVMBuildMul(gallivm->builder, address[sample_chan], four, "");
4695
4696 LLVMValueRef shifted_fmask =
4697 LLVMBuildLShr(gallivm->builder, fmask, sample_index4, "");
4698
4699 LLVMValueRef final_sample =
4700 LLVMBuildAnd(gallivm->builder, shifted_fmask, F, "");
4701
4702 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
4703 * resource descriptor is 0 (invalid),
4704 */
4705 LLVMValueRef fmask_desc =
4706 LLVMBuildBitCast(gallivm->builder, fmask_ptr,
4707 ctx->v8i32, "");
4708
4709 LLVMValueRef fmask_word1 =
4710 LLVMBuildExtractElement(gallivm->builder, fmask_desc,
4711 uint_bld->one, "");
4712
4713 LLVMValueRef word1_is_nonzero =
4714 LLVMBuildICmp(gallivm->builder, LLVMIntNE,
4715 fmask_word1, uint_bld->zero, "");
4716
4717 /* Replace the MSAA sample index. */
4718 address[sample_chan] =
4719 LLVMBuildSelect(gallivm->builder, word1_is_nonzero,
4720 final_sample, address[sample_chan], "");
4721 }
4722
4723 if (opcode == TGSI_OPCODE_TXF) {
4724 /* add tex offsets */
4725 if (inst->Texture.NumOffsets) {
4726 struct lp_build_context *uint_bld = &bld_base->uint_bld;
4727 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
4728 const struct tgsi_texture_offset *off = inst->TexOffsets;
4729
4730 assert(inst->Texture.NumOffsets == 1);
4731
4732 switch (target) {
4733 case TGSI_TEXTURE_3D:
4734 address[2] = lp_build_add(uint_bld, address[2],
4735 bld->immediates[off->Index][off->SwizzleZ]);
4736 /* fall through */
4737 case TGSI_TEXTURE_2D:
4738 case TGSI_TEXTURE_SHADOW2D:
4739 case TGSI_TEXTURE_RECT:
4740 case TGSI_TEXTURE_SHADOWRECT:
4741 case TGSI_TEXTURE_2D_ARRAY:
4742 case TGSI_TEXTURE_SHADOW2D_ARRAY:
4743 address[1] =
4744 lp_build_add(uint_bld, address[1],
4745 bld->immediates[off->Index][off->SwizzleY]);
4746 /* fall through */
4747 case TGSI_TEXTURE_1D:
4748 case TGSI_TEXTURE_SHADOW1D:
4749 case TGSI_TEXTURE_1D_ARRAY:
4750 case TGSI_TEXTURE_SHADOW1D_ARRAY:
4751 address[0] =
4752 lp_build_add(uint_bld, address[0],
4753 bld->immediates[off->Index][off->SwizzleX]);
4754 break;
4755 /* texture offsets do not apply to other texture targets */
4756 }
4757 }
4758 }
4759
4760 if (opcode == TGSI_OPCODE_TG4) {
4761 unsigned gather_comp = 0;
4762
4763 /* DMASK was repurposed for GATHER4. 4 components are always
4764 * returned and DMASK works like a swizzle - it selects
4765 * the component to fetch. The only valid DMASK values are
4766 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4767 * (red,red,red,red) etc.) The ISA document doesn't mention
4768 * this.
4769 */
4770
4771 /* Get the component index from src1.x for Gather4. */
4772 if (!tgsi_is_shadow_target(target)) {
4773 LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
4774 LLVMValueRef comp_imm;
4775 struct tgsi_src_register src1 = inst->Src[1].Register;
4776
4777 assert(src1.File == TGSI_FILE_IMMEDIATE);
4778
4779 comp_imm = imms[src1.Index][src1.SwizzleX];
4780 gather_comp = LLVMConstIntGetZExtValue(comp_imm);
4781 gather_comp = CLAMP(gather_comp, 0, 3);
4782 }
4783
4784 dmask = 1 << gather_comp;
4785 }
4786
4787 set_tex_fetch_args(ctx, emit_data, opcode, target, res_ptr,
4788 samp_ptr, address, count, dmask);
4789 }
4790
4791 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
4792 * incorrectly forces nearest filtering if the texture format is integer.
4793 * The only effect it has on Gather4, which always returns 4 texels for
4794 * bilinear filtering, is that the final coordinates are off by 0.5 of
4795 * the texel size.
4796 *
4797 * The workaround is to subtract 0.5 from the unnormalized coordinates,
4798 * or (0.5 / size) from the normalized coordinates.
4799 */
4800 static void si_lower_gather4_integer(struct si_shader_context *ctx,
4801 struct lp_build_emit_data *emit_data,
4802 const char *intr_name,
4803 unsigned coord_vgpr_index)
4804 {
4805 LLVMBuilderRef builder = ctx->gallivm.builder;
4806 LLVMValueRef coord = emit_data->args[0];
4807 LLVMValueRef half_texel[2];
4808 int c;
4809
4810 if (emit_data->inst->Texture.Texture == TGSI_TEXTURE_RECT ||
4811 emit_data->inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT) {
4812 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
4813 } else {
4814 struct tgsi_full_instruction txq_inst = {};
4815 struct lp_build_emit_data txq_emit_data = {};
4816
4817 /* Query the texture size. */
4818 txq_inst.Texture.Texture = emit_data->inst->Texture.Texture;
4819 txq_emit_data.inst = &txq_inst;
4820 txq_emit_data.dst_type = ctx->v4i32;
4821 set_tex_fetch_args(ctx, &txq_emit_data, TGSI_OPCODE_TXQ,
4822 txq_inst.Texture.Texture,
4823 emit_data->args[1], NULL,
4824 &ctx->soa.bld_base.uint_bld.zero,
4825 1, 0xf);
4826 txq_emit(NULL, &ctx->soa.bld_base, &txq_emit_data);
4827
4828 /* Compute -0.5 / size. */
4829 for (c = 0; c < 2; c++) {
4830 half_texel[c] =
4831 LLVMBuildExtractElement(builder, txq_emit_data.output[0],
4832 LLVMConstInt(ctx->i32, c, 0), "");
4833 half_texel[c] = LLVMBuildUIToFP(builder, half_texel[c], ctx->f32, "");
4834 half_texel[c] =
4835 lp_build_emit_llvm_unary(&ctx->soa.bld_base,
4836 TGSI_OPCODE_RCP, half_texel[c]);
4837 half_texel[c] = LLVMBuildFMul(builder, half_texel[c],
4838 LLVMConstReal(ctx->f32, -0.5), "");
4839 }
4840 }
4841
4842 for (c = 0; c < 2; c++) {
4843 LLVMValueRef tmp;
4844 LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index + c, 0);
4845
4846 tmp = LLVMBuildExtractElement(builder, coord, index, "");
4847 tmp = LLVMBuildBitCast(builder, tmp, ctx->f32, "");
4848 tmp = LLVMBuildFAdd(builder, tmp, half_texel[c], "");
4849 tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
4850 coord = LLVMBuildInsertElement(builder, coord, tmp, index, "");
4851 }
4852
4853 emit_data->args[0] = coord;
4854 emit_data->output[emit_data->chan] =
4855 lp_build_intrinsic(builder, intr_name, emit_data->dst_type,
4856 emit_data->args, emit_data->arg_count,
4857 LP_FUNC_ATTR_READNONE);
4858 }
4859
4860 static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
4861 struct lp_build_tgsi_context *bld_base,
4862 struct lp_build_emit_data *emit_data)
4863 {
4864 struct si_shader_context *ctx = si_shader_context(bld_base);
4865 struct lp_build_context *base = &bld_base->base;
4866 const struct tgsi_full_instruction *inst = emit_data->inst;
4867 unsigned opcode = inst->Instruction.Opcode;
4868 unsigned target = inst->Texture.Texture;
4869 char intr_name[127];
4870 bool has_offset = inst->Texture.NumOffsets > 0;
4871 bool is_shadow = tgsi_is_shadow_target(target);
4872 char type[64];
4873 const char *name = "llvm.SI.image.sample";
4874 const char *infix = "";
4875
4876 if (target == TGSI_TEXTURE_BUFFER) {
4877 emit_data->output[emit_data->chan] = lp_build_intrinsic(
4878 base->gallivm->builder,
4879 "llvm.SI.vs.load.input", emit_data->dst_type,
4880 emit_data->args, emit_data->arg_count,
4881 LP_FUNC_ATTR_READNONE);
4882 return;
4883 }
4884
4885 switch (opcode) {
4886 case TGSI_OPCODE_TXF:
4887 name = target == TGSI_TEXTURE_2D_MSAA ||
4888 target == TGSI_TEXTURE_2D_ARRAY_MSAA ?
4889 "llvm.SI.image.load" :
4890 "llvm.SI.image.load.mip";
4891 is_shadow = false;
4892 has_offset = false;
4893 break;
4894 case TGSI_OPCODE_LODQ:
4895 name = "llvm.SI.getlod";
4896 is_shadow = false;
4897 has_offset = false;
4898 break;
4899 case TGSI_OPCODE_TEX:
4900 case TGSI_OPCODE_TEX2:
4901 case TGSI_OPCODE_TXP:
4902 if (ctx->type != PIPE_SHADER_FRAGMENT)
4903 infix = ".lz";
4904 break;
4905 case TGSI_OPCODE_TXB:
4906 case TGSI_OPCODE_TXB2:
4907 assert(ctx->type == PIPE_SHADER_FRAGMENT);
4908 infix = ".b";
4909 break;
4910 case TGSI_OPCODE_TXL:
4911 case TGSI_OPCODE_TXL2:
4912 infix = ".l";
4913 break;
4914 case TGSI_OPCODE_TXD:
4915 infix = ".d";
4916 break;
4917 case TGSI_OPCODE_TG4:
4918 name = "llvm.SI.gather4";
4919 infix = ".lz";
4920 break;
4921 default:
4922 assert(0);
4923 return;
4924 }
4925
4926 /* Add the type and suffixes .c, .o if needed. */
4927 build_type_name_for_intr(LLVMTypeOf(emit_data->args[0]), type, sizeof(type));
4928 sprintf(intr_name, "%s%s%s%s.%s",
4929 name, is_shadow ? ".c" : "", infix,
4930 has_offset ? ".o" : "", type);
4931
4932 /* The hardware needs special lowering for Gather4 with integer formats. */
4933 if (opcode == TGSI_OPCODE_TG4) {
4934 struct tgsi_shader_info *info = &ctx->shader->selector->info;
4935 /* This will also work with non-constant indexing because of how
4936 * glsl_to_tgsi works and we intent to preserve that behavior.
4937 */
4938 const unsigned src_idx = 2;
4939 unsigned sampler = inst->Src[src_idx].Register.Index;
4940
4941 assert(inst->Src[src_idx].Register.File == TGSI_FILE_SAMPLER);
4942
4943 if (info->sampler_type[sampler] == TGSI_RETURN_TYPE_SINT ||
4944 info->sampler_type[sampler] == TGSI_RETURN_TYPE_UINT) {
4945 /* Texture coordinates start after:
4946 * {offset, bias, z-compare, derivatives}
4947 * Only the offset and z-compare can occur here.
4948 */
4949 si_lower_gather4_integer(ctx, emit_data, intr_name,
4950 (int)has_offset + (int)is_shadow);
4951 return;
4952 }
4953 }
4954
4955 emit_data->output[emit_data->chan] = lp_build_intrinsic(
4956 base->gallivm->builder, intr_name, emit_data->dst_type,
4957 emit_data->args, emit_data->arg_count,
4958 LP_FUNC_ATTR_READNONE);
4959 }
4960
4961 static void si_llvm_emit_txqs(
4962 const struct lp_build_tgsi_action *action,
4963 struct lp_build_tgsi_context *bld_base,
4964 struct lp_build_emit_data *emit_data)
4965 {
4966 struct si_shader_context *ctx = si_shader_context(bld_base);
4967 struct gallivm_state *gallivm = bld_base->base.gallivm;
4968 LLVMBuilderRef builder = gallivm->builder;
4969 LLVMValueRef res, samples;
4970 LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
4971
4972 tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
4973
4974
4975 /* Read the samples from the descriptor directly. */
4976 res = LLVMBuildBitCast(builder, res_ptr, ctx->v8i32, "");
4977 samples = LLVMBuildExtractElement(
4978 builder, res,
4979 lp_build_const_int32(gallivm, 3), "");
4980 samples = LLVMBuildLShr(builder, samples,
4981 lp_build_const_int32(gallivm, 16), "");
4982 samples = LLVMBuildAnd(builder, samples,
4983 lp_build_const_int32(gallivm, 0xf), "");
4984 samples = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1),
4985 samples, "");
4986
4987 emit_data->output[emit_data->chan] = samples;
4988 }
4989
4990 /*
4991 * SI implements derivatives using the local data store (LDS)
4992 * All writes to the LDS happen in all executing threads at
4993 * the same time. TID is the Thread ID for the current
4994 * thread and is a value between 0 and 63, representing
4995 * the thread's position in the wavefront.
4996 *
4997 * For the pixel shader threads are grouped into quads of four pixels.
4998 * The TIDs of the pixels of a quad are:
4999 *
5000 * +------+------+
5001 * |4n + 0|4n + 1|
5002 * +------+------+
5003 * |4n + 2|4n + 3|
5004 * +------+------+
5005 *
5006 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
5007 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
5008 * the current pixel's column, and masking with 0xfffffffe yields the TID
5009 * of the left pixel of the current pixel's row.
5010 *
5011 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
5012 * adding 2 yields the TID of the pixel below the top pixel.
5013 */
5014 /* masks for thread ID. */
5015 #define TID_MASK_TOP_LEFT 0xfffffffc
5016 #define TID_MASK_TOP 0xfffffffd
5017 #define TID_MASK_LEFT 0xfffffffe
5018
5019 static void si_llvm_emit_ddxy(
5020 const struct lp_build_tgsi_action *action,
5021 struct lp_build_tgsi_context *bld_base,
5022 struct lp_build_emit_data *emit_data)
5023 {
5024 struct si_shader_context *ctx = si_shader_context(bld_base);
5025 struct gallivm_state *gallivm = bld_base->base.gallivm;
5026 unsigned opcode = emit_data->info->opcode;
5027 LLVMValueRef thread_id, tl, trbl, tl_tid, trbl_tid, val, args[2];
5028 int idx;
5029 unsigned mask;
5030
5031 thread_id = get_thread_id(ctx);
5032
5033 if (opcode == TGSI_OPCODE_DDX_FINE)
5034 mask = TID_MASK_LEFT;
5035 else if (opcode == TGSI_OPCODE_DDY_FINE)
5036 mask = TID_MASK_TOP;
5037 else
5038 mask = TID_MASK_TOP_LEFT;
5039
5040 tl_tid = LLVMBuildAnd(gallivm->builder, thread_id,
5041 lp_build_const_int32(gallivm, mask), "");
5042
5043 /* for DDX we want to next X pixel, DDY next Y pixel. */
5044 idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;
5045 trbl_tid = LLVMBuildAdd(gallivm->builder, tl_tid,
5046 lp_build_const_int32(gallivm, idx), "");
5047
5048 val = LLVMBuildBitCast(gallivm->builder, emit_data->args[0], ctx->i32, "");
5049
5050 if (ctx->screen->has_ds_bpermute) {
5051 args[0] = LLVMBuildMul(gallivm->builder, tl_tid,
5052 lp_build_const_int32(gallivm, 4), "");
5053 args[1] = val;
5054 tl = lp_build_intrinsic(gallivm->builder,
5055 "llvm.amdgcn.ds.bpermute", ctx->i32,
5056 args, 2, LP_FUNC_ATTR_READNONE);
5057
5058 args[0] = LLVMBuildMul(gallivm->builder, trbl_tid,
5059 lp_build_const_int32(gallivm, 4), "");
5060 trbl = lp_build_intrinsic(gallivm->builder,
5061 "llvm.amdgcn.ds.bpermute", ctx->i32,
5062 args, 2, LP_FUNC_ATTR_READNONE);
5063 } else {
5064 LLVMValueRef store_ptr, load_ptr0, load_ptr1;
5065
5066 store_ptr = build_gep0(ctx, ctx->lds, thread_id);
5067 load_ptr0 = build_gep0(ctx, ctx->lds, tl_tid);
5068 load_ptr1 = build_gep0(ctx, ctx->lds, trbl_tid);
5069
5070 LLVMBuildStore(gallivm->builder, val, store_ptr);
5071 tl = LLVMBuildLoad(gallivm->builder, load_ptr0, "");
5072 trbl = LLVMBuildLoad(gallivm->builder, load_ptr1, "");
5073 }
5074
5075 tl = LLVMBuildBitCast(gallivm->builder, tl, ctx->f32, "");
5076 trbl = LLVMBuildBitCast(gallivm->builder, trbl, ctx->f32, "");
5077
5078 emit_data->output[emit_data->chan] =
5079 LLVMBuildFSub(gallivm->builder, trbl, tl, "");
5080 }
5081
5082 /*
5083 * this takes an I,J coordinate pair,
5084 * and works out the X and Y derivatives.
5085 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
5086 */
5087 static LLVMValueRef si_llvm_emit_ddxy_interp(
5088 struct lp_build_tgsi_context *bld_base,
5089 LLVMValueRef interp_ij)
5090 {
5091 struct si_shader_context *ctx = si_shader_context(bld_base);
5092 struct gallivm_state *gallivm = bld_base->base.gallivm;
5093 LLVMValueRef result[4], a;
5094 unsigned i;
5095
5096 for (i = 0; i < 2; i++) {
5097 a = LLVMBuildExtractElement(gallivm->builder, interp_ij,
5098 LLVMConstInt(ctx->i32, i, 0), "");
5099 result[i] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_DDX, a);
5100 result[2+i] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_DDY, a);
5101 }
5102
5103 return lp_build_gather_values(gallivm, result, 4);
5104 }
5105
5106 static void interp_fetch_args(
5107 struct lp_build_tgsi_context *bld_base,
5108 struct lp_build_emit_data *emit_data)
5109 {
5110 struct si_shader_context *ctx = si_shader_context(bld_base);
5111 struct gallivm_state *gallivm = bld_base->base.gallivm;
5112 const struct tgsi_full_instruction *inst = emit_data->inst;
5113
5114 if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) {
5115 /* offset is in second src, first two channels */
5116 emit_data->args[0] = lp_build_emit_fetch(bld_base,
5117 emit_data->inst, 1,
5118 TGSI_CHAN_X);
5119 emit_data->args[1] = lp_build_emit_fetch(bld_base,
5120 emit_data->inst, 1,
5121 TGSI_CHAN_Y);
5122 emit_data->arg_count = 2;
5123 } else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
5124 LLVMValueRef sample_position;
5125 LLVMValueRef sample_id;
5126 LLVMValueRef halfval = lp_build_const_float(gallivm, 0.5f);
5127
5128 /* fetch sample ID, then fetch its sample position,
5129 * and place into first two channels.
