projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: rename si_shader_create -> si_create_shader_variant for clarity
[mesa.git] / src / gallium / drivers / radeonsi / si_shader.c
1 /*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 #include <llvm/Config/llvm-config.h>
26
27 #include "util/u_memory.h"
28 #include "tgsi/tgsi_strings.h"
29 #include "tgsi/tgsi_from_mesa.h"
30
31 #include "ac_exp_param.h"
32 #include "ac_shader_util.h"
33 #include "ac_rtld.h"
34 #include "ac_llvm_util.h"
35 #include "si_shader_internal.h"
36 #include "si_pipe.h"
37 #include "sid.h"
38
39 #include "compiler/nir/nir.h"
40 #include "compiler/nir/nir_serialize.h"
41
42 static const char scratch_rsrc_dword0_symbol[] =
43 "SCRATCH_RSRC_DWORD0";
44
45 static const char scratch_rsrc_dword1_symbol[] =
46 "SCRATCH_RSRC_DWORD1";
47
48 static void si_llvm_emit_barrier(struct si_shader_context *ctx);
49
50 static void si_dump_shader_key(const struct si_shader *shader, FILE *f);
51
52 static void si_build_vs_prolog_function(struct si_shader_context *ctx,
53 union si_shader_part_key *key);
54 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
55 union si_shader_part_key *key);
56 static void si_build_ps_prolog_function(struct si_shader_context *ctx,
57 union si_shader_part_key *key);
58 static void si_build_ps_epilog_function(struct si_shader_context *ctx,
59 union si_shader_part_key *key);
60 static void si_fix_resource_usage(struct si_screen *sscreen,
61 struct si_shader *shader);
62
63 /* Ideally pass the sample mask input to the PS epilog as v14, which
64 * is its usual location, so that the shader doesn't have to add v_mov.
65 */
66 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 14
67
68 static bool llvm_type_is_64bit(struct si_shader_context *ctx,
69 LLVMTypeRef type)
70 {
71 if (type == ctx->ac.i64 || type == ctx->ac.f64)
72 return true;
73
74 return false;
75 }
76
77 /** Whether the shader runs as a combination of multiple API shaders */
78 static bool is_multi_part_shader(struct si_shader_context *ctx)
79 {
80 if (ctx->screen->info.chip_class <= GFX8)
81 return false;
82
83 return ctx->shader->key.as_ls ||
84 ctx->shader->key.as_es ||
85 ctx->type == PIPE_SHADER_TESS_CTRL ||
86 ctx->type == PIPE_SHADER_GEOMETRY;
87 }
88
89 /** Whether the shader runs on a merged HW stage (LSHS or ESGS) */
90 static bool is_merged_shader(struct si_shader_context *ctx)
91 {
92 return ctx->shader->key.as_ngg || is_multi_part_shader(ctx);
93 }
94
95 /**
96 * Returns a unique index for a per-patch semantic name and index. The index
97 * must be less than 32, so that a 32-bit bitmask of used inputs or outputs
98 * can be calculated.
99 */
100 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index)
101 {
102 switch (semantic_name) {
103 case TGSI_SEMANTIC_TESSOUTER:
104 return 0;
105 case TGSI_SEMANTIC_TESSINNER:
106 return 1;
107 case TGSI_SEMANTIC_PATCH:
108 assert(index < 30);
109 return 2 + index;
110
111 default:
112 assert(!"invalid semantic name");
113 return 0;
114 }
115 }
116
117 /**
118 * Returns a unique index for a semantic name and index. The index must be
119 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
120 * calculated.
121 */
122 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
123 unsigned is_varying)
124 {
125 switch (semantic_name) {
126 case TGSI_SEMANTIC_POSITION:
127 return 0;
128 case TGSI_SEMANTIC_GENERIC:
129 /* Since some shader stages use the the highest used IO index
130 * to determine the size to allocate for inputs/outputs
131 * (in LDS, tess and GS rings). GENERIC should be placed right
132 * after POSITION to make that size as small as possible.
133 */
134 if (index < SI_MAX_IO_GENERIC)
135 return 1 + index;
136
137 assert(!"invalid generic index");
138 return 0;
139 case TGSI_SEMANTIC_FOG:
140 return SI_MAX_IO_GENERIC + 1;
141 case TGSI_SEMANTIC_COLOR:
142 assert(index < 2);
143 return SI_MAX_IO_GENERIC + 2 + index;
144 case TGSI_SEMANTIC_BCOLOR:
145 assert(index < 2);
146 /* If it's a varying, COLOR and BCOLOR alias. */
147 if (is_varying)
148 return SI_MAX_IO_GENERIC + 2 + index;
149 else
150 return SI_MAX_IO_GENERIC + 4 + index;
151 case TGSI_SEMANTIC_TEXCOORD:
152 assert(index < 8);
153 return SI_MAX_IO_GENERIC + 6 + index;
154
155 /* These are rarely used between LS and HS or ES and GS. */
156 case TGSI_SEMANTIC_CLIPDIST:
157 assert(index < 2);
158 return SI_MAX_IO_GENERIC + 6 + 8 + index;
159 case TGSI_SEMANTIC_CLIPVERTEX:
160 return SI_MAX_IO_GENERIC + 6 + 8 + 2;
161 case TGSI_SEMANTIC_PSIZE:
162 return SI_MAX_IO_GENERIC + 6 + 8 + 3;
163
164 /* These can't be written by LS, HS, and ES. */
165 case TGSI_SEMANTIC_LAYER:
166 return SI_MAX_IO_GENERIC + 6 + 8 + 4;
167 case TGSI_SEMANTIC_VIEWPORT_INDEX:
168 return SI_MAX_IO_GENERIC + 6 + 8 + 5;
169 case TGSI_SEMANTIC_PRIMID:
170 STATIC_ASSERT(SI_MAX_IO_GENERIC + 6 + 8 + 6 <= 63);
171 return SI_MAX_IO_GENERIC + 6 + 8 + 6;
172 default:
173 fprintf(stderr, "invalid semantic name = %u\n", semantic_name);
174 assert(!"invalid semantic name");
175 return 0;
176 }
177 }
178
179 /**
180 * Get the value of a shader input parameter and extract a bitfield.
181 */
182 static LLVMValueRef unpack_llvm_param(struct si_shader_context *ctx,
183 LLVMValueRef value, unsigned rshift,
184 unsigned bitwidth)
185 {
186 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
187 value = ac_to_integer(&ctx->ac, value);
188
189 if (rshift)
190 value = LLVMBuildLShr(ctx->ac.builder, value,
191 LLVMConstInt(ctx->i32, rshift, 0), "");
192
193 if (rshift + bitwidth < 32) {
194 unsigned mask = (1 << bitwidth) - 1;
195 value = LLVMBuildAnd(ctx->ac.builder, value,
196 LLVMConstInt(ctx->i32, mask, 0), "");
197 }
198
199 return value;
200 }
201
202 LLVMValueRef si_unpack_param(struct si_shader_context *ctx,
203 struct ac_arg param, unsigned rshift,
204 unsigned bitwidth)
205 {
206 LLVMValueRef value = ac_get_arg(&ctx->ac, param);
207
208 return unpack_llvm_param(ctx, value, rshift, bitwidth);
209 }
210
211 static LLVMValueRef get_rel_patch_id(struct si_shader_context *ctx)
212 {
213 switch (ctx->type) {
214 case PIPE_SHADER_TESS_CTRL:
215 return si_unpack_param(ctx, ctx->args.tcs_rel_ids, 0, 8);
216
217 case PIPE_SHADER_TESS_EVAL:
218 return ac_get_arg(&ctx->ac, ctx->tes_rel_patch_id);
219
220 default:
221 assert(0);
222 return NULL;
223 }
224 }
225
226 /* Tessellation shaders pass outputs to the next shader using LDS.
227 *
228 * LS outputs = TCS inputs
229 * TCS outputs = TES inputs
230 *
231 * The LDS layout is:
232 * - TCS inputs for patch 0
233 * - TCS inputs for patch 1
234 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
235 * - ...
236 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
237 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
238 * - TCS outputs for patch 1
239 * - Per-patch TCS outputs for patch 1
240 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
241 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
242 * - ...
243 *
244 * All three shaders VS(LS), TCS, TES share the same LDS space.
245 */
246
247 static LLVMValueRef
248 get_tcs_in_patch_stride(struct si_shader_context *ctx)
249 {
250 return si_unpack_param(ctx, ctx->vs_state_bits, 8, 13);
251 }
252
253 static unsigned get_tcs_out_vertex_dw_stride_constant(struct si_shader_context *ctx)
254 {
255 assert(ctx->type == PIPE_SHADER_TESS_CTRL);
256
257 if (ctx->shader->key.mono.u.ff_tcs_inputs_to_copy)
258 return util_last_bit64(ctx->shader->key.mono.u.ff_tcs_inputs_to_copy) * 4;
259
260 return util_last_bit64(ctx->shader->selector->outputs_written) * 4;
261 }
262
263 static LLVMValueRef get_tcs_out_vertex_dw_stride(struct si_shader_context *ctx)
264 {
265 unsigned stride = get_tcs_out_vertex_dw_stride_constant(ctx);
266
267 return LLVMConstInt(ctx->i32, stride, 0);
268 }
269
270 static LLVMValueRef get_tcs_out_patch_stride(struct si_shader_context *ctx)
271 {
272 if (ctx->shader->key.mono.u.ff_tcs_inputs_to_copy)
273 return si_unpack_param(ctx, ctx->tcs_out_lds_layout, 0, 13);
274
275 const struct si_shader_info *info = &ctx->shader->selector->info;
276 unsigned tcs_out_vertices = info->properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
277 unsigned vertex_dw_stride = get_tcs_out_vertex_dw_stride_constant(ctx);
278 unsigned num_patch_outputs = util_last_bit64(ctx->shader->selector->patch_outputs_written);
279 unsigned patch_dw_stride = tcs_out_vertices * vertex_dw_stride +
280 num_patch_outputs * 4;
281 return LLVMConstInt(ctx->i32, patch_dw_stride, 0);
282 }
283
284 static LLVMValueRef
285 get_tcs_out_patch0_offset(struct si_shader_context *ctx)
286 {
287 return LLVMBuildMul(ctx->ac.builder,
288 si_unpack_param(ctx, ctx->tcs_out_lds_offsets, 0, 16),
289 LLVMConstInt(ctx->i32, 4, 0), "");
290 }
291
292 static LLVMValueRef
293 get_tcs_out_patch0_patch_data_offset(struct si_shader_context *ctx)
294 {
295 return LLVMBuildMul(ctx->ac.builder,
296 si_unpack_param(ctx, ctx->tcs_out_lds_offsets, 16, 16),
297 LLVMConstInt(ctx->i32, 4, 0), "");
298 }
299
300 static LLVMValueRef
301 get_tcs_in_current_patch_offset(struct si_shader_context *ctx)
302 {
303 LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
304 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
305
306 return LLVMBuildMul(ctx->ac.builder, patch_stride, rel_patch_id, "");
307 }
308
309 static LLVMValueRef
310 get_tcs_out_current_patch_offset(struct si_shader_context *ctx)
311 {
312 LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
313 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
314 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
315
316 return ac_build_imad(&ctx->ac, patch_stride, rel_patch_id, patch0_offset);
317 }
318
319 static LLVMValueRef
320 get_tcs_out_current_patch_data_offset(struct si_shader_context *ctx)
321 {
322 LLVMValueRef patch0_patch_data_offset =
323 get_tcs_out_patch0_patch_data_offset(ctx);
324 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
325 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
326
327 return ac_build_imad(&ctx->ac, patch_stride, rel_patch_id, patch0_patch_data_offset);
328 }
329
330 static LLVMValueRef get_num_tcs_out_vertices(struct si_shader_context *ctx)
331 {
332 unsigned tcs_out_vertices =
333 ctx->shader->selector ?
334 ctx->shader->selector->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT] : 0;
335
336 /* If !tcs_out_vertices, it's either the fixed-func TCS or the TCS epilog. */
337 if (ctx->type == PIPE_SHADER_TESS_CTRL && tcs_out_vertices)
338 return LLVMConstInt(ctx->i32, tcs_out_vertices, 0);
339
340 return si_unpack_param(ctx, ctx->tcs_offchip_layout, 6, 6);
341 }
342
343 static LLVMValueRef get_tcs_in_vertex_dw_stride(struct si_shader_context *ctx)
344 {
345 unsigned stride;
346
347 switch (ctx->type) {
348 case PIPE_SHADER_VERTEX:
349 stride = ctx->shader->selector->lshs_vertex_stride / 4;
350 return LLVMConstInt(ctx->i32, stride, 0);
351
352 case PIPE_SHADER_TESS_CTRL:
353 if (ctx->screen->info.chip_class >= GFX9 &&
354 ctx->shader->is_monolithic) {
355 stride = ctx->shader->key.part.tcs.ls->lshs_vertex_stride / 4;
356 return LLVMConstInt(ctx->i32, stride, 0);
357 }
358 return si_unpack_param(ctx, ctx->vs_state_bits, 24, 8);
359
360 default:
361 assert(0);
362 return NULL;
363 }
364 }
365
366 static LLVMValueRef unpack_sint16(struct si_shader_context *ctx,
367 LLVMValueRef i32, unsigned index)
368 {
369 assert(index <= 1);
370
371 if (index == 1)
372 return LLVMBuildAShr(ctx->ac.builder, i32,
373 LLVMConstInt(ctx->i32, 16, 0), "");
374
375 return LLVMBuildSExt(ctx->ac.builder,
376 LLVMBuildTrunc(ctx->ac.builder, i32,
377 ctx->ac.i16, ""),
378 ctx->i32, "");
379 }
380
381 void si_llvm_load_input_vs(
382 struct si_shader_context *ctx,
383 unsigned input_index,
384 LLVMValueRef out[4])
385 {
386 const struct si_shader_info *info = &ctx->shader->selector->info;
387 unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
388
389 if (vs_blit_property) {
390 LLVMValueRef vertex_id = ctx->abi.vertex_id;
391 LLVMValueRef sel_x1 = LLVMBuildICmp(ctx->ac.builder,
392 LLVMIntULE, vertex_id,
393 ctx->i32_1, "");
394 /* Use LLVMIntNE, because we have 3 vertices and only
395 * the middle one should use y2.
396 */
397 LLVMValueRef sel_y1 = LLVMBuildICmp(ctx->ac.builder,
398 LLVMIntNE, vertex_id,
399 ctx->i32_1, "");
400
401 unsigned param_vs_blit_inputs = ctx->vs_blit_inputs.arg_index;
402 if (input_index == 0) {
403 /* Position: */
404 LLVMValueRef x1y1 = LLVMGetParam(ctx->main_fn,
405 param_vs_blit_inputs);
406 LLVMValueRef x2y2 = LLVMGetParam(ctx->main_fn,
407 param_vs_blit_inputs + 1);
408
409 LLVMValueRef x1 = unpack_sint16(ctx, x1y1, 0);
410 LLVMValueRef y1 = unpack_sint16(ctx, x1y1, 1);
411 LLVMValueRef x2 = unpack_sint16(ctx, x2y2, 0);
412 LLVMValueRef y2 = unpack_sint16(ctx, x2y2, 1);
413
414 LLVMValueRef x = LLVMBuildSelect(ctx->ac.builder, sel_x1,
415 x1, x2, "");
416 LLVMValueRef y = LLVMBuildSelect(ctx->ac.builder, sel_y1,
417 y1, y2, "");
418
419 out[0] = LLVMBuildSIToFP(ctx->ac.builder, x, ctx->f32, "");
420 out[1] = LLVMBuildSIToFP(ctx->ac.builder, y, ctx->f32, "");
421 out[2] = LLVMGetParam(ctx->main_fn,
422 param_vs_blit_inputs + 2);
423 out[3] = ctx->ac.f32_1;
424 return;
425 }
426
427 /* Color or texture coordinates: */
428 assert(input_index == 1);
429
430 if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
431 for (int i = 0; i < 4; i++) {
432 out[i] = LLVMGetParam(ctx->main_fn,
433 param_vs_blit_inputs + 3 + i);
434 }
435 } else {
436 assert(vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD);
437 LLVMValueRef x1 = LLVMGetParam(ctx->main_fn,
438 param_vs_blit_inputs + 3);
439 LLVMValueRef y1 = LLVMGetParam(ctx->main_fn,
440 param_vs_blit_inputs + 4);
441 LLVMValueRef x2 = LLVMGetParam(ctx->main_fn,
442 param_vs_blit_inputs + 5);
443 LLVMValueRef y2 = LLVMGetParam(ctx->main_fn,
444 param_vs_blit_inputs + 6);
445
446 out[0] = LLVMBuildSelect(ctx->ac.builder, sel_x1,
447 x1, x2, "");
448 out[1] = LLVMBuildSelect(ctx->ac.builder, sel_y1,
449 y1, y2, "");
450 out[2] = LLVMGetParam(ctx->main_fn,
451 param_vs_blit_inputs + 7);
452 out[3] = LLVMGetParam(ctx->main_fn,
453 param_vs_blit_inputs + 8);
454 }
455 return;
456 }
457
458 unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs;
459 union si_vs_fix_fetch fix_fetch;
460 LLVMValueRef vb_desc;
461 LLVMValueRef vertex_index;
462 LLVMValueRef tmp;
463
464 if (input_index < num_vbos_in_user_sgprs) {
465 vb_desc = ac_get_arg(&ctx->ac, ctx->vb_descriptors[input_index]);
466 } else {
467 unsigned index= input_index - num_vbos_in_user_sgprs;
468 vb_desc = ac_build_load_to_sgpr(&ctx->ac,
469 ac_get_arg(&ctx->ac, ctx->vertex_buffers),
470 LLVMConstInt(ctx->i32, index, 0));
471 }
472
473 vertex_index = LLVMGetParam(ctx->main_fn,
474 ctx->vertex_index0.arg_index +
475 input_index);
476
477 /* Use the open-coded implementation for all loads of doubles and
478 * of dword-sized data that needs fixups. We need to insert conversion
479 * code anyway, and the amd/common code does it for us.
480 *
481 * Note: On LLVM <= 8, we can only open-code formats with
482 * channel size >= 4 bytes.
483 */
484 bool opencode = ctx->shader->key.mono.vs_fetch_opencode & (1 << input_index);
485 fix_fetch.bits = ctx->shader->key.mono.vs_fix_fetch[input_index].bits;
486 if (opencode ||
487 (fix_fetch.u.log_size == 3 && fix_fetch.u.format == AC_FETCH_FORMAT_FLOAT) ||
488 (fix_fetch.u.log_size == 2)) {
489 tmp = ac_build_opencoded_load_format(
490 &ctx->ac, fix_fetch.u.log_size, fix_fetch.u.num_channels_m1 + 1,
491 fix_fetch.u.format, fix_fetch.u.reverse, !opencode,
492 vb_desc, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0, 0, true);
493 for (unsigned i = 0; i < 4; ++i)
494 out[i] = LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->i32, i, false), "");
495 return;
496 }
497
498 /* Do multiple loads for special formats. */
499 unsigned required_channels = util_last_bit(info->input_usage_mask[input_index]);
500 LLVMValueRef fetches[4];
501 unsigned num_fetches;
502 unsigned fetch_stride;
503 unsigned channels_per_fetch;
504
505 if (fix_fetch.u.log_size <= 1 && fix_fetch.u.num_channels_m1 == 2) {
506 num_fetches = MIN2(required_channels, 3);
507 fetch_stride = 1 << fix_fetch.u.log_size;
508 channels_per_fetch = 1;
509 } else {
510 num_fetches = 1;
511 fetch_stride = 0;
512 channels_per_fetch = required_channels;
513 }
514
515 for (unsigned i = 0; i < num_fetches; ++i) {
516 LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
517 fetches[i] = ac_build_buffer_load_format(&ctx->ac, vb_desc, vertex_index, voffset,
518 channels_per_fetch, 0, true);
519 }
520
521 if (num_fetches == 1 && channels_per_fetch > 1) {
522 LLVMValueRef fetch = fetches[0];
523 for (unsigned i = 0; i < channels_per_fetch; ++i) {
524 tmp = LLVMConstInt(ctx->i32, i, false);
525 fetches[i] = LLVMBuildExtractElement(
526 ctx->ac.builder, fetch, tmp, "");
527 }
528 num_fetches = channels_per_fetch;
529 channels_per_fetch = 1;
530 }
531
532 for (unsigned i = num_fetches; i < 4; ++i)
533 fetches[i] = LLVMGetUndef(ctx->f32);
534
535 if (fix_fetch.u.log_size <= 1 && fix_fetch.u.num_channels_m1 == 2 &&
536 required_channels == 4) {
537 if (fix_fetch.u.format == AC_FETCH_FORMAT_UINT || fix_fetch.u.format == AC_FETCH_FORMAT_SINT)
538 fetches[3] = ctx->ac.i32_1;
539 else
540 fetches[3] = ctx->ac.f32_1;
541 } else if (fix_fetch.u.log_size == 3 &&
542 (fix_fetch.u.format == AC_FETCH_FORMAT_SNORM ||
543 fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED ||
544 fix_fetch.u.format == AC_FETCH_FORMAT_SINT) &&
545 required_channels == 4) {
546 /* For 2_10_10_10, the hardware returns an unsigned value;
547 * convert it to a signed one.
548 */
549 LLVMValueRef tmp = fetches[3];
550 LLVMValueRef c30 = LLVMConstInt(ctx->i32, 30, 0);
551
552 /* First, recover the sign-extended signed integer value. */
553 if (fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED)
554 tmp = LLVMBuildFPToUI(ctx->ac.builder, tmp, ctx->i32, "");
555 else
556 tmp = ac_to_integer(&ctx->ac, tmp);
557
558 /* For the integer-like cases, do a natural sign extension.
559 *
560 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
561 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
562 * exponent.
563 */
564 tmp = LLVMBuildShl(ctx->ac.builder, tmp,
565 fix_fetch.u.format == AC_FETCH_FORMAT_SNORM ?
566 LLVMConstInt(ctx->i32, 7, 0) : c30, "");
567 tmp = LLVMBuildAShr(ctx->ac.builder, tmp, c30, "");
568
569 /* Convert back to the right type. */
570 if (fix_fetch.u.format == AC_FETCH_FORMAT_SNORM) {
571 LLVMValueRef clamp;
572 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
573 tmp = LLVMBuildSIToFP(ctx->ac.builder, tmp, ctx->f32, "");
574 clamp = LLVMBuildFCmp(ctx->ac.builder, LLVMRealULT, tmp, neg_one, "");
575 tmp = LLVMBuildSelect(ctx->ac.builder, clamp, neg_one, tmp, "");
576 } else if (fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED) {
577 tmp = LLVMBuildSIToFP(ctx->ac.builder, tmp, ctx->f32, "");
578 }
579
580 fetches[3] = tmp;
581 }
582
583 for (unsigned i = 0; i < 4; ++i)
584 out[i] = ac_to_float(&ctx->ac, fetches[i]);
585 }
586
587 LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx,
588 unsigned swizzle)
589 {
590 if (swizzle > 0)
591 return ctx->i32_0;
592
593 switch (ctx->type) {
594 case PIPE_SHADER_VERTEX:
595 return ac_get_arg(&ctx->ac, ctx->vs_prim_id);
596 case PIPE_SHADER_TESS_CTRL:
597 return ac_get_arg(&ctx->ac, ctx->args.tcs_patch_id);
598 case PIPE_SHADER_TESS_EVAL:
599 return ac_get_arg(&ctx->ac, ctx->args.tes_patch_id);
600 case PIPE_SHADER_GEOMETRY:
601 return ac_get_arg(&ctx->ac, ctx->args.gs_prim_id);
602 default:
603 assert(0);
604 return ctx->i32_0;
605 }
606 }
607
608 static LLVMValueRef get_dw_address_from_generic_indices(struct si_shader_context *ctx,
609 LLVMValueRef vertex_dw_stride,
610 LLVMValueRef base_addr,
611 LLVMValueRef vertex_index,
612 LLVMValueRef param_index,
613 ubyte name, ubyte index)
614 {
615 if (vertex_dw_stride) {
616 base_addr = ac_build_imad(&ctx->ac, vertex_index,
617 vertex_dw_stride, base_addr);
618 }
619
620 if (param_index) {
621 base_addr = ac_build_imad(&ctx->ac, param_index,
622 LLVMConstInt(ctx->i32, 4, 0), base_addr);
623 }
624
625 int param = name == TGSI_SEMANTIC_PATCH ||
626 name == TGSI_SEMANTIC_TESSINNER ||
627 name == TGSI_SEMANTIC_TESSOUTER ?
628 si_shader_io_get_unique_index_patch(name, index) :
629 si_shader_io_get_unique_index(name, index, false);
630
631 /* Add the base address of the element. */
632 return LLVMBuildAdd(ctx->ac.builder, base_addr,
633 LLVMConstInt(ctx->i32, param * 4, 0), "");
634 }
635
636 /* The offchip buffer layout for TCS->TES is
637 *
638 * - attribute 0 of patch 0 vertex 0
639 * - attribute 0 of patch 0 vertex 1
640 * - attribute 0 of patch 0 vertex 2
641 * ...
642 * - attribute 0 of patch 1 vertex 0
643 * - attribute 0 of patch 1 vertex 1
644 * ...
645 * - attribute 1 of patch 0 vertex 0
646 * - attribute 1 of patch 0 vertex 1
647 * ...
648 * - per patch attribute 0 of patch 0
649 * - per patch attribute 0 of patch 1
650 * ...
651 *
652 * Note that every attribute has 4 components.
653 */
654 static LLVMValueRef get_tcs_tes_buffer_address(struct si_shader_context *ctx,
655 LLVMValueRef rel_patch_id,
656 LLVMValueRef vertex_index,
657 LLVMValueRef param_index)
658 {
659 LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
660 LLVMValueRef param_stride, constant16;
661
662 vertices_per_patch = get_num_tcs_out_vertices(ctx);
663 num_patches = si_unpack_param(ctx, ctx->tcs_offchip_layout, 0, 6);
664 total_vertices = LLVMBuildMul(ctx->ac.builder, vertices_per_patch,
665 num_patches, "");
666
667 constant16 = LLVMConstInt(ctx->i32, 16, 0);
668 if (vertex_index) {
669 base_addr = ac_build_imad(&ctx->ac, rel_patch_id,
670 vertices_per_patch, vertex_index);
671 param_stride = total_vertices;
672 } else {
673 base_addr = rel_patch_id;
674 param_stride = num_patches;
675 }
676
677 base_addr = ac_build_imad(&ctx->ac, param_index, param_stride, base_addr);
678 base_addr = LLVMBuildMul(ctx->ac.builder, base_addr, constant16, "");
679
680 if (!vertex_index) {
681 LLVMValueRef patch_data_offset =
682 si_unpack_param(ctx, ctx->tcs_offchip_layout, 12, 20);
683
684 base_addr = LLVMBuildAdd(ctx->ac.builder, base_addr,
685 patch_data_offset, "");
686 }
687 return base_addr;
688 }
689
690 static LLVMValueRef get_tcs_tes_buffer_address_from_generic_indices(
691 struct si_shader_context *ctx,
692 LLVMValueRef vertex_index,
693 LLVMValueRef param_index,
694 ubyte name, ubyte index)
695 {
696 unsigned param_index_base;
697
698 param_index_base = name == TGSI_SEMANTIC_PATCH ||
699 name == TGSI_SEMANTIC_TESSINNER ||
700 name == TGSI_SEMANTIC_TESSOUTER ?
701 si_shader_io_get_unique_index_patch(name, index) :
702 si_shader_io_get_unique_index(name, index, false);
703
704 if (param_index) {
705 param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
706 LLVMConstInt(ctx->i32, param_index_base, 0),
707 "");
708 } else {
709 param_index = LLVMConstInt(ctx->i32, param_index_base, 0);
710 }
711
712 return get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx),
713 vertex_index, param_index);
714 }
715
716 static LLVMValueRef si_build_gather_64bit(struct si_shader_context *ctx,
717 LLVMTypeRef type,
718 LLVMValueRef val1,
719 LLVMValueRef val2)
720 {
721 LLVMValueRef values[2] = {
722 ac_to_integer(&ctx->ac, val1),
723 ac_to_integer(&ctx->ac, val2),
724 };
725 LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, 2);
726 return LLVMBuildBitCast(ctx->ac.builder, result, type, "");
727 }
728
729 static LLVMValueRef buffer_load(struct si_shader_context *ctx,
730 LLVMTypeRef type, unsigned swizzle,
731 LLVMValueRef buffer, LLVMValueRef offset,
732 LLVMValueRef base, bool can_speculate)
733 {
734 LLVMValueRef value, value2;
735 LLVMTypeRef vec_type = LLVMVectorType(type, 4);
736
737 if (swizzle == ~0) {
738 value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
739 0, ac_glc, can_speculate, false);
740
741 return LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
742 }
743
744 if (!llvm_type_is_64bit(ctx, type)) {
745 value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
746 0, ac_glc, can_speculate, false);
747
748 value = LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
749 return LLVMBuildExtractElement(ctx->ac.builder, value,
750 LLVMConstInt(ctx->i32, swizzle, 0), "");
751 }
752
753 value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
754 swizzle * 4, ac_glc, can_speculate, false);
755
756 value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
757 swizzle * 4 + 4, ac_glc, can_speculate, false);
758
759 return si_build_gather_64bit(ctx, type, value, value2);
760 }
761
762 /**
763 * Load from LSHS LDS storage.