5130 */
5131 sample_id = lp_build_emit_fetch(bld_base,
5132 emit_data->inst, 1, TGSI_CHAN_X);
5133 sample_id = LLVMBuildBitCast(gallivm->builder, sample_id,
5134 ctx->i32, "");
5135 sample_position = load_sample_position(ctx, sample_id);
5136
5137 emit_data->args[0] = LLVMBuildExtractElement(gallivm->builder,
5138 sample_position,
5139 lp_build_const_int32(gallivm, 0), "");
5140
5141 emit_data->args[0] = LLVMBuildFSub(gallivm->builder, emit_data->args[0], halfval, "");
5142 emit_data->args[1] = LLVMBuildExtractElement(gallivm->builder,
5143 sample_position,
5144 lp_build_const_int32(gallivm, 1), "");
5145 emit_data->args[1] = LLVMBuildFSub(gallivm->builder, emit_data->args[1], halfval, "");
5146 emit_data->arg_count = 2;
5147 }
5148 }
5149
5150 static void build_interp_intrinsic(const struct lp_build_tgsi_action *action,
5151 struct lp_build_tgsi_context *bld_base,
5152 struct lp_build_emit_data *emit_data)
5153 {
5154 struct si_shader_context *ctx = si_shader_context(bld_base);
5155 struct si_shader *shader = ctx->shader;
5156 struct gallivm_state *gallivm = bld_base->base.gallivm;
5157 struct lp_build_context *uint = &bld_base->uint_bld;
5158 LLVMValueRef interp_param;
5159 const struct tgsi_full_instruction *inst = emit_data->inst;
5160 int input_index = inst->Src[0].Register.Index;
5161 int chan;
5162 int i;
5163 LLVMValueRef attr_number;
5164 LLVMValueRef params = LLVMGetParam(ctx->main_fn, SI_PARAM_PRIM_MASK);
5165 int interp_param_idx;
5166 unsigned interp = shader->selector->info.input_interpolate[input_index];
5167 unsigned location;
5168
5169 assert(inst->Src[0].Register.File == TGSI_FILE_INPUT);
5170
5171 if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
5172 inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE)
5173 location = TGSI_INTERPOLATE_LOC_CENTER;
5174 else
5175 location = TGSI_INTERPOLATE_LOC_CENTROID;
5176
5177 interp_param_idx = lookup_interp_param_index(interp, location);
5178 if (interp_param_idx == -1)
5179 return;
5180 else if (interp_param_idx)
5181 interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
5182 else
5183 interp_param = NULL;
5184
5185 attr_number = lp_build_const_int32(gallivm, input_index);
5186
5187 if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
5188 inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
5189 LLVMValueRef ij_out[2];
5190 LLVMValueRef ddxy_out = si_llvm_emit_ddxy_interp(bld_base, interp_param);
5191
5192 /*
5193 * take the I then J parameters, and the DDX/Y for it, and
5194 * calculate the IJ inputs for the interpolator.
5195 * temp1 = ddx * offset/sample.x + I;
5196 * interp_param.I = ddy * offset/sample.y + temp1;
5197 * temp1 = ddx * offset/sample.x + J;
5198 * interp_param.J = ddy * offset/sample.y + temp1;
5199 */
5200 for (i = 0; i < 2; i++) {
5201 LLVMValueRef ix_ll = lp_build_const_int32(gallivm, i);
5202 LLVMValueRef iy_ll = lp_build_const_int32(gallivm, i + 2);
5203 LLVMValueRef ddx_el = LLVMBuildExtractElement(gallivm->builder,
5204 ddxy_out, ix_ll, "");
5205 LLVMValueRef ddy_el = LLVMBuildExtractElement(gallivm->builder,
5206 ddxy_out, iy_ll, "");
5207 LLVMValueRef interp_el = LLVMBuildExtractElement(gallivm->builder,
5208 interp_param, ix_ll, "");
5209 LLVMValueRef temp1, temp2;
5210
5211 interp_el = LLVMBuildBitCast(gallivm->builder, interp_el,
5212 ctx->f32, "");
5213
5214 temp1 = LLVMBuildFMul(gallivm->builder, ddx_el, emit_data->args[0], "");
5215
5216 temp1 = LLVMBuildFAdd(gallivm->builder, temp1, interp_el, "");
5217
5218 temp2 = LLVMBuildFMul(gallivm->builder, ddy_el, emit_data->args[1], "");
5219
5220 ij_out[i] = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
5221 }
5222 interp_param = lp_build_gather_values(bld_base->base.gallivm, ij_out, 2);
5223 }
5224
5225 for (chan = 0; chan < 4; chan++) {
5226 LLVMValueRef llvm_chan;
5227 unsigned schan;
5228
5229 schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);
5230 llvm_chan = lp_build_const_int32(gallivm, schan);
5231
5232 if (interp_param) {
5233 interp_param = LLVMBuildBitCast(gallivm->builder,
5234 interp_param, LLVMVectorType(ctx->f32, 2), "");
5235 LLVMValueRef i = LLVMBuildExtractElement(
5236 gallivm->builder, interp_param, uint->zero, "");
5237 LLVMValueRef j = LLVMBuildExtractElement(
5238 gallivm->builder, interp_param, uint->one, "");
5239 emit_data->output[chan] = build_fs_interp(bld_base,
5240 llvm_chan, attr_number, params,
5241 i, j);
5242 } else {
5243 emit_data->output[chan] = build_fs_interp_mov(bld_base,
5244 lp_build_const_int32(gallivm, 2), /* P0 */
5245 llvm_chan, attr_number, params);
5246 }
5247 }
5248 }
5249
5250 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base,
5251 struct lp_build_emit_data *emit_data)
5252 {
5253 LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
5254 struct tgsi_src_register src0 = emit_data->inst->Src[0].Register;
5255 unsigned stream;
5256
5257 assert(src0.File == TGSI_FILE_IMMEDIATE);
5258
5259 stream = LLVMConstIntGetZExtValue(imms[src0.Index][src0.SwizzleX]) & 0x3;
5260 return stream;
5261 }
5262
5263 /* Emit one vertex from the geometry shader */
5264 static void si_llvm_emit_vertex(
5265 const struct lp_build_tgsi_action *action,
5266 struct lp_build_tgsi_context *bld_base,
5267 struct lp_build_emit_data *emit_data)
5268 {
5269 struct si_shader_context *ctx = si_shader_context(bld_base);
5270 struct lp_build_context *uint = &bld_base->uint_bld;
5271 struct si_shader *shader = ctx->shader;
5272 struct tgsi_shader_info *info = &shader->selector->info;
5273 struct gallivm_state *gallivm = bld_base->base.gallivm;
5274 struct lp_build_if_state if_state;
5275 LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
5276 SI_PARAM_GS2VS_OFFSET);
5277 LLVMValueRef gs_next_vertex;
5278 LLVMValueRef can_emit, kill;
5279 LLVMValueRef args[2];
5280 unsigned chan, offset;
5281 int i;
5282 unsigned stream;
5283
5284 stream = si_llvm_get_stream(bld_base, emit_data);
5285
5286 /* Write vertex attribute values to GSVS ring */
5287 gs_next_vertex = LLVMBuildLoad(gallivm->builder,
5288 ctx->gs_next_vertex[stream],
5289 "");
5290
5291 /* If this thread has already emitted the declared maximum number of
5292 * vertices, skip the write: excessive vertex emissions are not
5293 * supposed to have any effect.
5294 *
5295 * If the shader has no writes to memory, kill it instead. This skips
5296 * further memory loads and may allow LLVM to skip to the end
5297 * altogether.
5298 */
5299 can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULT, gs_next_vertex,
5300 lp_build_const_int32(gallivm,
5301 shader->selector->gs_max_out_vertices), "");
5302
5303 bool use_kill = !info->writes_memory;
5304 if (use_kill) {
5305 kill = lp_build_select(&bld_base->base, can_emit,
5306 lp_build_const_float(gallivm, 1.0f),
5307 lp_build_const_float(gallivm, -1.0f));
5308
5309 lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
5310 ctx->voidt, &kill, 1, 0);
5311 } else {
5312 lp_build_if(&if_state, gallivm, can_emit);
5313 }
5314
5315 offset = 0;
5316 for (i = 0; i < info->num_outputs; i++) {
5317 LLVMValueRef *out_ptr =
5318 ctx->soa.outputs[i];
5319
5320 for (chan = 0; chan < 4; chan++) {
5321 if (!(info->output_usagemask[i] & (1 << chan)) ||
5322 ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
5323 continue;
5324
5325 LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
5326 LLVMValueRef voffset =
5327 lp_build_const_int32(gallivm, offset *
5328 shader->selector->gs_max_out_vertices);
5329 offset++;
5330
5331 voffset = lp_build_add(uint, voffset, gs_next_vertex);
5332 voffset = lp_build_mul_imm(uint, voffset, 4);
5333
5334 out_val = LLVMBuildBitCast(gallivm->builder, out_val, ctx->i32, "");
5335
5336 build_tbuffer_store(ctx,
5337 ctx->gsvs_ring[stream],
5338 out_val, 1,
5339 voffset, soffset, 0,
5340 V_008F0C_BUF_DATA_FORMAT_32,
5341 V_008F0C_BUF_NUM_FORMAT_UINT,
5342 1, 0, 1, 1, 0);
5343 }
5344 }
5345
5346 gs_next_vertex = lp_build_add(uint, gs_next_vertex,
5347 lp_build_const_int32(gallivm, 1));
5348
5349 LLVMBuildStore(gallivm->builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
5350
5351 /* Signal vertex emission */
5352 args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS | (stream << 8));
5353 args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
5354 lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
5355 ctx->voidt, args, 2, 0);
5356
5357 if (!use_kill)
5358 lp_build_endif(&if_state);
5359 }
5360
5361 /* Cut one primitive from the geometry shader */
5362 static void si_llvm_emit_primitive(
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 gallivm_state *gallivm = bld_base->base.gallivm;
5369 LLVMValueRef args[2];
5370 unsigned stream;
5371
5372 /* Signal primitive cut */
5373 stream = si_llvm_get_stream(bld_base, emit_data);
5374 args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_CUT | SENDMSG_GS | (stream << 8));
5375 args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
5376 lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
5377 ctx->voidt, args, 2, 0);
5378 }
5379
5380 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
5381 struct lp_build_tgsi_context *bld_base,
5382 struct lp_build_emit_data *emit_data)
5383 {
5384 struct si_shader_context *ctx = si_shader_context(bld_base);
5385 struct gallivm_state *gallivm = bld_base->base.gallivm;
5386
5387 /* The real barrier instruction isn’t needed, because an entire patch
5388 * always fits into a single wave.
5389 */
5390 if (ctx->type == PIPE_SHADER_TESS_CTRL) {
5391 emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
5392 return;
5393 }
5394
5395 lp_build_intrinsic(gallivm->builder,
5396 HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
5397 : "llvm.AMDGPU.barrier.local",
5398 ctx->voidt, NULL, 0, 0);
5399 }
5400
5401 static const struct lp_build_tgsi_action tex_action = {
5402 .fetch_args = tex_fetch_args,
5403 .emit = build_tex_intrinsic,
5404 };
5405
5406 static const struct lp_build_tgsi_action interp_action = {
5407 .fetch_args = interp_fetch_args,
5408 .emit = build_interp_intrinsic,
5409 };
5410
5411 static void si_create_function(struct si_shader_context *ctx,
5412 const char *name,
5413 LLVMTypeRef *returns, unsigned num_returns,
5414 LLVMTypeRef *params, unsigned num_params,
5415 int last_sgpr)
5416 {
5417 int i;
5418
5419 si_llvm_create_func(ctx, name, returns, num_returns,
5420 params, num_params);
5421 si_llvm_shader_type(ctx->main_fn, ctx->type);
5422 ctx->return_value = LLVMGetUndef(ctx->return_type);
5423
5424 for (i = 0; i <= last_sgpr; ++i) {
5425 LLVMValueRef P = LLVMGetParam(ctx->main_fn, i);
5426
5427 /* The combination of:
5428 * - ByVal
5429 * - dereferenceable
5430 * - invariant.load
5431 * allows the optimization passes to move loads and reduces
5432 * SGPR spilling significantly.
5433 */
5434 if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
5435 lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_BYVAL);
5436 lp_add_attr_dereferenceable(P, UINT64_MAX);
5437 } else
5438 lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_INREG);
5439 }
5440
5441 if (ctx->screen->b.debug_flags & DBG_UNSAFE_MATH) {
5442 /* These were copied from some LLVM test. */
5443 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5444 "less-precise-fpmad",
5445 "true");
5446 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5447 "no-infs-fp-math",
5448 "true");
5449 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5450 "no-nans-fp-math",
5451 "true");
5452 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
5453 "unsafe-fp-math",
5454 "true");
5455 }
5456 }
5457
5458 static void create_meta_data(struct si_shader_context *ctx)
5459 {
5460 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
5461
5462 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(gallivm->context,
5463 "invariant.load", 14);
5464 ctx->range_md_kind = LLVMGetMDKindIDInContext(gallivm->context,
5465 "range", 5);
5466 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(gallivm->context,
5467 "amdgpu.uniform", 14);
5468
5469 ctx->empty_md = LLVMMDNodeInContext(gallivm->context, NULL, 0);
5470 }
5471
5472 static void declare_streamout_params(struct si_shader_context *ctx,
5473 struct pipe_stream_output_info *so,
5474 LLVMTypeRef *params, LLVMTypeRef i32,
5475 unsigned *num_params)
5476 {
5477 int i;
5478
5479 /* Streamout SGPRs. */
5480 if (so->num_outputs) {
5481 if (ctx->type != PIPE_SHADER_TESS_EVAL)
5482 params[ctx->param_streamout_config = (*num_params)++] = i32;
5483 else
5484 ctx->param_streamout_config = ctx->param_tess_offchip;
5485
5486 params[ctx->param_streamout_write_index = (*num_params)++] = i32;
5487 }
5488 /* A streamout buffer offset is loaded if the stride is non-zero. */
5489 for (i = 0; i < 4; i++) {
5490 if (!so->stride[i])
5491 continue;
5492
5493 params[ctx->param_streamout_offset[i] = (*num_params)++] = i32;
5494 }
5495 }
5496
5497 static unsigned llvm_get_type_size(LLVMTypeRef type)
5498 {
5499 LLVMTypeKind kind = LLVMGetTypeKind(type);
5500
5501 switch (kind) {
5502 case LLVMIntegerTypeKind:
5503 return LLVMGetIntTypeWidth(type) / 8;
5504 case LLVMFloatTypeKind:
5505 return 4;
5506 case LLVMPointerTypeKind:
5507 return 8;
5508 case LLVMVectorTypeKind:
5509 return LLVMGetVectorSize(type) *
5510 llvm_get_type_size(LLVMGetElementType(type));
5511 case LLVMArrayTypeKind:
5512 return LLVMGetArrayLength(type) *
5513 llvm_get_type_size(LLVMGetElementType(type));
5514 default:
5515 assert(0);
5516 return 0;
5517 }
5518 }
5519
5520 static void declare_tess_lds(struct si_shader_context *ctx)
5521 {
5522 struct gallivm_state *gallivm = &ctx->gallivm;
5523 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
5524 struct lp_build_context *uint = &bld_base->uint_bld;
5525
5526 unsigned lds_size = ctx->screen->b.chip_class >= CIK ? 65536 : 32768;
5527 ctx->lds = LLVMBuildIntToPtr(gallivm->builder, uint->zero,
5528 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), LOCAL_ADDR_SPACE),
5529 "tess_lds");
5530 }
5531
5532 static unsigned si_get_max_workgroup_size(struct si_shader *shader)
5533 {
5534 const unsigned *properties = shader->selector->info.properties;
5535 unsigned max_work_group_size =
5536 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
5537 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
5538 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
5539
5540 if (!max_work_group_size) {
5541 /* This is a variable group size compute shader,
5542 * compile it for the maximum possible group size.
5543 */
5544 max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
5545 }
5546 return max_work_group_size;
5547 }
5548
5549 static void create_function(struct si_shader_context *ctx)
5550 {
5551 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
5552 struct gallivm_state *gallivm = bld_base->base.gallivm;
5553 struct si_shader *shader = ctx->shader;
5554 LLVMTypeRef params[SI_NUM_PARAMS + SI_NUM_VERTEX_BUFFERS], v3i32;
5555 LLVMTypeRef returns[16+32*4];
5556 unsigned i, last_sgpr, num_params, num_return_sgprs;
5557 unsigned num_returns = 0;
5558 unsigned num_prolog_vgprs = 0;
5559
5560 v3i32 = LLVMVectorType(ctx->i32, 3);
5561
5562 params[SI_PARAM_RW_BUFFERS] = const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
5563 params[SI_PARAM_CONST_BUFFERS] = const_array(ctx->v16i8, SI_NUM_CONST_BUFFERS);
5564 params[SI_PARAM_SAMPLERS] = const_array(ctx->v8i32, SI_NUM_SAMPLERS);
5565 params[SI_PARAM_IMAGES] = const_array(ctx->v8i32, SI_NUM_IMAGES);
5566 params[SI_PARAM_SHADER_BUFFERS] = const_array(ctx->v4i32, SI_NUM_SHADER_BUFFERS);
5567
5568 switch (ctx->type) {
5569 case PIPE_SHADER_VERTEX:
5570 params[SI_PARAM_VERTEX_BUFFERS] = const_array(ctx->v16i8, SI_NUM_VERTEX_BUFFERS);
5571 params[SI_PARAM_BASE_VERTEX] = ctx->i32;
5572 params[SI_PARAM_START_INSTANCE] = ctx->i32;
5573 params[SI_PARAM_DRAWID] = ctx->i32;
5574 num_params = SI_PARAM_DRAWID+1;
5575
5576 if (shader->key.as_es) {
5577 params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
5578 } else if (shader->key.as_ls) {
5579 params[SI_PARAM_LS_OUT_LAYOUT] = ctx->i32;
5580 num_params = SI_PARAM_LS_OUT_LAYOUT+1;
5581 } else {
5582 if (shader->is_gs_copy_shader) {
5583 num_params = SI_PARAM_RW_BUFFERS+1;
5584 } else {
5585 params[SI_PARAM_VS_STATE_BITS] = ctx->i32;
5586 num_params = SI_PARAM_VS_STATE_BITS+1;
5587 }
5588
5589 /* The locations of the other parameters are assigned dynamically. */
5590 declare_streamout_params(ctx, &shader->selector->so,
5591 params, ctx->i32, &num_params);
5592 }
5593
5594 last_sgpr = num_params-1;
5595
5596 /* VGPRs */
5597 params[ctx->param_vertex_id = num_params++] = ctx->i32;
5598 params[ctx->param_rel_auto_id = num_params++] = ctx->i32;
5599 params[ctx->param_vs_prim_id = num_params++] = ctx->i32;
5600 params[ctx->param_instance_id = num_params++] = ctx->i32;
5601
5602 if (!shader->is_gs_copy_shader) {
5603 /* Vertex load indices. */
5604 ctx->param_vertex_index0 = num_params;
5605
5606 for (i = 0; i < shader->selector->info.num_inputs; i++)
5607 params[num_params++] = ctx->i32;
5608
5609 num_prolog_vgprs += shader->selector->info.num_inputs;
5610
5611 /* PrimitiveID output. */
5612 if (!shader->key.as_es && !shader->key.as_ls)
5613 for (i = 0; i <= VS_EPILOG_PRIMID_LOC; i++)
5614 returns[num_returns++] = ctx->f32;
5615 }
5616 break;
5617
5618 case PIPE_SHADER_TESS_CTRL:
5619 params[SI_PARAM_TCS_OFFCHIP_LAYOUT] = ctx->i32;
5620 params[SI_PARAM_TCS_OUT_OFFSETS] = ctx->i32;
5621 params[SI_PARAM_TCS_OUT_LAYOUT] = ctx->i32;
5622 params[SI_PARAM_TCS_IN_LAYOUT] = ctx->i32;
5623 params[ctx->param_oc_lds = SI_PARAM_TCS_OC_LDS] = ctx->i32;
5624 params[SI_PARAM_TESS_FACTOR_OFFSET] = ctx->i32;
5625 last_sgpr = SI_PARAM_TESS_FACTOR_OFFSET;
5626
5627 /* VGPRs */
5628 params[SI_PARAM_PATCH_ID] = ctx->i32;
5629 params[SI_PARAM_REL_IDS] = ctx->i32;
5630 num_params = SI_PARAM_REL_IDS+1;
5631
5632 /* SI_PARAM_TCS_OC_LDS and PARAM_TESS_FACTOR_OFFSET are
5633 * placed after the user SGPRs.