764 *
765 * \param type output value type
766 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
767 * \param dw_addr address in dwords
768 */
769 static LLVMValueRef lshs_lds_load(struct si_shader_context *ctx,
770 LLVMTypeRef type, unsigned swizzle,
771 LLVMValueRef dw_addr)
772 {
773 LLVMValueRef value;
774
775 if (swizzle == ~0) {
776 LLVMValueRef values[4];
777
778 for (unsigned chan = 0; chan < 4; chan++)
779 values[chan] = lshs_lds_load(ctx, type, chan, dw_addr);
780
781 return ac_build_gather_values(&ctx->ac, values, 4);
782 }
783
784 /* Split 64-bit loads. */
785 if (llvm_type_is_64bit(ctx, type)) {
786 LLVMValueRef lo, hi;
787
788 lo = lshs_lds_load(ctx, ctx->i32, swizzle, dw_addr);
789 hi = lshs_lds_load(ctx, ctx->i32, swizzle + 1, dw_addr);
790 return si_build_gather_64bit(ctx, type, lo, hi);
791 }
792
793 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
794 LLVMConstInt(ctx->i32, swizzle, 0), "");
795
796 value = ac_lds_load(&ctx->ac, dw_addr);
797
798 return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
799 }
800
801 /**
802 * Store to LSHS LDS storage.
803 *
804 * \param swizzle offset (typically 0..3)
805 * \param dw_addr address in dwords
806 * \param value value to store
807 */
808 static void lshs_lds_store(struct si_shader_context *ctx,
809 unsigned dw_offset_imm, LLVMValueRef dw_addr,
810 LLVMValueRef value)
811 {
812 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
813 LLVMConstInt(ctx->i32, dw_offset_imm, 0), "");
814
815 ac_lds_store(&ctx->ac, dw_addr, value);
816 }
817
818 enum si_tess_ring {
819 TCS_FACTOR_RING,
820 TESS_OFFCHIP_RING_TCS,
821 TESS_OFFCHIP_RING_TES,
822 };
823
824 static LLVMValueRef get_tess_ring_descriptor(struct si_shader_context *ctx,
825 enum si_tess_ring ring)
826 {
827 LLVMBuilderRef builder = ctx->ac.builder;
828 LLVMValueRef addr = ac_get_arg(&ctx->ac,
829 ring == TESS_OFFCHIP_RING_TES ?
830 ctx->tes_offchip_addr :
831 ctx->tcs_out_lds_layout);
832
833 /* TCS only receives high 13 bits of the address. */
834 if (ring == TESS_OFFCHIP_RING_TCS || ring == TCS_FACTOR_RING) {
835 addr = LLVMBuildAnd(builder, addr,
836 LLVMConstInt(ctx->i32, 0xfff80000, 0), "");
837 }
838
839 if (ring == TCS_FACTOR_RING) {
840 unsigned tf_offset = ctx->screen->tess_offchip_ring_size;
841 addr = LLVMBuildAdd(builder, addr,
842 LLVMConstInt(ctx->i32, tf_offset, 0), "");
843 }
844
845 uint32_t rsrc3 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
846 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
847 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
848 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
849
850 if (ctx->screen->info.chip_class >= GFX10)
851 rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
852 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
853 S_008F0C_RESOURCE_LEVEL(1);
854 else
855 rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
856 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
857
858 LLVMValueRef desc[4];
859 desc[0] = addr;
860 desc[1] = LLVMConstInt(ctx->i32,
861 S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0);
862 desc[2] = LLVMConstInt(ctx->i32, 0xffffffff, 0);
863 desc[3] = LLVMConstInt(ctx->i32, rsrc3, false);
864
865 return ac_build_gather_values(&ctx->ac, desc, 4);
866 }
867
868 static LLVMValueRef si_nir_load_tcs_varyings(struct ac_shader_abi *abi,
869 LLVMTypeRef type,
870 LLVMValueRef vertex_index,
871 LLVMValueRef param_index,
872 unsigned const_index,
873 unsigned location,
874 unsigned driver_location,
875 unsigned component,
876 unsigned num_components,
877 bool is_patch,
878 bool is_compact,
879 bool load_input)
880 {
881 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
882 struct si_shader_info *info = &ctx->shader->selector->info;
883 LLVMValueRef dw_addr, stride;
884 ubyte name, index;
885
886 driver_location = driver_location / 4;
887
888 if (load_input) {
889 name = info->input_semantic_name[driver_location];
890 index = info->input_semantic_index[driver_location];
891 } else {
892 name = info->output_semantic_name[driver_location];
893 index = info->output_semantic_index[driver_location];
894 }
895
896 assert((name == TGSI_SEMANTIC_PATCH ||
897 name == TGSI_SEMANTIC_TESSINNER ||
898 name == TGSI_SEMANTIC_TESSOUTER) == is_patch);
899
900 if (load_input) {
901 stride = get_tcs_in_vertex_dw_stride(ctx);
902 dw_addr = get_tcs_in_current_patch_offset(ctx);
903 } else {
904 if (is_patch) {
905 stride = NULL;
906 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
907 } else {
908 stride = get_tcs_out_vertex_dw_stride(ctx);
909 dw_addr = get_tcs_out_current_patch_offset(ctx);
910 }
911 }
912
913 if (!param_index) {
914 param_index = LLVMConstInt(ctx->i32, const_index, 0);
915 }
916
917 dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr,
918 vertex_index, param_index,
919 name, index);
920
921 LLVMValueRef value[4];
922 for (unsigned i = 0; i < num_components; i++) {
923 unsigned offset = i;
924 if (llvm_type_is_64bit(ctx, type))
925 offset *= 2;
926
927 offset += component;
928 value[i + component] = lshs_lds_load(ctx, type, offset, dw_addr);
929 }
930
931 return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
932 }
933
934 LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi,
935 LLVMTypeRef type,
936 LLVMValueRef vertex_index,
937 LLVMValueRef param_index,
938 unsigned const_index,
939 unsigned location,
940 unsigned driver_location,
941 unsigned component,
942 unsigned num_components,
943 bool is_patch,
944 bool is_compact,
945 bool load_input)
946 {
947 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
948 struct si_shader_info *info = &ctx->shader->selector->info;
949 LLVMValueRef base, addr;
950
951 driver_location = driver_location / 4;
952 ubyte name = info->input_semantic_name[driver_location];
953 ubyte index = info->input_semantic_index[driver_location];
954
955 assert((name == TGSI_SEMANTIC_PATCH ||
956 name == TGSI_SEMANTIC_TESSINNER ||
957 name == TGSI_SEMANTIC_TESSOUTER) == is_patch);
958
959 base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
960
961 if (!param_index) {
962 param_index = LLVMConstInt(ctx->i32, const_index, 0);
963 }
964
965 addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index,
966 param_index,
967 name, index);
968
969 /* TODO: This will generate rather ordinary llvm code, although it
970 * should be easy for the optimiser to fix up. In future we might want
971 * to refactor buffer_load().
972 */
973 LLVMValueRef value[4];
974 for (unsigned i = 0; i < num_components; i++) {
975 unsigned offset = i;
976 if (llvm_type_is_64bit(ctx, type)) {
977 offset *= 2;
978 if (offset == 4) {
979 ubyte name = info->input_semantic_name[driver_location + 1];
980 ubyte index = info->input_semantic_index[driver_location + 1];
981 addr = get_tcs_tes_buffer_address_from_generic_indices(ctx,
982 vertex_index,
983 param_index,
984 name, index);
985 }
986
987 offset = offset % 4;
988 }
989
990 offset += component;
991 value[i + component] = buffer_load(ctx, type, offset,
992 ctx->tess_offchip_ring, base, addr, true);
993 }
994
995 return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
996 }
997
998 static void si_nir_store_output_tcs(struct ac_shader_abi *abi,
999 const struct nir_variable *var,
1000 LLVMValueRef vertex_index,
1001 LLVMValueRef param_index,
1002 unsigned const_index,
1003 LLVMValueRef src,
1004 unsigned writemask)
1005 {
1006 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1007 struct si_shader_info *info = &ctx->shader->selector->info;
1008 const unsigned component = var->data.location_frac;
1009 unsigned driver_location = var->data.driver_location;
1010 LLVMValueRef dw_addr, stride;
1011 LLVMValueRef buffer, base, addr;
1012 LLVMValueRef values[8];
1013 bool skip_lds_store;
1014 bool is_tess_factor = false, is_tess_inner = false;
1015
1016 driver_location = driver_location / 4;
1017 ubyte name = info->output_semantic_name[driver_location];
1018 ubyte index = info->output_semantic_index[driver_location];
1019
1020 bool is_const = !param_index;
1021 if (!param_index)
1022 param_index = LLVMConstInt(ctx->i32, const_index, 0);
1023
1024 const bool is_patch = var->data.patch ||
1025 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
1026 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
1027
1028 assert((name == TGSI_SEMANTIC_PATCH ||
1029 name == TGSI_SEMANTIC_TESSINNER ||
1030 name == TGSI_SEMANTIC_TESSOUTER) == is_patch);
1031
1032 if (!is_patch) {
1033 stride = get_tcs_out_vertex_dw_stride(ctx);
1034 dw_addr = get_tcs_out_current_patch_offset(ctx);
1035 dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr,
1036 vertex_index, param_index,
1037 name, index);
1038
1039 skip_lds_store = !info->reads_pervertex_outputs;
1040 } else {
1041 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
1042 dw_addr = get_dw_address_from_generic_indices(ctx, NULL, dw_addr,
1043 vertex_index, param_index,
1044 name, index);
1045
1046 skip_lds_store = !info->reads_perpatch_outputs;
1047
1048 if (is_const && const_index == 0) {
1049 int name = info->output_semantic_name[driver_location];
1050
1051 /* Always write tess factors into LDS for the TCS epilog. */
1052 if (name == TGSI_SEMANTIC_TESSINNER ||
1053 name == TGSI_SEMANTIC_TESSOUTER) {
1054 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1055 skip_lds_store = !info->reads_tessfactor_outputs &&
1056 ctx->shader->selector->info.tessfactors_are_def_in_all_invocs;
1057 is_tess_factor = true;
1058 is_tess_inner = name == TGSI_SEMANTIC_TESSINNER;
1059 }
1060 }
1061 }
1062
1063 buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
1064
1065 base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
1066
1067 addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index,
1068 param_index, name, index);
1069
1070 for (unsigned chan = component; chan < 8; chan++) {
1071 if (!(writemask & (1 << chan)))
1072 continue;
1073 LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component);
1074
1075 unsigned buffer_store_offset = chan % 4;
1076 if (chan == 4) {
1077 ubyte name = info->output_semantic_name[driver_location + 1];
1078 ubyte index = info->output_semantic_index[driver_location + 1];
1079 addr = get_tcs_tes_buffer_address_from_generic_indices(ctx,
1080 vertex_index,
1081 param_index,
1082 name, index);
1083 }
1084
1085 /* Skip LDS stores if there is no LDS read of this output. */
1086 if (!skip_lds_store)
1087 lshs_lds_store(ctx, chan, dw_addr, value);
1088
1089 value = ac_to_integer(&ctx->ac, value);
1090 values[chan] = value;
1091
1092 if (writemask != 0xF && !is_tess_factor) {
1093 ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
1094 addr, base,
1095 4 * buffer_store_offset,
1096 ac_glc);
1097 }
1098
1099 /* Write tess factors into VGPRs for the epilog. */
1100 if (is_tess_factor &&
1101 ctx->shader->selector->info.tessfactors_are_def_in_all_invocs) {
1102 if (!is_tess_inner) {
1103 LLVMBuildStore(ctx->ac.builder, value, /* outer */
1104 ctx->invoc0_tess_factors[chan]);
1105 } else if (chan < 2) {
1106 LLVMBuildStore(ctx->ac.builder, value, /* inner */
1107 ctx->invoc0_tess_factors[4 + chan]);
1108 }
1109 }
1110 }
1111
1112 if (writemask == 0xF && !is_tess_factor) {
1113 LLVMValueRef value = ac_build_gather_values(&ctx->ac,
1114 values, 4);
1115 ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, addr,
1116 base, 0, ac_glc);
1117 }
1118 }
1119
1120 static LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi,
1121 unsigned input_index,
1122 unsigned vtx_offset_param,
1123 LLVMTypeRef type,
1124 unsigned swizzle)
1125 {
1126 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1127 struct si_shader *shader = ctx->shader;
1128 LLVMValueRef vtx_offset, soffset;
1129 struct si_shader_info *info = &shader->selector->info;
1130 unsigned semantic_name = info->input_semantic_name[input_index];
1131 unsigned semantic_index = info->input_semantic_index[input_index];
1132 unsigned param;
1133 LLVMValueRef value;
1134
1135 param = si_shader_io_get_unique_index(semantic_name, semantic_index, false);
1136
1137 /* GFX9 has the ESGS ring in LDS. */
1138 if (ctx->screen->info.chip_class >= GFX9) {
1139 unsigned index = vtx_offset_param;
1140
1141 switch (index / 2) {
1142 case 0:
1143 vtx_offset = si_unpack_param(ctx, ctx->gs_vtx01_offset,
1144 index % 2 ? 16 : 0, 16);
1145 break;
1146 case 1:
1147 vtx_offset = si_unpack_param(ctx, ctx->gs_vtx23_offset,
1148 index % 2 ? 16 : 0, 16);
1149 break;
1150 case 2:
1151 vtx_offset = si_unpack_param(ctx, ctx->gs_vtx45_offset,
1152 index % 2 ? 16 : 0, 16);
1153 break;
1154 default:
1155 assert(0);
1156 return NULL;
1157 }
1158
1159 unsigned offset = param * 4 + swizzle;
1160 vtx_offset = LLVMBuildAdd(ctx->ac.builder, vtx_offset,
1161 LLVMConstInt(ctx->i32, offset, false), "");
1162
1163 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->esgs_ring, vtx_offset);
1164 LLVMValueRef value = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1165 if (llvm_type_is_64bit(ctx, type)) {
1166 ptr = LLVMBuildGEP(ctx->ac.builder, ptr,
1167 &ctx->ac.i32_1, 1, "");
1168 LLVMValueRef values[2] = {
1169 value,
1170 LLVMBuildLoad(ctx->ac.builder, ptr, "")
1171 };
1172 value = ac_build_gather_values(&ctx->ac, values, 2);
1173 }
1174 return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
1175 }
1176
1177 /* GFX6: input load from the ESGS ring in memory. */
1178 if (swizzle == ~0) {
1179 LLVMValueRef values[4];
1180 unsigned chan;
1181 for (chan = 0; chan < 4; chan++) {
1182 values[chan] = si_llvm_load_input_gs(abi, input_index, vtx_offset_param,
1183 type, chan);
1184 }
1185 return ac_build_gather_values(&ctx->ac, values, 4);
1186 }
1187
1188 /* Get the vertex offset parameter on GFX6. */
1189 LLVMValueRef gs_vtx_offset = ac_get_arg(&ctx->ac,
1190 ctx->gs_vtx_offset[vtx_offset_param]);
1191
1192 vtx_offset = LLVMBuildMul(ctx->ac.builder, gs_vtx_offset,
1193 LLVMConstInt(ctx->i32, 4, 0), "");
1194
1195 soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0);
1196
1197 value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0,
1198 vtx_offset, soffset, 0, ac_glc, true, false);
1199 if (llvm_type_is_64bit(ctx, type)) {
1200 LLVMValueRef value2;
1201 soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0);
1202
1203 value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1,
1204 ctx->i32_0, vtx_offset, soffset,
1205 0, ac_glc, true, false);
1206 return si_build_gather_64bit(ctx, type, value, value2);
1207 }
1208 return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
1209 }
1210
1211 static LLVMValueRef si_nir_load_input_gs(struct ac_shader_abi *abi,
1212 unsigned location,
1213 unsigned driver_location,
1214 unsigned component,
1215 unsigned num_components,
1216 unsigned vertex_index,
1217 unsigned const_index,
1218 LLVMTypeRef type)
1219 {
1220 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1221
1222 LLVMValueRef value[4];
1223 for (unsigned i = 0; i < num_components; i++) {
1224 unsigned offset = i;
1225 if (llvm_type_is_64bit(ctx, type))
1226 offset *= 2;
1227
1228 offset += component;
1229 value[i + component] = si_llvm_load_input_gs(&ctx->abi, driver_location / 4 + const_index,
1230 vertex_index, type, offset);
1231 }
1232
1233 return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
1234 }
1235
1236 static LLVMValueRef si_build_fs_interp(struct si_shader_context *ctx,
1237 unsigned attr_index, unsigned chan,
1238 LLVMValueRef prim_mask,
1239 LLVMValueRef i, LLVMValueRef j)
1240 {
1241 if (i || j) {
1242 return ac_build_fs_interp(&ctx->ac,
1243 LLVMConstInt(ctx->i32, chan, 0),
1244 LLVMConstInt(ctx->i32, attr_index, 0),
1245 prim_mask, i, j);
1246 }
1247 return ac_build_fs_interp_mov(&ctx->ac,
1248 LLVMConstInt(ctx->i32, 2, 0), /* P0 */
1249 LLVMConstInt(ctx->i32, chan, 0),
1250 LLVMConstInt(ctx->i32, attr_index, 0),
1251 prim_mask);
1252 }
1253
1254 /**
1255 * Interpolate a fragment shader input.
1256 *
1257 * @param ctx context
1258 * @param input_index index of the input in hardware
1259 * @param semantic_name TGSI_SEMANTIC_*
1260 * @param semantic_index semantic index
1261 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1262 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1263 * @param interp_param interpolation weights (i,j)
1264 * @param prim_mask SI_PARAM_PRIM_MASK
1265 * @param face SI_PARAM_FRONT_FACE
1266 * @param result the return value (4 components)
1267 */
1268 static void interp_fs_color(struct si_shader_context *ctx,
1269 unsigned input_index,
1270 unsigned semantic_index,
1271 unsigned num_interp_inputs,
1272 unsigned colors_read_mask,
1273 LLVMValueRef interp_param,
1274 LLVMValueRef prim_mask,
1275 LLVMValueRef face,
1276 LLVMValueRef result[4])
1277 {
1278 LLVMValueRef i = NULL, j = NULL;
1279 unsigned chan;
1280
1281 /* fs.constant returns the param from the middle vertex, so it's not
1282 * really useful for flat shading. It's meant to be used for custom
1283 * interpolation (but the intrinsic can't fetch from the other two
1284 * vertices).
1285 *
1286 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1287 * to do the right thing. The only reason we use fs.constant is that
1288 * fs.interp cannot be used on integers, because they can be equal
1289 * to NaN.
1290 *
1291 * When interp is false we will use fs.constant or for newer llvm,
1292 * amdgcn.interp.mov.
1293 */
1294 bool interp = interp_param != NULL;
1295
1296 if (interp) {
1297 interp_param = LLVMBuildBitCast(ctx->ac.builder, interp_param,
1298 LLVMVectorType(ctx->f32, 2), "");
1299
1300 i = LLVMBuildExtractElement(ctx->ac.builder, interp_param,
1301 ctx->i32_0, "");
1302 j = LLVMBuildExtractElement(ctx->ac.builder, interp_param,
1303 ctx->i32_1, "");
1304 }
1305
1306 if (ctx->shader->key.part.ps.prolog.color_two_side) {
1307 LLVMValueRef is_face_positive;
1308
1309 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1310 * otherwise it's at offset "num_inputs".
1311 */
1312 unsigned back_attr_offset = num_interp_inputs;
1313 if (semantic_index == 1 && colors_read_mask & 0xf)
1314 back_attr_offset += 1;
1315
1316 is_face_positive = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
1317 face, ctx->i32_0, "");
1318
1319 for (chan = 0; chan < 4; chan++) {
1320 LLVMValueRef front, back;
1321
1322 front = si_build_fs_interp(ctx,
1323 input_index, chan,
1324 prim_mask, i, j);
1325 back = si_build_fs_interp(ctx,
1326 back_attr_offset, chan,
1327 prim_mask, i, j);
1328
1329 result[chan] = LLVMBuildSelect(ctx->ac.builder,
1330 is_face_positive,
1331 front,
1332 back,
1333 "");
1334 }
1335 } else {
1336 for (chan = 0; chan < 4; chan++) {
1337 result[chan] = si_build_fs_interp(ctx,
1338 input_index, chan,
1339 prim_mask, i, j);
1340 }
1341 }
1342 }
1343
1344 LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
1345 {
1346 return si_unpack_param(ctx, ctx->args.ancillary, 8, 4);
1347 }
1348
1349 static LLVMValueRef get_base_vertex(struct ac_shader_abi *abi)
1350 {
1351 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1352
1353 /* For non-indexed draws, the base vertex set by the driver
1354 * (for direct draws) or the CP (for indirect draws) is the
1355 * first vertex ID, but GLSL expects 0 to be returned.
1356 */
1357 LLVMValueRef vs_state = ac_get_arg(&ctx->ac,
1358 ctx->vs_state_bits);
1359 LLVMValueRef indexed;
1360
1361 indexed = LLVMBuildLShr(ctx->ac.builder, vs_state, ctx->i32_1, "");
1362 indexed = LLVMBuildTrunc(ctx->ac.builder, indexed, ctx->i1, "");
1363
1364 return LLVMBuildSelect(ctx->ac.builder, indexed,
1365 ac_get_arg(&ctx->ac, ctx->args.base_vertex),
1366 ctx->i32_0, "");
1367 }
1368
1369 static LLVMValueRef get_block_size(struct ac_shader_abi *abi)
1370 {
1371 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1372
1373 LLVMValueRef values[3];
1374 LLVMValueRef result;
1375 unsigned i;
1376 unsigned *properties = ctx->shader->selector->info.properties;
1377
1378 if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
1379 unsigned sizes[3] = {
1380 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
1381 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
1382 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
1383 };
1384
1385 for (i = 0; i < 3; ++i)
1386 values[i] = LLVMConstInt(ctx->i32, sizes[i], 0);
1387
1388 result = ac_build_gather_values(&ctx->ac, values, 3);
1389 } else {
1390 result = ac_get_arg(&ctx->ac, ctx->block_size);
1391 }
1392
1393 return result;
1394 }
1395
1396 /**
1397 * Load a dword from a constant buffer.
1398 */
1399 static LLVMValueRef buffer_load_const(struct si_shader_context *ctx,
1400 LLVMValueRef resource,
1401 LLVMValueRef offset)
1402 {
1403 return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL,
1404 0, 0, true, true);
1405 }
1406
1407 static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id)
1408 {
1409 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1410 LLVMValueRef desc = ac_get_arg(&ctx->ac, ctx->rw_buffers);
1411 LLVMValueRef buf_index = LLVMConstInt(ctx->i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
1412 LLVMValueRef resource = ac_build_load_to_sgpr(&ctx->ac, desc, buf_index);
1413
1414 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1415 LLVMValueRef offset0 = LLVMBuildMul(ctx->ac.builder, sample_id, LLVMConstInt(ctx->i32, 8, 0), "");
1416 LLVMValueRef offset1 = LLVMBuildAdd(ctx->ac.builder, offset0, LLVMConstInt(ctx->i32, 4, 0), "");
1417
1418 LLVMValueRef pos[4] = {
1419 buffer_load_const(ctx, resource, offset0),
1420 buffer_load_const(ctx, resource, offset1),
1421 LLVMConstReal(ctx->f32, 0),
1422 LLVMConstReal(ctx->f32, 0)
1423 };
1424
1425 return ac_build_gather_values(&ctx->ac, pos, 4);
1426 }
1427
1428 static LLVMValueRef load_sample_mask_in(struct ac_shader_abi *abi)
1429 {
1430 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1431 return ac_to_integer(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args.sample_coverage));
1432 }
1433
1434 static LLVMValueRef si_load_tess_coord(struct ac_shader_abi *abi)
1435 {
1436 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1437 LLVMValueRef coord[4] = {
1438 ac_get_arg(&ctx->ac, ctx->tes_u),
1439 ac_get_arg(&ctx->ac, ctx->tes_v),
1440 ctx->ac.f32_0,
1441 ctx->ac.f32_0
1442 };
1443
1444 /* For triangles, the vector should be (u, v, 1-u-v). */
1445 if (ctx->shader->selector->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] ==
1446 PIPE_PRIM_TRIANGLES) {
1447 coord[2] = LLVMBuildFSub(ctx->ac.builder, ctx->ac.f32_1,
1448 LLVMBuildFAdd(ctx->ac.builder,
1449 coord[0], coord[1], ""), "");
1450 }
1451 return ac_build_gather_values(&ctx->ac, coord, 4);
1452 }
1453
1454 static LLVMValueRef load_tess_level(struct si_shader_context *ctx,
1455 unsigned semantic_name)
1456 {
1457 LLVMValueRef base, addr;
1458
1459 int param = si_shader_io_get_unique_index_patch(semantic_name, 0);
1460
1461 base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
1462 addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL,
1463 LLVMConstInt(ctx->i32, param, 0));
1464
1465 return buffer_load(ctx, ctx->f32,
1466 ~0, ctx->tess_offchip_ring, base, addr, true);
1467
1468 }
1469
1470 static LLVMValueRef load_tess_level_default(struct si_shader_context *ctx,
1471 unsigned semantic_name)
1472 {
1473 LLVMValueRef buf, slot, val[4];
1474 int i, offset;
1475
1476 slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
1477 buf = ac_get_arg(&ctx->ac, ctx->rw_buffers);
1478 buf = ac_build_load_to_sgpr(&ctx->ac, buf, slot);
1479 offset = semantic_name == TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL ? 4 : 0;
1480
1481 for (i = 0; i < 4; i++)
1482 val[i] = buffer_load_const(ctx, buf,
1483 LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
1484 return ac_build_gather_values(&ctx->ac, val, 4);
1485 }
1486
1487 static LLVMValueRef si_load_tess_level(struct ac_shader_abi *abi,
1488 unsigned varying_id,
1489 bool load_default_state)
1490 {
1491 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1492 unsigned semantic_name;
1493
1494 if (load_default_state) {
1495 switch (varying_id) {
1496 case VARYING_SLOT_TESS_LEVEL_INNER:
1497 semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL;
1498 break;
1499 case VARYING_SLOT_TESS_LEVEL_OUTER:
1500 semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL;
1501 break;
1502 default:
1503 unreachable("unknown tess level");
1504 }
1505 return load_tess_level_default(ctx, semantic_name);
1506 }
1507
1508 switch (varying_id) {
1509 case VARYING_SLOT_TESS_LEVEL_INNER:
1510 semantic_name = TGSI_SEMANTIC_TESSINNER;
1511 break;
1512 case VARYING_SLOT_TESS_LEVEL_OUTER:
1513 semantic_name = TGSI_SEMANTIC_TESSOUTER;
1514 break;
1515 default:
1516 unreachable("unknown tess level");
1517 }
1518
1519 return load_tess_level(ctx, semantic_name);
1520
1521 }
1522
1523 static LLVMValueRef si_load_patch_vertices_in(struct ac_shader_abi *abi)
1524 {
1525 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1526 if (ctx->type == PIPE_SHADER_TESS_CTRL)
1527 return si_unpack_param(ctx, ctx->tcs_out_lds_layout, 13, 6);
1528 else if (ctx->type == PIPE_SHADER_TESS_EVAL)
1529 return get_num_tcs_out_vertices(ctx);
1530 else
1531 unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1532 }
1533
1534 void si_declare_compute_memory(struct si_shader_context *ctx)
1535 {
1536 struct si_shader_selector *sel = ctx->shader->selector;
1537 unsigned lds_size = sel->info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE];
1538
1539 LLVMTypeRef i8p = LLVMPointerType(ctx->i8, AC_ADDR_SPACE_LDS);
1540 LLVMValueRef var;
1541
1542 assert(!ctx->ac.lds);
1543
1544 var = LLVMAddGlobalInAddressSpace(ctx->ac.module,
1545 LLVMArrayType(ctx->i8, lds_size),
1546 "compute_lds",
1547 AC_ADDR_SPACE_LDS);
1548 LLVMSetAlignment(var, 64 * 1024);
1549
1550 ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, var, i8p, "");
1551 }
1552
1553 static LLVMValueRef load_const_buffer_desc_fast_path(struct si_shader_context *ctx)
1554 {
1555 LLVMValueRef ptr =
1556 ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers);
1557 struct si_shader_selector *sel = ctx->shader->selector;
1558
1559 /* Do the bounds checking with a descriptor, because
1560 * doing computation and manual bounds checking of 64-bit
1561 * addresses generates horrible VALU code with very high
1562 * VGPR usage and very low SIMD occupancy.