5634 */
5635 for (i = 0; i < SI_TCS_NUM_USER_SGPR + 2; i++)
5636 returns[num_returns++] = ctx->i32; /* SGPRs */
5637
5638 for (i = 0; i < 3; i++)
5639 returns[num_returns++] = ctx->f32; /* VGPRs */
5640 break;
5641
5642 case PIPE_SHADER_TESS_EVAL:
5643 params[SI_PARAM_TCS_OFFCHIP_LAYOUT] = ctx->i32;
5644 num_params = SI_PARAM_TCS_OFFCHIP_LAYOUT+1;
5645
5646 if (shader->key.as_es) {
5647 params[ctx->param_oc_lds = num_params++] = ctx->i32;
5648 params[ctx->param_tess_offchip = num_params++] = ctx->i32;
5649 params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
5650 } else {
5651 params[ctx->param_tess_offchip = num_params++] = ctx->i32;
5652 declare_streamout_params(ctx, &shader->selector->so,
5653 params, ctx->i32, &num_params);
5654 params[ctx->param_oc_lds = num_params++] = ctx->i32;
5655 }
5656 last_sgpr = num_params - 1;
5657
5658 /* VGPRs */
5659 params[ctx->param_tes_u = num_params++] = ctx->f32;
5660 params[ctx->param_tes_v = num_params++] = ctx->f32;
5661 params[ctx->param_tes_rel_patch_id = num_params++] = ctx->i32;
5662 params[ctx->param_tes_patch_id = num_params++] = ctx->i32;
5663
5664 /* PrimitiveID output. */
5665 if (!shader->key.as_es)
5666 for (i = 0; i <= VS_EPILOG_PRIMID_LOC; i++)
5667 returns[num_returns++] = ctx->f32;
5668 break;
5669
5670 case PIPE_SHADER_GEOMETRY:
5671 params[SI_PARAM_GS2VS_OFFSET] = ctx->i32;
5672 params[SI_PARAM_GS_WAVE_ID] = ctx->i32;
5673 last_sgpr = SI_PARAM_GS_WAVE_ID;
5674
5675 /* VGPRs */
5676 params[SI_PARAM_VTX0_OFFSET] = ctx->i32;
5677 params[SI_PARAM_VTX1_OFFSET] = ctx->i32;
5678 params[SI_PARAM_PRIMITIVE_ID] = ctx->i32;
5679 params[SI_PARAM_VTX2_OFFSET] = ctx->i32;
5680 params[SI_PARAM_VTX3_OFFSET] = ctx->i32;
5681 params[SI_PARAM_VTX4_OFFSET] = ctx->i32;
5682 params[SI_PARAM_VTX5_OFFSET] = ctx->i32;
5683 params[SI_PARAM_GS_INSTANCE_ID] = ctx->i32;
5684 num_params = SI_PARAM_GS_INSTANCE_ID+1;
5685 break;
5686
5687 case PIPE_SHADER_FRAGMENT:
5688 params[SI_PARAM_ALPHA_REF] = ctx->f32;
5689 params[SI_PARAM_PRIM_MASK] = ctx->i32;
5690 last_sgpr = SI_PARAM_PRIM_MASK;
5691 params[SI_PARAM_PERSP_SAMPLE] = ctx->v2i32;
5692 params[SI_PARAM_PERSP_CENTER] = ctx->v2i32;
5693 params[SI_PARAM_PERSP_CENTROID] = ctx->v2i32;
5694 params[SI_PARAM_PERSP_PULL_MODEL] = v3i32;
5695 params[SI_PARAM_LINEAR_SAMPLE] = ctx->v2i32;
5696 params[SI_PARAM_LINEAR_CENTER] = ctx->v2i32;
5697 params[SI_PARAM_LINEAR_CENTROID] = ctx->v2i32;
5698 params[SI_PARAM_LINE_STIPPLE_TEX] = ctx->f32;
5699 params[SI_PARAM_POS_X_FLOAT] = ctx->f32;
5700 params[SI_PARAM_POS_Y_FLOAT] = ctx->f32;
5701 params[SI_PARAM_POS_Z_FLOAT] = ctx->f32;
5702 params[SI_PARAM_POS_W_FLOAT] = ctx->f32;
5703 params[SI_PARAM_FRONT_FACE] = ctx->i32;
5704 shader->info.face_vgpr_index = 20;
5705 params[SI_PARAM_ANCILLARY] = ctx->i32;
5706 params[SI_PARAM_SAMPLE_COVERAGE] = ctx->f32;
5707 params[SI_PARAM_POS_FIXED_PT] = ctx->i32;
5708 num_params = SI_PARAM_POS_FIXED_PT+1;
5709
5710 /* Color inputs from the prolog. */
5711 if (shader->selector->info.colors_read) {
5712 unsigned num_color_elements =
5713 util_bitcount(shader->selector->info.colors_read);
5714
5715 assert(num_params + num_color_elements <= ARRAY_SIZE(params));
5716 for (i = 0; i < num_color_elements; i++)
5717 params[num_params++] = ctx->f32;
5718
5719 num_prolog_vgprs += num_color_elements;
5720 }
5721
5722 /* Outputs for the epilog. */
5723 num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
5724 num_returns =
5725 num_return_sgprs +
5726 util_bitcount(shader->selector->info.colors_written) * 4 +
5727 shader->selector->info.writes_z +
5728 shader->selector->info.writes_stencil +
5729 shader->selector->info.writes_samplemask +
5730 1 /* SampleMaskIn */;
5731
5732 num_returns = MAX2(num_returns,
5733 num_return_sgprs +
5734 PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
5735
5736 for (i = 0; i < num_return_sgprs; i++)
5737 returns[i] = ctx->i32;
5738 for (; i < num_returns; i++)
5739 returns[i] = ctx->f32;
5740 break;
5741
5742 case PIPE_SHADER_COMPUTE:
5743 params[SI_PARAM_GRID_SIZE] = v3i32;
5744 params[SI_PARAM_BLOCK_SIZE] = v3i32;
5745 params[SI_PARAM_BLOCK_ID] = v3i32;
5746 last_sgpr = SI_PARAM_BLOCK_ID;
5747
5748 params[SI_PARAM_THREAD_ID] = v3i32;
5749 num_params = SI_PARAM_THREAD_ID + 1;
5750 break;
5751 default:
5752 assert(0 && "unimplemented shader");
5753 return;
5754 }
5755
5756 assert(num_params <= ARRAY_SIZE(params));
5757
5758 si_create_function(ctx, "main", returns, num_returns, params,
5759 num_params, last_sgpr);
5760
5761 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5762 if (ctx->type == PIPE_SHADER_FRAGMENT &&
5763 ctx->separate_prolog) {
5764 si_llvm_add_attribute(ctx->main_fn,
5765 "InitialPSInputAddr",
5766 S_0286D0_PERSP_SAMPLE_ENA(1) |
5767 S_0286D0_PERSP_CENTER_ENA(1) |
5768 S_0286D0_PERSP_CENTROID_ENA(1) |
5769 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5770 S_0286D0_LINEAR_CENTER_ENA(1) |
5771 S_0286D0_LINEAR_CENTROID_ENA(1) |
5772 S_0286D0_FRONT_FACE_ENA(1) |
5773 S_0286D0_POS_FIXED_PT_ENA(1));
5774 } else if (ctx->type == PIPE_SHADER_COMPUTE) {
5775 si_llvm_add_attribute(ctx->main_fn,
5776 "amdgpu-max-work-group-size",
5777 si_get_max_workgroup_size(shader));
5778 }
5779
5780 shader->info.num_input_sgprs = 0;
5781 shader->info.num_input_vgprs = 0;
5782
5783 for (i = 0; i <= last_sgpr; ++i)
5784 shader->info.num_input_sgprs += llvm_get_type_size(params[i]) / 4;
5785
5786 for (; i < num_params; ++i)
5787 shader->info.num_input_vgprs += llvm_get_type_size(params[i]) / 4;
5788
5789 assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
5790 shader->info.num_input_vgprs -= num_prolog_vgprs;
5791
5792 if (!ctx->screen->has_ds_bpermute &&
5793 bld_base->info &&
5794 (bld_base->info->opcode_count[TGSI_OPCODE_DDX] > 0 ||
5795 bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0 ||
5796 bld_base->info->opcode_count[TGSI_OPCODE_DDX_FINE] > 0 ||
5797 bld_base->info->opcode_count[TGSI_OPCODE_DDY_FINE] > 0 ||
5798 bld_base->info->opcode_count[TGSI_OPCODE_INTERP_OFFSET] > 0 ||
5799 bld_base->info->opcode_count[TGSI_OPCODE_INTERP_SAMPLE] > 0))
5800 ctx->lds =
5801 LLVMAddGlobalInAddressSpace(gallivm->module,
5802 LLVMArrayType(ctx->i32, 64),
5803 "ddxy_lds",
5804 LOCAL_ADDR_SPACE);
5805
5806 if ((ctx->type == PIPE_SHADER_VERTEX && shader->key.as_ls) ||
5807 ctx->type == PIPE_SHADER_TESS_CTRL ||
5808 ctx->type == PIPE_SHADER_TESS_EVAL)
5809 declare_tess_lds(ctx);
5810 }
5811
5812 /**
5813 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5814 * for later use.
5815 */
5816 static void preload_ring_buffers(struct si_shader_context *ctx)
5817 {
5818 struct gallivm_state *gallivm =
5819 ctx->soa.bld_base.base.gallivm;
5820 LLVMBuilderRef builder = gallivm->builder;
5821
5822 LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
5823 SI_PARAM_RW_BUFFERS);
5824
5825 if ((ctx->type == PIPE_SHADER_VERTEX &&
5826 ctx->shader->key.as_es) ||
5827 (ctx->type == PIPE_SHADER_TESS_EVAL &&
5828 ctx->shader->key.as_es) ||
5829 ctx->type == PIPE_SHADER_GEOMETRY) {
5830 unsigned ring =
5831 ctx->type == PIPE_SHADER_GEOMETRY ? SI_GS_RING_ESGS
5832 : SI_ES_RING_ESGS;
5833 LLVMValueRef offset = lp_build_const_int32(gallivm, ring);
5834
5835 ctx->esgs_ring =
5836 build_indexed_load_const(ctx, buf_ptr, offset);
5837 }
5838
5839 if (ctx->shader->is_gs_copy_shader) {
5840 LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
5841
5842 ctx->gsvs_ring[0] =
5843 build_indexed_load_const(ctx, buf_ptr, offset);
5844 } else if (ctx->type == PIPE_SHADER_GEOMETRY) {
5845 const struct si_shader_selector *sel = ctx->shader->selector;
5846 struct lp_build_context *uint = &ctx->soa.bld_base.uint_bld;
5847 LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
5848 LLVMValueRef base_ring;
5849
5850 base_ring = build_indexed_load_const(ctx, buf_ptr, offset);
5851
5852 /* The conceptual layout of the GSVS ring is
5853 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5854 * but the real memory layout is swizzled across
5855 * threads:
5856 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5857 * t16v0c0 ..
5858 * Override the buffer descriptor accordingly.
5859 */
5860 LLVMTypeRef v2i64 = LLVMVectorType(ctx->i64, 2);
5861 uint64_t stream_offset = 0;
5862
5863 for (unsigned stream = 0; stream < 4; ++stream) {
5864 unsigned num_components;
5865 unsigned stride;
5866 unsigned num_records;
5867 LLVMValueRef ring, tmp;
5868
5869 num_components = sel->info.num_stream_output_components[stream];
5870 if (!num_components)
5871 continue;
5872
5873 stride = 4 * num_components * sel->gs_max_out_vertices;
5874
5875 /* Limit on the stride field for <= CIK. */
5876 assert(stride < (1 << 14));
5877
5878 num_records = 64;
5879 if (ctx->screen->b.chip_class >= VI)
5880 num_records *= stride;
5881
5882 ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
5883 tmp = LLVMBuildExtractElement(builder, ring, uint->zero, "");
5884 tmp = LLVMBuildAdd(builder, tmp,
5885 LLVMConstInt(ctx->i64,
5886 stream_offset, 0), "");
5887 stream_offset += stride * 64;
5888
5889 ring = LLVMBuildInsertElement(builder, ring, tmp, uint->zero, "");
5890 ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
5891 tmp = LLVMBuildExtractElement(builder, ring, uint->one, "");
5892 tmp = LLVMBuildOr(builder, tmp,
5893 LLVMConstInt(ctx->i32,
5894 S_008F04_STRIDE(stride) |
5895 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5896 ring = LLVMBuildInsertElement(builder, ring, tmp, uint->one, "");
5897 ring = LLVMBuildInsertElement(builder, ring,
5898 LLVMConstInt(ctx->i32, num_records, 0),
5899 LLVMConstInt(ctx->i32, 2, 0), "");
5900 ring = LLVMBuildInsertElement(builder, ring,
5901 LLVMConstInt(ctx->i32,
5902 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
5903 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
5904 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
5905 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
5906 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
5907 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
5908 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
5909 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5910 S_008F0C_ADD_TID_ENABLE(1),
5911 0),
5912 LLVMConstInt(ctx->i32, 3, 0), "");
5913 ring = LLVMBuildBitCast(builder, ring, ctx->v16i8, "");
5914
5915 ctx->gsvs_ring[stream] = ring;
5916 }
5917 }
5918 }
5919
5920 static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
5921 LLVMValueRef param_rw_buffers,
5922 unsigned param_pos_fixed_pt)
5923 {
5924 struct lp_build_tgsi_context *bld_base =
5925 &ctx->soa.bld_base;
5926 struct gallivm_state *gallivm = bld_base->base.gallivm;
5927 LLVMBuilderRef builder = gallivm->builder;
5928 LLVMValueRef slot, desc, offset, row, bit, address[2];
5929
5930 /* Use the fixed-point gl_FragCoord input.
5931 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5932 * per coordinate to get the repeating effect.
5933 */
5934 address[0] = unpack_param(ctx, param_pos_fixed_pt, 0, 5);
5935 address[1] = unpack_param(ctx, param_pos_fixed_pt, 16, 5);
5936
5937 /* Load the buffer descriptor. */
5938 slot = lp_build_const_int32(gallivm, SI_PS_CONST_POLY_STIPPLE);
5939 desc = build_indexed_load_const(ctx, param_rw_buffers, slot);
5940
5941 /* The stipple pattern is 32x32, each row has 32 bits. */
5942 offset = LLVMBuildMul(builder, address[1],
5943 LLVMConstInt(ctx->i32, 4, 0), "");
5944 row = buffer_load_const(ctx, desc, offset);
5945 row = LLVMBuildBitCast(builder, row, ctx->i32, "");
5946 bit = LLVMBuildLShr(builder, row, address[0], "");
5947 bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
5948
5949 /* The intrinsic kills the thread if arg < 0. */
5950 bit = LLVMBuildSelect(builder, bit, LLVMConstReal(ctx->f32, 0),
5951 LLVMConstReal(ctx->f32, -1), "");
5952 lp_build_intrinsic(builder, "llvm.AMDGPU.kill", ctx->voidt, &bit, 1, 0);
5953 }
5954
5955 void si_shader_binary_read_config(struct radeon_shader_binary *binary,
5956 struct si_shader_config *conf,
5957 unsigned symbol_offset)
5958 {
5959 unsigned i;
5960 const unsigned char *config =
5961 radeon_shader_binary_config_start(binary, symbol_offset);
5962 bool really_needs_scratch = false;
5963
5964 /* LLVM adds SGPR spills to the scratch size.
5965 * Find out if we really need the scratch buffer.
5966 */
5967 for (i = 0; i < binary->reloc_count; i++) {
5968 const struct radeon_shader_reloc *reloc = &binary->relocs[i];
5969
5970 if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name) ||
5971 !strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
5972 really_needs_scratch = true;
5973 break;
5974 }
5975 }
5976
5977 /* XXX: We may be able to emit some of these values directly rather than
5978 * extracting fields to be emitted later.