1563 */
1564 ptr = LLVMBuildPtrToInt(ctx->ac.builder, ptr, ctx->ac.intptr, "");
1565
1566 LLVMValueRef desc0, desc1;
1567 desc0 = ptr;
1568 desc1 = LLVMConstInt(ctx->i32,
1569 S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0);
1570
1571 uint32_t rsrc3 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
1572 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
1573 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
1574 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
1575
1576 if (ctx->screen->info.chip_class >= GFX10)
1577 rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
1578 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
1579 S_008F0C_RESOURCE_LEVEL(1);
1580 else
1581 rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
1582 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
1583
1584 LLVMValueRef desc_elems[] = {
1585 desc0,
1586 desc1,
1587 LLVMConstInt(ctx->i32, sel->info.constbuf0_num_slots * 16, 0),
1588 LLVMConstInt(ctx->i32, rsrc3, false)
1589 };
1590
1591 return ac_build_gather_values(&ctx->ac, desc_elems, 4);
1592 }
1593
1594 static LLVMValueRef load_ubo(struct ac_shader_abi *abi, LLVMValueRef index)
1595 {
1596 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1597 struct si_shader_selector *sel = ctx->shader->selector;
1598
1599 LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers);
1600
1601 if (sel->info.const_buffers_declared == 1 &&
1602 sel->info.shader_buffers_declared == 0) {
1603 return load_const_buffer_desc_fast_path(ctx);
1604 }
1605
1606 index = si_llvm_bound_index(ctx, index, ctx->num_const_buffers);
1607 index = LLVMBuildAdd(ctx->ac.builder, index,
1608 LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS, 0), "");
1609
1610 return ac_build_load_to_sgpr(&ctx->ac, ptr, index);
1611 }
1612
1613 static LLVMValueRef
1614 load_ssbo(struct ac_shader_abi *abi, LLVMValueRef index, bool write)
1615 {
1616 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
1617 LLVMValueRef rsrc_ptr = ac_get_arg(&ctx->ac,
1618 ctx->const_and_shader_buffers);
1619
1620 index = si_llvm_bound_index(ctx, index, ctx->num_shader_buffers);
1621 index = LLVMBuildSub(ctx->ac.builder,
1622 LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS - 1, 0),
1623 index, "");
1624
1625 return ac_build_load_to_sgpr(&ctx->ac, rsrc_ptr, index);
1626 }
1627
1628 /* Initialize arguments for the shader export intrinsic */
1629 static void si_llvm_init_export_args(struct si_shader_context *ctx,
1630 LLVMValueRef *values,
1631 unsigned target,
1632 struct ac_export_args *args)
1633 {
1634 LLVMValueRef f32undef = LLVMGetUndef(ctx->ac.f32);
1635 unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
1636 unsigned chan;
1637 bool is_int8, is_int10;
1638
1639 /* Default is 0xf. Adjusted below depending on the format. */
1640 args->enabled_channels = 0xf; /* writemask */
1641
1642 /* Specify whether the EXEC mask represents the valid mask */
1643 args->valid_mask = 0;
1644
1645 /* Specify whether this is the last export */
1646 args->done = 0;
1647
1648 /* Specify the target we are exporting */
1649 args->target = target;
1650
1651 if (ctx->type == PIPE_SHADER_FRAGMENT) {
1652 const struct si_shader_key *key = &ctx->shader->key;
1653 unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
1654 int cbuf = target - V_008DFC_SQ_EXP_MRT;
1655
1656 assert(cbuf >= 0 && cbuf < 8);
1657 spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
1658 is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
1659 is_int10 = (key->part.ps.epilog.color_is_int10 >> cbuf) & 0x1;
1660 }
1661
1662 args->compr = false;
1663 args->out[0] = f32undef;
1664 args->out[1] = f32undef;
1665 args->out[2] = f32undef;
1666 args->out[3] = f32undef;
1667
1668 LLVMValueRef (*packf)(struct ac_llvm_context *ctx, LLVMValueRef args[2]) = NULL;
1669 LLVMValueRef (*packi)(struct ac_llvm_context *ctx, LLVMValueRef args[2],
1670 unsigned bits, bool hi) = NULL;
1671
1672 switch (spi_shader_col_format) {
1673 case V_028714_SPI_SHADER_ZERO:
1674 args->enabled_channels = 0; /* writemask */
1675 args->target = V_008DFC_SQ_EXP_NULL;
1676 break;
1677
1678 case V_028714_SPI_SHADER_32_R:
1679 args->enabled_channels = 1; /* writemask */
1680 args->out[0] = values[0];
1681 break;
1682
1683 case V_028714_SPI_SHADER_32_GR:
1684 args->enabled_channels = 0x3; /* writemask */
1685 args->out[0] = values[0];
1686 args->out[1] = values[1];
1687 break;
1688
1689 case V_028714_SPI_SHADER_32_AR:
1690 if (ctx->screen->info.chip_class >= GFX10) {
1691 args->enabled_channels = 0x3; /* writemask */
1692 args->out[0] = values[0];
1693 args->out[1] = values[3];
1694 } else {
1695 args->enabled_channels = 0x9; /* writemask */
1696 args->out[0] = values[0];
1697 args->out[3] = values[3];
1698 }
1699 break;
1700
1701 case V_028714_SPI_SHADER_FP16_ABGR:
1702 packf = ac_build_cvt_pkrtz_f16;
1703 break;
1704
1705 case V_028714_SPI_SHADER_UNORM16_ABGR:
1706 packf = ac_build_cvt_pknorm_u16;
1707 break;
1708
1709 case V_028714_SPI_SHADER_SNORM16_ABGR:
1710 packf = ac_build_cvt_pknorm_i16;
1711 break;
1712
1713 case V_028714_SPI_SHADER_UINT16_ABGR:
1714 packi = ac_build_cvt_pk_u16;
1715 break;
1716
1717 case V_028714_SPI_SHADER_SINT16_ABGR:
1718 packi = ac_build_cvt_pk_i16;
1719 break;
1720
1721 case V_028714_SPI_SHADER_32_ABGR:
1722 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
1723 break;
1724 }
1725
1726 /* Pack f16 or norm_i16/u16. */
1727 if (packf) {
1728 for (chan = 0; chan < 2; chan++) {
1729 LLVMValueRef pack_args[2] = {
1730 values[2 * chan],
1731 values[2 * chan + 1]
1732 };
1733 LLVMValueRef packed;
1734
1735 packed = packf(&ctx->ac, pack_args);
1736 args->out[chan] = ac_to_float(&ctx->ac, packed);
1737 }
1738 args->compr = 1; /* COMPR flag */
1739 }
1740 /* Pack i16/u16. */
1741 if (packi) {
1742 for (chan = 0; chan < 2; chan++) {
1743 LLVMValueRef pack_args[2] = {
1744 ac_to_integer(&ctx->ac, values[2 * chan]),
1745 ac_to_integer(&ctx->ac, values[2 * chan + 1])
1746 };
1747 LLVMValueRef packed;
1748
1749 packed = packi(&ctx->ac, pack_args,
1750 is_int8 ? 8 : is_int10 ? 10 : 16,
1751 chan == 1);
1752 args->out[chan] = ac_to_float(&ctx->ac, packed);
1753 }
1754 args->compr = 1; /* COMPR flag */
1755 }
1756 }
1757
1758 static void si_alpha_test(struct si_shader_context *ctx, LLVMValueRef alpha)
1759 {
1760 if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
1761 static LLVMRealPredicate cond_map[PIPE_FUNC_ALWAYS + 1] = {
1762 [PIPE_FUNC_LESS] = LLVMRealOLT,
1763 [PIPE_FUNC_EQUAL] = LLVMRealOEQ,
1764 [PIPE_FUNC_LEQUAL] = LLVMRealOLE,
1765 [PIPE_FUNC_GREATER] = LLVMRealOGT,
1766 [PIPE_FUNC_NOTEQUAL] = LLVMRealONE,
1767 [PIPE_FUNC_GEQUAL] = LLVMRealOGE,
1768 };
1769 LLVMRealPredicate cond = cond_map[ctx->shader->key.part.ps.epilog.alpha_func];
1770 assert(cond);
1771
1772 LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
1773 SI_PARAM_ALPHA_REF);
1774 LLVMValueRef alpha_pass =
1775 LLVMBuildFCmp(ctx->ac.builder, cond, alpha, alpha_ref, "");
1776 ac_build_kill_if_false(&ctx->ac, alpha_pass);
1777 } else {
1778 ac_build_kill_if_false(&ctx->ac, ctx->i1false);
1779 }
1780 }
1781
1782 static LLVMValueRef si_scale_alpha_by_sample_mask(struct si_shader_context *ctx,
1783 LLVMValueRef alpha,
1784 unsigned samplemask_param)
1785 {
1786 LLVMValueRef coverage;
1787
1788 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
1789 coverage = LLVMGetParam(ctx->main_fn,
1790 samplemask_param);
1791 coverage = ac_to_integer(&ctx->ac, coverage);
1792
1793 coverage = ac_build_intrinsic(&ctx->ac, "llvm.ctpop.i32",
1794 ctx->i32,
1795 &coverage, 1, AC_FUNC_ATTR_READNONE);
1796
1797 coverage = LLVMBuildUIToFP(ctx->ac.builder, coverage,
1798 ctx->f32, "");
1799
1800 coverage = LLVMBuildFMul(ctx->ac.builder, coverage,
1801 LLVMConstReal(ctx->f32,
1802 1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");
1803
1804 return LLVMBuildFMul(ctx->ac.builder, alpha, coverage, "");
1805 }
1806
1807 static void si_llvm_emit_clipvertex(struct si_shader_context *ctx,
1808 struct ac_export_args *pos, LLVMValueRef *out_elts)
1809 {
1810 unsigned reg_index;
1811 unsigned chan;
1812 unsigned const_chan;
1813 LLVMValueRef base_elt;
1814 LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
1815 LLVMValueRef constbuf_index = LLVMConstInt(ctx->i32,
1816 SI_VS_CONST_CLIP_PLANES, 0);
1817 LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, ptr, constbuf_index);
1818
1819 for (reg_index = 0; reg_index < 2; reg_index ++) {
1820 struct ac_export_args *args = &pos[2 + reg_index];
1821
1822 args->out[0] =
1823 args->out[1] =
1824 args->out[2] =
1825 args->out[3] = LLVMConstReal(ctx->f32, 0.0f);
1826
1827 /* Compute dot products of position and user clip plane vectors */
1828 for (chan = 0; chan < 4; chan++) {
1829 for (const_chan = 0; const_chan < 4; const_chan++) {
1830 LLVMValueRef addr =
1831 LLVMConstInt(ctx->i32, ((reg_index * 4 + chan) * 4 +
1832 const_chan) * 4, 0);
1833 base_elt = buffer_load_const(ctx, const_resource,
1834 addr);
1835 args->out[chan] = ac_build_fmad(&ctx->ac, base_elt,
1836 out_elts[const_chan], args->out[chan]);
1837 }
1838 }
1839
1840 args->enabled_channels = 0xf;
1841 args->valid_mask = 0;
1842 args->done = 0;
1843 args->target = V_008DFC_SQ_EXP_POS + 2 + reg_index;
1844 args->compr = 0;
1845 }
1846 }
1847
1848 static void si_dump_streamout(struct pipe_stream_output_info *so)
1849 {
1850 unsigned i;
1851
1852 if (so->num_outputs)
1853 fprintf(stderr, "STREAMOUT\n");
1854
1855 for (i = 0; i < so->num_outputs; i++) {
1856 unsigned mask = ((1 << so->output[i].num_components) - 1) <<
1857 so->output[i].start_component;
1858 fprintf(stderr, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
1859 i, so->output[i].output_buffer,
1860 so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1,
1861 so->output[i].register_index,
1862 mask & 1 ? "x" : "",
1863 mask & 2 ? "y" : "",
1864 mask & 4 ? "z" : "",
1865 mask & 8 ? "w" : "");
1866 }
1867 }
1868
1869 void si_emit_streamout_output(struct si_shader_context *ctx,
1870 LLVMValueRef const *so_buffers,
1871 LLVMValueRef const *so_write_offsets,
1872 struct pipe_stream_output *stream_out,
1873 struct si_shader_output_values *shader_out)
1874 {
1875 unsigned buf_idx = stream_out->output_buffer;
1876 unsigned start = stream_out->start_component;
1877 unsigned num_comps = stream_out->num_components;
1878 LLVMValueRef out[4];
1879
1880 assert(num_comps && num_comps <= 4);
1881 if (!num_comps || num_comps > 4)
1882 return;
1883
1884 /* Load the output as int. */
1885 for (int j = 0; j < num_comps; j++) {
1886 assert(stream_out->stream == shader_out->vertex_stream[start + j]);
1887
1888 out[j] = ac_to_integer(&ctx->ac, shader_out->values[start + j]);
1889 }
1890
1891 /* Pack the output. */
1892 LLVMValueRef vdata = NULL;
1893
1894 switch (num_comps) {
1895 case 1: /* as i32 */
1896 vdata = out[0];
1897 break;
1898 case 2: /* as v2i32 */
1899 case 3: /* as v3i32 */
1900 if (ac_has_vec3_support(ctx->screen->info.chip_class, false)) {
1901 vdata = ac_build_gather_values(&ctx->ac, out, num_comps);
1902 break;
1903 }
1904 /* as v4i32 (aligned to 4) */
1905 out[3] = LLVMGetUndef(ctx->i32);
1906 /* fall through */
1907 case 4: /* as v4i32 */
1908 vdata = ac_build_gather_values(&ctx->ac, out, util_next_power_of_two(num_comps));
1909 break;
1910 }
1911
1912 ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf_idx],
1913 vdata, num_comps,
1914 so_write_offsets[buf_idx],
1915 ctx->i32_0,
1916 stream_out->dst_offset * 4, ac_glc | ac_slc);
1917 }
1918
1919 /**
1920 * Write streamout data to buffers for vertex stream @p stream (different
1921 * vertex streams can occur for GS copy shaders).
1922 */
1923 static void si_llvm_emit_streamout(struct si_shader_context *ctx,
1924 struct si_shader_output_values *outputs,
1925 unsigned noutput, unsigned stream)
1926 {
1927 struct si_shader_selector *sel = ctx->shader->selector;
1928 struct pipe_stream_output_info *so = &sel->so;
1929 LLVMBuilderRef builder = ctx->ac.builder;
1930 int i;
1931
1932 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
1933 LLVMValueRef so_vtx_count =
1934 si_unpack_param(ctx, ctx->streamout_config, 16, 7);
1935
1936 LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
1937
1938 /* can_emit = tid < so_vtx_count; */
1939 LLVMValueRef can_emit =
1940 LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
1941
1942 /* Emit the streamout code conditionally. This actually avoids
1943 * out-of-bounds buffer access. The hw tells us via the SGPR
1944 * (so_vtx_count) which threads are allowed to emit streamout data. */
1945 ac_build_ifcc(&ctx->ac, can_emit, 6501);
1946 {
1947 /* The buffer offset is computed as follows:
1948 * ByteOffset = streamout_offset[buffer_id]*4 +
1949 * (streamout_write_index + thread_id)*stride[buffer_id] +
1950 * attrib_offset
1951 */
1952
1953 LLVMValueRef so_write_index =
1954 ac_get_arg(&ctx->ac,
1955 ctx->streamout_write_index);
1956
1957 /* Compute (streamout_write_index + thread_id). */
1958 so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
1959
1960 /* Load the descriptor and compute the write offset for each
1961 * enabled buffer. */
1962 LLVMValueRef so_write_offset[4] = {};
1963 LLVMValueRef so_buffers[4];
1964 LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac,
1965 ctx->rw_buffers);
1966
1967 for (i = 0; i < 4; i++) {
1968 if (!so->stride[i])
1969 continue;
1970
1971 LLVMValueRef offset = LLVMConstInt(ctx->i32,
1972 SI_VS_STREAMOUT_BUF0 + i, 0);
1973
1974 so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
1975
1976 LLVMValueRef so_offset = ac_get_arg(&ctx->ac,
1977 ctx->streamout_offset[i]);
1978 so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
1979
1980 so_write_offset[i] = ac_build_imad(&ctx->ac, so_write_index,
1981 LLVMConstInt(ctx->i32, so->stride[i]*4, 0),
1982 so_offset);
1983 }
1984
1985 /* Write streamout data. */
1986 for (i = 0; i < so->num_outputs; i++) {
1987 unsigned reg = so->output[i].register_index;
1988
1989 if (reg >= noutput)
1990 continue;
1991
1992 if (stream != so->output[i].stream)
1993 continue;
1994
1995 si_emit_streamout_output(ctx, so_buffers, so_write_offset,
1996 &so->output[i], &outputs[reg]);
1997 }
1998 }
1999 ac_build_endif(&ctx->ac, 6501);
2000 }
2001
2002 static void si_export_param(struct si_shader_context *ctx, unsigned index,
2003 LLVMValueRef *values)
2004 {
2005 struct ac_export_args args;
2006
2007 si_llvm_init_export_args(ctx, values,
2008 V_008DFC_SQ_EXP_PARAM + index, &args);
2009 ac_build_export(&ctx->ac, &args);
2010 }
2011
2012 static void si_build_param_exports(struct si_shader_context *ctx,
2013 struct si_shader_output_values *outputs,
2014 unsigned noutput)
2015 {
2016 struct si_shader *shader = ctx->shader;
2017 unsigned param_count = 0;
2018
2019 for (unsigned i = 0; i < noutput; i++) {
2020 unsigned semantic_name = outputs[i].semantic_name;
2021 unsigned semantic_index = outputs[i].semantic_index;
2022
2023 if (outputs[i].vertex_stream[0] != 0 &&
2024 outputs[i].vertex_stream[1] != 0 &&
2025 outputs[i].vertex_stream[2] != 0 &&
2026 outputs[i].vertex_stream[3] != 0)
2027 continue;
2028
2029 switch (semantic_name) {
2030 case TGSI_SEMANTIC_LAYER:
2031 case TGSI_SEMANTIC_VIEWPORT_INDEX:
2032 case TGSI_SEMANTIC_CLIPDIST:
2033 case TGSI_SEMANTIC_COLOR:
2034 case TGSI_SEMANTIC_BCOLOR:
2035 case TGSI_SEMANTIC_PRIMID:
2036 case TGSI_SEMANTIC_FOG:
2037 case TGSI_SEMANTIC_TEXCOORD:
2038 case TGSI_SEMANTIC_GENERIC:
2039 break;
2040 default:
2041 continue;
2042 }
2043
2044 if ((semantic_name != TGSI_SEMANTIC_GENERIC ||
2045 semantic_index < SI_MAX_IO_GENERIC) &&
2046 shader->key.opt.kill_outputs &
2047 (1ull << si_shader_io_get_unique_index(semantic_name,
2048 semantic_index, true)))
2049 continue;
2050
2051 si_export_param(ctx, param_count, outputs[i].values);
2052
2053 assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
2054 shader->info.vs_output_param_offset[i] = param_count++;
2055 }
2056
2057 shader->info.nr_param_exports = param_count;
2058 }
2059
2060 /**
2061 * Vertex color clamping.
2062 *
2063 * This uses a state constant loaded in a user data SGPR and
2064 * an IF statement is added that clamps all colors if the constant
2065 * is true.
2066 */
2067 static void si_vertex_color_clamping(struct si_shader_context *ctx,
2068 struct si_shader_output_values *outputs,
2069 unsigned noutput)
2070 {
2071 LLVMValueRef addr[SI_MAX_VS_OUTPUTS][4];
2072 bool has_colors = false;
2073
2074 /* Store original colors to alloca variables. */
2075 for (unsigned i = 0; i < noutput; i++) {
2076 if (outputs[i].semantic_name != TGSI_SEMANTIC_COLOR &&
2077 outputs[i].semantic_name != TGSI_SEMANTIC_BCOLOR)
2078 continue;
2079
2080 for (unsigned j = 0; j < 4; j++) {
2081 addr[i][j] = ac_build_alloca_undef(&ctx->ac, ctx->f32, "");
2082 LLVMBuildStore(ctx->ac.builder, outputs[i].values[j], addr[i][j]);
2083 }
2084 has_colors = true;
2085 }
2086
2087 if (!has_colors)
2088 return;
2089
2090 /* The state is in the first bit of the user SGPR. */
2091 LLVMValueRef cond = ac_get_arg(&ctx->ac, ctx->vs_state_bits);
2092 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->i1, "");
2093
2094 ac_build_ifcc(&ctx->ac, cond, 6502);
2095
2096 /* Store clamped colors to alloca variables within the conditional block. */
2097 for (unsigned i = 0; i < noutput; i++) {
2098 if (outputs[i].semantic_name != TGSI_SEMANTIC_COLOR &&
2099 outputs[i].semantic_name != TGSI_SEMANTIC_BCOLOR)
2100 continue;
2101
2102 for (unsigned j = 0; j < 4; j++) {
2103 LLVMBuildStore(ctx->ac.builder,
2104 ac_build_clamp(&ctx->ac, outputs[i].values[j]),
2105 addr[i][j]);
2106 }
2107 }
2108 ac_build_endif(&ctx->ac, 6502);
2109
2110 /* Load clamped colors */
2111 for (unsigned i = 0; i < noutput; i++) {
2112 if (outputs[i].semantic_name != TGSI_SEMANTIC_COLOR &&
2113 outputs[i].semantic_name != TGSI_SEMANTIC_BCOLOR)
2114 continue;
2115
2116 for (unsigned j = 0; j < 4; j++) {
2117 outputs[i].values[j] =
2118 LLVMBuildLoad(ctx->ac.builder, addr[i][j], "");
2119 }
2120 }
2121 }
2122
2123 /* Generate export instructions for hardware VS shader stage or NGG GS stage
2124 * (position and parameter data only).
2125 */
2126 void si_llvm_export_vs(struct si_shader_context *ctx,
2127 struct si_shader_output_values *outputs,
2128 unsigned noutput)
2129 {
2130 struct si_shader *shader = ctx->shader;
2131 struct ac_export_args pos_args[4] = {};
2132 LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL, viewport_index_value = NULL;
2133 unsigned pos_idx;
2134 int i;
2135
2136 si_vertex_color_clamping(ctx, outputs, noutput);
2137
2138 /* Build position exports. */
2139 for (i = 0; i < noutput; i++) {
2140 switch (outputs[i].semantic_name) {
2141 case TGSI_SEMANTIC_POSITION:
2142 si_llvm_init_export_args(ctx, outputs[i].values,
2143 V_008DFC_SQ_EXP_POS, &pos_args[0]);
2144 break;
2145 case TGSI_SEMANTIC_PSIZE:
2146 psize_value = outputs[i].values[0];
2147 break;
2148 case TGSI_SEMANTIC_LAYER:
2149 layer_value = outputs[i].values[0];
2150 break;
2151 case TGSI_SEMANTIC_VIEWPORT_INDEX:
2152 viewport_index_value = outputs[i].values[0];
2153 break;
2154 case TGSI_SEMANTIC_EDGEFLAG:
2155 edgeflag_value = outputs[i].values[0];
2156 break;
2157 case TGSI_SEMANTIC_CLIPDIST:
2158 if (!shader->key.opt.clip_disable) {
2159 unsigned index = 2 + outputs[i].semantic_index;
2160 si_llvm_init_export_args(ctx, outputs[i].values,
2161 V_008DFC_SQ_EXP_POS + index,
2162 &pos_args[index]);
2163 }
2164 break;
2165 case TGSI_SEMANTIC_CLIPVERTEX:
2166 if (!shader->key.opt.clip_disable) {
2167 si_llvm_emit_clipvertex(ctx, pos_args,
2168 outputs[i].values);
2169 }
2170 break;
2171 }
2172 }
2173
2174 /* We need to add the position output manually if it's missing. */
2175 if (!pos_args[0].out[0]) {
2176 pos_args[0].enabled_channels = 0xf; /* writemask */
2177 pos_args[0].valid_mask = 0; /* EXEC mask */
2178 pos_args[0].done = 0; /* last export? */
2179 pos_args[0].target = V_008DFC_SQ_EXP_POS;
2180 pos_args[0].compr = 0; /* COMPR flag */
2181 pos_args[0].out[0] = ctx->ac.f32_0; /* X */
2182 pos_args[0].out[1] = ctx->ac.f32_0; /* Y */
2183 pos_args[0].out[2] = ctx->ac.f32_0; /* Z */
2184 pos_args[0].out[3] = ctx->ac.f32_1; /* W */
2185 }
2186
2187 bool pos_writes_edgeflag = shader->selector->info.writes_edgeflag &&
2188 !shader->key.as_ngg;
2189
2190 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2191 if (shader->selector->info.writes_psize ||
2192 pos_writes_edgeflag ||
2193 shader->selector->info.writes_viewport_index ||
2194 shader->selector->info.writes_layer) {
2195 pos_args[1].enabled_channels = shader->selector->info.writes_psize |
2196 (pos_writes_edgeflag << 1) |
2197 (shader->selector->info.writes_layer << 2);
2198
2199 pos_args[1].valid_mask = 0; /* EXEC mask */
2200 pos_args[1].done = 0; /* last export? */
2201 pos_args[1].target = V_008DFC_SQ_EXP_POS + 1;
2202 pos_args[1].compr = 0; /* COMPR flag */
2203 pos_args[1].out[0] = ctx->ac.f32_0; /* X */
2204 pos_args[1].out[1] = ctx->ac.f32_0; /* Y */
2205 pos_args[1].out[2] = ctx->ac.f32_0; /* Z */
2206 pos_args[1].out[3] = ctx->ac.f32_0; /* W */
2207
2208 if (shader->selector->info.writes_psize)
2209 pos_args[1].out[0] = psize_value;
2210
2211 if (pos_writes_edgeflag) {
2212 /* The output is a float, but the hw expects an integer
2213 * with the first bit containing the edge flag. */
2214 edgeflag_value = LLVMBuildFPToUI(ctx->ac.builder,
2215 edgeflag_value,
2216 ctx->i32, "");
2217 edgeflag_value = ac_build_umin(&ctx->ac,
2218 edgeflag_value,
2219 ctx->i32_1);
2220
2221 /* The LLVM intrinsic expects a float. */
2222 pos_args[1].out[1] = ac_to_float(&ctx->ac, edgeflag_value);
2223 }
2224
2225 if (ctx->screen->info.chip_class >= GFX9) {
2226 /* GFX9 has the layer in out.z[10:0] and the viewport
2227 * index in out.z[19:16].
2228 */
2229 if (shader->selector->info.writes_layer)
2230 pos_args[1].out[2] = layer_value;
2231
2232 if (shader->selector->info.writes_viewport_index) {
2233 LLVMValueRef v = viewport_index_value;
2234
2235 v = ac_to_integer(&ctx->ac, v);
2236 v = LLVMBuildShl(ctx->ac.builder, v,
2237 LLVMConstInt(ctx->i32, 16, 0), "");
2238 v = LLVMBuildOr(ctx->ac.builder, v,
2239 ac_to_integer(&ctx->ac, pos_args[1].out[2]), "");
2240 pos_args[1].out[2] = ac_to_float(&ctx->ac, v);
2241 pos_args[1].enabled_channels |= 1 << 2;
2242 }
2243 } else {
2244 if (shader->selector->info.writes_layer)
2245 pos_args[1].out[2] = layer_value;
2246
2247 if (shader->selector->info.writes_viewport_index) {
2248 pos_args[1].out[3] = viewport_index_value;
2249 pos_args[1].enabled_channels |= 1 << 3;
2250 }
2251 }
2252 }
2253
2254 for (i = 0; i < 4; i++)
2255 if (pos_args[i].out[0])
2256 shader->info.nr_pos_exports++;
2257
2258 /* Navi10-14 skip POS0 exports if EXEC=0 and DONE=0, causing a hang.
2259 * Setting valid_mask=1 prevents it and has no other effect.
2260 */
2261 if (ctx->screen->info.family == CHIP_NAVI10 ||
2262 ctx->screen->info.family == CHIP_NAVI12 ||
2263 ctx->screen->info.family == CHIP_NAVI14)
2264 pos_args[0].valid_mask = 1;
2265
2266 pos_idx = 0;
2267 for (i = 0; i < 4; i++) {
2268 if (!pos_args[i].out[0])
2269 continue;
2270
2271 /* Specify the target we are exporting */
2272 pos_args[i].target = V_008DFC_SQ_EXP_POS + pos_idx++;
2273
2274 if (pos_idx == shader->info.nr_pos_exports)
2275 /* Specify that this is the last export */
2276 pos_args[i].done = 1;
2277
2278 ac_build_export(&ctx->ac, &pos_args[i]);
2279 }
2280
2281 /* Build parameter exports. */
2282 si_build_param_exports(ctx, outputs, noutput);
2283 }
2284
2285 /**
2286 * Forward all outputs from the vertex shader to the TES. This is only used
2287 * for the fixed function TCS.
2288 */
2289 static void si_copy_tcs_inputs(struct si_shader_context *ctx)
2290 {
2291 LLVMValueRef invocation_id, buffer, buffer_offset;
2292 LLVMValueRef lds_vertex_stride, lds_base;
2293 uint64_t inputs;
2294
2295 invocation_id = si_unpack_param(ctx, ctx->args.tcs_rel_ids, 8, 5);
2296 buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
2297 buffer_offset = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
2298
2299 lds_vertex_stride = get_tcs_in_vertex_dw_stride(ctx);
2300 lds_base = get_tcs_in_current_patch_offset(ctx);
2301 lds_base = ac_build_imad(&ctx->ac, invocation_id, lds_vertex_stride,
2302 lds_base);
2303
2304 inputs = ctx->shader->key.mono.u.ff_tcs_inputs_to_copy;
2305 while (inputs) {
2306 unsigned i = u_bit_scan64(&inputs);
2307
2308 LLVMValueRef lds_ptr = LLVMBuildAdd(ctx->ac.builder, lds_base,
2309 LLVMConstInt(ctx->i32, 4 * i, 0),
2310 "");
2311
2312 LLVMValueRef buffer_addr = get_tcs_tes_buffer_address(ctx,
2313 get_rel_patch_id(ctx),
2314 invocation_id,
2315 LLVMConstInt(ctx->i32, i, 0));
2316
2317 LLVMValueRef value = lshs_lds_load(ctx, ctx->ac.i32, ~0, lds_ptr);
2318
2319 ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buffer_addr,
2320 buffer_offset, 0, ac_glc);
2321 }
2322 }
2323
2324 static void si_write_tess_factors(struct si_shader_context *ctx,
2325 LLVMValueRef rel_patch_id,
2326 LLVMValueRef invocation_id,
2327 LLVMValueRef tcs_out_current_patch_data_offset,
2328 LLVMValueRef invoc0_tf_outer[4],
2329 LLVMValueRef invoc0_tf_inner[2])
2330 {
2331 struct si_shader *shader = ctx->shader;
2332 unsigned tess_inner_index, tess_outer_index;
2333 LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
2334 LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
2335 unsigned stride, outer_comps, inner_comps, i, offset;
2336
2337 /* Add a barrier before loading tess factors from LDS. */
2338 if (!shader->key.part.tcs.epilog.invoc0_tess_factors_are_def)
2339 si_llvm_emit_barrier(ctx);
2340
2341 /* Do this only for invocation 0, because the tess levels are per-patch,
2342 * not per-vertex.
2343 *
2344 * This can't jump, because invocation 0 executes this. It should
2345 * at least mask out the loads and stores for other invocations.
2346 */
2347 ac_build_ifcc(&ctx->ac,
2348 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
2349 invocation_id, ctx->i32_0, ""), 6503);
2350
2351 /* Determine the layout of one tess factor element in the buffer. */
2352 switch (shader->key.part.tcs.epilog.prim_mode) {
2353 case PIPE_PRIM_LINES:
2354 stride = 2; /* 2 dwords, 1 vec2 store */
2355 outer_comps = 2;
2356 inner_comps = 0;
2357 break;
2358 case PIPE_PRIM_TRIANGLES:
2359 stride = 4; /* 4 dwords, 1 vec4 store */
2360 outer_comps = 3;
2361 inner_comps = 1;
2362 break;
2363 case PIPE_PRIM_QUADS:
2364 stride = 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2365 outer_comps = 4;
2366 inner_comps = 2;
2367 break;
2368 default:
2369 assert(0);
2370 return;
2371 }
2372
2373 for (i = 0; i < 4; i++) {
2374 inner[i] = LLVMGetUndef(ctx->i32);
2375 outer[i] = LLVMGetUndef(ctx->i32);
2376 }
2377
2378 if (shader->key.part.tcs.epilog.invoc0_tess_factors_are_def) {
2379 /* Tess factors are in VGPRs. */
2380 for (i = 0; i < outer_comps; i++)
2381 outer[i] = out[i] = invoc0_tf_outer[i];
2382 for (i = 0; i < inner_comps; i++)
2383 inner[i] = out[outer_comps+i] = invoc0_tf_inner[i];
2384 } else {
2385 /* Load tess_inner and tess_outer from LDS.
2386 * Any invocation can write them, so we can't get them from a temporary.
2387 */
2388 tess_inner_index = si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER, 0);
2389 tess_outer_index = si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER, 0);
2390
2391 lds_base = tcs_out_current_patch_data_offset;
2392 lds_inner = LLVMBuildAdd(ctx->ac.builder, lds_base,
2393 LLVMConstInt(ctx->i32,
2394 tess_inner_index * 4, 0), "");
2395 lds_outer = LLVMBuildAdd(ctx->ac.builder, lds_base,
2396 LLVMConstInt(ctx->i32,
2397 tess_outer_index * 4, 0), "");
2398
2399 for (i = 0; i < outer_comps; i++) {
2400 outer[i] = out[i] =
2401 lshs_lds_load(ctx, ctx->ac.i32, i, lds_outer);
2402 }
2403 for (i = 0; i < inner_comps; i++) {
2404 inner[i] = out[outer_comps+i] =
2405 lshs_lds_load(ctx, ctx->ac.i32, i, lds_inner);
2406 }
2407 }
2408
2409 if (shader->key.part.tcs.epilog.prim_mode == PIPE_PRIM_LINES) {
2410 /* For isolines, the hardware expects tess factors in the
2411 * reverse order from what NIR specifies.
2412 */
2413 LLVMValueRef tmp = out[0];
2414 out[0] = out[1];
2415 out[1] = tmp;
2416 }
2417
2418 /* Convert the outputs to vectors for stores. */
2419 vec0 = ac_build_gather_values(&ctx->ac, out, MIN2(stride, 4));
2420 vec1 = NULL;
2421
2422 if (stride > 4)
2423 vec1 = ac_build_gather_values(&ctx->ac, out+4, stride - 4);
2424
2425 /* Get the buffer. */
2426 buffer = get_tess_ring_descriptor(ctx, TCS_FACTOR_RING);
2427
2428 /* Get the offset. */
2429 tf_base = ac_get_arg(&ctx->ac,
2430 ctx->tcs_factor_offset);
2431 byteoffset = LLVMBuildMul(ctx->ac.builder, rel_patch_id,
2432 LLVMConstInt(ctx->i32, 4 * stride, 0), "");
2433
2434 ac_build_ifcc(&ctx->ac,
2435 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
2436 rel_patch_id, ctx->i32_0, ""), 6504);
2437
2438 /* Store the dynamic HS control word. */
2439 offset = 0;
2440 if (ctx->screen->info.chip_class <= GFX8) {
2441 ac_build_buffer_store_dword(&ctx->ac, buffer,
2442 LLVMConstInt(ctx->i32, 0x80000000, 0),
2443 1, ctx->i32_0, tf_base,
2444 offset, ac_glc);
2445 offset += 4;
2446 }
2447
2448 ac_build_endif(&ctx->ac, 6504);
2449
2450 /* Store the tessellation factors. */
2451 ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
2452 MIN2(stride, 4), byteoffset, tf_base,
2453 offset, ac_glc);
2454 offset += 16;
2455 if (vec1)
2456 ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
2457 stride - 4, byteoffset, tf_base,
2458 offset, ac_glc);
2459
2460 /* Store the tess factors into the offchip buffer if TES reads them. */
2461 if (shader->key.part.tcs.epilog.tes_reads_tess_factors) {
2462 LLVMValueRef buf, base, inner_vec, outer_vec, tf_outer_offset;
2463 LLVMValueRef tf_inner_offset;
2464 unsigned param_outer, param_inner;
2465
2466 buf = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
2467 base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
2468
2469 param_outer = si_shader_io_get_unique_index_patch(
2470 TGSI_SEMANTIC_TESSOUTER, 0);
2471 tf_outer_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
2472 LLVMConstInt(ctx->i32, param_outer, 0));
2473
2474 unsigned outer_vec_size =
2475 ac_has_vec3_support(ctx->screen->info.chip_class, false) ?
2476 outer_comps : util_next_power_of_two(outer_comps);
2477 outer_vec = ac_build_gather_values(&ctx->ac, outer, outer_vec_size);
2478
2479 ac_build_buffer_store_dword(&ctx->ac, buf, outer_vec,
2480 outer_comps, tf_outer_offset,
2481 base, 0, ac_glc);
2482 if (inner_comps) {
2483 param_inner = si_shader_io_get_unique_index_patch(
2484 TGSI_SEMANTIC_TESSINNER, 0);
2485 tf_inner_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
2486 LLVMConstInt(ctx->i32, param_inner, 0));
2487
2488 inner_vec = inner_comps == 1 ? inner[0] :
2489 ac_build_gather_values(&ctx->ac, inner, inner_comps);
2490 ac_build_buffer_store_dword(&ctx->ac, buf, inner_vec,
2491 inner_comps, tf_inner_offset,
2492 base, 0, ac_glc);
2493 }
2494 }
2495
2496 ac_build_endif(&ctx->ac, 6503);
2497 }
2498
2499 static LLVMValueRef
2500 si_insert_input_ret(struct si_shader_context *ctx, LLVMValueRef ret,
2501 struct ac_arg param, unsigned return_index)
2502 {
2503 return LLVMBuildInsertValue(ctx->ac.builder, ret,
2504 ac_get_arg(&ctx->ac, param),
2505 return_index, "");
2506 }
2507
2508 static LLVMValueRef
2509 si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueRef ret,
2510 struct ac_arg param, unsigned return_index)
2511 {
2512 LLVMBuilderRef builder = ctx->ac.builder;
2513 LLVMValueRef p = ac_get_arg(&ctx->ac, param);
2514
2515 return LLVMBuildInsertValue(builder, ret,
2516 ac_to_float(&ctx->ac, p),
2517 return_index, "");
2518 }
2519
2520 static LLVMValueRef
2521 si_insert_input_ptr(struct si_shader_context *ctx, LLVMValueRef ret,
2522 struct ac_arg param, unsigned return_index)
2523 {
2524 LLVMBuilderRef builder = ctx->ac.builder;
2525 LLVMValueRef ptr = ac_get_arg(&ctx->ac, param);
2526 ptr = LLVMBuildPtrToInt(builder, ptr, ctx->i32, "");
2527 return LLVMBuildInsertValue(builder, ret, ptr, return_index, "");
2528 }
2529
2530 /* This only writes the tessellation factor levels. */
2531 static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi *abi,
2532 unsigned max_outputs,
2533 LLVMValueRef *addrs)
2534 {
2535 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
2536 LLVMBuilderRef builder = ctx->ac.builder;
2537 LLVMValueRef rel_patch_id, invocation_id, tf_lds_offset;
2538
2539 si_copy_tcs_inputs(ctx);
2540
2541 rel_patch_id = get_rel_patch_id(ctx);
2542 invocation_id = si_unpack_param(ctx, ctx->args.tcs_rel_ids, 8, 5);
2543 tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
2544
2545 if (ctx->screen->info.chip_class >= GFX9) {
2546 LLVMBasicBlockRef blocks[2] = {
2547 LLVMGetInsertBlock(builder),
2548 ctx->merged_wrap_if_entry_block
2549 };
2550 LLVMValueRef values[2];
2551
2552 ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
2553
2554 values[0] = rel_patch_id;
2555 values[1] = LLVMGetUndef(ctx->i32);
2556 rel_patch_id = ac_build_phi(&ctx->ac, ctx->i32, 2, values, blocks);
2557
2558 values[0] = tf_lds_offset;
2559 values[1] = LLVMGetUndef(ctx->i32);
2560 tf_lds_offset = ac_build_phi(&ctx->ac, ctx->i32, 2, values, blocks);
2561
2562 values[0] = invocation_id;
2563 values[1] = ctx->i32_1; /* cause the epilog to skip threads */
2564 invocation_id = ac_build_phi(&ctx->ac, ctx->i32, 2, values, blocks);
2565 }
2566
2567 /* Return epilog parameters from this function. */
2568 LLVMValueRef ret = ctx->return_value;
2569 unsigned vgpr;
2570
2571 if (ctx->screen->info.chip_class >= GFX9) {
2572 ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_layout,
2573 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
2574 ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_layout,
2575 8 + GFX9_SGPR_TCS_OUT_LAYOUT);
2576 /* Tess offchip and tess factor offsets are at the beginning. */
2577 ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_offset, 2);
2578 ret = si_insert_input_ret(ctx, ret, ctx->tcs_factor_offset, 4);
2579 vgpr = 8 + GFX9_SGPR_TCS_OUT_LAYOUT + 1;
2580 } else {
2581 ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_layout,
2582 GFX6_SGPR_TCS_OFFCHIP_LAYOUT);
2583 ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_layout,
2584 GFX6_SGPR_TCS_OUT_LAYOUT);
2585 /* Tess offchip and tess factor offsets are after user SGPRs. */
2586 ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_offset,
2587 GFX6_TCS_NUM_USER_SGPR);
2588 ret = si_insert_input_ret(ctx, ret, ctx->tcs_factor_offset,
2589 GFX6_TCS_NUM_USER_SGPR + 1);
2590 vgpr = GFX6_TCS_NUM_USER_SGPR + 2;
2591 }
2592
2593 /* VGPRs */
2594 rel_patch_id = ac_to_float(&ctx->ac, rel_patch_id);
2595 invocation_id = ac_to_float(&ctx->ac, invocation_id);
2596 tf_lds_offset = ac_to_float(&ctx->ac, tf_lds_offset);
2597
2598 /* Leave a hole corresponding to the two input VGPRs. This ensures that
2599 * the invocation_id output does not alias the tcs_rel_ids input,
2600 * which saves a V_MOV on gfx9.
2601 */
2602 vgpr += 2;
2603
2604 ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
2605 ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");
2606
2607 if (ctx->shader->selector->info.tessfactors_are_def_in_all_invocs) {
2608 vgpr++; /* skip the tess factor LDS offset */
2609 for (unsigned i = 0; i < 6; i++) {
2610 LLVMValueRef value =
2611 LLVMBuildLoad(builder, ctx->invoc0_tess_factors[i], "");
2612 value = ac_to_float(&ctx->ac, value);
2613 ret = LLVMBuildInsertValue(builder, ret, value, vgpr++, "");
2614 }
2615 } else {
2616 ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
2617 }
2618 ctx->return_value = ret;
2619 }
2620
2621 /* Pass TCS inputs from LS to TCS on GFX9. */
2622 static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
2623 {
2624 LLVMValueRef ret = ctx->return_value;
2625
2626 ret = si_insert_input_ptr(ctx, ret, ctx->other_const_and_shader_buffers, 0);
2627 ret = si_insert_input_ptr(ctx, ret, ctx->other_samplers_and_images, 1);
2628 ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_offset, 2);
2629 ret = si_insert_input_ret(ctx, ret, ctx->merged_wave_info, 3);
2630 ret = si_insert_input_ret(ctx, ret, ctx->tcs_factor_offset, 4);
2631 ret = si_insert_input_ret(ctx, ret, ctx->merged_scratch_offset, 5);
2632
2633 ret = si_insert_input_ptr(ctx, ret, ctx->rw_buffers,
2634 8 + SI_SGPR_RW_BUFFERS);
2635 ret = si_insert_input_ptr(ctx, ret,
2636 ctx->bindless_samplers_and_images,
2637 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
2638
2639 ret = si_insert_input_ret(ctx, ret, ctx->vs_state_bits,
2640 8 + SI_SGPR_VS_STATE_BITS);
2641
2642 ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_layout,
2643 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
2644 ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_offsets,
2645 8 + GFX9_SGPR_TCS_OUT_OFFSETS);
2646 ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_layout,
2647 8 + GFX9_SGPR_TCS_OUT_LAYOUT);
2648
2649 unsigned vgpr = 8 + GFX9_TCS_NUM_USER_SGPR;
2650 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
2651 ac_to_float(&ctx->ac,
2652 ac_get_arg(&ctx->ac, ctx->args.tcs_patch_id)),
2653 vgpr++, "");
2654 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
2655 ac_to_float(&ctx->ac,
2656 ac_get_arg(&ctx->ac, ctx->args.tcs_rel_ids)),
2657 vgpr++, "");
2658 ctx->return_value = ret;
2659 }
2660
2661 /* Pass GS inputs from ES to GS on GFX9. */
2662 static void si_set_es_return_value_for_gs(struct si_shader_context *ctx)
2663 {
2664 LLVMValueRef ret = ctx->return_value;
2665
2666 ret = si_insert_input_ptr(ctx, ret, ctx->other_const_and_shader_buffers, 0);
2667 ret = si_insert_input_ptr(ctx, ret, ctx->other_samplers_and_images, 1);
2668 if (ctx->shader->key.as_ngg)
2669 ret = si_insert_input_ptr(ctx, ret, ctx->gs_tg_info, 2);
2670 else
2671 ret = si_insert_input_ret(ctx, ret, ctx->gs2vs_offset, 2);
2672 ret = si_insert_input_ret(ctx, ret, ctx->merged_wave_info, 3);
2673 ret = si_insert_input_ret(ctx, ret, ctx->merged_scratch_offset, 5);
2674
2675 ret = si_insert_input_ptr(ctx, ret, ctx->rw_buffers,
2676 8 + SI_SGPR_RW_BUFFERS);
2677 ret = si_insert_input_ptr(ctx, ret,
2678 ctx->bindless_samplers_and_images,
2679 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
2680 if (ctx->screen->use_ngg) {
2681 ret = si_insert_input_ptr(ctx, ret, ctx->vs_state_bits,
2682 8 + SI_SGPR_VS_STATE_BITS);
2683 }
2684
2685 unsigned vgpr;
2686 if (ctx->type == PIPE_SHADER_VERTEX)
2687 vgpr = 8 + GFX9_VSGS_NUM_USER_SGPR;
2688 else
2689 vgpr = 8 + GFX9_TESGS_NUM_USER_SGPR;
2690
2691 ret = si_insert_input_ret_float(ctx, ret, ctx->gs_vtx01_offset, vgpr++);
2692 ret = si_insert_input_ret_float(ctx, ret, ctx->gs_vtx23_offset, vgpr++);
2693 ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_prim_id, vgpr++);
2694 ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_invocation_id, vgpr++);
2695 ret = si_insert_input_ret_float(ctx, ret, ctx->gs_vtx45_offset, vgpr++);
2696 ctx->return_value = ret;
2697 }
2698
2699 static void si_llvm_emit_ls_epilogue(struct ac_shader_abi *abi,
2700 unsigned max_outputs,
2701 LLVMValueRef *addrs)
2702 {
2703 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
2704 struct si_shader *shader = ctx->shader;
2705 struct si_shader_info *info = &shader->selector->info;
2706 unsigned i, chan;
2707 LLVMValueRef vertex_id = ac_get_arg(&ctx->ac, ctx->rel_auto_id);
2708 LLVMValueRef vertex_dw_stride = get_tcs_in_vertex_dw_stride(ctx);
2709 LLVMValueRef base_dw_addr = LLVMBuildMul(ctx->ac.builder, vertex_id,
2710 vertex_dw_stride, "");
2711
2712 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2713 * its inputs from it. */
2714 for (i = 0; i < info->num_outputs; i++) {
2715 unsigned name = info->output_semantic_name[i];
2716 unsigned index = info->output_semantic_index[i];
2717
2718 /* The ARB_shader_viewport_layer_array spec contains the
2719 * following issue:
2720 *
2721 * 2) What happens if gl_ViewportIndex or gl_Layer is
2722 * written in the vertex shader and a geometry shader is
2723 * present?
2724 *
2725 * RESOLVED: The value written by the last vertex processing
2726 * stage is used. If the last vertex processing stage
2727 * (vertex, tessellation evaluation or geometry) does not
2728 * statically assign to gl_ViewportIndex or gl_Layer, index
2729 * or layer zero is assumed.
2730 *
2731 * So writes to those outputs in VS-as-LS are simply ignored.
2732 */
2733 if (name == TGSI_SEMANTIC_LAYER ||
2734 name == TGSI_SEMANTIC_VIEWPORT_INDEX)
2735 continue;
2736
2737 int param = si_shader_io_get_unique_index(name, index, false);
2738 LLVMValueRef dw_addr = LLVMBuildAdd(ctx->ac.builder, base_dw_addr,
2739 LLVMConstInt(ctx->i32, param * 4, 0), "");
2740
2741 for (chan = 0; chan < 4; chan++) {
2742 if (!(info->output_usagemask[i] & (1 << chan)))
2743 continue;
2744
2745 lshs_lds_store(ctx, chan, dw_addr,
2746 LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], ""));
2747 }
2748 }
2749
2750 if (ctx->screen->info.chip_class >= GFX9)
2751 si_set_ls_return_value_for_tcs(ctx);
2752 }
2753
2754 static void si_llvm_emit_es_epilogue(struct ac_shader_abi *abi,
2755 unsigned max_outputs,
2756 LLVMValueRef *addrs)
2757 {
2758 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
2759 struct si_shader *es = ctx->shader;
2760 struct si_shader_info *info = &es->selector->info;
2761 LLVMValueRef lds_base = NULL;
2762 unsigned chan;
2763 int i;
2764
2765 if (ctx->screen->info.chip_class >= GFX9 && info->num_outputs) {
2766 unsigned itemsize_dw = es->selector->esgs_itemsize / 4;
2767 LLVMValueRef vertex_idx = ac_get_thread_id(&ctx->ac);
2768 LLVMValueRef wave_idx = si_unpack_param(ctx, ctx->merged_wave_info, 24, 4);
2769 vertex_idx = LLVMBuildOr(ctx->ac.builder, vertex_idx,
2770 LLVMBuildMul(ctx->ac.builder, wave_idx,
2771 LLVMConstInt(ctx->i32, ctx->ac.wave_size, false), ""), "");
2772 lds_base = LLVMBuildMul(ctx->ac.builder, vertex_idx,
2773 LLVMConstInt(ctx->i32, itemsize_dw, 0), "");
2774 }
2775
2776 for (i = 0; i < info->num_outputs; i++) {
2777 int param;
2778
2779 if (info->output_semantic_name[i] == TGSI_SEMANTIC_VIEWPORT_INDEX ||
2780 info->output_semantic_name[i] == TGSI_SEMANTIC_LAYER)
2781 continue;
2782
2783 param = si_shader_io_get_unique_index(info->output_semantic_name[i],
2784 info->output_semantic_index[i], false);
2785
2786 for (chan = 0; chan < 4; chan++) {
2787 if (!(info->output_usagemask[i] & (1 << chan)))
2788 continue;
2789
2790 LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
2791 out_val = ac_to_integer(&ctx->ac, out_val);
2792
2793 /* GFX9 has the ESGS ring in LDS. */
2794 if (ctx->screen->info.chip_class >= GFX9) {
2795 LLVMValueRef idx = LLVMConstInt(ctx->i32, param * 4 + chan, false);
2796 idx = LLVMBuildAdd(ctx->ac.builder, lds_base, idx, "");
2797 ac_build_indexed_store(&ctx->ac, ctx->esgs_ring, idx, out_val);
2798 continue;
2799 }
2800
2801 ac_build_buffer_store_dword(&ctx->ac,
2802 ctx->esgs_ring,
2803 out_val, 1, NULL,
2804 ac_get_arg(&ctx->ac, ctx->es2gs_offset),
2805 (4 * param + chan) * 4,
2806 ac_glc | ac_slc | ac_swizzled);
2807 }
2808 }
2809
2810 if (ctx->screen->info.chip_class >= GFX9)
2811 si_set_es_return_value_for_gs(ctx);
2812 }
2813
2814 static LLVMValueRef si_get_gs_wave_id(struct si_shader_context *ctx)
2815 {
2816 if (ctx->screen->info.chip_class >= GFX9)
2817 return si_unpack_param(ctx, ctx->merged_wave_info, 16, 8);
2818 else
2819 return ac_get_arg(&ctx->ac, ctx->gs_wave_id);
2820 }
2821
2822 static void emit_gs_epilogue(struct si_shader_context *ctx)
2823 {
2824 if (ctx->shader->key.as_ngg) {
2825 gfx10_ngg_gs_emit_epilogue(ctx);
2826 return;
2827 }
2828
2829 if (ctx->screen->info.chip_class >= GFX10)
2830 LLVMBuildFence(ctx->ac.builder, LLVMAtomicOrderingRelease, false, "");
2831
2832 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
2833 si_get_gs_wave_id(ctx));
2834
2835 if (ctx->screen->info.chip_class >= GFX9)
2836 ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
2837 }
2838
2839 static void si_llvm_emit_gs_epilogue(struct ac_shader_abi *abi,
2840 unsigned max_outputs,
2841 LLVMValueRef *addrs)
2842 {
2843 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
2844 struct si_shader_info UNUSED *info = &ctx->shader->selector->info;
2845
2846 assert(info->num_outputs <= max_outputs);
2847
2848 emit_gs_epilogue(ctx);
2849 }
2850
2851 static void si_llvm_emit_vs_epilogue(struct ac_shader_abi *abi,
2852 unsigned max_outputs,
2853 LLVMValueRef *addrs)
2854 {
2855 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
2856 struct si_shader_info *info = &ctx->shader->selector->info;
2857 struct si_shader_output_values *outputs = NULL;
2858 int i,j;
2859
2860 assert(!ctx->shader->is_gs_copy_shader);
2861 assert(info->num_outputs <= max_outputs);
2862
2863 outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
2864
2865 for (i = 0; i < info->num_outputs; i++) {
2866 outputs[i].semantic_name = info->output_semantic_name[i];
2867 outputs[i].semantic_index = info->output_semantic_index[i];
2868
2869 for (j = 0; j < 4; j++) {
2870 outputs[i].values[j] =
2871 LLVMBuildLoad(ctx->ac.builder,
2872 addrs[4 * i + j],
2873 "");
2874 outputs[i].vertex_stream[j] =
2875 (info->output_streams[i] >> (2 * j)) & 3;
2876 }
2877 }
2878
2879 if (!ctx->screen->use_ngg_streamout &&
2880 ctx->shader->selector->so.num_outputs)
2881 si_llvm_emit_streamout(ctx, outputs, i, 0);
2882
2883 /* Export PrimitiveID. */
2884 if (ctx->shader->key.mono.u.vs_export_prim_id) {
2885 outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
2886 outputs[i].semantic_index = 0;
2887 outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
2888 for (j = 1; j < 4; j++)
2889 outputs[i].values[j] = LLVMConstReal(ctx->f32, 0);
2890
2891 memset(outputs[i].vertex_stream, 0,
2892 sizeof(outputs[i].vertex_stream));
2893 i++;
2894 }
2895
2896 si_llvm_export_vs(ctx, outputs, i);
2897 FREE(outputs);
2898 }
2899
2900 static void si_llvm_emit_prim_discard_cs_epilogue(struct ac_shader_abi *abi,
2901 unsigned max_outputs,
2902 LLVMValueRef *addrs)
2903 {
2904 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
2905 struct si_shader_info *info = &ctx->shader->selector->info;
2906 LLVMValueRef pos[4] = {};
2907
2908 assert(info->num_outputs <= max_outputs);
2909
2910 for (unsigned i = 0; i < info->num_outputs; i++) {
2911 if (info->output_semantic_name[i] != TGSI_SEMANTIC_POSITION)
2912 continue;
2913
2914 for (unsigned chan = 0; chan < 4; chan++)
2915 pos[chan] = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
2916 break;
2917 }
2918 assert(pos[0] != NULL);
2919
2920 /* Return the position output. */
2921 LLVMValueRef ret = ctx->return_value;
2922 for (unsigned chan = 0; chan < 4; chan++)
2923 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, pos[chan], chan, "");
2924 ctx->return_value = ret;
2925 }
2926
2927 struct si_ps_exports {
2928 unsigned num;
2929 struct ac_export_args args[10];
2930 };
2931
2932 static void si_export_mrt_z(struct si_shader_context *ctx,
2933 LLVMValueRef depth, LLVMValueRef stencil,
2934 LLVMValueRef samplemask, struct si_ps_exports *exp)
2935 {
2936 struct ac_export_args args;
2937
2938 ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &args);
2939
2940 memcpy(&exp->args[exp->num++], &args, sizeof(args));
2941 }
2942
2943 static void si_export_mrt_color(struct si_shader_context *ctx,
2944 LLVMValueRef *color, unsigned index,
2945 unsigned samplemask_param,
2946 bool is_last, struct si_ps_exports *exp)
2947 {
2948 int i;
2949
2950 /* Clamp color */
2951 if (ctx->shader->key.part.ps.epilog.clamp_color)
2952 for (i = 0; i < 4; i++)
2953 color[i] = ac_build_clamp(&ctx->ac, color[i]);
2954
2955 /* Alpha to one */
2956 if (ctx->shader->key.part.ps.epilog.alpha_to_one)
2957 color[3] = ctx->ac.f32_1;
2958
2959 /* Alpha test */
2960 if (index == 0 &&
2961 ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
2962 si_alpha_test(ctx, color[3]);
2963
2964 /* Line & polygon smoothing */
2965 if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
2966 color[3] = si_scale_alpha_by_sample_mask(ctx, color[3],
2967 samplemask_param);
2968
2969 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2970 if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
2971 struct ac_export_args args[8];
2972 int c, last = -1;
2973
2974 /* Get the export arguments, also find out what the last one is. */
2975 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
2976 si_llvm_init_export_args(ctx, color,
2977 V_008DFC_SQ_EXP_MRT + c, &args[c]);
2978 if (args[c].enabled_channels)
2979 last = c;
2980 }
2981
2982 /* Emit all exports. */
2983 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
2984 if (is_last && last == c) {
2985 args[c].valid_mask = 1; /* whether the EXEC mask is valid */
2986 args[c].done = 1; /* DONE bit */
2987 } else if (!args[c].enabled_channels)
2988 continue; /* unnecessary NULL export */
2989
2990 memcpy(&exp->args[exp->num++], &args[c], sizeof(args[c]));
2991 }
2992 } else {
2993 struct ac_export_args args;
2994
2995 /* Export */
2996 si_llvm_init_export_args(ctx, color, V_008DFC_SQ_EXP_MRT + index,
2997 &args);
2998 if (is_last) {
2999 args.valid_mask = 1; /* whether the EXEC mask is valid */
3000 args.done = 1; /* DONE bit */
3001 } else if (!args.enabled_channels)
3002 return; /* unnecessary NULL export */
3003
3004 memcpy(&exp->args[exp->num++], &args, sizeof(args));
3005 }
3006 }
3007
3008 static void si_emit_ps_exports(struct si_shader_context *ctx,
3009 struct si_ps_exports *exp)
3010 {
3011 for (unsigned i = 0; i < exp->num; i++)
3012 ac_build_export(&ctx->ac, &exp->args[i]);
3013 }
3014
3015 /**
3016 * Return PS outputs in this order:
3017 *
3018 * v[0:3] = color0.xyzw
3019 * v[4:7] = color1.xyzw
3020 * ...