5979 */
5980
5981 for (i = 0; i < binary->config_size_per_symbol; i+= 8) {
5982 unsigned reg = util_le32_to_cpu(*(uint32_t*)(config + i));
5983 unsigned value = util_le32_to_cpu(*(uint32_t*)(config + i + 4));
5984 switch (reg) {
5985 case R_00B028_SPI_SHADER_PGM_RSRC1_PS:
5986 case R_00B128_SPI_SHADER_PGM_RSRC1_VS:
5987 case R_00B228_SPI_SHADER_PGM_RSRC1_GS:
5988 case R_00B848_COMPUTE_PGM_RSRC1:
5989 conf->num_sgprs = MAX2(conf->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
5990 conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
5991 conf->float_mode = G_00B028_FLOAT_MODE(value);
5992 conf->rsrc1 = value;
5993 break;
5994 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
5995 conf->lds_size = MAX2(conf->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
5996 break;
5997 case R_00B84C_COMPUTE_PGM_RSRC2:
5998 conf->lds_size = MAX2(conf->lds_size, G_00B84C_LDS_SIZE(value));
5999 conf->rsrc2 = value;
6000 break;
6001 case R_0286CC_SPI_PS_INPUT_ENA:
6002 conf->spi_ps_input_ena = value;
6003 break;
6004 case R_0286D0_SPI_PS_INPUT_ADDR:
6005 conf->spi_ps_input_addr = value;
6006 break;
6007 case R_0286E8_SPI_TMPRING_SIZE:
6008 case R_00B860_COMPUTE_TMPRING_SIZE:
6009 /* WAVESIZE is in units of 256 dwords. */
6010 if (really_needs_scratch)
6011 conf->scratch_bytes_per_wave =
6012 G_00B860_WAVESIZE(value) * 256 * 4;
6013 break;
6014 case 0x4: /* SPILLED_SGPRS */
6015 conf->spilled_sgprs = value;
6016 break;
6017 case 0x8: /* SPILLED_VGPRS */
6018 conf->spilled_vgprs = value;
6019 break;
6020 default:
6021 {
6022 static bool printed;
6023
6024 if (!printed) {
6025 fprintf(stderr, "Warning: LLVM emitted unknown "
6026 "config register: 0x%x\n", reg);
6027 printed = true;
6028 }
6029 }
6030 break;
6031 }
6032 }
6033
6034 if (!conf->spi_ps_input_addr)
6035 conf->spi_ps_input_addr = conf->spi_ps_input_ena;
6036 }
6037
6038 void si_shader_apply_scratch_relocs(struct si_context *sctx,
6039 struct si_shader *shader,
6040 struct si_shader_config *config,
6041 uint64_t scratch_va)
6042 {
6043 unsigned i;
6044 uint32_t scratch_rsrc_dword0 = scratch_va;
6045 uint32_t scratch_rsrc_dword1 =
6046 S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
6047
6048 /* Enable scratch coalescing if LLVM sets ELEMENT_SIZE & INDEX_STRIDE
6049 * correctly.
6050 */
6051 if (HAVE_LLVM >= 0x0309)
6052 scratch_rsrc_dword1 |= S_008F04_SWIZZLE_ENABLE(1);
6053 else
6054 scratch_rsrc_dword1 |=
6055 S_008F04_STRIDE(config->scratch_bytes_per_wave / 64);
6056
6057 for (i = 0 ; i < shader->binary.reloc_count; i++) {
6058 const struct radeon_shader_reloc *reloc =
6059 &shader->binary.relocs[i];
6060 if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name)) {
6061 util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset,
6062 &scratch_rsrc_dword0, 4);
6063 } else if (!strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
6064 util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset,
6065 &scratch_rsrc_dword1, 4);
6066 }
6067 }
6068 }
6069
6070 static unsigned si_get_shader_binary_size(struct si_shader *shader)
6071 {
6072 unsigned size = shader->binary.code_size;
6073
6074 if (shader->prolog)
6075 size += shader->prolog->binary.code_size;
6076 if (shader->epilog)
6077 size += shader->epilog->binary.code_size;
6078 return size;
6079 }
6080
6081 int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader)
6082 {
6083 const struct radeon_shader_binary *prolog =
6084 shader->prolog ? &shader->prolog->binary : NULL;
6085 const struct radeon_shader_binary *epilog =
6086 shader->epilog ? &shader->epilog->binary : NULL;
6087 const struct radeon_shader_binary *mainb = &shader->binary;
6088 unsigned bo_size = si_get_shader_binary_size(shader) +
6089 (!epilog ? mainb->rodata_size : 0);
6090 unsigned char *ptr;
6091
6092 assert(!prolog || !prolog->rodata_size);
6093 assert((!prolog && !epilog) || !mainb->rodata_size);
6094 assert(!epilog || !epilog->rodata_size);
6095
6096 r600_resource_reference(&shader->bo, NULL);
6097 shader->bo = (struct r600_resource*)
6098 pipe_buffer_create(&sscreen->b.b, 0,
6099 PIPE_USAGE_IMMUTABLE, bo_size);
6100 if (!shader->bo)
6101 return -ENOMEM;
6102
6103 /* Upload. */
6104 ptr = sscreen->b.ws->buffer_map(shader->bo->buf, NULL,
6105 PIPE_TRANSFER_READ_WRITE);
6106
6107 if (prolog) {
6108 util_memcpy_cpu_to_le32(ptr, prolog->code, prolog->code_size);
6109 ptr += prolog->code_size;
6110 }
6111
6112 util_memcpy_cpu_to_le32(ptr, mainb->code, mainb->code_size);
6113 ptr += mainb->code_size;
6114
6115 if (epilog)
6116 util_memcpy_cpu_to_le32(ptr, epilog->code, epilog->code_size);
6117 else if (mainb->rodata_size > 0)
6118 util_memcpy_cpu_to_le32(ptr, mainb->rodata, mainb->rodata_size);
6119
6120 sscreen->b.ws->buffer_unmap(shader->bo->buf);
6121 return 0;
6122 }
6123
6124 static void si_shader_dump_disassembly(const struct radeon_shader_binary *binary,
6125 struct pipe_debug_callback *debug,
6126 const char *name, FILE *file)
6127 {
6128 char *line, *p;
6129 unsigned i, count;
6130
6131 if (binary->disasm_string) {
6132 fprintf(file, "Shader %s disassembly:\n", name);
6133 fprintf(file, "%s", binary->disasm_string);
6134
6135 if (debug && debug->debug_message) {
6136 /* Very long debug messages are cut off, so send the
6137 * disassembly one line at a time. This causes more
6138 * overhead, but on the plus side it simplifies
6139 * parsing of resulting logs.
6140 */
6141 pipe_debug_message(debug, SHADER_INFO,
6142 "Shader Disassembly Begin");
6143
6144 line = binary->disasm_string;
6145 while (*line) {
6146 p = util_strchrnul(line, '\n');
6147 count = p - line;
6148
6149 if (count) {
6150 pipe_debug_message(debug, SHADER_INFO,
6151 "%.*s", count, line);
6152 }
6153
6154 if (!*p)
6155 break;
6156 line = p + 1;
6157 }
6158
6159 pipe_debug_message(debug, SHADER_INFO,
6160 "Shader Disassembly End");
6161 }
6162 } else {
6163 fprintf(file, "Shader %s binary:\n", name);
6164 for (i = 0; i < binary->code_size; i += 4) {
6165 fprintf(file, "@0x%x: %02x%02x%02x%02x\n", i,
6166 binary->code[i + 3], binary->code[i + 2],
6167 binary->code[i + 1], binary->code[i]);
6168 }
6169 }
6170 }
6171
6172 static void si_shader_dump_stats(struct si_screen *sscreen,
6173 struct si_shader *shader,
6174 struct pipe_debug_callback *debug,
6175 unsigned processor,
6176 FILE *file)
6177 {
6178 struct si_shader_config *conf = &shader->config;
6179 unsigned num_inputs = shader->selector ? shader->selector->info.num_inputs : 0;
6180 unsigned code_size = si_get_shader_binary_size(shader);
6181 unsigned lds_increment = sscreen->b.chip_class >= CIK ? 512 : 256;
6182 unsigned lds_per_wave = 0;
6183 unsigned max_simd_waves = 10;
6184
6185 /* Compute LDS usage for PS. */
6186 switch (processor) {
6187 case PIPE_SHADER_FRAGMENT:
6188 /* The minimum usage per wave is (num_inputs * 48). The maximum
6189 * usage is (num_inputs * 48 * 16).
6190 * We can get anything in between and it varies between waves.
6191 *
6192 * The 48 bytes per input for a single primitive is equal to
6193 * 4 bytes/component * 4 components/input * 3 points.
6194 *
6195 * Other stages don't know the size at compile time or don't
6196 * allocate LDS per wave, but instead they do it per thread group.
6197 */
6198 lds_per_wave = conf->lds_size * lds_increment +
6199 align(num_inputs * 48, lds_increment);
6200 break;
6201 case PIPE_SHADER_COMPUTE:
6202 if (shader->selector) {
6203 unsigned max_workgroup_size =
6204 si_get_max_workgroup_size(shader);
6205 lds_per_wave = (conf->lds_size * lds_increment) /
6206 DIV_ROUND_UP(max_workgroup_size, 64);
6207 }
6208 break;
6209 }
6210
6211 /* Compute the per-SIMD wave counts. */
6212 if (conf->num_sgprs) {
6213 if (sscreen->b.chip_class >= VI)
6214 max_simd_waves = MIN2(max_simd_waves, 800 / conf->num_sgprs);
6215 else
6216 max_simd_waves = MIN2(max_simd_waves, 512 / conf->num_sgprs);
6217 }
6218
6219 if (conf->num_vgprs)
6220 max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
6221
6222 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
6223 * 16KB makes some SIMDs unoccupied). */
6224 if (lds_per_wave)
6225 max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
6226
6227 if (file != stderr ||
6228 r600_can_dump_shader(&sscreen->b, processor)) {
6229 if (processor == PIPE_SHADER_FRAGMENT) {
6230 fprintf(file, "*** SHADER CONFIG ***\n"
6231 "SPI_PS_INPUT_ADDR = 0x%04x\n"
6232 "SPI_PS_INPUT_ENA = 0x%04x\n",
6233 conf->spi_ps_input_addr, conf->spi_ps_input_ena);
6234 }
6235
6236 fprintf(file, "*** SHADER STATS ***\n"
6237 "SGPRS: %d\n"
6238 "VGPRS: %d\n"
6239 "Spilled SGPRs: %d\n"
6240 "Spilled VGPRs: %d\n"
6241 "Private memory VGPRs: %d\n"
6242 "Code Size: %d bytes\n"
6243 "LDS: %d blocks\n"
6244 "Scratch: %d bytes per wave\n"
6245 "Max Waves: %d\n"
6246 "********************\n\n\n",
6247 conf->num_sgprs, conf->num_vgprs,
6248 conf->spilled_sgprs, conf->spilled_vgprs,
6249 conf->private_mem_vgprs, code_size,
6250 conf->lds_size, conf->scratch_bytes_per_wave,
6251 max_simd_waves);
6252 }
6253
6254 pipe_debug_message(debug, SHADER_INFO,
6255 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
6256 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
6257 "Spilled VGPRs: %d PrivMem VGPRs: %d",
6258 conf->num_sgprs, conf->num_vgprs, code_size,
6259 conf->lds_size, conf->scratch_bytes_per_wave,
6260 max_simd_waves, conf->spilled_sgprs,
6261 conf->spilled_vgprs, conf->private_mem_vgprs);
6262 }
6263
6264 static const char *si_get_shader_name(struct si_shader *shader,
6265 unsigned processor)
6266 {
6267 switch (processor) {
6268 case PIPE_SHADER_VERTEX:
6269 if (shader->key.as_es)
6270 return "Vertex Shader as ES";
6271 else if (shader->key.as_ls)
6272 return "Vertex Shader as LS";
6273 else
6274 return "Vertex Shader as VS";
6275 case PIPE_SHADER_TESS_CTRL:
6276 return "Tessellation Control Shader";
6277 case PIPE_SHADER_TESS_EVAL:
6278 if (shader->key.as_es)
6279 return "Tessellation Evaluation Shader as ES";
6280 else
6281 return "Tessellation Evaluation Shader as VS";
6282 case PIPE_SHADER_GEOMETRY:
6283 if (shader->is_gs_copy_shader)
6284 return "GS Copy Shader as VS";
6285 else
6286 return "Geometry Shader";
6287 case PIPE_SHADER_FRAGMENT:
6288 return "Pixel Shader";
6289 case PIPE_SHADER_COMPUTE:
6290 return "Compute Shader";
6291 default:
6292 return "Unknown Shader";
6293 }
6294 }
6295
6296 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
6297 struct pipe_debug_callback *debug, unsigned processor,
6298 FILE *file)
6299 {
6300 if (file != stderr ||
6301 r600_can_dump_shader(&sscreen->b, processor))
6302 si_dump_shader_key(processor, &shader->key, file);
6303
6304 if (file != stderr && shader->binary.llvm_ir_string) {
6305 fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
6306 si_get_shader_name(shader, processor));
6307 fprintf(file, "%s\n", shader->binary.llvm_ir_string);
6308 }
6309
6310 if (file != stderr ||
6311 (r600_can_dump_shader(&sscreen->b, processor) &&
6312 !(sscreen->b.debug_flags & DBG_NO_ASM))) {
6313 fprintf(file, "\n%s:\n", si_get_shader_name(shader, processor));
6314
6315 if (shader->prolog)
6316 si_shader_dump_disassembly(&shader->prolog->binary,
6317 debug, "prolog", file);
6318
6319 si_shader_dump_disassembly(&shader->binary, debug, "main", file);
6320
6321 if (shader->epilog)
6322 si_shader_dump_disassembly(&shader->epilog->binary,
6323 debug, "epilog", file);
6324 fprintf(file, "\n");
6325 }
6326
6327 si_shader_dump_stats(sscreen, shader, debug, processor, file);
6328 }
6329
6330 int si_compile_llvm(struct si_screen *sscreen,
6331 struct radeon_shader_binary *binary,
6332 struct si_shader_config *conf,
6333 LLVMTargetMachineRef tm,
6334 LLVMModuleRef mod,
6335 struct pipe_debug_callback *debug,
6336 unsigned processor,
6337 const char *name)
6338 {
6339 int r = 0;
6340 unsigned count = p_atomic_inc_return(&sscreen->b.num_compilations);
6341
6342 if (r600_can_dump_shader(&sscreen->b, processor)) {
6343 fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
6344
6345 if (!(sscreen->b.debug_flags & (DBG_NO_IR | DBG_PREOPT_IR))) {
6346 fprintf(stderr, "%s LLVM IR:\n\n", name);
6347 LLVMDumpModule(mod);
6348 fprintf(stderr, "\n");
6349 }
6350 }
6351
6352 if (sscreen->record_llvm_ir) {
6353 char *ir = LLVMPrintModuleToString(mod);
6354 binary->llvm_ir_string = strdup(ir);
6355 LLVMDisposeMessage(ir);
6356 }
6357
6358 if (!si_replace_shader(count, binary)) {
6359 r = si_llvm_compile(mod, binary, tm, debug);
6360 if (r)
6361 return r;
6362 }
6363
6364 si_shader_binary_read_config(binary, conf, 0);
6365
6366 /* Enable 64-bit and 16-bit denormals, because there is no performance
6367 * cost.
6368 *
6369 * If denormals are enabled, all floating-point output modifiers are
6370 * ignored.
6371 *
6372 * Don't enable denormals for 32-bit floats, because:
6373 * - Floating-point output modifiers would be ignored by the hw.
6374 * - Some opcodes don't support denormals, such as v_mad_f32. We would
6375 * have to stop using those.
6376 * - SI & CI would be very slow.
6377 */
6378 conf->float_mode |= V_00B028_FP_64_DENORMS;
6379
6380 FREE(binary->config);
6381 FREE(binary->global_symbol_offsets);
6382 binary->config = NULL;
6383 binary->global_symbol_offsets = NULL;
6384
6385 /* Some shaders can't have rodata because their binaries can be
6386 * concatenated.