3021 * vN+0 = Depth
3022 * vN+1 = Stencil
3023 * vN+2 = SampleMask
3024 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3025 *
3026 * The alpha-ref SGPR is returned via its original location.
3027 */
3028 static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi,
3029 unsigned max_outputs,
3030 LLVMValueRef *addrs)
3031 {
3032 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
3033 struct si_shader *shader = ctx->shader;
3034 struct si_shader_info *info = &shader->selector->info;
3035 LLVMBuilderRef builder = ctx->ac.builder;
3036 unsigned i, j, first_vgpr, vgpr;
3037
3038 LLVMValueRef color[8][4] = {};
3039 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
3040 LLVMValueRef ret;
3041
3042 if (ctx->postponed_kill)
3043 ac_build_kill_if_false(&ctx->ac, LLVMBuildLoad(builder, ctx->postponed_kill, ""));
3044
3045 /* Read the output values. */
3046 for (i = 0; i < info->num_outputs; i++) {
3047 unsigned semantic_name = info->output_semantic_name[i];
3048 unsigned semantic_index = info->output_semantic_index[i];
3049
3050 switch (semantic_name) {
3051 case TGSI_SEMANTIC_COLOR:
3052 assert(semantic_index < 8);
3053 for (j = 0; j < 4; j++) {
3054 LLVMValueRef ptr = addrs[4 * i + j];
3055 LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
3056 color[semantic_index][j] = result;
3057 }
3058 break;
3059 case TGSI_SEMANTIC_POSITION:
3060 depth = LLVMBuildLoad(builder,
3061 addrs[4 * i + 0], "");
3062 break;
3063 case TGSI_SEMANTIC_STENCIL:
3064 stencil = LLVMBuildLoad(builder,
3065 addrs[4 * i + 0], "");
3066 break;
3067 case TGSI_SEMANTIC_SAMPLEMASK:
3068 samplemask = LLVMBuildLoad(builder,
3069 addrs[4 * i + 0], "");
3070 break;
3071 default:
3072 fprintf(stderr, "Warning: GFX6 unhandled fs output type:%d\n",
3073 semantic_name);
3074 }
3075 }
3076
3077 /* Fill the return structure. */
3078 ret = ctx->return_value;
3079
3080 /* Set SGPRs. */
3081 ret = LLVMBuildInsertValue(builder, ret,
3082 ac_to_integer(&ctx->ac,
3083 LLVMGetParam(ctx->main_fn,
3084 SI_PARAM_ALPHA_REF)),
3085 SI_SGPR_ALPHA_REF, "");
3086
3087 /* Set VGPRs */
3088 first_vgpr = vgpr = SI_SGPR_ALPHA_REF + 1;
3089 for (i = 0; i < ARRAY_SIZE(color); i++) {
3090 if (!color[i][0])
3091 continue;
3092
3093 for (j = 0; j < 4; j++)
3094 ret = LLVMBuildInsertValue(builder, ret, color[i][j], vgpr++, "");
3095 }
3096 if (depth)
3097 ret = LLVMBuildInsertValue(builder, ret, depth, vgpr++, "");
3098 if (stencil)
3099 ret = LLVMBuildInsertValue(builder, ret, stencil, vgpr++, "");
3100 if (samplemask)
3101 ret = LLVMBuildInsertValue(builder, ret, samplemask, vgpr++, "");
3102
3103 /* Add the input sample mask for smoothing at the end. */
3104 if (vgpr < first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC)
3105 vgpr = first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC;
3106 ret = LLVMBuildInsertValue(builder, ret,
3107 LLVMGetParam(ctx->main_fn,
3108 SI_PARAM_SAMPLE_COVERAGE), vgpr++, "");
3109
3110 ctx->return_value = ret;
3111 }
3112
3113 /* Emit one vertex from the geometry shader */
3114 static void si_llvm_emit_vertex(struct ac_shader_abi *abi,
3115 unsigned stream,
3116 LLVMValueRef *addrs)
3117 {
3118 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
3119
3120 if (ctx->shader->key.as_ngg) {
3121 gfx10_ngg_gs_emit_vertex(ctx, stream, addrs);
3122 return;
3123 }
3124
3125 struct si_shader_info *info = &ctx->shader->selector->info;
3126 struct si_shader *shader = ctx->shader;
3127 LLVMValueRef soffset = ac_get_arg(&ctx->ac, ctx->gs2vs_offset);
3128 LLVMValueRef gs_next_vertex;
3129 LLVMValueRef can_emit;
3130 unsigned chan, offset;
3131 int i;
3132
3133 /* Write vertex attribute values to GSVS ring */
3134 gs_next_vertex = LLVMBuildLoad(ctx->ac.builder,
3135 ctx->gs_next_vertex[stream],
3136 "");
3137
3138 /* If this thread has already emitted the declared maximum number of
3139 * vertices, skip the write: excessive vertex emissions are not
3140 * supposed to have any effect.
3141 *
3142 * If the shader has no writes to memory, kill it instead. This skips
3143 * further memory loads and may allow LLVM to skip to the end
3144 * altogether.
3145 */
3146 can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, gs_next_vertex,
3147 LLVMConstInt(ctx->i32,
3148 shader->selector->gs_max_out_vertices, 0), "");
3149
3150 bool use_kill = !info->writes_memory;
3151 if (use_kill) {
3152 ac_build_kill_if_false(&ctx->ac, can_emit);
3153 } else {
3154 ac_build_ifcc(&ctx->ac, can_emit, 6505);
3155 }
3156
3157 offset = 0;
3158 for (i = 0; i < info->num_outputs; i++) {
3159 for (chan = 0; chan < 4; chan++) {
3160 if (!(info->output_usagemask[i] & (1 << chan)) ||
3161 ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
3162 continue;
3163
3164 LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
3165 LLVMValueRef voffset =
3166 LLVMConstInt(ctx->i32, offset *
3167 shader->selector->gs_max_out_vertices, 0);
3168 offset++;
3169
3170 voffset = LLVMBuildAdd(ctx->ac.builder, voffset, gs_next_vertex, "");
3171 voffset = LLVMBuildMul(ctx->ac.builder, voffset,
3172 LLVMConstInt(ctx->i32, 4, 0), "");
3173
3174 out_val = ac_to_integer(&ctx->ac, out_val);
3175
3176 ac_build_buffer_store_dword(&ctx->ac,
3177 ctx->gsvs_ring[stream],
3178 out_val, 1,
3179 voffset, soffset, 0,
3180 ac_glc | ac_slc | ac_swizzled);
3181 }
3182 }
3183
3184 gs_next_vertex = LLVMBuildAdd(ctx->ac.builder, gs_next_vertex, ctx->i32_1, "");
3185 LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
3186
3187 /* Signal vertex emission if vertex data was written. */
3188 if (offset) {
3189 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
3190 si_get_gs_wave_id(ctx));
3191 }
3192
3193 if (!use_kill)
3194 ac_build_endif(&ctx->ac, 6505);
3195 }
3196
3197 /* Cut one primitive from the geometry shader */
3198 static void si_llvm_emit_primitive(struct ac_shader_abi *abi,
3199 unsigned stream)
3200 {
3201 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
3202
3203 if (ctx->shader->key.as_ngg) {
3204 LLVMBuildStore(ctx->ac.builder, ctx->ac.i32_0, ctx->gs_curprim_verts[stream]);
3205 return;
3206 }
3207
3208 /* Signal primitive cut */
3209 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
3210 si_get_gs_wave_id(ctx));
3211 }
3212
3213 static void si_llvm_emit_barrier(struct si_shader_context *ctx)
3214 {
3215 /* GFX6 only (thanks to a hw bug workaround):
3216 * The real barrier instruction isn’t needed, because an entire patch
3217 * always fits into a single wave.
3218 */
3219 if (ctx->screen->info.chip_class == GFX6 &&
3220 ctx->type == PIPE_SHADER_TESS_CTRL) {
3221 ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
3222 return;
3223 }
3224
3225 ac_build_s_barrier(&ctx->ac);
3226 }
3227
3228 void si_create_function(struct si_shader_context *ctx,
3229 const char *name,
3230 LLVMTypeRef *returns, unsigned num_returns,
3231 unsigned max_workgroup_size)
3232 {
3233 si_llvm_create_func(ctx, name, returns, num_returns);
3234 ctx->return_value = LLVMGetUndef(ctx->return_type);
3235
3236 if (ctx->screen->info.address32_hi) {
3237 ac_llvm_add_target_dep_function_attr(ctx->main_fn,
3238 "amdgpu-32bit-address-high-bits",
3239 ctx->screen->info.address32_hi);
3240 }
3241
3242 LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
3243 "no-signed-zeros-fp-math",
3244 "true");
3245
3246 ac_llvm_set_workgroup_size(ctx->main_fn, max_workgroup_size);
3247 }
3248
3249 static void declare_streamout_params(struct si_shader_context *ctx,
3250 struct pipe_stream_output_info *so)
3251 {
3252 if (ctx->screen->use_ngg_streamout) {
3253 if (ctx->type == PIPE_SHADER_TESS_EVAL)
3254 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
3255 return;
3256 }
3257
3258 /* Streamout SGPRs. */
3259 if (so->num_outputs) {
3260 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_config);
3261 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_write_index);
3262 } else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
3263 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
3264 }
3265
3266 /* A streamout buffer offset is loaded if the stride is non-zero. */
3267 for (int i = 0; i < 4; i++) {
3268 if (!so->stride[i])
3269 continue;
3270
3271 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_offset[i]);
3272 }
3273 }
3274
3275 static unsigned si_get_max_workgroup_size(const struct si_shader *shader)
3276 {
3277 switch (shader->selector->type) {
3278 case PIPE_SHADER_VERTEX:
3279 case PIPE_SHADER_TESS_EVAL:
3280 return shader->key.as_ngg ? 128 : 0;
3281
3282 case PIPE_SHADER_TESS_CTRL:
3283 /* Return this so that LLVM doesn't remove s_barrier
3284 * instructions on chips where we use s_barrier. */
3285 return shader->selector->screen->info.chip_class >= GFX7 ? 128 : 0;
3286
3287 case PIPE_SHADER_GEOMETRY:
3288 return shader->selector->screen->info.chip_class >= GFX9 ? 128 : 0;
3289
3290 case PIPE_SHADER_COMPUTE:
3291 break; /* see below */
3292
3293 default:
3294 return 0;
3295 }
3296
3297 const unsigned *properties = shader->selector->info.properties;
3298 unsigned max_work_group_size =
3299 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
3300 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
3301 properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
3302
3303 if (!max_work_group_size) {
3304 /* This is a variable group size compute shader,
3305 * compile it for the maximum possible group size.
3306 */
3307 max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
3308 }
3309 return max_work_group_size;
3310 }
3311
3312 static void declare_const_and_shader_buffers(struct si_shader_context *ctx,
3313 bool assign_params)
3314 {
3315 enum ac_arg_type const_shader_buf_type;
3316
3317 if (ctx->shader->selector->info.const_buffers_declared == 1 &&
3318 ctx->shader->selector->info.shader_buffers_declared == 0)
3319 const_shader_buf_type = AC_ARG_CONST_FLOAT_PTR;
3320 else
3321 const_shader_buf_type = AC_ARG_CONST_DESC_PTR;
3322
3323 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, const_shader_buf_type,
3324 assign_params ? &ctx->const_and_shader_buffers :
3325 &ctx->other_const_and_shader_buffers);
3326 }
3327
3328 static void declare_samplers_and_images(struct si_shader_context *ctx,
3329 bool assign_params)
3330 {
3331 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
3332 assign_params ? &ctx->samplers_and_images :
3333 &ctx->other_samplers_and_images);
3334 }
3335
3336 static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
3337 bool assign_params)
3338 {
3339 declare_const_and_shader_buffers(ctx, assign_params);
3340 declare_samplers_and_images(ctx, assign_params);
3341 }
3342
3343 static void declare_global_desc_pointers(struct si_shader_context *ctx)
3344 {
3345 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
3346 &ctx->rw_buffers);
3347 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
3348 &ctx->bindless_samplers_and_images);
3349 }
3350
3351 static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx)
3352 {
3353 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
3354 if (!ctx->shader->is_gs_copy_shader) {
3355 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.base_vertex);
3356 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.start_instance);
3357 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.draw_id);
3358 }
3359 }
3360
3361 static void declare_vb_descriptor_input_sgprs(struct si_shader_context *ctx)
3362 {
3363 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, &ctx->vertex_buffers);
3364
3365 unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs;
3366 if (num_vbos_in_user_sgprs) {
3367 unsigned user_sgprs = ctx->args.num_sgprs_used;
3368
3369 if (is_merged_shader(ctx))
3370 user_sgprs -= 8;
3371 assert(user_sgprs <= SI_SGPR_VS_VB_DESCRIPTOR_FIRST);
3372
3373 /* Declare unused SGPRs to align VB descriptors to 4 SGPRs (hw requirement). */
3374 for (unsigned i = user_sgprs; i < SI_SGPR_VS_VB_DESCRIPTOR_FIRST; i++)
3375 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
3376
3377 assert(num_vbos_in_user_sgprs <= ARRAY_SIZE(ctx->vb_descriptors));
3378 for (unsigned i = 0; i < num_vbos_in_user_sgprs; i++)
3379 ac_add_arg(&ctx->args, AC_ARG_SGPR, 4, AC_ARG_INT, &ctx->vb_descriptors[i]);
3380 }
3381 }
3382
3383 static void declare_vs_input_vgprs(struct si_shader_context *ctx,
3384 unsigned *num_prolog_vgprs)
3385 {
3386 struct si_shader *shader = ctx->shader;
3387
3388 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.vertex_id);
3389 if (shader->key.as_ls) {
3390 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->rel_auto_id);
3391 if (ctx->screen->info.chip_class >= GFX10) {
3392 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
3393 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
3394 } else {
3395 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
3396 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
3397 }
3398 } else if (ctx->screen->info.chip_class >= GFX10) {
3399 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
3400 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
3401 &ctx->vs_prim_id); /* user vgpr or PrimID (legacy) */
3402 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
3403 } else {
3404 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
3405 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->vs_prim_id);
3406 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
3407 }
3408
3409 if (!shader->is_gs_copy_shader) {
3410 /* Vertex load indices. */
3411 if (shader->selector->info.num_inputs) {
3412 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
3413 &ctx->vertex_index0);
3414 for (unsigned i = 1; i < shader->selector->info.num_inputs; i++)
3415 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL);
3416 }
3417 *num_prolog_vgprs += shader->selector->info.num_inputs;
3418 }
3419 }
3420
3421 static void declare_vs_blit_inputs(struct si_shader_context *ctx,
3422 unsigned vs_blit_property)
3423 {
3424 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
3425 &ctx->vs_blit_inputs); /* i16 x1, y1 */
3426 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* i16 x1, y1 */
3427 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* depth */
3428
3429 if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
3430 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color0 */
3431 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color1 */
3432 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color2 */
3433 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color3 */
3434 } else if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD) {
3435 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x1 */
3436 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y1 */
3437 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x2 */
3438 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y2 */
3439 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.z */
3440 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.w */
3441 }
3442 }
3443
3444 static void declare_tes_input_vgprs(struct si_shader_context *ctx)
3445 {
3446 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_u);
3447 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_v);
3448 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->tes_rel_patch_id);
3449 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tes_patch_id);
3450 }
3451
3452 enum {
3453 /* Convenient merged shader definitions. */
3454 SI_SHADER_MERGED_VERTEX_TESSCTRL = PIPE_SHADER_TYPES,
3455 SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY,
3456 };
3457
3458 static void add_arg_checked(struct ac_shader_args *args,
3459 enum ac_arg_regfile file,
3460 unsigned registers, enum ac_arg_type type,
3461 struct ac_arg *arg,
3462 unsigned idx)
3463 {
3464 assert(args->arg_count == idx);
3465 ac_add_arg(args, file, registers, type, arg);
3466 }
3467
3468 static void create_function(struct si_shader_context *ctx)
3469 {
3470 struct si_shader *shader = ctx->shader;
3471 LLVMTypeRef returns[AC_MAX_ARGS];
3472 unsigned i, num_return_sgprs;
3473 unsigned num_returns = 0;
3474 unsigned num_prolog_vgprs = 0;
3475 unsigned type = ctx->type;
3476 unsigned vs_blit_property =
3477 shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
3478
3479 memset(&ctx->args, 0, sizeof(ctx->args));
3480
3481 /* Set MERGED shaders. */
3482 if (ctx->screen->info.chip_class >= GFX9) {
3483 if (shader->key.as_ls || type == PIPE_SHADER_TESS_CTRL)
3484 type = SI_SHADER_MERGED_VERTEX_TESSCTRL; /* LS or HS */
3485 else if (shader->key.as_es || shader->key.as_ngg || type == PIPE_SHADER_GEOMETRY)
3486 type = SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY;
3487 }
3488
3489 switch (type) {
3490 case PIPE_SHADER_VERTEX:
3491 declare_global_desc_pointers(ctx);
3492
3493 if (vs_blit_property) {
3494 declare_vs_blit_inputs(ctx, vs_blit_property);
3495
3496 /* VGPRs */
3497 declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
3498 break;
3499 }
3500
3501 declare_per_stage_desc_pointers(ctx, true);
3502 declare_vs_specific_input_sgprs(ctx);
3503 if (!shader->is_gs_copy_shader)
3504 declare_vb_descriptor_input_sgprs(ctx);
3505
3506 if (shader->key.as_es) {
3507 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
3508 &ctx->es2gs_offset);
3509 } else if (shader->key.as_ls) {
3510 /* no extra parameters */
3511 } else {
3512 /* The locations of the other parameters are assigned dynamically. */
3513 declare_streamout_params(ctx, &shader->selector->so);
3514 }
3515
3516 /* VGPRs */
3517 declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
3518
3519 /* Return values */
3520 if (shader->key.opt.vs_as_prim_discard_cs) {
3521 for (i = 0; i < 4; i++)
3522 returns[num_returns++] = ctx->f32; /* VGPRs */
3523 }
3524 break;
3525
3526 case PIPE_SHADER_TESS_CTRL: /* GFX6-GFX8 */
3527 declare_global_desc_pointers(ctx);
3528 declare_per_stage_desc_pointers(ctx, true);
3529 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
3530 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets);
3531 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout);
3532 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
3533 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
3534 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
3535
3536 /* VGPRs */
3537 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
3538 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
3539
3540 /* param_tcs_offchip_offset and param_tcs_factor_offset are
3541 * placed after the user SGPRs.
3542 */
3543 for (i = 0; i < GFX6_TCS_NUM_USER_SGPR + 2; i++)
3544 returns[num_returns++] = ctx->i32; /* SGPRs */
3545 for (i = 0; i < 11; i++)
3546 returns[num_returns++] = ctx->f32; /* VGPRs */
3547 break;
3548
3549 case SI_SHADER_MERGED_VERTEX_TESSCTRL:
3550 /* Merged stages have 8 system SGPRs at the beginning. */
3551 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
3552 declare_per_stage_desc_pointers(ctx,
3553 ctx->type == PIPE_SHADER_TESS_CTRL);
3554 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
3555 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
3556 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
3557 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
3558 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
3559 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
3560
3561 declare_global_desc_pointers(ctx);
3562 declare_per_stage_desc_pointers(ctx,
3563 ctx->type == PIPE_SHADER_VERTEX);
3564 declare_vs_specific_input_sgprs(ctx);
3565
3566 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
3567 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets);
3568 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout);
3569 declare_vb_descriptor_input_sgprs(ctx);
3570
3571 /* VGPRs (first TCS, then VS) */
3572 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
3573 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
3574
3575 if (ctx->type == PIPE_SHADER_VERTEX) {
3576 declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
3577
3578 /* LS return values are inputs to the TCS main shader part. */
3579 for (i = 0; i < 8 + GFX9_TCS_NUM_USER_SGPR; i++)
3580 returns[num_returns++] = ctx->i32; /* SGPRs */
3581 for (i = 0; i < 2; i++)
3582 returns[num_returns++] = ctx->f32; /* VGPRs */
3583 } else {
3584 /* TCS return values are inputs to the TCS epilog.
3585 *
3586 * param_tcs_offchip_offset, param_tcs_factor_offset,
3587 * param_tcs_offchip_layout, and param_rw_buffers
3588 * should be passed to the epilog.
3589 */
3590 for (i = 0; i <= 8 + GFX9_SGPR_TCS_OUT_LAYOUT; i++)
3591 returns[num_returns++] = ctx->i32; /* SGPRs */
3592 for (i = 0; i < 11; i++)
3593 returns[num_returns++] = ctx->f32; /* VGPRs */
3594 }
3595 break;
3596
3597 case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
3598 /* Merged stages have 8 system SGPRs at the beginning. */
3599 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
3600 declare_per_stage_desc_pointers(ctx,
3601 ctx->type == PIPE_SHADER_GEOMETRY);
3602
3603 if (ctx->shader->key.as_ngg)
3604 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_tg_info);
3605 else
3606 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
3607
3608 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
3609 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
3610 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
3611 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
3612 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
3613
3614 declare_global_desc_pointers(ctx);
3615 if (ctx->type != PIPE_SHADER_VERTEX || !vs_blit_property) {
3616 declare_per_stage_desc_pointers(ctx,
3617 (ctx->type == PIPE_SHADER_VERTEX ||
3618 ctx->type == PIPE_SHADER_TESS_EVAL));
3619 }
3620
3621 if (ctx->type == PIPE_SHADER_VERTEX) {
3622 if (vs_blit_property)
3623 declare_vs_blit_inputs(ctx, vs_blit_property);
3624 else
3625 declare_vs_specific_input_sgprs(ctx);
3626 } else {
3627 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
3628 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
3629 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tes_offchip_addr);
3630 /* Declare as many input SGPRs as the VS has. */
3631 }
3632
3633 if (ctx->type == PIPE_SHADER_VERTEX)
3634 declare_vb_descriptor_input_sgprs(ctx);
3635
3636 /* VGPRs (first GS, then VS/TES) */
3637 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx01_offset);
3638 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx23_offset);
3639 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
3640 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
3641 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx45_offset);
3642
3643 if (ctx->type == PIPE_SHADER_VERTEX) {
3644 declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
3645 } else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
3646 declare_tes_input_vgprs(ctx);
3647 }
3648
3649 if (ctx->shader->key.as_es &&
3650 (ctx->type == PIPE_SHADER_VERTEX ||
3651 ctx->type == PIPE_SHADER_TESS_EVAL)) {
3652 unsigned num_user_sgprs;
3653
3654 if (ctx->type == PIPE_SHADER_VERTEX)
3655 num_user_sgprs = GFX9_VSGS_NUM_USER_SGPR;
3656 else
3657 num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
3658
3659 /* ES return values are inputs to GS. */
3660 for (i = 0; i < 8 + num_user_sgprs; i++)
3661 returns[num_returns++] = ctx->i32; /* SGPRs */
3662 for (i = 0; i < 5; i++)
3663 returns[num_returns++] = ctx->f32; /* VGPRs */
3664 }
3665 break;
3666
3667 case PIPE_SHADER_TESS_EVAL:
3668 declare_global_desc_pointers(ctx);
3669 declare_per_stage_desc_pointers(ctx, true);
3670 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
3671 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
3672 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tes_offchip_addr);
3673
3674 if (shader->key.as_es) {
3675 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
3676 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
3677 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->es2gs_offset);
3678 } else {
3679 declare_streamout_params(ctx, &shader->selector->so);
3680 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
3681 }
3682
3683 /* VGPRs */
3684 declare_tes_input_vgprs(ctx);
3685 break;
3686
3687 case PIPE_SHADER_GEOMETRY:
3688 declare_global_desc_pointers(ctx);
3689 declare_per_stage_desc_pointers(ctx, true);
3690 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
3691 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_wave_id);
3692
3693 /* VGPRs */
3694 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[0]);
3695 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[1]);
3696 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
3697 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[2]);
3698 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[3]);
3699 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[4]);
3700 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[5]);
3701 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
3702 break;
3703
3704 case PIPE_SHADER_FRAGMENT:
3705 declare_global_desc_pointers(ctx);
3706 declare_per_stage_desc_pointers(ctx, true);
3707 add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL,
3708 SI_PARAM_ALPHA_REF);
3709 add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
3710 &ctx->args.prim_mask, SI_PARAM_PRIM_MASK);
3711
3712 add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.persp_sample,
3713 SI_PARAM_PERSP_SAMPLE);
3714 add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
3715 &ctx->args.persp_center, SI_PARAM_PERSP_CENTER);
3716 add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
3717 &ctx->args.persp_centroid, SI_PARAM_PERSP_CENTROID);
3718 add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT,
3719 NULL, SI_PARAM_PERSP_PULL_MODEL);
3720 add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
3721 &ctx->args.linear_sample, SI_PARAM_LINEAR_SAMPLE);
3722 add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
3723 &ctx->args.linear_center, SI_PARAM_LINEAR_CENTER);
3724 add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
3725 &ctx->args.linear_centroid, SI_PARAM_LINEAR_CENTROID);
3726 add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_FLOAT,
3727 NULL, SI_PARAM_LINE_STIPPLE_TEX);
3728 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
3729 &ctx->args.frag_pos[0], SI_PARAM_POS_X_FLOAT);
3730 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
3731 &ctx->args.frag_pos[1], SI_PARAM_POS_Y_FLOAT);
3732 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
3733 &ctx->args.frag_pos[2], SI_PARAM_POS_Z_FLOAT);
3734 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
3735 &ctx->args.frag_pos[3], SI_PARAM_POS_W_FLOAT);
3736 shader->info.face_vgpr_index = ctx->args.num_vgprs_used;
3737 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
3738 &ctx->args.front_face, SI_PARAM_FRONT_FACE);
3739 shader->info.ancillary_vgpr_index = ctx->args.num_vgprs_used;
3740 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
3741 &ctx->args.ancillary, SI_PARAM_ANCILLARY);
3742 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
3743 &ctx->args.sample_coverage, SI_PARAM_SAMPLE_COVERAGE);
3744 add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
3745 &ctx->pos_fixed_pt, SI_PARAM_POS_FIXED_PT);
3746
3747 /* Color inputs from the prolog. */
3748 if (shader->selector->info.colors_read) {
3749 unsigned num_color_elements =
3750 util_bitcount(shader->selector->info.colors_read);
3751
3752 for (i = 0; i < num_color_elements; i++)
3753 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
3754
3755 num_prolog_vgprs += num_color_elements;
3756 }
3757
3758 /* Outputs for the epilog. */
3759 num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
3760 num_returns =
3761 num_return_sgprs +
3762 util_bitcount(shader->selector->info.colors_written) * 4 +
3763 shader->selector->info.writes_z +
3764 shader->selector->info.writes_stencil +
3765 shader->selector->info.writes_samplemask +
3766 1 /* SampleMaskIn */;
3767
3768 num_returns = MAX2(num_returns,
3769 num_return_sgprs +
3770 PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
3771
3772 for (i = 0; i < num_return_sgprs; i++)
3773 returns[i] = ctx->i32;
3774 for (; i < num_returns; i++)
3775 returns[i] = ctx->f32;
3776 break;
3777
3778 case PIPE_SHADER_COMPUTE:
3779 declare_global_desc_pointers(ctx);
3780 declare_per_stage_desc_pointers(ctx, true);
3781 if (shader->selector->info.uses_grid_size)
3782 ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT,
3783 &ctx->args.num_work_groups);
3784 if (shader->selector->info.uses_block_size &&
3785 shader->selector->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0)
3786 ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT, &ctx->block_size);
3787
3788 unsigned cs_user_data_dwords =
3789 shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
3790 if (cs_user_data_dwords) {
3791 ac_add_arg(&ctx->args, AC_ARG_SGPR, cs_user_data_dwords, AC_ARG_INT,
3792 &ctx->cs_user_data);
3793 }
3794
3795 /* Hardware SGPRs. */
3796 for (i = 0; i < 3; i++) {
3797 if (shader->selector->info.uses_block_id[i]) {
3798 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
3799 &ctx->args.workgroup_ids[i]);
3800 }
3801 }
3802 if (shader->selector->info.uses_subgroup_info)
3803 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.tg_size);
3804
3805 /* Hardware VGPRs. */
3806 ac_add_arg(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT,
3807 &ctx->args.local_invocation_ids);
3808 break;
3809 default:
3810 assert(0 && "unimplemented shader");
3811 return;
3812 }
3813
3814 si_create_function(ctx, "main", returns, num_returns,
3815 si_get_max_workgroup_size(shader));
3816
3817 /* Reserve register locations for VGPR inputs the PS prolog may need. */
3818 if (ctx->type == PIPE_SHADER_FRAGMENT && !ctx->shader->is_monolithic) {
3819 ac_llvm_add_target_dep_function_attr(ctx->main_fn,
3820 "InitialPSInputAddr",
3821 S_0286D0_PERSP_SAMPLE_ENA(1) |
3822 S_0286D0_PERSP_CENTER_ENA(1) |
3823 S_0286D0_PERSP_CENTROID_ENA(1) |
3824 S_0286D0_LINEAR_SAMPLE_ENA(1) |
3825 S_0286D0_LINEAR_CENTER_ENA(1) |
3826 S_0286D0_LINEAR_CENTROID_ENA(1) |
3827 S_0286D0_FRONT_FACE_ENA(1) |
3828 S_0286D0_ANCILLARY_ENA(1) |
3829 S_0286D0_POS_FIXED_PT_ENA(1));
3830 }
3831
3832 shader->info.num_input_sgprs = ctx->args.num_sgprs_used;
3833 shader->info.num_input_vgprs = ctx->args.num_vgprs_used;
3834
3835 assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
3836 shader->info.num_input_vgprs -= num_prolog_vgprs;
3837
3838 if (shader->key.as_ls || ctx->type == PIPE_SHADER_TESS_CTRL) {
3839 if (USE_LDS_SYMBOLS && LLVM_VERSION_MAJOR >= 9) {
3840 /* The LSHS size is not known until draw time, so we append it
3841 * at the end of whatever LDS use there may be in the rest of
3842 * the shader (currently none, unless LLVM decides to do its
3843 * own LDS-based lowering).