6387 */
6388 if (binary->rodata_size &&
6389 (processor == PIPE_SHADER_VERTEX ||
6390 processor == PIPE_SHADER_TESS_CTRL ||
6391 processor == PIPE_SHADER_TESS_EVAL ||
6392 processor == PIPE_SHADER_FRAGMENT)) {
6393 fprintf(stderr, "radeonsi: The shader can't have rodata.");
6394 return -EINVAL;
6395 }
6396
6397 return r;
6398 }
6399
6400 static void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
6401 {
6402 if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
6403 LLVMBuildRetVoid(ctx->gallivm.builder);
6404 else
6405 LLVMBuildRet(ctx->gallivm.builder, ret);
6406 }
6407
6408 /* Generate code for the hardware VS shader stage to go with a geometry shader */
6409 struct si_shader *
6410 si_generate_gs_copy_shader(struct si_screen *sscreen,
6411 LLVMTargetMachineRef tm,
6412 struct si_shader_selector *gs_selector,
6413 struct pipe_debug_callback *debug)
6414 {
6415 struct si_shader_context ctx;
6416 struct si_shader *shader;
6417 struct gallivm_state *gallivm = &ctx.gallivm;
6418 LLVMBuilderRef builder;
6419 struct lp_build_tgsi_context *bld_base = &ctx.soa.bld_base;
6420 struct lp_build_context *uint = &bld_base->uint_bld;
6421 struct si_shader_output_values *outputs;
6422 struct tgsi_shader_info *gsinfo = &gs_selector->info;
6423 LLVMValueRef args[9];
6424 int i, r;
6425
6426 outputs = MALLOC(gsinfo->num_outputs * sizeof(outputs[0]));
6427
6428 if (!outputs)
6429 return NULL;
6430
6431 shader = CALLOC_STRUCT(si_shader);
6432 if (!shader) {
6433 FREE(outputs);
6434 return NULL;
6435 }
6436
6437
6438 shader->selector = gs_selector;
6439 shader->is_gs_copy_shader = true;
6440
6441 si_init_shader_ctx(&ctx, sscreen, shader, tm);
6442 ctx.type = PIPE_SHADER_VERTEX;
6443
6444 builder = gallivm->builder;
6445
6446 create_meta_data(&ctx);
6447 create_function(&ctx);
6448 preload_ring_buffers(&ctx);
6449
6450 args[0] = ctx.gsvs_ring[0];
6451 args[1] = lp_build_mul_imm(uint,
6452 LLVMGetParam(ctx.main_fn,
6453 ctx.param_vertex_id),
6454 4);
6455 args[3] = uint->zero;
6456 args[4] = uint->one; /* OFFEN */
6457 args[5] = uint->zero; /* IDXEN */
6458 args[6] = uint->one; /* GLC */
6459 args[7] = uint->one; /* SLC */
6460 args[8] = uint->zero; /* TFE */
6461
6462 /* Fetch the vertex stream ID.*/
6463 LLVMValueRef stream_id;
6464
6465 if (gs_selector->so.num_outputs)
6466 stream_id = unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
6467 else
6468 stream_id = uint->zero;
6469
6470 /* Fill in output information. */
6471 for (i = 0; i < gsinfo->num_outputs; ++i) {
6472 outputs[i].semantic_name = gsinfo->output_semantic_name[i];
6473 outputs[i].semantic_index = gsinfo->output_semantic_index[i];
6474
6475 for (int chan = 0; chan < 4; chan++) {
6476 outputs[i].vertex_stream[chan] =
6477 (gsinfo->output_streams[i] >> (2 * chan)) & 3;
6478 }
6479 }
6480
6481 LLVMBasicBlockRef end_bb;
6482 LLVMValueRef switch_inst;
6483
6484 end_bb = LLVMAppendBasicBlockInContext(gallivm->context, ctx.main_fn, "end");
6485 switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);
6486
6487 for (int stream = 0; stream < 4; stream++) {
6488 LLVMBasicBlockRef bb;
6489 unsigned offset;
6490
6491 if (!gsinfo->num_stream_output_components[stream])
6492 continue;
6493
6494 if (stream > 0 && !gs_selector->so.num_outputs)
6495 continue;
6496
6497 bb = LLVMInsertBasicBlockInContext(gallivm->context, end_bb, "out");
6498 LLVMAddCase(switch_inst, lp_build_const_int32(gallivm, stream), bb);
6499 LLVMPositionBuilderAtEnd(builder, bb);
6500
6501 /* Fetch vertex data from GSVS ring */
6502 offset = 0;
6503 for (i = 0; i < gsinfo->num_outputs; ++i) {
6504 for (unsigned chan = 0; chan < 4; chan++) {
6505 if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
6506 outputs[i].vertex_stream[chan] != stream) {
6507 outputs[i].values[chan] = ctx.soa.bld_base.base.undef;
6508 continue;
6509 }
6510
6511 args[2] = lp_build_const_int32(
6512 gallivm,
6513 offset * gs_selector->gs_max_out_vertices * 16 * 4);
6514 offset++;
6515
6516 outputs[i].values[chan] =
6517 LLVMBuildBitCast(gallivm->builder,
6518 lp_build_intrinsic(gallivm->builder,
6519 "llvm.SI.buffer.load.dword.i32.i32",
6520 ctx.i32, args, 9,
6521 LP_FUNC_ATTR_READONLY),
6522 ctx.f32, "");
6523 }
6524 }
6525
6526 /* Streamout and exports. */
6527 if (gs_selector->so.num_outputs) {
6528 si_llvm_emit_streamout(&ctx, outputs,
6529 gsinfo->num_outputs,
6530 stream);
6531 }
6532
6533 if (stream == 0)
6534 si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
6535
6536 LLVMBuildBr(builder, end_bb);
6537 }
6538
6539 LLVMPositionBuilderAtEnd(builder, end_bb);
6540
6541 LLVMBuildRetVoid(gallivm->builder);
6542
6543 /* Dump LLVM IR before any optimization passes */
6544 if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
6545 r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
6546 LLVMDumpModule(bld_base->base.gallivm->module);
6547
6548 si_llvm_finalize_module(&ctx,
6549 r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_GEOMETRY));
6550
6551 r = si_compile_llvm(sscreen, &ctx.shader->binary,
6552 &ctx.shader->config, ctx.tm,
6553 bld_base->base.gallivm->module,
6554 debug, PIPE_SHADER_GEOMETRY,
6555 "GS Copy Shader");
6556 if (!r) {
6557 if (r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
6558 fprintf(stderr, "GS Copy Shader:\n");
6559 si_shader_dump(sscreen, ctx.shader, debug,
6560 PIPE_SHADER_GEOMETRY, stderr);
6561 r = si_shader_binary_upload(sscreen, ctx.shader);
6562 }
6563
6564 si_llvm_dispose(&ctx);
6565
6566 FREE(outputs);
6567
6568 if (r != 0) {
6569 FREE(shader);
6570 shader = NULL;
6571 }
6572 return shader;
6573 }
6574
6575 static void si_dump_shader_key(unsigned shader, struct si_shader_key *key,
6576 FILE *f)
6577 {
6578 int i;
6579
6580 fprintf(f, "SHADER KEY\n");
6581
6582 switch (shader) {
6583 case PIPE_SHADER_VERTEX:
6584 fprintf(f, " part.vs.prolog.instance_divisors = {");
6585 for (i = 0; i < ARRAY_SIZE(key->part.vs.prolog.instance_divisors); i++)
6586 fprintf(f, !i ? "%u" : ", %u",
6587 key->part.vs.prolog.instance_divisors[i]);
6588 fprintf(f, "}\n");
6589 fprintf(f, " part.vs.epilog.export_prim_id = %u\n", key->part.vs.epilog.export_prim_id);
6590 fprintf(f, " as_es = %u\n", key->as_es);
6591 fprintf(f, " as_ls = %u\n", key->as_ls);
6592 fprintf(f, " mono.vs.fix_fetch = 0x%x\n", key->mono.vs.fix_fetch);
6593 break;
6594
6595 case PIPE_SHADER_TESS_CTRL:
6596 fprintf(f, " part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
6597 fprintf(f, " mono.tcs.inputs_to_copy = 0x%"PRIx64"\n", key->mono.tcs.inputs_to_copy);
6598 break;
6599
6600 case PIPE_SHADER_TESS_EVAL:
6601 fprintf(f, " part.tes.epilog.export_prim_id = %u\n", key->part.tes.epilog.export_prim_id);
6602 fprintf(f, " as_es = %u\n", key->as_es);
6603 break;
6604
6605 case PIPE_SHADER_GEOMETRY:
6606 fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
6607 break;
6608
6609 case PIPE_SHADER_COMPUTE:
6610 break;
6611
6612 case PIPE_SHADER_FRAGMENT:
6613 fprintf(f, " part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
6614 fprintf(f, " part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
6615 fprintf(f, " part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
6616 fprintf(f, " part.ps.prolog.force_persp_sample_interp = %u\n", key->part.ps.prolog.force_persp_sample_interp);
6617 fprintf(f, " part.ps.prolog.force_linear_sample_interp = %u\n", key->part.ps.prolog.force_linear_sample_interp);
6618 fprintf(f, " part.ps.prolog.force_persp_center_interp = %u\n", key->part.ps.prolog.force_persp_center_interp);
6619 fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
6620 fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
6621 fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
6622 fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
6623 fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
6624 fprintf(f, " part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
6625 fprintf(f, " part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
6626 fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
6627 fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
6628 fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
6629 break;
6630
6631 default:
6632 assert(0);
6633 }
6634
6635 if ((shader == PIPE_SHADER_GEOMETRY ||
6636 shader == PIPE_SHADER_TESS_EVAL ||
6637 shader == PIPE_SHADER_VERTEX) &&
6638 !key->as_es && !key->as_ls) {
6639 fprintf(f, " opt.hw_vs.kill_outputs = 0x%"PRIx64"\n", key->opt.hw_vs.kill_outputs);
6640 fprintf(f, " opt.hw_vs.kill_outputs2 = 0x%x\n", key->opt.hw_vs.kill_outputs2);
6641 fprintf(f, " opt.hw_vs.clip_disable = %u\n", key->opt.hw_vs.clip_disable);
6642 }
6643 }
6644
6645 static void si_init_shader_ctx(struct si_shader_context *ctx,
6646 struct si_screen *sscreen,
6647 struct si_shader *shader,
6648 LLVMTargetMachineRef tm)
6649 {
6650 struct lp_build_tgsi_context *bld_base;
6651 struct lp_build_tgsi_action tmpl = {};
6652
6653 si_llvm_context_init(ctx, sscreen, shader, tm,
6654 (shader && shader->selector) ? &shader->selector->info : NULL,
6655 (shader && shader->selector) ? shader->selector->tokens : NULL);
6656
6657 bld_base = &ctx->soa.bld_base;
6658 bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
6659
6660 bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID] = interp_action;
6661 bld_base->op_actions[TGSI_OPCODE_INTERP_SAMPLE] = interp_action;
6662 bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET] = interp_action;
6663
6664 bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
6665 bld_base->op_actions[TGSI_OPCODE_TEX2] = tex_action;
6666 bld_base->op_actions[TGSI_OPCODE_TXB] = tex_action;
6667 bld_base->op_actions[TGSI_OPCODE_TXB2] = tex_action;
6668 bld_base->op_actions[TGSI_OPCODE_TXD] = tex_action;
6669 bld_base->op_actions[TGSI_OPCODE_TXF] = tex_action;
6670 bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action;
6671 bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action;
6672 bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
6673 bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = txq_fetch_args;
6674 bld_base->op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
6675 bld_base->op_actions[TGSI_OPCODE_TG4] = tex_action;
6676 bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action;
6677 bld_base->op_actions[TGSI_OPCODE_TXQS].emit = si_llvm_emit_txqs;
6678
6679 bld_base->op_actions[TGSI_OPCODE_LOAD].fetch_args = load_fetch_args;
6680 bld_base->op_actions[TGSI_OPCODE_LOAD].emit = load_emit;
6681 bld_base->op_actions[TGSI_OPCODE_STORE].fetch_args = store_fetch_args;
6682 bld_base->op_actions[TGSI_OPCODE_STORE].emit = store_emit;
6683 bld_base->op_actions[TGSI_OPCODE_RESQ].fetch_args = resq_fetch_args;
6684 bld_base->op_actions[TGSI_OPCODE_RESQ].emit = resq_emit;
6685
6686 tmpl.fetch_args = atomic_fetch_args;
6687 tmpl.emit = atomic_emit;
6688 bld_base->op_actions[TGSI_OPCODE_ATOMUADD] = tmpl;
6689 bld_base->op_actions[TGSI_OPCODE_ATOMUADD].intr_name = "add";
6690 bld_base->op_actions[TGSI_OPCODE_ATOMXCHG] = tmpl;
6691 bld_base->op_actions[TGSI_OPCODE_ATOMXCHG].intr_name = "swap";
6692 bld_base->op_actions[TGSI_OPCODE_ATOMCAS] = tmpl;
6693 bld_base->op_actions[TGSI_OPCODE_ATOMCAS].intr_name = "cmpswap";
6694 bld_base->op_actions[TGSI_OPCODE_ATOMAND] = tmpl;
6695 bld_base->op_actions[TGSI_OPCODE_ATOMAND].intr_name = "and";
6696 bld_base->op_actions[TGSI_OPCODE_ATOMOR] = tmpl;
6697 bld_base->op_actions[TGSI_OPCODE_ATOMOR].intr_name = "or";
6698 bld_base->op_actions[TGSI_OPCODE_ATOMXOR] = tmpl;
6699 bld_base->op_actions[TGSI_OPCODE_ATOMXOR].intr_name = "xor";
6700 bld_base->op_actions[TGSI_OPCODE_ATOMUMIN] = tmpl;
6701 bld_base->op_actions[TGSI_OPCODE_ATOMUMIN].intr_name = "umin";
6702 bld_base->op_actions[TGSI_OPCODE_ATOMUMAX] = tmpl;
6703 bld_base->op_actions[TGSI_OPCODE_ATOMUMAX].intr_name = "umax";
6704 bld_base->op_actions[TGSI_OPCODE_ATOMIMIN] = tmpl;
6705 bld_base->op_actions[TGSI_OPCODE_ATOMIMIN].intr_name = "smin";
6706 bld_base->op_actions[TGSI_OPCODE_ATOMIMAX] = tmpl;
6707 bld_base->op_actions[TGSI_OPCODE_ATOMIMAX].intr_name = "smax";
6708
6709 bld_base->op_actions[TGSI_OPCODE_MEMBAR].emit = membar_emit;
6710
6711 bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
6712 bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
6713 bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
6714 bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;
6715
6716 bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
6717 bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
6718 bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
6719 }
6720
6721 /* Return true if the PARAM export has been eliminated. */
6722 static bool si_eliminate_const_output(struct si_shader_context *ctx,
6723 LLVMValueRef inst, unsigned offset)
6724 {
6725 struct si_shader *shader = ctx->shader;
6726 unsigned num_outputs = shader->selector->info.num_outputs;
6727 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
6728 bool is_zero[4] = {}, is_one[4] = {};
6729
6730 for (i = 0; i < 4; i++) {
6731 LLVMBool loses_info;
6732 LLVMValueRef p = LLVMGetOperand(inst, 5 + i);
6733
6734 /* It's a constant expression. Undef outputs are eliminated too. */
6735 if (LLVMIsUndef(p)) {
6736 is_zero[i] = true;
6737 is_one[i] = true;
6738 } else if (LLVMIsAConstantFP(p)) {
6739 double a = LLVMConstRealGetDouble(p, &loses_info);
6740
6741 if (a == 0)
6742 is_zero[i] = true;
6743 else if (a == 1)
6744 is_one[i] = true;
6745 else
6746 return false; /* other constant */
6747 } else
6748 return false;
6749 }
6750
6751 /* Only certain combinations of 0 and 1 can be eliminated. */
6752 if (is_zero[0] && is_zero[1] && is_zero[2])
6753 default_val = is_zero[3] ? 0 : 1;
6754 else if (is_one[0] && is_one[1] && is_one[2])
6755 default_val = is_zero[3] ? 2 : 3;
6756 else
6757 return false;
6758
6759 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
6760 LLVMInstructionEraseFromParent(inst);
6761
6762 /* Change OFFSET to DEFAULT_VAL. */
6763 for (i = 0; i < num_outputs; i++) {
6764 if (shader->info.vs_output_param_offset[i] == offset) {
6765 shader->info.vs_output_param_offset[i] =
6766 EXP_PARAM_DEFAULT_VAL_0000 + default_val;
6767 break;
6768 }
6769 }
6770 return true;
6771 }
6772
6773 struct si_vs_exports {
6774 unsigned num;
6775 unsigned offset[SI_MAX_VS_OUTPUTS];
6776 LLVMValueRef inst[SI_MAX_VS_OUTPUTS];
6777 };
6778
6779 static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx)
6780 {
6781 struct si_shader *shader = ctx->shader;
6782 struct tgsi_shader_info *info = &shader->selector->info;
6783 LLVMBasicBlockRef bb;
6784 struct si_vs_exports exports;
6785 bool removed_any = false;
6786
6787 exports.num = 0;
6788
6789 if (ctx->type == PIPE_SHADER_FRAGMENT ||
6790 ctx->type == PIPE_SHADER_COMPUTE ||
6791 shader->key.as_es ||
6792 shader->key.as_ls)
6793 return;
6794
6795 /* Process all LLVM instructions. */
6796 bb = LLVMGetFirstBasicBlock(ctx->main_fn);
6797 while (bb) {
6798 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
6799
6800 while (inst) {
6801 LLVMValueRef cur = inst;
6802 inst = LLVMGetNextInstruction(inst);
6803
6804 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
6805 continue;
6806
6807 LLVMValueRef callee = lp_get_called_value(cur);
6808
6809 if (!lp_is_function(callee))
6810 continue;
6811
6812 const char *name = LLVMGetValueName(callee);
6813 unsigned num_args = LLVMCountParams(callee);
6814
6815 /* Check if this is an export instruction. */
6816 if (num_args != 9 || strcmp(name, "llvm.SI.export"))
6817 continue;
6818
6819 LLVMValueRef arg = LLVMGetOperand(cur, 3);
6820 unsigned target = LLVMConstIntGetZExtValue(arg);
6821
6822 if (target < V_008DFC_SQ_EXP_PARAM)
6823 continue;
6824
6825 target -= V_008DFC_SQ_EXP_PARAM;
6826
6827 /* Eliminate constant value PARAM exports. */
6828 if (si_eliminate_const_output(ctx, cur, target)) {
6829 removed_any = true;
6830 } else {
6831 exports.offset[exports.num] = target;
6832 exports.inst[exports.num] = cur;
6833 exports.num++;
6834 }
6835 }
6836 bb = LLVMGetNextBasicBlock(bb);
6837 }
6838
6839 /* Remove holes in export memory due to removed PARAM exports.
6840 * This is done by renumbering all PARAM exports.
6841 */
6842 if (removed_any) {
6843 ubyte current_offset[SI_MAX_VS_OUTPUTS];
6844 unsigned new_count = 0;
6845 unsigned out, i;
6846
6847 /* Make a copy of the offsets. We need the old version while
6848 * we are modifying some of them. */
6849 assert(sizeof(current_offset) ==
6850 sizeof(shader->info.vs_output_param_offset));
6851 memcpy(current_offset, shader->info.vs_output_param_offset,
6852 sizeof(current_offset));
6853
6854 for (i = 0; i < exports.num; i++) {
6855 unsigned offset = exports.offset[i];
6856
6857 for (out = 0; out < info->num_outputs; out++) {
6858 if (current_offset[out] != offset)
6859 continue;
6860
6861 LLVMSetOperand(exports.inst[i], 3,
6862 LLVMConstInt(ctx->i32,
6863 V_008DFC_SQ_EXP_PARAM + new_count, 0));
6864 shader->info.vs_output_param_offset[out] = new_count;
6865 new_count++;
6866 break;
6867 }
6868 }
6869 shader->info.nr_param_exports = new_count;
6870 }
6871 }
6872
6873 static void si_count_scratch_private_memory(struct si_shader_context *ctx)
6874 {
6875 ctx->shader->config.private_mem_vgprs = 0;
6876
6877 /* Process all LLVM instructions. */
6878 LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(ctx->main_fn);
6879 while (bb) {
6880 LLVMValueRef next = LLVMGetFirstInstruction(bb);
6881
6882 while (next) {
6883 LLVMValueRef inst = next;
6884 next = LLVMGetNextInstruction(next);
6885
6886 if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
6887 continue;
6888
6889 LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
6890 /* No idea why LLVM aligns allocas to 4 elements. */
6891 unsigned alignment = LLVMGetAlignment(inst);
6892 unsigned dw_size = align(llvm_get_type_size(type) / 4, alignment);
6893 ctx->shader->config.private_mem_vgprs += dw_size;
6894 }
6895 bb = LLVMGetNextBasicBlock(bb);
6896 }
6897 }
6898
6899 static bool si_compile_tgsi_main(struct si_shader_context *ctx,
6900 struct si_shader *shader)
6901 {
6902 struct si_shader_selector *sel = shader->selector;
6903 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
6904
6905 switch (ctx->type) {
6906 case PIPE_SHADER_VERTEX:
6907 ctx->load_input = declare_input_vs;
6908 if (shader->key.as_ls)
6909 bld_base->emit_epilogue = si_llvm_emit_ls_epilogue;
6910 else if (shader->key.as_es)
6911 bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
6912 else
6913 bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
6914 break;
6915 case PIPE_SHADER_TESS_CTRL:
6916 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tcs;
6917 bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = fetch_output_tcs;
6918 bld_base->emit_store = store_output_tcs;
6919 bld_base->emit_epilogue = si_llvm_emit_tcs_epilogue;
6920 break;
6921 case PIPE_SHADER_TESS_EVAL:
6922 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tes;
6923 if (shader->key.as_es)
6924 bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
6925 else
6926 bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
6927 break;
6928 case PIPE_SHADER_GEOMETRY:
6929 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_gs;
6930 bld_base->emit_epilogue = si_llvm_emit_gs_epilogue;
6931 break;
6932 case PIPE_SHADER_FRAGMENT:
6933 ctx->load_input = declare_input_fs;
6934 bld_base->emit_epilogue = si_llvm_return_fs_outputs;
6935 break;
6936 case PIPE_SHADER_COMPUTE:
6937 ctx->declare_memory_region = declare_compute_memory;
6938 break;
6939 default:
6940 assert(!"Unsupported shader type");
6941 return false;
6942 }
6943
6944 create_meta_data(ctx);
6945 create_function(ctx);
6946 preload_ring_buffers(ctx);
6947
6948 if (ctx->type == PIPE_SHADER_GEOMETRY) {
6949 int i;
6950 for (i = 0; i < 4; i++) {
6951 ctx->gs_next_vertex[i] =
6952 lp_build_alloca(bld_base->base.gallivm,
6953 ctx->i32, "");
6954 }
6955 }
6956
6957 if (!lp_build_tgsi_llvm(bld_base, sel->tokens)) {
6958 fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n");
6959 return false;
6960 }
6961
6962 si_llvm_build_ret(ctx, ctx->return_value);
6963 return true;
6964 }
6965
6966 /**
6967 * Compute the VS prolog key, which contains all the information needed to
6968 * build the VS prolog function, and set shader->info bits where needed.