3844 */
3845 ctx->ac.lds = LLVMAddGlobalInAddressSpace(
3846 ctx->ac.module, LLVMArrayType(ctx->i32, 0),
3847 "__lds_end", AC_ADDR_SPACE_LDS);
3848 LLVMSetAlignment(ctx->ac.lds, 256);
3849 } else {
3850 ac_declare_lds_as_pointer(&ctx->ac);
3851 }
3852 }
3853
3854 /* Unlike radv, we override these arguments in the prolog, so to the
3855 * API shader they appear as normal arguments.
3856 */
3857 if (ctx->type == PIPE_SHADER_VERTEX) {
3858 ctx->abi.vertex_id = ac_get_arg(&ctx->ac, ctx->args.vertex_id);
3859 ctx->abi.instance_id = ac_get_arg(&ctx->ac, ctx->args.instance_id);
3860 } else if (ctx->type == PIPE_SHADER_FRAGMENT) {
3861 ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args.persp_centroid);
3862 ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args.linear_centroid);
3863 }
3864 }
3865
3866 /* Ensure that the esgs ring is declared.
3867 *
3868 * We declare it with 64KB alignment as a hint that the
3869 * pointer value will always be 0.
3870 */
3871 static void declare_esgs_ring(struct si_shader_context *ctx)
3872 {
3873 if (ctx->esgs_ring)
3874 return;
3875
3876 assert(!LLVMGetNamedGlobal(ctx->ac.module, "esgs_ring"));
3877
3878 ctx->esgs_ring = LLVMAddGlobalInAddressSpace(
3879 ctx->ac.module, LLVMArrayType(ctx->i32, 0),
3880 "esgs_ring",
3881 AC_ADDR_SPACE_LDS);
3882 LLVMSetLinkage(ctx->esgs_ring, LLVMExternalLinkage);
3883 LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
3884 }
3885
3886 /**
3887 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
3888 * for later use.
3889 */
3890 static void preload_ring_buffers(struct si_shader_context *ctx)
3891 {
3892 LLVMBuilderRef builder = ctx->ac.builder;
3893
3894 LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
3895
3896 if (ctx->shader->key.as_es || ctx->type == PIPE_SHADER_GEOMETRY) {
3897 if (ctx->screen->info.chip_class <= GFX8) {
3898 unsigned ring =
3899 ctx->type == PIPE_SHADER_GEOMETRY ? SI_GS_RING_ESGS
3900 : SI_ES_RING_ESGS;
3901 LLVMValueRef offset = LLVMConstInt(ctx->i32, ring, 0);
3902
3903 ctx->esgs_ring =
3904 ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
3905 } else {
3906 if (USE_LDS_SYMBOLS && LLVM_VERSION_MAJOR >= 9) {
3907 /* Declare the ESGS ring as an explicit LDS symbol. */
3908 declare_esgs_ring(ctx);
3909 } else {
3910 ac_declare_lds_as_pointer(&ctx->ac);
3911 ctx->esgs_ring = ctx->ac.lds;
3912 }
3913 }
3914 }
3915
3916 if (ctx->shader->is_gs_copy_shader) {
3917 LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
3918
3919 ctx->gsvs_ring[0] =
3920 ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
3921 } else if (ctx->type == PIPE_SHADER_GEOMETRY) {
3922 const struct si_shader_selector *sel = ctx->shader->selector;
3923 LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
3924 LLVMValueRef base_ring;
3925
3926 base_ring = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
3927
3928 /* The conceptual layout of the GSVS ring is
3929 * v0c0 .. vLv0 v0c1 .. vLc1 ..
3930 * but the real memory layout is swizzled across
3931 * threads:
3932 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
3933 * t16v0c0 ..
3934 * Override the buffer descriptor accordingly.
3935 */
3936 LLVMTypeRef v2i64 = LLVMVectorType(ctx->i64, 2);
3937 uint64_t stream_offset = 0;
3938
3939 for (unsigned stream = 0; stream < 4; ++stream) {
3940 unsigned num_components;
3941 unsigned stride;
3942 unsigned num_records;
3943 LLVMValueRef ring, tmp;
3944
3945 num_components = sel->info.num_stream_output_components[stream];
3946 if (!num_components)
3947 continue;
3948
3949 stride = 4 * num_components * sel->gs_max_out_vertices;
3950
3951 /* Limit on the stride field for <= GFX7. */
3952 assert(stride < (1 << 14));
3953
3954 num_records = ctx->ac.wave_size;
3955
3956 ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
3957 tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_0, "");
3958 tmp = LLVMBuildAdd(builder, tmp,
3959 LLVMConstInt(ctx->i64,
3960 stream_offset, 0), "");
3961 stream_offset += stride * ctx->ac.wave_size;
3962
3963 ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_0, "");
3964 ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
3965 tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_1, "");
3966 tmp = LLVMBuildOr(builder, tmp,
3967 LLVMConstInt(ctx->i32,
3968 S_008F04_STRIDE(stride) |
3969 S_008F04_SWIZZLE_ENABLE(1), 0), "");
3970 ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_1, "");
3971 ring = LLVMBuildInsertElement(builder, ring,
3972 LLVMConstInt(ctx->i32, num_records, 0),
3973 LLVMConstInt(ctx->i32, 2, 0), "");
3974
3975 uint32_t rsrc3 =
3976 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
3977 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3978 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
3979 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
3980 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
3981 S_008F0C_ADD_TID_ENABLE(1);
3982
3983 if (ctx->ac.chip_class >= GFX10) {
3984 rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
3985 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
3986 S_008F0C_RESOURCE_LEVEL(1);
3987 } else {
3988 rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3989 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
3990 S_008F0C_ELEMENT_SIZE(1); /* element_size = 4 (bytes) */
3991 }
3992
3993 ring = LLVMBuildInsertElement(builder, ring,
3994 LLVMConstInt(ctx->i32, rsrc3, false),
3995 LLVMConstInt(ctx->i32, 3, 0), "");
3996
3997 ctx->gsvs_ring[stream] = ring;
3998 }
3999 } else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
4000 ctx->tess_offchip_ring = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TES);
4001 }
4002 }
4003
4004 static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
4005 LLVMValueRef param_rw_buffers,
4006 struct ac_arg param_pos_fixed_pt)
4007 {
4008 LLVMBuilderRef builder = ctx->ac.builder;
4009 LLVMValueRef slot, desc, offset, row, bit, address[2];
4010
4011 /* Use the fixed-point gl_FragCoord input.
4012 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
4013 * per coordinate to get the repeating effect.
4014 */
4015 address[0] = si_unpack_param(ctx, param_pos_fixed_pt, 0, 5);
4016 address[1] = si_unpack_param(ctx, param_pos_fixed_pt, 16, 5);
4017
4018 /* Load the buffer descriptor. */
4019 slot = LLVMConstInt(ctx->i32, SI_PS_CONST_POLY_STIPPLE, 0);
4020 desc = ac_build_load_to_sgpr(&ctx->ac, param_rw_buffers, slot);
4021
4022 /* The stipple pattern is 32x32, each row has 32 bits. */
4023 offset = LLVMBuildMul(builder, address[1],
4024 LLVMConstInt(ctx->i32, 4, 0), "");
4025 row = buffer_load_const(ctx, desc, offset);
4026 row = ac_to_integer(&ctx->ac, row);
4027 bit = LLVMBuildLShr(builder, row, address[0], "");
4028 bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
4029 ac_build_kill_if_false(&ctx->ac, bit);
4030 }
4031
4032 /* For the UMR disassembler. */
4033 #define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
4034 #define DEBUGGER_NUM_MARKERS 5
4035
4036 static bool si_shader_binary_open(struct si_screen *screen,
4037 struct si_shader *shader,
4038 struct ac_rtld_binary *rtld)
4039 {
4040 const struct si_shader_selector *sel = shader->selector;
4041 const char *part_elfs[5];
4042 size_t part_sizes[5];
4043 unsigned num_parts = 0;
4044
4045 #define add_part(shader_or_part) \
4046 if (shader_or_part) { \
4047 part_elfs[num_parts] = (shader_or_part)->binary.elf_buffer; \
4048 part_sizes[num_parts] = (shader_or_part)->binary.elf_size; \
4049 num_parts++; \
4050 }
4051
4052 add_part(shader->prolog);
4053 add_part(shader->previous_stage);
4054 add_part(shader->prolog2);
4055 add_part(shader);
4056 add_part(shader->epilog);
4057
4058 #undef add_part
4059
4060 struct ac_rtld_symbol lds_symbols[2];
4061 unsigned num_lds_symbols = 0;
4062
4063 if (sel && screen->info.chip_class >= GFX9 && !shader->is_gs_copy_shader &&
4064 (sel->type == PIPE_SHADER_GEOMETRY || shader->key.as_ngg)) {
4065 /* We add this symbol even on LLVM <= 8 to ensure that
4066 * shader->config.lds_size is set correctly below.
4067 */
4068 struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
4069 sym->name = "esgs_ring";
4070 sym->size = shader->gs_info.esgs_ring_size;
4071 sym->align = 64 * 1024;
4072 }
4073
4074 if (shader->key.as_ngg && sel->type == PIPE_SHADER_GEOMETRY) {
4075 struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
4076 sym->name = "ngg_emit";
4077 sym->size = shader->ngg.ngg_emit_size * 4;
4078 sym->align = 4;
4079 }
4080
4081 bool ok = ac_rtld_open(rtld, (struct ac_rtld_open_info){
4082 .info = &screen->info,
4083 .options = {
4084 .halt_at_entry = screen->options.halt_shaders,
4085 },
4086 .shader_type = tgsi_processor_to_shader_stage(sel->type),
4087 .wave_size = si_get_shader_wave_size(shader),
4088 .num_parts = num_parts,
4089 .elf_ptrs = part_elfs,
4090 .elf_sizes = part_sizes,
4091 .num_shared_lds_symbols = num_lds_symbols,
4092 .shared_lds_symbols = lds_symbols });
4093
4094 if (rtld->lds_size > 0) {
4095 unsigned alloc_granularity = screen->info.chip_class >= GFX7 ? 512 : 256;
4096 shader->config.lds_size =
4097 align(rtld->lds_size, alloc_granularity) / alloc_granularity;
4098 }
4099
4100 return ok;
4101 }
4102
4103 static unsigned si_get_shader_binary_size(struct si_screen *screen, struct si_shader *shader)
4104 {
4105 struct ac_rtld_binary rtld;
4106 si_shader_binary_open(screen, shader, &rtld);
4107 return rtld.exec_size;
4108 }
4109
4110 static bool si_get_external_symbol(void *data, const char *name, uint64_t *value)
4111 {
4112 uint64_t *scratch_va = data;
4113
4114 if (!strcmp(scratch_rsrc_dword0_symbol, name)) {
4115 *value = (uint32_t)*scratch_va;
4116 return true;
4117 }
4118 if (!strcmp(scratch_rsrc_dword1_symbol, name)) {
4119 /* Enable scratch coalescing. */
4120 *value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) |
4121 S_008F04_SWIZZLE_ENABLE(1);
4122 return true;
4123 }
4124
4125 return false;
4126 }
4127
4128 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
4129 uint64_t scratch_va)
4130 {
4131 struct ac_rtld_binary binary;
4132 if (!si_shader_binary_open(sscreen, shader, &binary))
4133 return false;
4134
4135 si_resource_reference(&shader->bo, NULL);
4136 shader->bo = si_aligned_buffer_create(&sscreen->b,
4137 sscreen->info.cpdma_prefetch_writes_memory ?
4138 0 : SI_RESOURCE_FLAG_READ_ONLY,
4139 PIPE_USAGE_IMMUTABLE,
4140 align(binary.rx_size, SI_CPDMA_ALIGNMENT),
4141 256);
4142 if (!shader->bo)
4143 return false;
4144
4145 /* Upload. */
4146 struct ac_rtld_upload_info u = {};
4147 u.binary = &binary;
4148 u.get_external_symbol = si_get_external_symbol;
4149 u.cb_data = &scratch_va;
4150 u.rx_va = shader->bo->gpu_address;
4151 u.rx_ptr = sscreen->ws->buffer_map(shader->bo->buf, NULL,
4152 PIPE_TRANSFER_READ_WRITE |
4153 PIPE_TRANSFER_UNSYNCHRONIZED |
4154 RADEON_TRANSFER_TEMPORARY);
4155 if (!u.rx_ptr)
4156 return false;
4157
4158 bool ok = ac_rtld_upload(&u);
4159
4160 sscreen->ws->buffer_unmap(shader->bo->buf);
4161 ac_rtld_close(&binary);
4162
4163 return ok;
4164 }
4165
4166 static void si_shader_dump_disassembly(struct si_screen *screen,
4167 const struct si_shader_binary *binary,
4168 enum pipe_shader_type shader_type,
4169 unsigned wave_size,
4170 struct pipe_debug_callback *debug,
4171 const char *name, FILE *file)
4172 {
4173 struct ac_rtld_binary rtld_binary;
4174
4175 if (!ac_rtld_open(&rtld_binary, (struct ac_rtld_open_info){
4176 .info = &screen->info,
4177 .shader_type = tgsi_processor_to_shader_stage(shader_type),
4178 .wave_size = wave_size,
4179 .num_parts = 1,
4180 .elf_ptrs = &binary->elf_buffer,
4181 .elf_sizes = &binary->elf_size }))
4182 return;
4183
4184 const char *disasm;
4185 size_t nbytes;
4186
4187 if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm", &disasm, &nbytes))
4188 goto out;
4189
4190 if (nbytes > INT_MAX)
4191 goto out;
4192
4193 if (debug && debug->debug_message) {
4194 /* Very long debug messages are cut off, so send the
4195 * disassembly one line at a time. This causes more
4196 * overhead, but on the plus side it simplifies
4197 * parsing of resulting logs.
4198 */
4199 pipe_debug_message(debug, SHADER_INFO,
4200 "Shader Disassembly Begin");
4201
4202 uint64_t line = 0;
4203 while (line < nbytes) {
4204 int count = nbytes - line;
4205 const char *nl = memchr(disasm + line, '\n', nbytes - line);
4206 if (nl)
4207 count = nl - (disasm + line);
4208
4209 if (count) {
4210 pipe_debug_message(debug, SHADER_INFO,
4211 "%.*s", count, disasm + line);
4212 }
4213
4214 line += count + 1;
4215 }
4216
4217 pipe_debug_message(debug, SHADER_INFO,
4218 "Shader Disassembly End");
4219 }
4220
4221 if (file) {
4222 fprintf(file, "Shader %s disassembly:\n", name);
4223 fprintf(file, "%*s", (int)nbytes, disasm);
4224 }
4225
4226 out:
4227 ac_rtld_close(&rtld_binary);
4228 }
4229
4230 static void si_calculate_max_simd_waves(struct si_shader *shader)
4231 {
4232 struct si_screen *sscreen = shader->selector->screen;
4233 struct ac_shader_config *conf = &shader->config;
4234 unsigned num_inputs = shader->selector->info.num_inputs;
4235 unsigned lds_increment = sscreen->info.chip_class >= GFX7 ? 512 : 256;
4236 unsigned lds_per_wave = 0;
4237 unsigned max_simd_waves;
4238
4239 max_simd_waves = sscreen->info.max_wave64_per_simd;
4240
4241 /* Compute LDS usage for PS. */
4242 switch (shader->selector->type) {
4243 case PIPE_SHADER_FRAGMENT:
4244 /* The minimum usage per wave is (num_inputs * 48). The maximum
4245 * usage is (num_inputs * 48 * 16).
4246 * We can get anything in between and it varies between waves.
4247 *
4248 * The 48 bytes per input for a single primitive is equal to
4249 * 4 bytes/component * 4 components/input * 3 points.
4250 *
4251 * Other stages don't know the size at compile time or don't
4252 * allocate LDS per wave, but instead they do it per thread group.
4253 */
4254 lds_per_wave = conf->lds_size * lds_increment +
4255 align(num_inputs * 48, lds_increment);
4256 break;
4257 case PIPE_SHADER_COMPUTE:
4258 if (shader->selector) {
4259 unsigned max_workgroup_size =
4260 si_get_max_workgroup_size(shader);
4261 lds_per_wave = (conf->lds_size * lds_increment) /
4262 DIV_ROUND_UP(max_workgroup_size,
4263 sscreen->compute_wave_size);
4264 }
4265 break;
4266 default:;
4267 }
4268
4269 /* Compute the per-SIMD wave counts. */
4270 if (conf->num_sgprs) {
4271 max_simd_waves =
4272 MIN2(max_simd_waves,
4273 sscreen->info.num_physical_sgprs_per_simd / conf->num_sgprs);
4274 }
4275
4276 if (conf->num_vgprs) {
4277 /* Always print wave limits as Wave64, so that we can compare
4278 * Wave32 and Wave64 with shader-db fairly. */
4279 unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd;
4280 max_simd_waves = MIN2(max_simd_waves, max_vgprs / conf->num_vgprs);
4281 }
4282
4283 /* LDS is 64KB per CU (4 SIMDs) on GFX6-9, which is 16KB per SIMD (usage above
4284 * 16KB makes some SIMDs unoccupied).
4285 *
4286 * LDS is 128KB in WGP mode and 64KB in CU mode. Assume the WGP mode is used.
4287 */
4288 unsigned max_lds_size = sscreen->info.chip_class >= GFX10 ? 128*1024 : 64*1024;
4289 unsigned max_lds_per_simd = max_lds_size / 4;
4290 if (lds_per_wave)
4291 max_simd_waves = MIN2(max_simd_waves, max_lds_per_simd / lds_per_wave);
4292
4293 shader->info.max_simd_waves = max_simd_waves;
4294 }
4295
4296 void si_shader_dump_stats_for_shader_db(struct si_screen *screen,
4297 struct si_shader *shader,
4298 struct pipe_debug_callback *debug)
4299 {
4300 const struct ac_shader_config *conf = &shader->config;
4301
4302 if (screen->options.debug_disassembly)
4303 si_shader_dump_disassembly(screen, &shader->binary,
4304 shader->selector->type,
4305 si_get_shader_wave_size(shader),
4306 debug, "main", NULL);
4307
4308 pipe_debug_message(debug, SHADER_INFO,
4309 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
4310 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
4311 "Spilled VGPRs: %d PrivMem VGPRs: %d",
4312 conf->num_sgprs, conf->num_vgprs,
4313 si_get_shader_binary_size(screen, shader),
4314 conf->lds_size, conf->scratch_bytes_per_wave,
4315 shader->info.max_simd_waves, conf->spilled_sgprs,
4316 conf->spilled_vgprs, shader->info.private_mem_vgprs);
4317 }
4318
4319 static void si_shader_dump_stats(struct si_screen *sscreen,
4320 struct si_shader *shader,
4321 FILE *file,
4322 bool check_debug_option)
4323 {
4324 const struct ac_shader_config *conf = &shader->config;
4325
4326 if (!check_debug_option ||
4327 si_can_dump_shader(sscreen, shader->selector->type)) {
4328 if (shader->selector->type == PIPE_SHADER_FRAGMENT) {
4329 fprintf(file, "*** SHADER CONFIG ***\n"
4330 "SPI_PS_INPUT_ADDR = 0x%04x\n"
4331 "SPI_PS_INPUT_ENA = 0x%04x\n",
4332 conf->spi_ps_input_addr, conf->spi_ps_input_ena);
4333 }
4334
4335 fprintf(file, "*** SHADER STATS ***\n"
4336 "SGPRS: %d\n"
4337 "VGPRS: %d\n"
4338 "Spilled SGPRs: %d\n"
4339 "Spilled VGPRs: %d\n"
4340 "Private memory VGPRs: %d\n"
4341 "Code Size: %d bytes\n"
4342 "LDS: %d blocks\n"
4343 "Scratch: %d bytes per wave\n"
4344 "Max Waves: %d\n"
4345 "********************\n\n\n",
4346 conf->num_sgprs, conf->num_vgprs,
4347 conf->spilled_sgprs, conf->spilled_vgprs,
4348 shader->info.private_mem_vgprs,
4349 si_get_shader_binary_size(sscreen, shader),
4350 conf->lds_size, conf->scratch_bytes_per_wave,
4351 shader->info.max_simd_waves);
4352 }
4353 }
4354
4355 const char *si_get_shader_name(const struct si_shader *shader)
4356 {
4357 switch (shader->selector->type) {
4358 case PIPE_SHADER_VERTEX:
4359 if (shader->key.as_es)
4360 return "Vertex Shader as ES";
4361 else if (shader->key.as_ls)
4362 return "Vertex Shader as LS";
4363 else if (shader->key.opt.vs_as_prim_discard_cs)
4364 return "Vertex Shader as Primitive Discard CS";
4365 else if (shader->key.as_ngg)
4366 return "Vertex Shader as ESGS";
4367 else
4368 return "Vertex Shader as VS";
4369 case PIPE_SHADER_TESS_CTRL:
4370 return "Tessellation Control Shader";
4371 case PIPE_SHADER_TESS_EVAL:
4372 if (shader->key.as_es)
4373 return "Tessellation Evaluation Shader as ES";
4374 else if (shader->key.as_ngg)
4375 return "Tessellation Evaluation Shader as ESGS";
4376 else
4377 return "Tessellation Evaluation Shader as VS";
4378 case PIPE_SHADER_GEOMETRY:
4379 if (shader->is_gs_copy_shader)
4380 return "GS Copy Shader as VS";
4381 else
4382 return "Geometry Shader";
4383 case PIPE_SHADER_FRAGMENT:
4384 return "Pixel Shader";
4385 case PIPE_SHADER_COMPUTE:
4386 return "Compute Shader";
4387 default:
4388 return "Unknown Shader";
4389 }
4390 }
4391
4392 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
4393 struct pipe_debug_callback *debug,
4394 FILE *file, bool check_debug_option)
4395 {
4396 enum pipe_shader_type shader_type = shader->selector->type;
4397
4398 if (!check_debug_option ||
4399 si_can_dump_shader(sscreen, shader_type))
4400 si_dump_shader_key(shader, file);
4401
4402 if (!check_debug_option && shader->binary.llvm_ir_string) {
4403 if (shader->previous_stage &&
4404 shader->previous_stage->binary.llvm_ir_string) {
4405 fprintf(file, "\n%s - previous stage - LLVM IR:\n\n",
4406 si_get_shader_name(shader));
4407 fprintf(file, "%s\n", shader->previous_stage->binary.llvm_ir_string);
4408 }
4409
4410 fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
4411 si_get_shader_name(shader));
4412 fprintf(file, "%s\n", shader->binary.llvm_ir_string);
4413 }
4414
4415 if (!check_debug_option ||
4416 (si_can_dump_shader(sscreen, shader_type) &&
4417 !(sscreen->debug_flags & DBG(NO_ASM)))) {
4418 unsigned wave_size = si_get_shader_wave_size(shader);
4419
4420 fprintf(file, "\n%s:\n", si_get_shader_name(shader));
4421
4422 if (shader->prolog)
4423 si_shader_dump_disassembly(sscreen, &shader->prolog->binary,
4424 shader_type, wave_size, debug, "prolog", file);
4425 if (shader->previous_stage)
4426 si_shader_dump_disassembly(sscreen, &shader->previous_stage->binary,
4427 shader_type, wave_size, debug, "previous stage", file);
4428 if (shader->prolog2)
4429 si_shader_dump_disassembly(sscreen, &shader->prolog2->binary,
4430 shader_type, wave_size, debug, "prolog2", file);
4431
4432 si_shader_dump_disassembly(sscreen, &shader->binary, shader_type,
4433 wave_size, debug, "main", file);
4434
4435 if (shader->epilog)
4436 si_shader_dump_disassembly(sscreen, &shader->epilog->binary,
4437 shader_type, wave_size, debug, "epilog", file);
4438 fprintf(file, "\n");
4439 }
4440
4441 si_shader_dump_stats(sscreen, shader, file, check_debug_option);
4442 }
4443
4444 static int si_compile_llvm(struct si_screen *sscreen,
4445 struct si_shader_binary *binary,
4446 struct ac_shader_config *conf,
4447 struct ac_llvm_compiler *compiler,
4448 LLVMModuleRef mod,
4449 struct pipe_debug_callback *debug,
4450 enum pipe_shader_type shader_type,
4451 unsigned wave_size,
4452 const char *name,
4453 bool less_optimized)
4454 {
4455 unsigned count = p_atomic_inc_return(&sscreen->num_compilations);
4456
4457 if (si_can_dump_shader(sscreen, shader_type)) {
4458 fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
4459
4460 if (!(sscreen->debug_flags & (DBG(NO_IR) | DBG(PREOPT_IR)))) {
4461 fprintf(stderr, "%s LLVM IR:\n\n", name);
4462 ac_dump_module(mod);
4463 fprintf(stderr, "\n");
4464 }
4465 }
4466
4467 if (sscreen->record_llvm_ir) {
4468 char *ir = LLVMPrintModuleToString(mod);
4469 binary->llvm_ir_string = strdup(ir);
4470 LLVMDisposeMessage(ir);
4471 }
4472
4473 if (!si_replace_shader(count, binary)) {
4474 unsigned r = si_llvm_compile(mod, binary, compiler, debug,
4475 less_optimized, wave_size);
4476 if (r)
4477 return r;
4478 }
4479
4480 struct ac_rtld_binary rtld;
4481 if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){
4482 .info = &sscreen->info,
4483 .shader_type = tgsi_processor_to_shader_stage(shader_type),
4484 .wave_size = wave_size,
4485 .num_parts = 1,
4486 .elf_ptrs = &binary->elf_buffer,
4487 .elf_sizes = &binary->elf_size }))
4488 return -1;
4489
4490 bool ok = ac_rtld_read_config(&rtld, conf);
4491 ac_rtld_close(&rtld);
4492 if (!ok)
4493 return -1;
4494
4495 /* Enable 64-bit and 16-bit denormals, because there is no performance
4496 * cost.
4497 *
4498 * If denormals are enabled, all floating-point output modifiers are
4499 * ignored.
4500 *
4501 * Don't enable denormals for 32-bit floats, because:
4502 * - Floating-point output modifiers would be ignored by the hw.
4503 * - Some opcodes don't support denormals, such as v_mad_f32. We would
4504 * have to stop using those.
4505 * - GFX6 & GFX7 would be very slow.