6969 */
6970 static void si_get_vs_prolog_key(struct si_shader *shader,
6971 union si_shader_part_key *key)
6972 {
6973 struct tgsi_shader_info *info = &shader->selector->info;
6974
6975 memset(key, 0, sizeof(*key));
6976 key->vs_prolog.states = shader->key.part.vs.prolog;
6977 key->vs_prolog.num_input_sgprs = shader->info.num_input_sgprs;
6978 key->vs_prolog.last_input = MAX2(1, info->num_inputs) - 1;
6979
6980 /* Set the instanceID flag. */
6981 for (unsigned i = 0; i < info->num_inputs; i++)
6982 if (key->vs_prolog.states.instance_divisors[i])
6983 shader->info.uses_instanceid = true;
6984 }
6985
6986 /**
6987 * Compute the VS epilog key, which contains all the information needed to
6988 * build the VS epilog function, and set the PrimitiveID output offset.
6989 */
6990 static void si_get_vs_epilog_key(struct si_shader *shader,
6991 struct si_vs_epilog_bits *states,
6992 union si_shader_part_key *key)
6993 {
6994 memset(key, 0, sizeof(*key));
6995 key->vs_epilog.states = *states;
6996
6997 /* Set up the PrimitiveID output. */
6998 if (shader->key.part.vs.epilog.export_prim_id) {
6999 unsigned index = shader->selector->info.num_outputs;
7000 unsigned offset = shader->info.nr_param_exports++;
7001
7002 key->vs_epilog.prim_id_param_offset = offset;
7003 assert(index < ARRAY_SIZE(shader->info.vs_output_param_offset));
7004 shader->info.vs_output_param_offset[index] = offset;
7005 }
7006 }
7007
7008 /**
7009 * Compute the PS prolog key, which contains all the information needed to
7010 * build the PS prolog function, and set related bits in shader->config.
7011 */
7012 static void si_get_ps_prolog_key(struct si_shader *shader,
7013 union si_shader_part_key *key,
7014 bool separate_prolog)
7015 {
7016 struct tgsi_shader_info *info = &shader->selector->info;
7017
7018 memset(key, 0, sizeof(*key));
7019 key->ps_prolog.states = shader->key.part.ps.prolog;
7020 key->ps_prolog.colors_read = info->colors_read;
7021 key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
7022 key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
7023 key->ps_prolog.wqm = info->uses_derivatives &&
7024 (key->ps_prolog.colors_read ||
7025 key->ps_prolog.states.force_persp_sample_interp ||
7026 key->ps_prolog.states.force_linear_sample_interp ||
7027 key->ps_prolog.states.force_persp_center_interp ||
7028 key->ps_prolog.states.force_linear_center_interp ||
7029 key->ps_prolog.states.bc_optimize_for_persp ||
7030 key->ps_prolog.states.bc_optimize_for_linear);
7031
7032 if (info->colors_read) {
7033 unsigned *color = shader->selector->color_attr_index;
7034
7035 if (shader->key.part.ps.prolog.color_two_side) {
7036 /* BCOLORs are stored after the last input. */
7037 key->ps_prolog.num_interp_inputs = info->num_inputs;
7038 key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
7039 shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
7040 }
7041
7042 for (unsigned i = 0; i < 2; i++) {
7043 unsigned interp = info->input_interpolate[color[i]];
7044 unsigned location = info->input_interpolate_loc[color[i]];
7045
7046 if (!(info->colors_read & (0xf << i*4)))
7047 continue;
7048
7049 key->ps_prolog.color_attr_index[i] = color[i];
7050
7051 if (shader->key.part.ps.prolog.flatshade_colors &&
7052 interp == TGSI_INTERPOLATE_COLOR)
7053 interp = TGSI_INTERPOLATE_CONSTANT;
7054
7055 switch (interp) {
7056 case TGSI_INTERPOLATE_CONSTANT:
7057 key->ps_prolog.color_interp_vgpr_index[i] = -1;
7058 break;
7059 case TGSI_INTERPOLATE_PERSPECTIVE:
7060 case TGSI_INTERPOLATE_COLOR:
7061 /* Force the interpolation location for colors here. */
7062 if (shader->key.part.ps.prolog.force_persp_sample_interp)
7063 location = TGSI_INTERPOLATE_LOC_SAMPLE;
7064 if (shader->key.part.ps.prolog.force_persp_center_interp)
7065 location = TGSI_INTERPOLATE_LOC_CENTER;
7066
7067 switch (location) {
7068 case TGSI_INTERPOLATE_LOC_SAMPLE:
7069 key->ps_prolog.color_interp_vgpr_index[i] = 0;
7070 shader->config.spi_ps_input_ena |=
7071 S_0286CC_PERSP_SAMPLE_ENA(1);
7072 break;
7073 case TGSI_INTERPOLATE_LOC_CENTER:
7074 key->ps_prolog.color_interp_vgpr_index[i] = 2;
7075 shader->config.spi_ps_input_ena |=
7076 S_0286CC_PERSP_CENTER_ENA(1);
7077 break;
7078 case TGSI_INTERPOLATE_LOC_CENTROID:
7079 key->ps_prolog.color_interp_vgpr_index[i] = 4;
7080 shader->config.spi_ps_input_ena |=
7081 S_0286CC_PERSP_CENTROID_ENA(1);
7082 break;
7083 default:
7084 assert(0);
7085 }
7086 break;
7087 case TGSI_INTERPOLATE_LINEAR:
7088 /* Force the interpolation location for colors here. */
7089 if (shader->key.part.ps.prolog.force_linear_sample_interp)
7090 location = TGSI_INTERPOLATE_LOC_SAMPLE;
7091 if (shader->key.part.ps.prolog.force_linear_center_interp)
7092 location = TGSI_INTERPOLATE_LOC_CENTER;
7093
7094 /* The VGPR assignment for non-monolithic shaders
7095 * works because InitialPSInputAddr is set on the
7096 * main shader and PERSP_PULL_MODEL is never used.
7097 */
7098 switch (location) {
7099 case TGSI_INTERPOLATE_LOC_SAMPLE:
7100 key->ps_prolog.color_interp_vgpr_index[i] =
7101 separate_prolog ? 6 : 9;
7102 shader->config.spi_ps_input_ena |=
7103 S_0286CC_LINEAR_SAMPLE_ENA(1);
7104 break;
7105 case TGSI_INTERPOLATE_LOC_CENTER:
7106 key->ps_prolog.color_interp_vgpr_index[i] =
7107 separate_prolog ? 8 : 11;
7108 shader->config.spi_ps_input_ena |=
7109 S_0286CC_LINEAR_CENTER_ENA(1);
7110 break;
7111 case TGSI_INTERPOLATE_LOC_CENTROID:
7112 key->ps_prolog.color_interp_vgpr_index[i] =
7113 separate_prolog ? 10 : 13;
7114 shader->config.spi_ps_input_ena |=
7115 S_0286CC_LINEAR_CENTROID_ENA(1);
7116 break;
7117 default:
7118 assert(0);
7119 }
7120 break;
7121 default:
7122 assert(0);
7123 }
7124 }
7125 }
7126 }
7127
7128 /**
7129 * Check whether a PS prolog is required based on the key.
7130 */
7131 static bool si_need_ps_prolog(const union si_shader_part_key *key)
7132 {
7133 return key->ps_prolog.colors_read ||
7134 key->ps_prolog.states.force_persp_sample_interp ||
7135 key->ps_prolog.states.force_linear_sample_interp ||
7136 key->ps_prolog.states.force_persp_center_interp ||
7137 key->ps_prolog.states.force_linear_center_interp ||
7138 key->ps_prolog.states.bc_optimize_for_persp ||
7139 key->ps_prolog.states.bc_optimize_for_linear ||
7140 key->ps_prolog.states.poly_stipple;
7141 }
7142
7143 /**
7144 * Compute the PS epilog key, which contains all the information needed to
7145 * build the PS epilog function.
7146 */
7147 static void si_get_ps_epilog_key(struct si_shader *shader,
7148 union si_shader_part_key *key)
7149 {
7150 struct tgsi_shader_info *info = &shader->selector->info;
7151 memset(key, 0, sizeof(*key));
7152 key->ps_epilog.colors_written = info->colors_written;
7153 key->ps_epilog.writes_z = info->writes_z;
7154 key->ps_epilog.writes_stencil = info->writes_stencil;
7155 key->ps_epilog.writes_samplemask = info->writes_samplemask;
7156 key->ps_epilog.states = shader->key.part.ps.epilog;
7157 }
7158
7159 /**
7160 * Build the GS prolog function. Rotate the input vertices for triangle strips
7161 * with adjacency.
7162 */
7163 static void si_build_gs_prolog_function(struct si_shader_context *ctx,
7164 union si_shader_part_key *key)
7165 {
7166 const unsigned num_sgprs = SI_GS_NUM_USER_SGPR + 2;
7167 const unsigned num_vgprs = 8;
7168 struct gallivm_state *gallivm = &ctx->gallivm;
7169 LLVMBuilderRef builder = gallivm->builder;
7170 LLVMTypeRef params[32];
7171 LLVMTypeRef returns[32];
7172 LLVMValueRef func, ret;
7173
7174 for (unsigned i = 0; i < num_sgprs; ++i) {
7175 params[i] = ctx->i32;
7176 returns[i] = ctx->i32;
7177 }
7178
7179 for (unsigned i = 0; i < num_vgprs; ++i) {
7180 params[num_sgprs + i] = ctx->i32;
7181 returns[num_sgprs + i] = ctx->f32;
7182 }
7183
7184 /* Create the function. */
7185 si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs,
7186 params, num_sgprs + num_vgprs, num_sgprs - 1);
7187 func = ctx->main_fn;
7188
7189 /* Copy inputs to outputs. This should be no-op, as the registers match,
7190 * but it will prevent the compiler from overwriting them unintentionally.
7191 */
7192 ret = ctx->return_value;
7193 for (unsigned i = 0; i < num_sgprs; i++) {
7194 LLVMValueRef p = LLVMGetParam(func, i);
7195 ret = LLVMBuildInsertValue(builder, ret, p, i, "");
7196 }
7197 for (unsigned i = 0; i < num_vgprs; i++) {
7198 LLVMValueRef p = LLVMGetParam(func, num_sgprs + i);
7199 p = LLVMBuildBitCast(builder, p, ctx->f32, "");
7200 ret = LLVMBuildInsertValue(builder, ret, p, num_sgprs + i, "");
7201 }
7202
7203 if (key->gs_prolog.states.tri_strip_adj_fix) {
7204 /* Remap the input vertices for every other primitive. */
7205 const unsigned vtx_params[6] = {
7206 num_sgprs,
7207 num_sgprs + 1,
7208 num_sgprs + 3,
7209 num_sgprs + 4,
7210 num_sgprs + 5,
7211 num_sgprs + 6
7212 };
7213 LLVMValueRef prim_id, rotate;
7214
7215 prim_id = LLVMGetParam(func, num_sgprs + 2);
7216 rotate = LLVMBuildTrunc(builder, prim_id, ctx->i1, "");
7217
7218 for (unsigned i = 0; i < 6; ++i) {
7219 LLVMValueRef base, rotated, actual;
7220 base = LLVMGetParam(func, vtx_params[i]);
7221 rotated = LLVMGetParam(func, vtx_params[(i + 4) % 6]);
7222 actual = LLVMBuildSelect(builder, rotate, rotated, base, "");
7223 actual = LLVMBuildBitCast(builder, actual, ctx->f32, "");
7224 ret = LLVMBuildInsertValue(builder, ret, actual, vtx_params[i], "");
7225 }
7226 }
7227
7228 LLVMBuildRet(builder, ret);
7229 }
7230
7231 /**
7232 * Given a list of shader part functions, build a wrapper function that
7233 * runs them in sequence to form a monolithic shader.
7234 */
7235 static void si_build_wrapper_function(struct si_shader_context *ctx,
7236 LLVMValueRef *parts,
7237 unsigned num_parts,
7238 unsigned main_part)
7239 {
7240 struct gallivm_state *gallivm = &ctx->gallivm;
7241 LLVMBuilderRef builder = ctx->gallivm.builder;
7242 /* PS epilog has one arg per color component */
7243 LLVMTypeRef param_types[48];
7244 LLVMValueRef out[48];
7245 LLVMTypeRef function_type;
7246 unsigned num_params;
7247 unsigned num_out;
7248 MAYBE_UNUSED unsigned num_out_sgpr; /* used in debug checks */
7249 unsigned num_sgprs, num_vgprs;
7250 unsigned last_sgpr_param;
7251 unsigned gprs;
7252
7253 for (unsigned i = 0; i < num_parts; ++i) {
7254 lp_add_function_attr(parts[i], -1, LP_FUNC_ATTR_ALWAYSINLINE);
7255 LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
7256 }
7257
7258 /* The parameters of the wrapper function correspond to those of the
7259 * first part in terms of SGPRs and VGPRs, but we use the types of the
7260 * main part to get the right types. This is relevant for the
7261 * dereferenceable attribute on descriptor table pointers.
7262 */
7263 num_sgprs = 0;
7264 num_vgprs = 0;
7265
7266 function_type = LLVMGetElementType(LLVMTypeOf(parts[0]));
7267 num_params = LLVMCountParamTypes(function_type);
7268
7269 for (unsigned i = 0; i < num_params; ++i) {
7270 LLVMValueRef param = LLVMGetParam(parts[0], i);
7271
7272 if (ac_is_sgpr_param(param)) {
7273 assert(num_vgprs == 0);
7274 num_sgprs += llvm_get_type_size(LLVMTypeOf(param)) / 4;
7275 } else {
7276 num_vgprs += llvm_get_type_size(LLVMTypeOf(param)) / 4;
7277 }
7278 }
7279 assert(num_vgprs + num_sgprs <= ARRAY_SIZE(param_types));
7280
7281 num_params = 0;
7282 last_sgpr_param = 0;
7283 gprs = 0;
7284 while (gprs < num_sgprs + num_vgprs) {
7285 LLVMValueRef param = LLVMGetParam(parts[main_part], num_params);
7286 unsigned size;
7287
7288 param_types[num_params] = LLVMTypeOf(param);
7289 if (gprs < num_sgprs)
7290 last_sgpr_param = num_params;
7291 size = llvm_get_type_size(param_types[num_params]) / 4;
7292 num_params++;
7293
7294 assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
7295 assert(gprs + size <= num_sgprs + num_vgprs &&
7296 (gprs >= num_sgprs || gprs + size <= num_sgprs));
7297
7298 gprs += size;
7299 }
7300
7301 si_create_function(ctx, "wrapper", NULL, 0, param_types, num_params, last_sgpr_param);
7302
7303 /* Record the arguments of the function as if they were an output of
7304 * a previous part.
7305 */
7306 num_out = 0;
7307 num_out_sgpr = 0;
7308
7309 for (unsigned i = 0; i < num_params; ++i) {
7310 LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
7311 LLVMTypeRef param_type = LLVMTypeOf(param);
7312 LLVMTypeRef out_type = i <= last_sgpr_param ? ctx->i32 : ctx->f32;
7313 unsigned size = llvm_get_type_size(param_type) / 4;
7314
7315 if (size == 1) {
7316 if (param_type != out_type)
7317 param = LLVMBuildBitCast(builder, param, out_type, "");
7318 out[num_out++] = param;
7319 } else {
7320 LLVMTypeRef vector_type = LLVMVectorType(out_type, size);
7321
7322 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
7323 param = LLVMBuildPtrToInt(builder, param, ctx->i64, "");
7324 param_type = ctx->i64;
7325 }
7326
7327 if (param_type != vector_type)
7328 param = LLVMBuildBitCast(builder, param, vector_type, "");
7329
7330 for (unsigned j = 0; j < size; ++j)
7331 out[num_out++] = LLVMBuildExtractElement(
7332 builder, param, LLVMConstInt(ctx->i32, j, 0), "");
7333 }
7334
7335 if (i <= last_sgpr_param)
7336 num_out_sgpr = num_out;
7337 }
7338
7339 /* Now chain the parts. */
7340 for (unsigned part = 0; part < num_parts; ++part) {
7341 LLVMValueRef in[48];
7342 LLVMValueRef ret;
7343 LLVMTypeRef ret_type;
7344 unsigned out_idx = 0;
7345
7346 num_params = LLVMCountParams(parts[part]);
7347 assert(num_params <= ARRAY_SIZE(param_types));
7348
7349 /* Derive arguments for the next part from outputs of the
7350 * previous one.
7351 */
7352 for (unsigned param_idx = 0; param_idx < num_params; ++param_idx) {
7353 LLVMValueRef param;
7354 LLVMTypeRef param_type;
7355 bool is_sgpr;
7356 unsigned param_size;
7357 LLVMValueRef arg = NULL;
7358
7359 param = LLVMGetParam(parts[part], param_idx);
7360 param_type = LLVMTypeOf(param);
7361 param_size = llvm_get_type_size(param_type) / 4;
7362 is_sgpr = ac_is_sgpr_param(param);
7363
7364 if (is_sgpr) {
7365 #if HAVE_LLVM < 0x0400
7366 LLVMRemoveAttribute(param, LLVMByValAttribute);
7367 #else
7368 unsigned kind_id = LLVMGetEnumAttributeKindForName("byval", 5);
7369 LLVMRemoveEnumAttributeAtIndex(parts[part], param_idx + 1, kind_id);
7370 #endif
7371 lp_add_function_attr(parts[part], param_idx + 1, LP_FUNC_ATTR_INREG);
7372 }
7373
7374 assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));
7375 assert(is_sgpr || out_idx >= num_out_sgpr);
7376
7377 if (param_size == 1)
7378 arg = out[out_idx];
7379 else
7380 arg = lp_build_gather_values(gallivm, &out[out_idx], param_size);
7381
7382 if (LLVMTypeOf(arg) != param_type) {
7383 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
7384 arg = LLVMBuildBitCast(builder, arg, ctx->i64, "");
7385 arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
7386 } else {
7387 arg = LLVMBuildBitCast(builder, arg, param_type, "");
7388 }
7389 }
7390
7391 in[param_idx] = arg;
7392 out_idx += param_size;
7393 }
7394
7395 ret = LLVMBuildCall(builder, parts[part], in, num_params, "");
7396 ret_type = LLVMTypeOf(ret);
7397
7398 /* Extract the returned GPRs. */
7399 num_out = 0;
7400 num_out_sgpr = 0;
7401
7402 if (LLVMGetTypeKind(ret_type) != LLVMVoidTypeKind) {
7403 assert(LLVMGetTypeKind(ret_type) == LLVMStructTypeKind);
7404
7405 unsigned ret_size = LLVMCountStructElementTypes(ret_type);
7406
7407 for (unsigned i = 0; i < ret_size; ++i) {
7408 LLVMValueRef val =
7409 LLVMBuildExtractValue(builder, ret, i, "");
7410
7411 out[num_out++] = val;
7412
7413 if (LLVMTypeOf(val) == ctx->i32) {
7414 assert(num_out_sgpr + 1 == num_out);
7415 num_out_sgpr = num_out;
7416 }
7417 }
7418 }
7419 }
7420
7421 LLVMBuildRetVoid(builder);
7422 }
7423
7424 int si_compile_tgsi_shader(struct si_screen *sscreen,
7425 LLVMTargetMachineRef tm,
7426 struct si_shader *shader,
7427 bool is_monolithic,
7428 struct pipe_debug_callback *debug)
7429 {
7430 struct si_shader_selector *sel = shader->selector;
7431 struct si_shader_context ctx;
7432 struct lp_build_tgsi_context *bld_base;
7433 LLVMModuleRef mod;
7434 int r = -1;
7435
7436 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
7437 * conversion fails. */
7438 if (r600_can_dump_shader(&sscreen->b, sel->info.processor) &&
7439 !(sscreen->b.debug_flags & DBG_NO_TGSI)) {
7440 tgsi_dump(sel->tokens, 0);
7441 si_dump_streamout(&sel->so);
7442 }
7443
7444 si_init_shader_ctx(&ctx, sscreen, shader, tm);
7445 ctx.separate_prolog = !is_monolithic;
7446
7447 memset(shader->info.vs_output_param_offset, EXP_PARAM_UNDEFINED,
7448 sizeof(shader->info.vs_output_param_offset));
7449
7450 shader->info.uses_instanceid = sel->info.uses_instanceid;
7451
7452 bld_base = &ctx.soa.bld_base;
7453 ctx.load_system_value = declare_system_value;
7454
7455 if (!si_compile_tgsi_main(&ctx, shader)) {
7456 si_llvm_dispose(&ctx);
7457 return -1;
7458 }
7459
7460 if (is_monolithic && ctx.type == PIPE_SHADER_VERTEX) {
7461 LLVMValueRef parts[3];
7462 bool need_prolog;
7463 bool need_epilog;
7464
7465 need_prolog = sel->info.num_inputs;
7466 need_epilog = !shader->key.as_es && !shader->key.as_ls;
7467
7468 parts[need_prolog ? 1 : 0] = ctx.main_fn;
7469
7470 if (need_prolog) {
7471 union si_shader_part_key prolog_key;
7472 si_get_vs_prolog_key(shader, &prolog_key);
7473 si_build_vs_prolog_function(&ctx, &prolog_key);
7474 parts[0] = ctx.main_fn;
7475 }
7476
7477 if (need_epilog) {
7478 union si_shader_part_key epilog_key;
7479 si_get_vs_epilog_key(shader, &shader->key.part.vs.epilog, &epilog_key);
7480 si_build_vs_epilog_function(&ctx, &epilog_key);
7481 parts[need_prolog ? 2 : 1] = ctx.main_fn;
7482 }
7483
7484 si_build_wrapper_function(&ctx, parts, 1 + need_prolog + need_epilog,
7485 need_prolog ? 1 : 0);
7486 } else if (is_monolithic && ctx.type == PIPE_SHADER_TESS_CTRL) {
7487 LLVMValueRef parts[2];
7488 union si_shader_part_key epilog_key;
7489
7490 parts[0] = ctx.main_fn;
7491
7492 memset(&epilog_key, 0, sizeof(epilog_key));
7493 epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
7494 si_build_tcs_epilog_function(&ctx, &epilog_key);
7495 parts[1] = ctx.main_fn;
7496
7497 si_build_wrapper_function(&ctx, parts, 2, 0);
7498 } else if (is_monolithic && ctx.type == PIPE_SHADER_TESS_EVAL &&
7499 !shader->key.as_es) {
7500 LLVMValueRef parts[2];
7501 union si_shader_part_key epilog_key;
7502
7503 parts[0] = ctx.main_fn;
7504
7505 si_get_vs_epilog_key(shader, &shader->key.part.tes.epilog, &epilog_key);
7506 si_build_vs_epilog_function(&ctx, &epilog_key);
7507 parts[1] = ctx.main_fn;
7508
7509 si_build_wrapper_function(&ctx, parts, 2, 0);
7510 } else if (is_monolithic && ctx.type == PIPE_SHADER_GEOMETRY) {
7511 LLVMValueRef parts[2];
7512 union si_shader_part_key prolog_key;
7513
7514 parts[1] = ctx.main_fn;
7515
7516 memset(&prolog_key, 0, sizeof(prolog_key));
7517 prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
7518 si_build_gs_prolog_function(&ctx, &prolog_key);
7519 parts[0] = ctx.main_fn;
7520
7521 si_build_wrapper_function(&ctx, parts, 2, 1);
7522 } else if (is_monolithic && ctx.type == PIPE_SHADER_FRAGMENT) {
7523 LLVMValueRef parts[3];
7524 union si_shader_part_key prolog_key;
7525 union si_shader_part_key epilog_key;
7526 bool need_prolog;
7527
7528 si_get_ps_prolog_key(shader, &prolog_key, false);
7529 need_prolog = si_need_ps_prolog(&prolog_key);
7530
7531 parts[need_prolog ? 1 : 0] = ctx.main_fn;
7532
7533 if (need_prolog) {
7534 si_build_ps_prolog_function(&ctx, &prolog_key);
7535 parts[0] = ctx.main_fn;
7536 }
7537
7538 si_get_ps_epilog_key(shader, &epilog_key);
7539 si_build_ps_epilog_function(&ctx, &epilog_key);
7540 parts[need_prolog ? 2 : 1] = ctx.main_fn;
7541
7542 si_build_wrapper_function(&ctx, parts, need_prolog ? 3 : 2, need_prolog ? 1 : 0);
7543 }
7544
7545 mod = bld_base->base.gallivm->module;
7546
7547 /* Dump LLVM IR before any optimization passes */
7548 if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
7549 r600_can_dump_shader(&sscreen->b, ctx.type))
7550 LLVMDumpModule(mod);
7551
7552 si_llvm_finalize_module(&ctx,
7553 r600_extra_shader_checks(&sscreen->b, ctx.type));
7554
7555 /* Post-optimization transformations and analysis. */
7556 si_eliminate_const_vs_outputs(&ctx);
7557
7558 if ((debug && debug->debug_message) ||
7559 r600_can_dump_shader(&sscreen->b, ctx.type))
7560 si_count_scratch_private_memory(&ctx);
7561
7562 /* Compile to bytecode. */
7563 r = si_compile_llvm(sscreen, &shader->binary, &shader->config, tm,
7564 mod, debug, ctx.type, "TGSI shader");
7565 si_llvm_dispose(&ctx);
7566 if (r) {
7567 fprintf(stderr, "LLVM failed to compile shader\n");
7568 return r;
7569 }
7570
7571 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
7572 * LLVM 3.9svn has this bug.
7573 */
7574 if (sel->type == PIPE_SHADER_COMPUTE) {
7575 unsigned wave_size = 64;
7576 unsigned max_vgprs = 256;
7577 unsigned max_sgprs = sscreen->b.chip_class >= VI ? 800 : 512;
7578 unsigned max_sgprs_per_wave = 128;
7579 unsigned max_block_threads = si_get_max_workgroup_size(shader);
7580 unsigned min_waves_per_cu = DIV_ROUND_UP(max_block_threads, wave_size);
7581 unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);
7582
7583 max_vgprs = max_vgprs / min_waves_per_simd;
7584 max_sgprs = MIN2(max_sgprs / min_waves_per_simd, max_sgprs_per_wave);
7585
7586 if (shader->config.num_sgprs > max_sgprs ||
7587 shader->config.num_vgprs > max_vgprs) {
7588 fprintf(stderr, "LLVM failed to compile a shader correctly: "
7589 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
7590 shader->config.num_sgprs, shader->config.num_vgprs,
7591 max_sgprs, max_vgprs);
7592
7593 /* Just terminate the process, because dependent
7594 * shaders can hang due to bad input data, but use
7595 * the env var to allow shader-db to work.
7596 */
7597 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
7598 abort();
7599 }
7600 }
7601
7602 /* Add the scratch offset to input SGPRs. */
7603 if (shader->config.scratch_bytes_per_wave)
7604 shader->info.num_input_sgprs += 1; /* scratch byte offset */
7605
7606 /* Calculate the number of fragment input VGPRs. */
7607 if (ctx.type == PIPE_SHADER_FRAGMENT) {
7608 shader->info.num_input_vgprs = 0;
7609 shader->info.face_vgpr_index = -1;
7610
7611 if (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_addr))
7612 shader->info.num_input_vgprs += 2;
7613 if (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr))
7614 shader->info.num_input_vgprs += 2;
7615 if (G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_addr))
7616 shader->info.num_input_vgprs += 2;
7617 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader->config.spi_ps_input_addr))
7618 shader->info.num_input_vgprs += 3;
7619 if (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_addr))
7620 shader->info.num_input_vgprs += 2;
7621 if (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr))
7622 shader->info.num_input_vgprs += 2;
7623 if (G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_addr))
7624 shader->info.num_input_vgprs += 2;
7625 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader->config.spi_ps_input_addr))
7626 shader->info.num_input_vgprs += 1;
7627 if (G_0286CC_POS_X_FLOAT_ENA(shader->config.spi_ps_input_addr))
7628 shader->info.num_input_vgprs += 1;
7629 if (G_0286CC_POS_Y_FLOAT_ENA(shader->config.spi_ps_input_addr))
7630 shader->info.num_input_vgprs += 1;
7631 if (G_0286CC_POS_Z_FLOAT_ENA(shader->config.spi_ps_input_addr))
7632 shader->info.num_input_vgprs += 1;
7633 if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_addr))
7634 shader->info.num_input_vgprs += 1;
7635 if (G_0286CC_FRONT_FACE_ENA(shader->config.spi_ps_input_addr)) {
7636 shader->info.face_vgpr_index = shader->info.num_input_vgprs;
7637 shader->info.num_input_vgprs += 1;
7638 }
7639 if (G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr))
7640 shader->info.num_input_vgprs += 1;
7641 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader->config.spi_ps_input_addr))
7642 shader->info.num_input_vgprs += 1;
7643 if (G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr))
7644 shader->info.num_input_vgprs += 1;
7645 }
7646
7647 return 0;
7648 }
7649
7650 /**
7651 * Create, compile and return a shader part (prolog or epilog).
7652 *
7653 * \param sscreen screen
7654 * \param list list of shader parts of the same category
7655 * \param type shader type
7656 * \param key shader part key
7657 * \param prolog whether the part being requested is a prolog
7658 * \param tm LLVM target machine
7659 * \param debug debug callback
7660 * \param build the callback responsible for building the main function
7661 * \return non-NULL on success
7662 */
7663 static struct si_shader_part *
7664 si_get_shader_part(struct si_screen *sscreen,
7665 struct si_shader_part **list,
7666 enum pipe_shader_type type,
7667 bool prolog,
7668 union si_shader_part_key *key,
7669 LLVMTargetMachineRef tm,
7670 struct pipe_debug_callback *debug,
7671 void (*build)(struct si_shader_context *,
7672 union si_shader_part_key *),
7673 const char *name)
7674 {
7675 struct si_shader_part *result;
7676
7677 pipe_mutex_lock(sscreen->shader_parts_mutex);
7678
7679 /* Find existing. */
7680 for (result = *list; result; result = result->next) {
7681 if (memcmp(&result->key, key, sizeof(*key)) == 0) {
7682 pipe_mutex_unlock(sscreen->shader_parts_mutex);
7683 return result;
7684 }
7685 }
7686
7687 /* Compile a new one. */
7688 result = CALLOC_STRUCT(si_shader_part);
7689 result->key = *key;
7690
7691 struct si_shader shader = {};
7692 struct si_shader_context ctx;
7693 struct gallivm_state *gallivm = &ctx.gallivm;
7694
7695 si_init_shader_ctx(&ctx, sscreen, &shader, tm);
7696 ctx.type = type;
7697
7698 switch (type) {
7699 case PIPE_SHADER_VERTEX:
7700 break;
7701 case PIPE_SHADER_TESS_CTRL:
7702 assert(!prolog);
7703 shader.key.part.tcs.epilog = key->tcs_epilog.states;
7704 break;
7705 case PIPE_SHADER_GEOMETRY:
7706 assert(prolog);
7707 break;
7708 case PIPE_SHADER_FRAGMENT:
7709 if (prolog)
7710 shader.key.part.ps.prolog = key->ps_prolog.states;
7711 else
7712 shader.key.part.ps.epilog = key->ps_epilog.states;
7713 break;
7714 default:
7715 unreachable("bad shader part");
7716 }
7717
7718 build(&ctx, key);
7719
7720 /* Compile. */
7721 si_llvm_finalize_module(&ctx,
7722 r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
7723
7724 if (si_compile_llvm(sscreen, &result->binary, &result->config, tm,
7725 gallivm->module, debug, ctx.type, name)) {
7726 FREE(result);
7727 result = NULL;
7728 goto out;
7729 }
7730
7731 result->next = *list;
7732 *list = result;
7733
7734 out:
7735 si_llvm_dispose(&ctx);
7736 pipe_mutex_unlock(sscreen->shader_parts_mutex);
7737 return result;
7738 }
7739
7740 /**
7741 * Build the vertex shader prolog function.
7742 *
7743 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7744 * All inputs are returned unmodified. The vertex load indices are
7745 * stored after them, which will be used by the API VS for fetching inputs.
7746 *
7747 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7748 * input_v0,
7749 * input_v1,
7750 * input_v2,
7751 * input_v3,
7752 * (VertexID + BaseVertex),
7753 * (InstanceID + StartInstance),
7754 * (InstanceID / 2 + StartInstance)
7755 */
7756 static void si_build_vs_prolog_function(struct si_shader_context *ctx,
7757 union si_shader_part_key *key)
7758 {
7759 struct gallivm_state *gallivm = &ctx->gallivm;
7760 LLVMTypeRef *params, *returns;
7761 LLVMValueRef ret, func;
7762 int last_sgpr, num_params, num_returns, i;
7763
7764 ctx->param_vertex_id = key->vs_prolog.num_input_sgprs;
7765 ctx->param_instance_id = key->vs_prolog.num_input_sgprs + 3;
7766
7767 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7768 params = alloca((key->vs_prolog.num_input_sgprs + 4) *
7769 sizeof(LLVMTypeRef));
7770 returns = alloca((key->vs_prolog.num_input_sgprs + 4 +
7771 key->vs_prolog.last_input + 1) *
7772 sizeof(LLVMTypeRef));
7773 num_params = 0;
7774 num_returns = 0;
7775
7776 /* Declare input and output SGPRs. */
7777 num_params = 0;
7778 for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
7779 params[num_params++] = ctx->i32;
7780 returns[num_returns++] = ctx->i32;
7781 }
7782 last_sgpr = num_params - 1;
7783
7784 /* 4 preloaded VGPRs (outputs must be floats) */
7785 for (i = 0; i < 4; i++) {
7786 params[num_params++] = ctx->i32;
7787 returns[num_returns++] = ctx->f32;
7788 }
7789
7790 /* Vertex load indices. */
7791 for (i = 0; i <= key->vs_prolog.last_input; i++)
7792 returns[num_returns++] = ctx->f32;
7793
7794 /* Create the function. */
7795 si_create_function(ctx, "vs_prolog", returns, num_returns, params,
7796 num_params, last_sgpr);
7797 func = ctx->main_fn;
7798
7799 /* Copy inputs to outputs. This should be no-op, as the registers match,
7800 * but it will prevent the compiler from overwriting them unintentionally.
7801 */
7802 ret = ctx->return_value;
7803 for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
7804 LLVMValueRef p = LLVMGetParam(func, i);
7805 ret = LLVMBuildInsertValue(gallivm->builder, ret, p, i, "");
7806 }
7807 for (i = num_params - 4; i < num_params; i++) {
7808 LLVMValueRef p = LLVMGetParam(func, i);
7809 p = LLVMBuildBitCast(gallivm->builder, p, ctx->f32, "");
7810 ret = LLVMBuildInsertValue(gallivm->builder, ret, p, i, "");
7811 }
7812
7813 /* Compute vertex load indices from instance divisors. */
7814 for (i = 0; i <= key->vs_prolog.last_input; i++) {
7815 unsigned divisor = key->vs_prolog.states.instance_divisors[i];
7816 LLVMValueRef index;
7817
7818 if (divisor) {
7819 /* InstanceID / Divisor + StartInstance */
7820 index = get_instance_index_for_fetch(ctx,
7821 SI_SGPR_START_INSTANCE,
7822 divisor);
7823 } else {
7824 /* VertexID + BaseVertex */
7825 index = LLVMBuildAdd(gallivm->builder,
7826 LLVMGetParam(func, ctx->param_vertex_id),
7827 LLVMGetParam(func, SI_SGPR_BASE_VERTEX), "");
7828 }
7829
7830 index = LLVMBuildBitCast(gallivm->builder, index, ctx->f32, "");
7831 ret = LLVMBuildInsertValue(gallivm->builder, ret, index,
7832 num_params++, "");
7833 }
7834
7835 si_llvm_build_ret(ctx, ret);
7836 }
7837
7838 /**
7839 * Build the vertex shader epilog function. This is also used by the tessellation
7840 * evaluation shader compiled as VS.
7841 *
7842 * The input is PrimitiveID.
7843 *
7844 * If PrimitiveID is required by the pixel shader, export it.
7845 * Otherwise, do nothing.
7846 */
7847 static void si_build_vs_epilog_function(struct si_shader_context *ctx,
7848 union si_shader_part_key *key)
7849 {
7850 struct gallivm_state *gallivm = &ctx->gallivm;
7851 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
7852 LLVMTypeRef params[5];
7853 int num_params, i;
7854
7855 /* Declare input VGPRs. */
7856 num_params = key->vs_epilog.states.export_prim_id ?