4506 */
4507 conf->float_mode |= V_00B028_FP_64_DENORMS;
4508
4509 return 0;
4510 }
4511
4512 static void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
4513 {
4514 if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
4515 LLVMBuildRetVoid(ctx->ac.builder);
4516 else
4517 LLVMBuildRet(ctx->ac.builder, ret);
4518 }
4519
4520 /* Generate code for the hardware VS shader stage to go with a geometry shader */
4521 struct si_shader *
4522 si_generate_gs_copy_shader(struct si_screen *sscreen,
4523 struct ac_llvm_compiler *compiler,
4524 struct si_shader_selector *gs_selector,
4525 struct pipe_debug_callback *debug)
4526 {
4527 struct si_shader_context ctx;
4528 struct si_shader *shader;
4529 LLVMBuilderRef builder;
4530 struct si_shader_output_values outputs[SI_MAX_VS_OUTPUTS];
4531 struct si_shader_info *gsinfo = &gs_selector->info;
4532 int i;
4533
4534
4535 shader = CALLOC_STRUCT(si_shader);
4536 if (!shader)
4537 return NULL;
4538
4539 /* We can leave the fence as permanently signaled because the GS copy
4540 * shader only becomes visible globally after it has been compiled. */
4541 util_queue_fence_init(&shader->ready);
4542
4543 shader->selector = gs_selector;
4544 shader->is_gs_copy_shader = true;
4545
4546 si_llvm_context_init(&ctx, sscreen, compiler,
4547 si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false),
4548 64);
4549 ctx.shader = shader;
4550 ctx.type = PIPE_SHADER_VERTEX;
4551
4552 builder = ctx.ac.builder;
4553
4554 create_function(&ctx);
4555 preload_ring_buffers(&ctx);
4556
4557 LLVMValueRef voffset =
4558 LLVMBuildMul(ctx.ac.builder, ctx.abi.vertex_id,
4559 LLVMConstInt(ctx.i32, 4, 0), "");
4560
4561 /* Fetch the vertex stream ID.*/
4562 LLVMValueRef stream_id;
4563
4564 if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs)
4565 stream_id = si_unpack_param(&ctx, ctx.streamout_config, 24, 2);
4566 else
4567 stream_id = ctx.i32_0;
4568
4569 /* Fill in output information. */
4570 for (i = 0; i < gsinfo->num_outputs; ++i) {
4571 outputs[i].semantic_name = gsinfo->output_semantic_name[i];
4572 outputs[i].semantic_index = gsinfo->output_semantic_index[i];
4573
4574 for (int chan = 0; chan < 4; chan++) {
4575 outputs[i].vertex_stream[chan] =
4576 (gsinfo->output_streams[i] >> (2 * chan)) & 3;
4577 }
4578 }
4579
4580 LLVMBasicBlockRef end_bb;
4581 LLVMValueRef switch_inst;
4582
4583 end_bb = LLVMAppendBasicBlockInContext(ctx.ac.context, ctx.main_fn, "end");
4584 switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);
4585
4586 for (int stream = 0; stream < 4; stream++) {
4587 LLVMBasicBlockRef bb;
4588 unsigned offset;
4589
4590 if (!gsinfo->num_stream_output_components[stream])
4591 continue;
4592
4593 if (stream > 0 && !gs_selector->so.num_outputs)
4594 continue;
4595
4596 bb = LLVMInsertBasicBlockInContext(ctx.ac.context, end_bb, "out");
4597 LLVMAddCase(switch_inst, LLVMConstInt(ctx.i32, stream, 0), bb);
4598 LLVMPositionBuilderAtEnd(builder, bb);
4599
4600 /* Fetch vertex data from GSVS ring */
4601 offset = 0;
4602 for (i = 0; i < gsinfo->num_outputs; ++i) {
4603 for (unsigned chan = 0; chan < 4; chan++) {
4604 if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
4605 outputs[i].vertex_stream[chan] != stream) {
4606 outputs[i].values[chan] = LLVMGetUndef(ctx.f32);
4607 continue;
4608 }
4609
4610 LLVMValueRef soffset = LLVMConstInt(ctx.i32,
4611 offset * gs_selector->gs_max_out_vertices * 16 * 4, 0);
4612 offset++;
4613
4614 outputs[i].values[chan] =
4615 ac_build_buffer_load(&ctx.ac,
4616 ctx.gsvs_ring[0], 1,
4617 ctx.i32_0, voffset,
4618 soffset, 0, ac_glc | ac_slc,
4619 true, false);
4620 }
4621 }
4622
4623 /* Streamout and exports. */
4624 if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs) {
4625 si_llvm_emit_streamout(&ctx, outputs,
4626 gsinfo->num_outputs,
4627 stream);
4628 }
4629
4630 if (stream == 0)
4631 si_llvm_export_vs(&ctx, outputs, gsinfo->num_outputs);
4632
4633 LLVMBuildBr(builder, end_bb);
4634 }
4635
4636 LLVMPositionBuilderAtEnd(builder, end_bb);
4637
4638 LLVMBuildRetVoid(ctx.ac.builder);
4639
4640 ctx.type = PIPE_SHADER_GEOMETRY; /* override for shader dumping */
4641 si_llvm_optimize_module(&ctx);
4642
4643 bool ok = false;
4644 if (si_compile_llvm(sscreen, &ctx.shader->binary,
4645 &ctx.shader->config, ctx.compiler,
4646 ctx.ac.module,
4647 debug, PIPE_SHADER_GEOMETRY, ctx.ac.wave_size,
4648 "GS Copy Shader", false) == 0) {
4649 if (si_can_dump_shader(sscreen, PIPE_SHADER_GEOMETRY))
4650 fprintf(stderr, "GS Copy Shader:\n");
4651 si_shader_dump(sscreen, ctx.shader, debug, stderr, true);
4652
4653 if (!ctx.shader->config.scratch_bytes_per_wave)
4654 ok = si_shader_binary_upload(sscreen, ctx.shader, 0);
4655 else
4656 ok = true;
4657 }
4658
4659 si_llvm_dispose(&ctx);
4660
4661 if (!ok) {
4662 FREE(shader);
4663 shader = NULL;
4664 } else {
4665 si_fix_resource_usage(sscreen, shader);
4666 }
4667 return shader;
4668 }
4669
4670 static void si_dump_shader_key_vs(const struct si_shader_key *key,
4671 const struct si_vs_prolog_bits *prolog,
4672 const char *prefix, FILE *f)
4673 {
4674 fprintf(f, " %s.instance_divisor_is_one = %u\n",
4675 prefix, prolog->instance_divisor_is_one);
4676 fprintf(f, " %s.instance_divisor_is_fetched = %u\n",
4677 prefix, prolog->instance_divisor_is_fetched);
4678 fprintf(f, " %s.unpack_instance_id_from_vertex_id = %u\n",
4679 prefix, prolog->unpack_instance_id_from_vertex_id);
4680 fprintf(f, " %s.ls_vgpr_fix = %u\n",
4681 prefix, prolog->ls_vgpr_fix);
4682
4683 fprintf(f, " mono.vs.fetch_opencode = %x\n", key->mono.vs_fetch_opencode);
4684 fprintf(f, " mono.vs.fix_fetch = {");
4685 for (int i = 0; i < SI_MAX_ATTRIBS; i++) {
4686 union si_vs_fix_fetch fix = key->mono.vs_fix_fetch[i];
4687 if (i)
4688 fprintf(f, ", ");
4689 if (!fix.bits)
4690 fprintf(f, "0");
4691 else
4692 fprintf(f, "%u.%u.%u.%u", fix.u.reverse, fix.u.log_size,
4693 fix.u.num_channels_m1, fix.u.format);
4694 }
4695 fprintf(f, "}\n");
4696 }
4697
4698 static void si_dump_shader_key(const struct si_shader *shader, FILE *f)
4699 {
4700 const struct si_shader_key *key = &shader->key;
4701 enum pipe_shader_type shader_type = shader->selector->type;
4702
4703 fprintf(f, "SHADER KEY\n");
4704
4705 switch (shader_type) {
4706 case PIPE_SHADER_VERTEX:
4707 si_dump_shader_key_vs(key, &key->part.vs.prolog,
4708 "part.vs.prolog", f);
4709 fprintf(f, " as_es = %u\n", key->as_es);
4710 fprintf(f, " as_ls = %u\n", key->as_ls);
4711 fprintf(f, " as_ngg = %u\n", key->as_ngg);
4712 fprintf(f, " mono.u.vs_export_prim_id = %u\n",
4713 key->mono.u.vs_export_prim_id);
4714 fprintf(f, " opt.vs_as_prim_discard_cs = %u\n",
4715 key->opt.vs_as_prim_discard_cs);
4716 fprintf(f, " opt.cs_prim_type = %s\n",
4717 tgsi_primitive_names[key->opt.cs_prim_type]);
4718 fprintf(f, " opt.cs_indexed = %u\n",
4719 key->opt.cs_indexed);
4720 fprintf(f, " opt.cs_instancing = %u\n",
4721 key->opt.cs_instancing);
4722 fprintf(f, " opt.cs_primitive_restart = %u\n",
4723 key->opt.cs_primitive_restart);
4724 fprintf(f, " opt.cs_provoking_vertex_first = %u\n",
4725 key->opt.cs_provoking_vertex_first);
4726 fprintf(f, " opt.cs_need_correct_orientation = %u\n",
4727 key->opt.cs_need_correct_orientation);
4728 fprintf(f, " opt.cs_cull_front = %u\n",
4729 key->opt.cs_cull_front);
4730 fprintf(f, " opt.cs_cull_back = %u\n",
4731 key->opt.cs_cull_back);
4732 fprintf(f, " opt.cs_cull_z = %u\n",
4733 key->opt.cs_cull_z);
4734 fprintf(f, " opt.cs_halfz_clip_space = %u\n",
4735 key->opt.cs_halfz_clip_space);
4736 break;
4737
4738 case PIPE_SHADER_TESS_CTRL:
4739 if (shader->selector->screen->info.chip_class >= GFX9) {
4740 si_dump_shader_key_vs(key, &key->part.tcs.ls_prolog,
4741 "part.tcs.ls_prolog", f);
4742 }
4743 fprintf(f, " part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
4744 fprintf(f, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64"\n", key->mono.u.ff_tcs_inputs_to_copy);
4745 break;
4746
4747 case PIPE_SHADER_TESS_EVAL:
4748 fprintf(f, " as_es = %u\n", key->as_es);
4749 fprintf(f, " as_ngg = %u\n", key->as_ngg);
4750 fprintf(f, " mono.u.vs_export_prim_id = %u\n",
4751 key->mono.u.vs_export_prim_id);
4752 break;
4753
4754 case PIPE_SHADER_GEOMETRY:
4755 if (shader->is_gs_copy_shader)
4756 break;
4757
4758 if (shader->selector->screen->info.chip_class >= GFX9 &&
4759 key->part.gs.es->type == PIPE_SHADER_VERTEX) {
4760 si_dump_shader_key_vs(key, &key->part.gs.vs_prolog,
4761 "part.gs.vs_prolog", f);
4762 }
4763 fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
4764 fprintf(f, " part.gs.prolog.gfx9_prev_is_vs = %u\n", key->part.gs.prolog.gfx9_prev_is_vs);
4765 fprintf(f, " as_ngg = %u\n", key->as_ngg);
4766 break;
4767
4768 case PIPE_SHADER_COMPUTE:
4769 break;
4770
4771 case PIPE_SHADER_FRAGMENT:
4772 fprintf(f, " part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
4773 fprintf(f, " part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
4774 fprintf(f, " part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
4775 fprintf(f, " part.ps.prolog.force_persp_sample_interp = %u\n", key->part.ps.prolog.force_persp_sample_interp);
4776 fprintf(f, " part.ps.prolog.force_linear_sample_interp = %u\n", key->part.ps.prolog.force_linear_sample_interp);
4777 fprintf(f, " part.ps.prolog.force_persp_center_interp = %u\n", key->part.ps.prolog.force_persp_center_interp);
4778 fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
4779 fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
4780 fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
4781 fprintf(f, " part.ps.prolog.samplemask_log_ps_iter = %u\n", key->part.ps.prolog.samplemask_log_ps_iter);
4782 fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
4783 fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
4784 fprintf(f, " part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
4785 fprintf(f, " part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
4786 fprintf(f, " part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
4787 fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
4788 fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
4789 fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
4790 fprintf(f, " mono.u.ps.interpolate_at_sample_force_center = %u\n", key->mono.u.ps.interpolate_at_sample_force_center);
4791 fprintf(f, " mono.u.ps.fbfetch_msaa = %u\n", key->mono.u.ps.fbfetch_msaa);
4792 fprintf(f, " mono.u.ps.fbfetch_is_1D = %u\n", key->mono.u.ps.fbfetch_is_1D);
4793 fprintf(f, " mono.u.ps.fbfetch_layered = %u\n", key->mono.u.ps.fbfetch_layered);
4794 break;
4795
4796 default:
4797 assert(0);
4798 }
4799
4800 if ((shader_type == PIPE_SHADER_GEOMETRY ||
4801 shader_type == PIPE_SHADER_TESS_EVAL ||
4802 shader_type == PIPE_SHADER_VERTEX) &&
4803 !key->as_es && !key->as_ls) {
4804 fprintf(f, " opt.kill_outputs = 0x%"PRIx64"\n", key->opt.kill_outputs);
4805 fprintf(f, " opt.clip_disable = %u\n", key->opt.clip_disable);
4806 }
4807 }
4808
4809 static void si_optimize_vs_outputs(struct si_shader_context *ctx)
4810 {
4811 struct si_shader *shader = ctx->shader;
4812 struct si_shader_info *info = &shader->selector->info;
4813
4814 if ((ctx->type != PIPE_SHADER_VERTEX &&
4815 ctx->type != PIPE_SHADER_TESS_EVAL) ||
4816 shader->key.as_ls ||
4817 shader->key.as_es)
4818 return;
4819
4820 ac_optimize_vs_outputs(&ctx->ac,
4821 ctx->main_fn,
4822 shader->info.vs_output_param_offset,
4823 info->num_outputs,
4824 &shader->info.nr_param_exports);
4825 }
4826
4827 static void si_init_exec_from_input(struct si_shader_context *ctx,
4828 struct ac_arg param, unsigned bitoffset)
4829 {
4830 LLVMValueRef args[] = {
4831 ac_get_arg(&ctx->ac, param),
4832 LLVMConstInt(ctx->i32, bitoffset, 0),
4833 };
4834 ac_build_intrinsic(&ctx->ac,
4835 "llvm.amdgcn.init.exec.from.input",
4836 ctx->voidt, args, 2, AC_FUNC_ATTR_CONVERGENT);
4837 }
4838
4839 static bool si_vs_needs_prolog(const struct si_shader_selector *sel,
4840 const struct si_vs_prolog_bits *key)
4841 {
4842 /* VGPR initialization fixup for Vega10 and Raven is always done in the
4843 * VS prolog. */
4844 return sel->vs_needs_prolog ||
4845 key->ls_vgpr_fix ||
4846 key->unpack_instance_id_from_vertex_id;
4847 }
4848
4849 LLVMValueRef si_is_es_thread(struct si_shader_context *ctx)
4850 {
4851 /* Return true if the current thread should execute an ES thread. */
4852 return LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
4853 ac_get_thread_id(&ctx->ac),
4854 si_unpack_param(ctx, ctx->merged_wave_info, 0, 8), "");
4855 }
4856
4857 LLVMValueRef si_is_gs_thread(struct si_shader_context *ctx)
4858 {
4859 /* Return true if the current thread should execute a GS thread. */
4860 return LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
4861 ac_get_thread_id(&ctx->ac),
4862 si_unpack_param(ctx, ctx->merged_wave_info, 8, 8), "");
4863 }
4864
4865 static void si_llvm_emit_kill(struct ac_shader_abi *abi, LLVMValueRef visible)
4866 {
4867 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
4868 LLVMBuilderRef builder = ctx->ac.builder;
4869
4870 if (ctx->shader->selector->force_correct_derivs_after_kill) {
4871 /* Kill immediately while maintaining WQM. */
4872 ac_build_kill_if_false(&ctx->ac,
4873 ac_build_wqm_vote(&ctx->ac, visible));
4874
4875 LLVMValueRef mask = LLVMBuildLoad(builder, ctx->postponed_kill, "");
4876 mask = LLVMBuildAnd(builder, mask, visible, "");
4877 LLVMBuildStore(builder, mask, ctx->postponed_kill);
4878 return;
4879 }
4880
4881 ac_build_kill_if_false(&ctx->ac, visible);
4882 }
4883
4884 static bool si_build_main_function(struct si_shader_context *ctx,
4885 struct nir_shader *nir, bool free_nir)
4886 {
4887 struct si_shader *shader = ctx->shader;
4888 struct si_shader_selector *sel = shader->selector;
4889
4890 // TODO clean all this up!
4891 switch (ctx->type) {
4892 case PIPE_SHADER_VERTEX:
4893 if (shader->key.as_ls)
4894 ctx->abi.emit_outputs = si_llvm_emit_ls_epilogue;
4895 else if (shader->key.as_es)
4896 ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
4897 else if (shader->key.opt.vs_as_prim_discard_cs)
4898 ctx->abi.emit_outputs = si_llvm_emit_prim_discard_cs_epilogue;
4899 else if (shader->key.as_ngg)
4900 ctx->abi.emit_outputs = gfx10_emit_ngg_epilogue;
4901 else
4902 ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue;
4903 ctx->abi.load_base_vertex = get_base_vertex;
4904 break;
4905 case PIPE_SHADER_TESS_CTRL:
4906 ctx->abi.load_tess_varyings = si_nir_load_tcs_varyings;
4907 ctx->abi.load_tess_level = si_load_tess_level;
4908 ctx->abi.store_tcs_outputs = si_nir_store_output_tcs;
4909 ctx->abi.emit_outputs = si_llvm_emit_tcs_epilogue;
4910 ctx->abi.load_patch_vertices_in = si_load_patch_vertices_in;
4911 break;
4912 case PIPE_SHADER_TESS_EVAL:
4913 ctx->abi.load_tess_varyings = si_nir_load_input_tes;
4914 ctx->abi.load_tess_coord = si_load_tess_coord;
4915 ctx->abi.load_tess_level = si_load_tess_level;
4916 ctx->abi.load_patch_vertices_in = si_load_patch_vertices_in;
4917 if (shader->key.as_es)
4918 ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
4919 else if (shader->key.as_ngg)
4920 ctx->abi.emit_outputs = gfx10_emit_ngg_epilogue;
4921 else
4922 ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue;
4923 break;
4924 case PIPE_SHADER_GEOMETRY:
4925 ctx->abi.load_inputs = si_nir_load_input_gs;
4926 ctx->abi.emit_vertex = si_llvm_emit_vertex;
4927 ctx->abi.emit_primitive = si_llvm_emit_primitive;
4928 ctx->abi.emit_outputs = si_llvm_emit_gs_epilogue;
4929 break;
4930 case PIPE_SHADER_FRAGMENT:
4931 ctx->abi.emit_outputs = si_llvm_return_fs_outputs;
4932 ctx->abi.load_sample_position = load_sample_position;
4933 ctx->abi.load_sample_mask_in = load_sample_mask_in;
4934 ctx->abi.emit_fbfetch = si_nir_emit_fbfetch;
4935 ctx->abi.emit_kill = si_llvm_emit_kill;
4936 break;
4937 case PIPE_SHADER_COMPUTE:
4938 ctx->abi.load_local_group_size = get_block_size;
4939 break;
4940 default:
4941 assert(!"Unsupported shader type");
4942 return false;
4943 }
4944
4945 ctx->abi.load_ubo = load_ubo;
4946 ctx->abi.load_ssbo = load_ssbo;
4947
4948 create_function(ctx);
4949 preload_ring_buffers(ctx);
4950
4951 if (ctx->type == PIPE_SHADER_TESS_CTRL &&
4952 sel->info.tessfactors_are_def_in_all_invocs) {
4953 for (unsigned i = 0; i < 6; i++) {
4954 ctx->invoc0_tess_factors[i] =
4955 ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
4956 }
4957 }
4958
4959 if (ctx->type == PIPE_SHADER_GEOMETRY) {
4960 for (unsigned i = 0; i < 4; i++) {
4961 ctx->gs_next_vertex[i] =
4962 ac_build_alloca(&ctx->ac, ctx->i32, "");
4963 }
4964 if (shader->key.as_ngg) {
4965 for (unsigned i = 0; i < 4; ++i) {
4966 ctx->gs_curprim_verts[i] =
4967 ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
4968 ctx->gs_generated_prims[i] =
4969 ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
4970 }
4971
4972 unsigned scratch_size = 8;
4973 if (sel->so.num_outputs)
4974 scratch_size = 44;
4975
4976 LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, scratch_size);
4977 ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
4978 ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
4979 LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(ai32));
4980 LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
4981
4982 ctx->gs_ngg_emit = LLVMAddGlobalInAddressSpace(ctx->ac.module,
4983 LLVMArrayType(ctx->i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
4984 LLVMSetLinkage(ctx->gs_ngg_emit, LLVMExternalLinkage);
4985 LLVMSetAlignment(ctx->gs_ngg_emit, 4);
4986 }
4987 }
4988
4989 if (ctx->type != PIPE_SHADER_GEOMETRY &&
4990 (shader->key.as_ngg && !shader->key.as_es)) {
4991 /* Unconditionally declare scratch space base for streamout and
4992 * vertex compaction. Whether space is actually allocated is
4993 * determined during linking / PM4 creation.
4994 *
4995 * Add an extra dword per vertex to ensure an odd stride, which
4996 * avoids bank conflicts for SoA accesses.
4997 */
4998 if (!gfx10_is_ngg_passthrough(shader))
4999 declare_esgs_ring(ctx);
5000
5001 /* This is really only needed when streamout and / or vertex
5002 * compaction is enabled.
5003 */
5004 if (sel->so.num_outputs && !ctx->gs_ngg_scratch) {
5005 LLVMTypeRef asi32 = LLVMArrayType(ctx->i32, 8);
5006 ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
5007 asi32, "ngg_scratch", AC_ADDR_SPACE_LDS);
5008 LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(asi32));
5009 LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
5010 }
5011 }
5012
5013 /* For GFX9 merged shaders:
5014 * - Set EXEC for the first shader. If the prolog is present, set
5015 * EXEC there instead.
5016 * - Add a barrier before the second shader.
5017 * - In the second shader, reset EXEC to ~0 and wrap the main part in
5018 * an if-statement. This is required for correctness in geometry
5019 * shaders, to ensure that empty GS waves do not send GS_EMIT and
5020 * GS_CUT messages.
5021 *
5022 * For monolithic merged shaders, the first shader is wrapped in an
5023 * if-block together with its prolog in si_build_wrapper_function.
5024 *
5025 * NGG vertex and tess eval shaders running as the last
5026 * vertex/geometry stage handle execution explicitly using
5027 * if-statements.
5028 */
5029 if (ctx->screen->info.chip_class >= GFX9) {
5030 if (!shader->is_monolithic &&
5031 (shader->key.as_es || shader->key.as_ls) &&
5032 (ctx->type == PIPE_SHADER_TESS_EVAL ||
5033 (ctx->type == PIPE_SHADER_VERTEX &&
5034 !si_vs_needs_prolog(sel, &shader->key.part.vs.prolog)))) {
5035 si_init_exec_from_input(ctx,
5036 ctx->merged_wave_info, 0);
5037 } else if (ctx->type == PIPE_SHADER_TESS_CTRL ||
5038 ctx->type == PIPE_SHADER_GEOMETRY ||
5039 (shader->key.as_ngg && !shader->key.as_es)) {
5040 LLVMValueRef thread_enabled;
5041 bool nested_barrier;
5042
5043 if (!shader->is_monolithic ||
5044 (ctx->type == PIPE_SHADER_TESS_EVAL &&
5045 (shader->key.as_ngg && !shader->key.as_es)))
5046 ac_init_exec_full_mask(&ctx->ac);
5047
5048 if (ctx->type == PIPE_SHADER_TESS_CTRL ||
5049 ctx->type == PIPE_SHADER_GEOMETRY) {
5050 if (ctx->type == PIPE_SHADER_GEOMETRY && shader->key.as_ngg) {
5051 gfx10_ngg_gs_emit_prologue(ctx);
5052 nested_barrier = false;
5053 } else {
5054 nested_barrier = true;
5055 }
5056
5057 thread_enabled = si_is_gs_thread(ctx);
5058 } else {
5059 thread_enabled = si_is_es_thread(ctx);
5060 nested_barrier = false;
5061 }
5062
5063 ctx->merged_wrap_if_entry_block = LLVMGetInsertBlock(ctx->ac.builder);
5064 ctx->merged_wrap_if_label = 11500;
5065 ac_build_ifcc(&ctx->ac, thread_enabled, ctx->merged_wrap_if_label);
5066
5067 if (nested_barrier) {
5068 /* Execute a barrier before the second shader in
5069 * a merged shader.
5070 *
5071 * Execute the barrier inside the conditional block,
5072 * so that empty waves can jump directly to s_endpgm,
5073 * which will also signal the barrier.
5074 *
5075 * This is possible in gfx9, because an empty wave
5076 * for the second shader does not participate in
5077 * the epilogue. With NGG, empty waves may still
5078 * be required to export data (e.g. GS output vertices),
5079 * so we cannot let them exit early.
5080 *
5081 * If the shader is TCS and the TCS epilog is present
5082 * and contains a barrier, it will wait there and then
5083 * reach s_endpgm.
5084 */
5085 si_llvm_emit_barrier(ctx);
5086 }
5087 }
5088 }
5089
5090 if (sel->force_correct_derivs_after_kill) {
5091 ctx->postponed_kill = ac_build_alloca_undef(&ctx->ac, ctx->i1, "");
5092 /* true = don't kill. */
5093 LLVMBuildStore(ctx->ac.builder, ctx->i1true,
5094 ctx->postponed_kill);
5095 }
5096
5097 bool success = si_nir_build_llvm(ctx, nir);
5098 if (free_nir)
5099 ralloc_free(nir);
5100 if (!success) {
5101 fprintf(stderr, "Failed to translate shader from NIR to LLVM\n");
5102 return false;
5103 }
5104
5105 si_llvm_build_ret(ctx, ctx->return_value);
5106 return true;
5107 }
5108
5109 /**
5110 * Compute the VS prolog key, which contains all the information needed to
5111 * build the VS prolog function, and set shader->info bits where needed.
5112 *
5113 * \param info Shader info of the vertex shader.
5114 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
5115 * \param prolog_key Key of the VS prolog
5116 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
5117 * \param key Output shader part key.
5118 */
5119 static void si_get_vs_prolog_key(const struct si_shader_info *info,
5120 unsigned num_input_sgprs,
5121 const struct si_vs_prolog_bits *prolog_key,
5122 struct si_shader *shader_out,
5123 union si_shader_part_key *key)
5124 {
5125 memset(key, 0, sizeof(*key));
5126 key->vs_prolog.states = *prolog_key;
5127 key->vs_prolog.num_input_sgprs = num_input_sgprs;
5128 key->vs_prolog.num_inputs = info->num_inputs;
5129 key->vs_prolog.as_ls = shader_out->key.as_ls;
5130 key->vs_prolog.as_es = shader_out->key.as_es;
5131 key->vs_prolog.as_ngg = shader_out->key.as_ngg;
5132
5133 if (shader_out->selector->type == PIPE_SHADER_TESS_CTRL) {
5134 key->vs_prolog.as_ls = 1;
5135 key->vs_prolog.num_merged_next_stage_vgprs = 2;
5136 } else if (shader_out->selector->type == PIPE_SHADER_GEOMETRY) {
5137 key->vs_prolog.as_es = 1;
5138 key->vs_prolog.num_merged_next_stage_vgprs = 5;
5139 } else if (shader_out->key.as_ngg) {
5140 key->vs_prolog.num_merged_next_stage_vgprs = 5;
5141 }
5142
5143 /* Enable loading the InstanceID VGPR. */
5144 uint16_t input_mask = u_bit_consecutive(0, info->num_inputs);
5145
5146 if ((key->vs_prolog.states.instance_divisor_is_one |
5147 key->vs_prolog.states.instance_divisor_is_fetched) & input_mask)
5148 shader_out->info.uses_instanceid = true;
5149 }
5150
5151 /**
5152 * Compute the PS prolog key, which contains all the information needed to
5153 * build the PS prolog function, and set related bits in shader->config.