7857 (VS_EPILOG_PRIMID_LOC + 1) : 0;
7858 assert(num_params <= ARRAY_SIZE(params));
7859
7860 for (i = 0; i < num_params; i++)
7861 params[i] = ctx->f32;
7862
7863 /* Create the function. */
7864 si_create_function(ctx, "vs_epilog", NULL, 0, params, num_params, -1);
7865
7866 /* Emit exports. */
7867 if (key->vs_epilog.states.export_prim_id) {
7868 struct lp_build_context *base = &bld_base->base;
7869 struct lp_build_context *uint = &bld_base->uint_bld;
7870 LLVMValueRef args[9];
7871
7872 args[0] = lp_build_const_int32(base->gallivm, 0x0); /* enabled channels */
7873 args[1] = uint->zero; /* whether the EXEC mask is valid */
7874 args[2] = uint->zero; /* DONE bit */
7875 args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_PARAM +
7876 key->vs_epilog.prim_id_param_offset);
7877 args[4] = uint->zero; /* COMPR flag (0 = 32-bit export) */
7878 args[5] = LLVMGetParam(ctx->main_fn,
7879 VS_EPILOG_PRIMID_LOC); /* X */
7880 args[6] = base->undef; /* Y */
7881 args[7] = base->undef; /* Z */
7882 args[8] = base->undef; /* W */
7883
7884 lp_build_intrinsic(base->gallivm->builder, "llvm.SI.export",
7885 LLVMVoidTypeInContext(base->gallivm->context),
7886 args, 9, 0);
7887 }
7888
7889 LLVMBuildRetVoid(gallivm->builder);
7890 }
7891
7892 /**
7893 * Create & compile a vertex shader epilog. This a helper used by VS and TES.
7894 */
7895 static bool si_get_vs_epilog(struct si_screen *sscreen,
7896 LLVMTargetMachineRef tm,
7897 struct si_shader *shader,
7898 struct pipe_debug_callback *debug,
7899 struct si_vs_epilog_bits *states)
7900 {
7901 union si_shader_part_key epilog_key;
7902
7903 si_get_vs_epilog_key(shader, states, &epilog_key);
7904
7905 shader->epilog = si_get_shader_part(sscreen, &sscreen->vs_epilogs,
7906 PIPE_SHADER_VERTEX, true,
7907 &epilog_key, tm, debug,
7908 si_build_vs_epilog_function,
7909 "Vertex Shader Epilog");
7910 return shader->epilog != NULL;
7911 }
7912
7913 /**
7914 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7915 */
7916 static bool si_shader_select_vs_parts(struct si_screen *sscreen,
7917 LLVMTargetMachineRef tm,
7918 struct si_shader *shader,
7919 struct pipe_debug_callback *debug)
7920 {
7921 struct tgsi_shader_info *info = &shader->selector->info;
7922 union si_shader_part_key prolog_key;
7923
7924 /* Get the prolog. */
7925 si_get_vs_prolog_key(shader, &prolog_key);
7926
7927 /* The prolog is a no-op if there are no inputs. */
7928 if (info->num_inputs) {
7929 shader->prolog =
7930 si_get_shader_part(sscreen, &sscreen->vs_prologs,
7931 PIPE_SHADER_VERTEX, true,
7932 &prolog_key, tm, debug,
7933 si_build_vs_prolog_function,
7934 "Vertex Shader Prolog");
7935 if (!shader->prolog)
7936 return false;
7937 }
7938
7939 /* Get the epilog. */
7940 if (!shader->key.as_es && !shader->key.as_ls &&
7941 !si_get_vs_epilog(sscreen, tm, shader, debug,
7942 &shader->key.part.vs.epilog))
7943 return false;
7944
7945 return true;
7946 }
7947
7948 /**
7949 * Select and compile (or reuse) TES parts (epilog).
7950 */
7951 static bool si_shader_select_tes_parts(struct si_screen *sscreen,
7952 LLVMTargetMachineRef tm,
7953 struct si_shader *shader,
7954 struct pipe_debug_callback *debug)
7955 {
7956 if (shader->key.as_es)
7957 return true;
7958
7959 /* TES compiled as VS. */
7960 return si_get_vs_epilog(sscreen, tm, shader, debug,
7961 &shader->key.part.tes.epilog);
7962 }
7963
7964 /**
7965 * Compile the TCS epilog function. This writes tesselation factors to memory
7966 * based on the output primitive type of the tesselator (determined by TES).
7967 */
7968 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
7969 union si_shader_part_key *key)
7970 {
7971 struct gallivm_state *gallivm = &ctx->gallivm;
7972 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
7973 LLVMTypeRef params[16];
7974 LLVMValueRef func;
7975 int last_sgpr, num_params;
7976
7977 /* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
7978 params[SI_PARAM_RW_BUFFERS] = const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
7979 params[SI_PARAM_CONST_BUFFERS] = ctx->i64;
7980 params[SI_PARAM_SAMPLERS] = ctx->i64;
7981 params[SI_PARAM_IMAGES] = ctx->i64;
7982 params[SI_PARAM_SHADER_BUFFERS] = ctx->i64;
7983 params[SI_PARAM_TCS_OFFCHIP_LAYOUT] = ctx->i32;
7984 params[SI_PARAM_TCS_OUT_OFFSETS] = ctx->i32;
7985 params[SI_PARAM_TCS_OUT_LAYOUT] = ctx->i32;
7986 params[SI_PARAM_TCS_IN_LAYOUT] = ctx->i32;
7987 params[ctx->param_oc_lds = SI_PARAM_TCS_OC_LDS] = ctx->i32;
7988 params[SI_PARAM_TESS_FACTOR_OFFSET] = ctx->i32;
7989 last_sgpr = SI_PARAM_TESS_FACTOR_OFFSET;
7990 num_params = last_sgpr + 1;
7991
7992 params[num_params++] = ctx->i32; /* patch index within the wave (REL_PATCH_ID) */
7993 params[num_params++] = ctx->i32; /* invocation ID within the patch */
7994 params[num_params++] = ctx->i32; /* LDS offset where tess factors should be loaded from */
7995
7996 /* Create the function. */
7997 si_create_function(ctx, "tcs_epilog", NULL, 0, params, num_params, last_sgpr);
7998 declare_tess_lds(ctx);
7999 func = ctx->main_fn;
8000
8001 si_write_tess_factors(bld_base,
8002 LLVMGetParam(func, last_sgpr + 1),
8003 LLVMGetParam(func, last_sgpr + 2),
8004 LLVMGetParam(func, last_sgpr + 3));
8005
8006 LLVMBuildRetVoid(gallivm->builder);
8007 }
8008
8009 /**
8010 * Select and compile (or reuse) TCS parts (epilog).
8011 */
8012 static bool si_shader_select_tcs_parts(struct si_screen *sscreen,
8013 LLVMTargetMachineRef tm,
8014 struct si_shader *shader,
8015 struct pipe_debug_callback *debug)
8016 {
8017 union si_shader_part_key epilog_key;
8018
8019 /* Get the epilog. */
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->soa.bld_base;
8308 LLVMTypeRef params[16+8*4+3];
8309 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
8310 int last_sgpr, num_params, i;
8311 struct si_ps_exports exp = {};
8312
8313 /* Declare input SGPRs. */
8314 params[SI_PARAM_RW_BUFFERS] = ctx->i64;
8315 params[SI_PARAM_CONST_BUFFERS] = ctx->i64;
8316 params[SI_PARAM_SAMPLERS] = ctx->i64;
8317 params[SI_PARAM_IMAGES] = ctx->i64;
8318 params[SI_PARAM_SHADER_BUFFERS] = ctx->i64;
8319 params[SI_PARAM_ALPHA_REF] = ctx->f32;
8320 last_sgpr = SI_PARAM_ALPHA_REF;
8321
8322 /* Declare input VGPRs. */
8323 num_params = (last_sgpr + 1) +
8324 util_bitcount(key->ps_epilog.colors_written) * 4 +
8325 key->ps_epilog.writes_z +
8326 key->ps_epilog.writes_stencil +
8327 key->ps_epilog.writes_samplemask;
8328
8329 num_params = MAX2(num_params,
8330 last_sgpr + 1 + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
8331
8332 assert(num_params <= ARRAY_SIZE(params));
8333
8334 for (i = last_sgpr + 1; i < num_params; i++)
8335 params[i] = ctx->f32;
8336
8337 /* Create the function. */
8338 si_create_function(ctx, "ps_epilog", NULL, 0, params, num_params, last_sgpr);
8339 /* Disable elimination of unused inputs. */
8340 si_llvm_add_attribute(ctx->main_fn,
8341 "InitialPSInputAddr", 0xffffff);
8342
8343 /* Process colors. */
8344 unsigned vgpr = last_sgpr + 1;
8345 unsigned colors_written = key->ps_epilog.colors_written;
8346 int last_color_export = -1;
8347
8348 /* Find the last color export. */
8349 if (!key->ps_epilog.writes_z &&
8350 !key->ps_epilog.writes_stencil &&
8351 !key->ps_epilog.writes_samplemask) {
8352 unsigned spi_format = key->ps_epilog.states.spi_shader_col_format;
8353
8354 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
8355 if (colors_written == 0x1 && key->ps_epilog.states.last_cbuf > 0) {
8356 /* Just set this if any of the colorbuffers are enabled. */
8357 if (spi_format &
8358 ((1llu << (4 * (key->ps_epilog.states.last_cbuf + 1))) - 1))
8359 last_color_export = 0;
8360 } else {
8361 for (i = 0; i < 8; i++)
8362 if (colors_written & (1 << i) &&
8363 (spi_format >> (i * 4)) & 0xf)
8364 last_color_export = i;
8365 }
8366 }
8367
8368 while (colors_written) {
8369 LLVMValueRef color[4];
8370 int mrt = u_bit_scan(&colors_written);
8371
8372 for (i = 0; i < 4; i++)
8373 color[i] = LLVMGetParam(ctx->main_fn, vgpr++);
8374
8375 si_export_mrt_color(bld_base, color, mrt,
8376 num_params - 1,
8377 mrt == last_color_export, &exp);
8378 }
8379
8380 /* Process depth, stencil, samplemask. */
8381 if (key->ps_epilog.writes_z)
8382 depth = LLVMGetParam(ctx->main_fn, vgpr++);
8383 if (key->ps_epilog.writes_stencil)
8384 stencil = LLVMGetParam(ctx->main_fn, vgpr++);
8385 if (key->ps_epilog.writes_samplemask)
8386 samplemask = LLVMGetParam(ctx->main_fn, vgpr++);
8387
8388 if (depth || stencil || samplemask)
8389 si_export_mrt_z(bld_base, depth, stencil, samplemask, &exp);
8390 else if (last_color_export == -1)
8391 si_export_null(bld_base);
8392
8393 if (exp.num)
8394 si_emit_ps_exports(ctx, &exp);
8395
8396 /* Compile. */
8397 LLVMBuildRetVoid(gallivm->builder);
8398 }
8399
8400 /**
8401 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
8402 */
8403 static bool si_shader_select_ps_parts(struct si_screen *sscreen,
8404 LLVMTargetMachineRef tm,
8405 struct si_shader *shader,
8406 struct pipe_debug_callback *debug)
8407 {
8408 union si_shader_part_key prolog_key;
8409 union si_shader_part_key epilog_key;
8410
8411 /* Get the prolog. */
8412 si_get_ps_prolog_key(shader, &prolog_key, true);
8413
8414 /* The prolog is a no-op if these aren't set. */
8415 if (si_need_ps_prolog(&prolog_key)) {
8416 shader->prolog =
8417 si_get_shader_part(sscreen, &sscreen->ps_prologs,
8418 PIPE_SHADER_FRAGMENT, true,
8419 &prolog_key, tm, debug,
8420 si_build_ps_prolog_function,
8421 "Fragment Shader Prolog");
8422 if (!shader->prolog)
8423 return false;
8424 }
8425
8426 /* Get the epilog. */
8427 si_get_ps_epilog_key(shader, &epilog_key);
8428
8429 shader->epilog =
8430 si_get_shader_part(sscreen, &sscreen->ps_epilogs,
8431 PIPE_SHADER_FRAGMENT, false,
8432 &epilog_key, tm, debug,
8433 si_build_ps_epilog_function,
8434 "Fragment Shader Epilog");
8435 if (!shader->epilog)
8436 return false;
8437
8438 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
8439 if (shader->key.part.ps.prolog.poly_stipple) {
8440 shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
8441 assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
8442 }
8443
8444 /* Set up the enable bits for per-sample shading if needed. */
8445 if (shader->key.part.ps.prolog.force_persp_sample_interp &&
8446 (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
8447 G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8448 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
8449 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
8450 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
8451 }
8452 if (shader->key.part.ps.prolog.force_linear_sample_interp &&
8453 (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
8454 G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8455 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
8456 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
8457 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
8458 }
8459 if (shader->key.part.ps.prolog.force_persp_center_interp &&
8460 (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
8461 G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8462 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
8463 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
8464 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
8465 }
8466 if (shader->key.part.ps.prolog.force_linear_center_interp &&
8467 (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
8468 G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
8469 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
8470 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
8471 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
8472 }
8473
8474 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
8475 if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_ena) &&
8476 !(shader->config.spi_ps_input_ena & 0xf)) {
8477 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
8478 assert(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr));
8479 }
8480
8481 /* At least one pair of interpolation weights must be enabled. */
8482 if (!(shader->config.spi_ps_input_ena & 0x7f)) {
8483 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
8484 assert(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr));
8485 }
8486
8487 /* The sample mask input is always enabled, because the API shader always
8488 * passes it through to the epilog. Disable it here if it's unused.
8489 */
8490 if (!shader->key.part.ps.epilog.poly_line_smoothing &&
8491 !shader->selector->info.reads_samplemask)
8492 shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;
8493
8494 return true;
8495 }
8496
8497 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
8498 unsigned *lds_size)
8499 {
8500 /* SPI barrier management bug:
8501 * Make sure we have at least 4k of LDS in use to avoid the bug.
8502 * It applies to workgroup sizes of more than one wavefront.
8503 */
8504 if (sscreen->b.family == CHIP_BONAIRE ||
8505 sscreen->b.family == CHIP_KABINI ||
8506 sscreen->b.family == CHIP_MULLINS)
8507 *lds_size = MAX2(*lds_size, 8);
8508 }
8509
8510 static void si_fix_resource_usage(struct si_screen *sscreen,
8511 struct si_shader *shader)
8512 {
8513 unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */
8514
8515 shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
8516
8517 if (shader->selector->type == PIPE_SHADER_COMPUTE &&
8518 si_get_max_workgroup_size(shader) > 64) {
8519 si_multiwave_lds_size_workaround(sscreen,
8520 &shader->config.lds_size);
8521 }
8522 }
8523
8524 int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
8525 struct si_shader *shader,
8526 struct pipe_debug_callback *debug)
8527 {
8528 struct si_shader_selector *sel = shader->selector;
8529 struct si_shader *mainp = sel->main_shader_part;
8530 int r;
8531
8532 /* LS, ES, VS are compiled on demand if the main part hasn't been
8533 * compiled for that stage.
8534 *
8535 * Vertex shaders are compiled on demand when a vertex fetch
8536 * workaround must be applied.
8537 */
8538 if (shader->is_monolithic) {
8539 /* Monolithic shader (compiled as a whole, has many variants,
8540 * may take a long time to compile).
8541 */
8542 r = si_compile_tgsi_shader(sscreen, tm, shader, true, debug);
8543 if (r)
8544 return r;
8545 } else {
8546 /* The shader consists of 2-3 parts:
8547 *
8548 * - the middle part is the user shader, it has 1 variant only
8549 * and it was compiled during the creation of the shader
8550 * selector
8551 * - the prolog part is inserted at the beginning
8552 * - the epilog part is inserted at the end
8553 *
8554 * The prolog and epilog have many (but simple) variants.
8555 */
8556
8557 /* Copy the compiled TGSI shader data over. */
8558 shader->is_binary_shared = true;
8559 shader->binary = mainp->binary;
8560 shader->config = mainp->config;
8561 shader->info.num_input_sgprs = mainp->info.num_input_sgprs;
8562 shader->info.num_input_vgprs = mainp->info.num_input_vgprs;
8563 shader->info.face_vgpr_index = mainp->info.face_vgpr_index;
8564 memcpy(shader->info.vs_output_param_offset,
8565 mainp->info.vs_output_param_offset,
8566 sizeof(mainp->info.vs_output_param_offset));
8567 shader->info.uses_instanceid = mainp->info.uses_instanceid;
8568 shader->info.nr_pos_exports = mainp->info.nr_pos_exports;
8569 shader->info.nr_param_exports = mainp->info.nr_param_exports;
8570
8571 /* Select prologs and/or epilogs. */
8572 switch (sel->type) {
8573 case PIPE_SHADER_VERTEX:
8574 if (!si_shader_select_vs_parts(sscreen, tm, shader, debug))
8575 return -1;
8576 break;
8577 case PIPE_SHADER_TESS_CTRL:
8578 if (!si_shader_select_tcs_parts(sscreen, tm, shader, debug))
8579 return -1;
8580 break;
8581 case PIPE_SHADER_TESS_EVAL:
8582 if (!si_shader_select_tes_parts(sscreen, tm, shader, debug))
8583 return -1;
8584 break;
8585 case PIPE_SHADER_GEOMETRY:
8586 if (!si_shader_select_gs_parts(sscreen, tm, shader, debug))
8587 return -1;
8588 break;
8589 case PIPE_SHADER_FRAGMENT:
8590 if (!si_shader_select_ps_parts(sscreen, tm, shader, debug))
8591 return -1;
8592
8593 /* Make sure we have at least as many VGPRs as there
8594 * are allocated inputs.
8595 */
8596 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8597 shader->info.num_input_vgprs);
8598 break;
8599 }
8600
8601 /* Update SGPR and VGPR counts. */
8602 if (shader->prolog) {
8603 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
8604 shader->prolog->config.num_sgprs);
8605 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8606 shader->prolog->config.num_vgprs);
8607 }
8608 if (shader->epilog) {
8609 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
8610 shader->epilog->config.num_sgprs);
8611 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
8612 shader->epilog->config.num_vgprs);
8613 }
8614 }
8615
8616 si_fix_resource_usage(sscreen, shader);
8617 si_shader_dump(sscreen, shader, debug, sel->info.processor,
8618 stderr);
8619
8620 /* Upload. */
8621 r = si_shader_binary_upload(sscreen, shader);
8622 if (r) {
8623 fprintf(stderr, "LLVM failed to upload shader\n");
8624 return r;
8625 }
8626
8627 return 0;
8628 }
8629
8630 void si_shader_destroy(struct si_shader *shader)
8631 {
8632 if (shader->scratch_bo)
8633 r600_resource_reference(&shader->scratch_bo, NULL);
8634
8635 r600_resource_reference(&shader->bo, NULL);
8636
8637 if (!shader->is_binary_shared)
8638 radeon_shader_binary_clean(&shader->binary);
8639
8640 free(shader->shader_log);
8641 }