5154 */
5155 static void si_get_ps_prolog_key(struct si_shader *shader,
5156 union si_shader_part_key *key,
5157 bool separate_prolog)
5158 {
5159 struct si_shader_info *info = &shader->selector->info;
5160
5161 memset(key, 0, sizeof(*key));
5162 key->ps_prolog.states = shader->key.part.ps.prolog;
5163 key->ps_prolog.colors_read = info->colors_read;
5164 key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
5165 key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
5166 key->ps_prolog.wqm = info->uses_derivatives &&
5167 (key->ps_prolog.colors_read ||
5168 key->ps_prolog.states.force_persp_sample_interp ||
5169 key->ps_prolog.states.force_linear_sample_interp ||
5170 key->ps_prolog.states.force_persp_center_interp ||
5171 key->ps_prolog.states.force_linear_center_interp ||
5172 key->ps_prolog.states.bc_optimize_for_persp ||
5173 key->ps_prolog.states.bc_optimize_for_linear);
5174 key->ps_prolog.ancillary_vgpr_index = shader->info.ancillary_vgpr_index;
5175
5176 if (info->colors_read) {
5177 unsigned *color = shader->selector->color_attr_index;
5178
5179 if (shader->key.part.ps.prolog.color_two_side) {
5180 /* BCOLORs are stored after the last input. */
5181 key->ps_prolog.num_interp_inputs = info->num_inputs;
5182 key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
5183 if (separate_prolog)
5184 shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
5185 }
5186
5187 for (unsigned i = 0; i < 2; i++) {
5188 unsigned interp = info->input_interpolate[color[i]];
5189 unsigned location = info->input_interpolate_loc[color[i]];
5190
5191 if (!(info->colors_read & (0xf << i*4)))
5192 continue;
5193
5194 key->ps_prolog.color_attr_index[i] = color[i];
5195
5196 if (shader->key.part.ps.prolog.flatshade_colors &&
5197 interp == TGSI_INTERPOLATE_COLOR)
5198 interp = TGSI_INTERPOLATE_CONSTANT;
5199
5200 switch (interp) {
5201 case TGSI_INTERPOLATE_CONSTANT:
5202 key->ps_prolog.color_interp_vgpr_index[i] = -1;
5203 break;
5204 case TGSI_INTERPOLATE_PERSPECTIVE:
5205 case TGSI_INTERPOLATE_COLOR:
5206 /* Force the interpolation location for colors here. */
5207 if (shader->key.part.ps.prolog.force_persp_sample_interp)
5208 location = TGSI_INTERPOLATE_LOC_SAMPLE;
5209 if (shader->key.part.ps.prolog.force_persp_center_interp)
5210 location = TGSI_INTERPOLATE_LOC_CENTER;
5211
5212 switch (location) {
5213 case TGSI_INTERPOLATE_LOC_SAMPLE:
5214 key->ps_prolog.color_interp_vgpr_index[i] = 0;
5215 if (separate_prolog) {
5216 shader->config.spi_ps_input_ena |=
5217 S_0286CC_PERSP_SAMPLE_ENA(1);
5218 }
5219 break;
5220 case TGSI_INTERPOLATE_LOC_CENTER:
5221 key->ps_prolog.color_interp_vgpr_index[i] = 2;
5222 if (separate_prolog) {
5223 shader->config.spi_ps_input_ena |=
5224 S_0286CC_PERSP_CENTER_ENA(1);
5225 }
5226 break;
5227 case TGSI_INTERPOLATE_LOC_CENTROID:
5228 key->ps_prolog.color_interp_vgpr_index[i] = 4;
5229 if (separate_prolog) {
5230 shader->config.spi_ps_input_ena |=
5231 S_0286CC_PERSP_CENTROID_ENA(1);
5232 }
5233 break;
5234 default:
5235 assert(0);
5236 }
5237 break;
5238 case TGSI_INTERPOLATE_LINEAR:
5239 /* Force the interpolation location for colors here. */
5240 if (shader->key.part.ps.prolog.force_linear_sample_interp)
5241 location = TGSI_INTERPOLATE_LOC_SAMPLE;
5242 if (shader->key.part.ps.prolog.force_linear_center_interp)
5243 location = TGSI_INTERPOLATE_LOC_CENTER;
5244
5245 /* The VGPR assignment for non-monolithic shaders
5246 * works because InitialPSInputAddr is set on the
5247 * main shader and PERSP_PULL_MODEL is never used.
5248 */
5249 switch (location) {
5250 case TGSI_INTERPOLATE_LOC_SAMPLE:
5251 key->ps_prolog.color_interp_vgpr_index[i] =
5252 separate_prolog ? 6 : 9;
5253 if (separate_prolog) {
5254 shader->config.spi_ps_input_ena |=
5255 S_0286CC_LINEAR_SAMPLE_ENA(1);
5256 }
5257 break;
5258 case TGSI_INTERPOLATE_LOC_CENTER:
5259 key->ps_prolog.color_interp_vgpr_index[i] =
5260 separate_prolog ? 8 : 11;
5261 if (separate_prolog) {
5262 shader->config.spi_ps_input_ena |=
5263 S_0286CC_LINEAR_CENTER_ENA(1);
5264 }
5265 break;
5266 case TGSI_INTERPOLATE_LOC_CENTROID:
5267 key->ps_prolog.color_interp_vgpr_index[i] =
5268 separate_prolog ? 10 : 13;
5269 if (separate_prolog) {
5270 shader->config.spi_ps_input_ena |=
5271 S_0286CC_LINEAR_CENTROID_ENA(1);
5272 }
5273 break;
5274 default:
5275 assert(0);
5276 }
5277 break;
5278 default:
5279 assert(0);
5280 }
5281 }
5282 }
5283 }
5284
5285 /**
5286 * Check whether a PS prolog is required based on the key.
5287 */
5288 static bool si_need_ps_prolog(const union si_shader_part_key *key)
5289 {
5290 return key->ps_prolog.colors_read ||
5291 key->ps_prolog.states.force_persp_sample_interp ||
5292 key->ps_prolog.states.force_linear_sample_interp ||
5293 key->ps_prolog.states.force_persp_center_interp ||
5294 key->ps_prolog.states.force_linear_center_interp ||
5295 key->ps_prolog.states.bc_optimize_for_persp ||
5296 key->ps_prolog.states.bc_optimize_for_linear ||
5297 key->ps_prolog.states.poly_stipple ||
5298 key->ps_prolog.states.samplemask_log_ps_iter;
5299 }
5300
5301 /**
5302 * Compute the PS epilog key, which contains all the information needed to
5303 * build the PS epilog function.
5304 */
5305 static void si_get_ps_epilog_key(struct si_shader *shader,
5306 union si_shader_part_key *key)
5307 {
5308 struct si_shader_info *info = &shader->selector->info;
5309 memset(key, 0, sizeof(*key));
5310 key->ps_epilog.colors_written = info->colors_written;
5311 key->ps_epilog.writes_z = info->writes_z;
5312 key->ps_epilog.writes_stencil = info->writes_stencil;
5313 key->ps_epilog.writes_samplemask = info->writes_samplemask;
5314 key->ps_epilog.states = shader->key.part.ps.epilog;
5315 }
5316
5317 /**
5318 * Build the GS prolog function. Rotate the input vertices for triangle strips
5319 * with adjacency.
5320 */
5321 static void si_build_gs_prolog_function(struct si_shader_context *ctx,
5322 union si_shader_part_key *key)
5323 {
5324 unsigned num_sgprs, num_vgprs;
5325 LLVMBuilderRef builder = ctx->ac.builder;
5326 LLVMTypeRef returns[AC_MAX_ARGS];
5327 LLVMValueRef func, ret;
5328
5329 memset(&ctx->args, 0, sizeof(ctx->args));
5330
5331 if (ctx->screen->info.chip_class >= GFX9) {
5332 if (key->gs_prolog.states.gfx9_prev_is_vs)
5333 num_sgprs = 8 + GFX9_VSGS_NUM_USER_SGPR;
5334 else
5335 num_sgprs = 8 + GFX9_TESGS_NUM_USER_SGPR;
5336 num_vgprs = 5; /* ES inputs are not needed by GS */
5337 } else {
5338 num_sgprs = GFX6_GS_NUM_USER_SGPR + 2;
5339 num_vgprs = 8;
5340 }
5341
5342 for (unsigned i = 0; i < num_sgprs; ++i) {
5343 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
5344 returns[i] = ctx->i32;
5345 }
5346
5347 for (unsigned i = 0; i < num_vgprs; ++i) {
5348 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL);
5349 returns[num_sgprs + i] = ctx->f32;
5350 }
5351
5352 /* Create the function. */
5353 si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs,
5354 0);
5355 func = ctx->main_fn;
5356
5357 /* Set the full EXEC mask for the prolog, because we are only fiddling
5358 * with registers here. The main shader part will set the correct EXEC
5359 * mask.
5360 */
5361 if (ctx->screen->info.chip_class >= GFX9 && !key->gs_prolog.is_monolithic)
5362 ac_init_exec_full_mask(&ctx->ac);
5363
5364 /* Copy inputs to outputs. This should be no-op, as the registers match,
5365 * but it will prevent the compiler from overwriting them unintentionally.
5366 */
5367 ret = ctx->return_value;
5368 for (unsigned i = 0; i < num_sgprs; i++) {
5369 LLVMValueRef p = LLVMGetParam(func, i);
5370 ret = LLVMBuildInsertValue(builder, ret, p, i, "");
5371 }
5372 for (unsigned i = 0; i < num_vgprs; i++) {
5373 LLVMValueRef p = LLVMGetParam(func, num_sgprs + i);
5374 p = ac_to_float(&ctx->ac, p);
5375 ret = LLVMBuildInsertValue(builder, ret, p, num_sgprs + i, "");
5376 }
5377
5378 if (key->gs_prolog.states.tri_strip_adj_fix) {
5379 /* Remap the input vertices for every other primitive. */
5380 const struct ac_arg gfx6_vtx_params[6] = {
5381 { .used = true, .arg_index = num_sgprs },
5382 { .used = true, .arg_index = num_sgprs + 1 },
5383 { .used = true, .arg_index = num_sgprs + 3 },
5384 { .used = true, .arg_index = num_sgprs + 4 },
5385 { .used = true, .arg_index = num_sgprs + 5 },
5386 { .used = true, .arg_index = num_sgprs + 6 },
5387 };
5388 const struct ac_arg gfx9_vtx_params[3] = {
5389 { .used = true, .arg_index = num_sgprs },
5390 { .used = true, .arg_index = num_sgprs + 1 },
5391 { .used = true, .arg_index = num_sgprs + 4 },
5392 };
5393 LLVMValueRef vtx_in[6], vtx_out[6];
5394 LLVMValueRef prim_id, rotate;
5395
5396 if (ctx->screen->info.chip_class >= GFX9) {
5397 for (unsigned i = 0; i < 3; i++) {
5398 vtx_in[i*2] = si_unpack_param(ctx, gfx9_vtx_params[i], 0, 16);
5399 vtx_in[i*2+1] = si_unpack_param(ctx, gfx9_vtx_params[i], 16, 16);
5400 }
5401 } else {
5402 for (unsigned i = 0; i < 6; i++)
5403 vtx_in[i] = ac_get_arg(&ctx->ac, gfx6_vtx_params[i]);
5404 }
5405
5406 prim_id = LLVMGetParam(func, num_sgprs + 2);
5407 rotate = LLVMBuildTrunc(builder, prim_id, ctx->i1, "");
5408
5409 for (unsigned i = 0; i < 6; ++i) {
5410 LLVMValueRef base, rotated;
5411 base = vtx_in[i];
5412 rotated = vtx_in[(i + 4) % 6];
5413 vtx_out[i] = LLVMBuildSelect(builder, rotate, rotated, base, "");
5414 }
5415
5416 if (ctx->screen->info.chip_class >= GFX9) {
5417 for (unsigned i = 0; i < 3; i++) {
5418 LLVMValueRef hi, out;
5419
5420 hi = LLVMBuildShl(builder, vtx_out[i*2+1],
5421 LLVMConstInt(ctx->i32, 16, 0), "");
5422 out = LLVMBuildOr(builder, vtx_out[i*2], hi, "");
5423 out = ac_to_float(&ctx->ac, out);
5424 ret = LLVMBuildInsertValue(builder, ret, out,
5425 gfx9_vtx_params[i].arg_index, "");
5426 }
5427 } else {
5428 for (unsigned i = 0; i < 6; i++) {
5429 LLVMValueRef out;
5430
5431 out = ac_to_float(&ctx->ac, vtx_out[i]);
5432 ret = LLVMBuildInsertValue(builder, ret, out,
5433 gfx6_vtx_params[i].arg_index, "");
5434 }
5435 }
5436 }
5437
5438 LLVMBuildRet(builder, ret);
5439 }
5440
5441 /**
5442 * Given a list of shader part functions, build a wrapper function that
5443 * runs them in sequence to form a monolithic shader.
5444 */
5445 static void si_build_wrapper_function(struct si_shader_context *ctx,
5446 LLVMValueRef *parts,
5447 unsigned num_parts,
5448 unsigned main_part,
5449 unsigned next_shader_first_part)
5450 {
5451 LLVMBuilderRef builder = ctx->ac.builder;
5452 /* PS epilog has one arg per color component; gfx9 merged shader
5453 * prologs need to forward 40 SGPRs.
5454 */
5455 LLVMValueRef initial[AC_MAX_ARGS], out[AC_MAX_ARGS];
5456 LLVMTypeRef function_type;
5457 unsigned num_first_params;
5458 unsigned num_out, initial_num_out;
5459 ASSERTED unsigned num_out_sgpr; /* used in debug checks */
5460 ASSERTED unsigned initial_num_out_sgpr; /* used in debug checks */
5461 unsigned num_sgprs, num_vgprs;
5462 unsigned gprs;
5463
5464 memset(&ctx->args, 0, sizeof(ctx->args));
5465
5466 for (unsigned i = 0; i < num_parts; ++i) {
5467 ac_add_function_attr(ctx->ac.context, parts[i], -1,
5468 AC_FUNC_ATTR_ALWAYSINLINE);
5469 LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
5470 }
5471
5472 /* The parameters of the wrapper function correspond to those of the
5473 * first part in terms of SGPRs and VGPRs, but we use the types of the
5474 * main part to get the right types. This is relevant for the
5475 * dereferenceable attribute on descriptor table pointers.
5476 */
5477 num_sgprs = 0;
5478 num_vgprs = 0;
5479
5480 function_type = LLVMGetElementType(LLVMTypeOf(parts[0]));
5481 num_first_params = LLVMCountParamTypes(function_type);
5482
5483 for (unsigned i = 0; i < num_first_params; ++i) {
5484 LLVMValueRef param = LLVMGetParam(parts[0], i);
5485
5486 if (ac_is_sgpr_param(param)) {
5487 assert(num_vgprs == 0);
5488 num_sgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
5489 } else {
5490 num_vgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
5491 }
5492 }
5493
5494 gprs = 0;
5495 while (gprs < num_sgprs + num_vgprs) {
5496 LLVMValueRef param = LLVMGetParam(parts[main_part], ctx->args.arg_count);
5497 LLVMTypeRef type = LLVMTypeOf(param);
5498 unsigned size = ac_get_type_size(type) / 4;
5499
5500 /* This is going to get casted anyways, so we don't have to
5501 * have the exact same type. But we do have to preserve the
5502 * pointer-ness so that LLVM knows about it.
5503 */
5504 enum ac_arg_type arg_type = AC_ARG_INT;
5505 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
5506 arg_type = AC_ARG_CONST_PTR;
5507 }
5508
5509 ac_add_arg(&ctx->args, gprs < num_sgprs ? AC_ARG_SGPR : AC_ARG_VGPR,
5510 size, arg_type, NULL);
5511
5512 assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
5513 assert(gprs + size <= num_sgprs + num_vgprs &&
5514 (gprs >= num_sgprs || gprs + size <= num_sgprs));
5515
5516 gprs += size;
5517 }
5518
5519 /* Prepare the return type. */
5520 unsigned num_returns = 0;
5521 LLVMTypeRef returns[AC_MAX_ARGS], last_func_type, return_type;
5522
5523 last_func_type = LLVMGetElementType(LLVMTypeOf(parts[num_parts - 1]));
5524 return_type = LLVMGetReturnType(last_func_type);
5525
5526 switch (LLVMGetTypeKind(return_type)) {
5527 case LLVMStructTypeKind:
5528 num_returns = LLVMCountStructElementTypes(return_type);
5529 assert(num_returns <= ARRAY_SIZE(returns));
5530 LLVMGetStructElementTypes(return_type, returns);
5531 break;
5532 case LLVMVoidTypeKind:
5533 break;
5534 default:
5535 unreachable("unexpected type");
5536 }
5537
5538 si_create_function(ctx, "wrapper", returns, num_returns,
5539 si_get_max_workgroup_size(ctx->shader));
5540
5541 if (is_merged_shader(ctx))
5542 ac_init_exec_full_mask(&ctx->ac);
5543
5544 /* Record the arguments of the function as if they were an output of
5545 * a previous part.
5546 */
5547 num_out = 0;
5548 num_out_sgpr = 0;
5549
5550 for (unsigned i = 0; i < ctx->args.arg_count; ++i) {
5551 LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
5552 LLVMTypeRef param_type = LLVMTypeOf(param);
5553 LLVMTypeRef out_type = ctx->args.args[i].file == AC_ARG_SGPR ? ctx->i32 : ctx->f32;
5554 unsigned size = ac_get_type_size(param_type) / 4;
5555
5556 if (size == 1) {
5557 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
5558 param = LLVMBuildPtrToInt(builder, param, ctx->i32, "");
5559 param_type = ctx->i32;
5560 }
5561
5562 if (param_type != out_type)
5563 param = LLVMBuildBitCast(builder, param, out_type, "");
5564 out[num_out++] = param;
5565 } else {
5566 LLVMTypeRef vector_type = LLVMVectorType(out_type, size);
5567
5568 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
5569 param = LLVMBuildPtrToInt(builder, param, ctx->i64, "");
5570 param_type = ctx->i64;
5571 }
5572
5573 if (param_type != vector_type)
5574 param = LLVMBuildBitCast(builder, param, vector_type, "");
5575
5576 for (unsigned j = 0; j < size; ++j)
5577 out[num_out++] = LLVMBuildExtractElement(
5578 builder, param, LLVMConstInt(ctx->i32, j, 0), "");
5579 }
5580
5581 if (ctx->args.args[i].file == AC_ARG_SGPR)
5582 num_out_sgpr = num_out;
5583 }
5584
5585 memcpy(initial, out, sizeof(out));
5586 initial_num_out = num_out;
5587 initial_num_out_sgpr = num_out_sgpr;
5588
5589 /* Now chain the parts. */
5590 LLVMValueRef ret = NULL;
5591 for (unsigned part = 0; part < num_parts; ++part) {
5592 LLVMValueRef in[AC_MAX_ARGS];
5593 LLVMTypeRef ret_type;
5594 unsigned out_idx = 0;
5595 unsigned num_params = LLVMCountParams(parts[part]);
5596
5597 /* Merged shaders are executed conditionally depending
5598 * on the number of enabled threads passed in the input SGPRs. */
5599 if (is_multi_part_shader(ctx) && part == 0) {
5600 LLVMValueRef ena, count = initial[3];
5601
5602 count = LLVMBuildAnd(builder, count,
5603 LLVMConstInt(ctx->i32, 0x7f, 0), "");
5604 ena = LLVMBuildICmp(builder, LLVMIntULT,
5605 ac_get_thread_id(&ctx->ac), count, "");
5606 ac_build_ifcc(&ctx->ac, ena, 6506);
5607 }
5608
5609 /* Derive arguments for the next part from outputs of the
5610 * previous one.
5611 */
5612 for (unsigned param_idx = 0; param_idx < num_params; ++param_idx) {
5613 LLVMValueRef param;
5614 LLVMTypeRef param_type;
5615 bool is_sgpr;
5616 unsigned param_size;
5617 LLVMValueRef arg = NULL;
5618
5619 param = LLVMGetParam(parts[part], param_idx);
5620 param_type = LLVMTypeOf(param);
5621 param_size = ac_get_type_size(param_type) / 4;
5622 is_sgpr = ac_is_sgpr_param(param);
5623
5624 if (is_sgpr) {
5625 ac_add_function_attr(ctx->ac.context, parts[part],
5626 param_idx + 1, AC_FUNC_ATTR_INREG);
5627 } else if (out_idx < num_out_sgpr) {
5628 /* Skip returned SGPRs the current part doesn't
5629 * declare on the input. */
5630 out_idx = num_out_sgpr;
5631 }
5632
5633 assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));
5634
5635 if (param_size == 1)
5636 arg = out[out_idx];
5637 else
5638 arg = ac_build_gather_values(&ctx->ac, &out[out_idx], param_size);
5639
5640 if (LLVMTypeOf(arg) != param_type) {
5641 if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
5642 if (LLVMGetPointerAddressSpace(param_type) ==
5643 AC_ADDR_SPACE_CONST_32BIT) {
5644 arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
5645 arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
5646 } else {
5647 arg = LLVMBuildBitCast(builder, arg, ctx->i64, "");
5648 arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
5649 }
5650 } else {
5651 arg = LLVMBuildBitCast(builder, arg, param_type, "");
5652 }
5653 }
5654
5655 in[param_idx] = arg;
5656 out_idx += param_size;
5657 }
5658
5659 ret = ac_build_call(&ctx->ac, parts[part], in, num_params);
5660
5661 if (is_multi_part_shader(ctx) &&
5662 part + 1 == next_shader_first_part) {
5663 ac_build_endif(&ctx->ac, 6506);
5664
5665 /* The second half of the merged shader should use
5666 * the inputs from the toplevel (wrapper) function,
5667 * not the return value from the last call.
5668 *
5669 * That's because the last call was executed condi-
5670 * tionally, so we can't consume it in the main
5671 * block.
5672 */
5673 memcpy(out, initial, sizeof(initial));
5674 num_out = initial_num_out;
5675 num_out_sgpr = initial_num_out_sgpr;
5676 continue;
5677 }
5678
5679 /* Extract the returned GPRs. */
5680 ret_type = LLVMTypeOf(ret);
5681 num_out = 0;
5682 num_out_sgpr = 0;
5683
5684 if (LLVMGetTypeKind(ret_type) != LLVMVoidTypeKind) {
5685 assert(LLVMGetTypeKind(ret_type) == LLVMStructTypeKind);
5686
5687 unsigned ret_size = LLVMCountStructElementTypes(ret_type);
5688
5689 for (unsigned i = 0; i < ret_size; ++i) {
5690 LLVMValueRef val =
5691 LLVMBuildExtractValue(builder, ret, i, "");
5692
5693 assert(num_out < ARRAY_SIZE(out));
5694 out[num_out++] = val;
5695
5696 if (LLVMTypeOf(val) == ctx->i32) {
5697 assert(num_out_sgpr + 1 == num_out);
5698 num_out_sgpr = num_out;
5699 }
5700 }
5701 }
5702 }
5703
5704 /* Return the value from the last part. */
5705 if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
5706 LLVMBuildRetVoid(builder);
5707 else
5708 LLVMBuildRet(builder, ret);
5709 }
5710
5711 static bool si_should_optimize_less(struct ac_llvm_compiler *compiler,
5712 struct si_shader_selector *sel)
5713 {
5714 if (!compiler->low_opt_passes)
5715 return false;
5716
5717 /* Assume a slow CPU. */
5718 assert(!sel->screen->info.has_dedicated_vram &&
5719 sel->screen->info.chip_class <= GFX8);
5720
5721 /* For a crazy dEQP test containing 2597 memory opcodes, mostly
5722 * buffer stores. */
5723 return sel->type == PIPE_SHADER_COMPUTE &&
5724 sel->info.num_memory_instructions > 1000;
5725 }
5726
5727 static struct nir_shader *get_nir_shader(struct si_shader_selector *sel,
5728 bool *free_nir)
5729 {
5730 *free_nir = false;
5731
5732 if (sel->nir) {
5733 return sel->nir;
5734 } else if (sel->nir_binary) {
5735 struct pipe_screen *screen = &sel->screen->b;
5736 const void *options =
5737 screen->get_compiler_options(screen, PIPE_SHADER_IR_NIR,
5738 sel->type);
5739
5740 struct blob_reader blob_reader;
5741 blob_reader_init(&blob_reader, sel->nir_binary, sel->nir_size);
5742 *free_nir = true;
5743 return nir_deserialize(NULL, options, &blob_reader);
5744 }
5745 return NULL;
5746 }
5747
5748 int si_compile_shader(struct si_screen *sscreen,
5749 struct ac_llvm_compiler *compiler,
5750 struct si_shader *shader,
5751 struct pipe_debug_callback *debug)
5752 {
5753 struct si_shader_selector *sel = shader->selector;
5754 struct si_shader_context ctx;
5755 bool free_nir;
5756 struct nir_shader *nir = get_nir_shader(sel, &free_nir);
5757 int r = -1;
5758
5759 /* Dump NIR before doing NIR->LLVM conversion in case the
5760 * conversion fails. */
5761 if (si_can_dump_shader(sscreen, sel->type) &&
5762 !(sscreen->debug_flags & DBG(NO_NIR))) {
5763 nir_print_shader(nir, stderr);
5764 si_dump_streamout(&sel->so);
5765 }
5766
5767 si_llvm_context_init(&ctx, sscreen, compiler, si_get_shader_wave_size(shader), 64);
5768 si_llvm_context_set_ir(&ctx, shader);
5769
5770 memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
5771 sizeof(shader->info.vs_output_param_offset));
5772
5773 shader->info.uses_instanceid = sel->info.uses_instanceid;
5774
5775 if (!si_build_main_function(&ctx, nir, free_nir)) {
5776 si_llvm_dispose(&ctx);
5777 return -1;
5778 }
5779
5780 if (shader->is_monolithic && ctx.type == PIPE_SHADER_VERTEX) {
5781 LLVMValueRef parts[2];
5782 bool need_prolog = si_vs_needs_prolog(sel, &shader->key.part.vs.prolog);
5783
5784 parts[1] = ctx.main_fn;
5785
5786 if (need_prolog) {
5787 union si_shader_part_key prolog_key;
5788 si_get_vs_prolog_key(&sel->info,
5789 shader->info.num_input_sgprs,
5790 &shader->key.part.vs.prolog,
5791 shader, &prolog_key);
5792 prolog_key.vs_prolog.is_monolithic = true;
5793 si_build_vs_prolog_function(&ctx, &prolog_key);
5794 parts[0] = ctx.main_fn;
5795 }
5796
5797 si_build_wrapper_function(&ctx, parts + !need_prolog,
5798 1 + need_prolog, need_prolog, 0);
5799
5800 if (ctx.shader->key.opt.vs_as_prim_discard_cs)
5801 si_build_prim_discard_compute_shader(&ctx);
5802 } else if (shader->is_monolithic && ctx.type == PIPE_SHADER_TESS_CTRL) {
5803 if (sscreen->info.chip_class >= GFX9) {
5804 struct si_shader_selector *ls = shader->key.part.tcs.ls;
5805 LLVMValueRef parts[4];
5806 bool vs_needs_prolog =
5807 si_vs_needs_prolog(ls, &shader->key.part.tcs.ls_prolog);
5808
5809 /* TCS main part */
5810 parts[2] = ctx.main_fn;
5811
5812 /* TCS epilog */
5813 union si_shader_part_key tcs_epilog_key;
5814 memset(&tcs_epilog_key, 0, sizeof(tcs_epilog_key));
5815 tcs_epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
5816 si_build_tcs_epilog_function(&ctx, &tcs_epilog_key);
5817 parts[3] = ctx.main_fn;
5818
5819 /* VS as LS main part */
5820 nir = get_nir_shader(ls, &free_nir);
5821 struct si_shader shader_ls = {};
5822 shader_ls.selector = ls;
5823 shader_ls.key.as_ls = 1;
5824 shader_ls.key.mono = shader->key.mono;
5825 shader_ls.key.opt = shader->key.opt;
5826 shader_ls.is_monolithic = true;
5827 si_llvm_context_set_ir(&ctx, &shader_ls);
5828
5829 if (!si_build_main_function(&ctx, nir, free_nir)) {
5830 si_llvm_dispose(&ctx);
5831 return -1;
5832 }
5833 shader->info.uses_instanceid |= ls->info.uses_instanceid;
5834 parts[1] = ctx.main_fn;
5835
5836 /* LS prolog */
5837 if (vs_needs_prolog) {
5838 union si_shader_part_key vs_prolog_key;
5839 si_get_vs_prolog_key(&ls->info,
5840 shader_ls.info.num_input_sgprs,
5841 &shader->key.part.tcs.ls_prolog,
5842 shader, &vs_prolog_key);
5843 vs_prolog_key.vs_prolog.is_monolithic = true;
5844 si_build_vs_prolog_function(&ctx, &vs_prolog_key);
5845 parts[0] = ctx.main_fn;
5846 }
5847
5848 /* Reset the shader context. */
5849 ctx.shader = shader;
5850 ctx.type = PIPE_SHADER_TESS_CTRL;
5851
5852 si_build_wrapper_function(&ctx,
5853 parts + !vs_needs_prolog,
5854 4 - !vs_needs_prolog, vs_needs_prolog,
5855 vs_needs_prolog ? 2 : 1);
5856 } else {
5857 LLVMValueRef parts[2];
5858 union si_shader_part_key epilog_key;
5859
5860 parts[0] = ctx.main_fn;
5861
5862 memset(&epilog_key, 0, sizeof(epilog_key));
5863 epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
5864 si_build_tcs_epilog_function(&ctx, &epilog_key);
5865 parts[1] = ctx.main_fn;
5866
5867 si_build_wrapper_function(&ctx, parts, 2, 0, 0);
5868 }
5869 } else if (shader->is_monolithic && ctx.type == PIPE_SHADER_GEOMETRY) {
5870 if (ctx.screen->info.chip_class >= GFX9) {
5871 struct si_shader_selector *es = shader->key.part.gs.es;
5872 LLVMValueRef es_prolog = NULL;
5873 LLVMValueRef es_main = NULL;
5874 LLVMValueRef gs_prolog = NULL;
5875 LLVMValueRef gs_main = ctx.main_fn;
5876
5877 /* GS prolog */
5878 union si_shader_part_key gs_prolog_key;
5879 memset(&gs_prolog_key, 0, sizeof(gs_prolog_key));
5880 gs_prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
5881 gs_prolog_key.gs_prolog.is_monolithic = true;
5882 gs_prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
5883 si_build_gs_prolog_function(&ctx, &gs_prolog_key);
5884 gs_prolog = ctx.main_fn;
5885
5886 /* ES main part */
5887 nir = get_nir_shader(es, &free_nir);
5888 struct si_shader shader_es = {};
5889 shader_es.selector = es;
5890 shader_es.key.as_es = 1;
5891 shader_es.key.as_ngg = shader->key.as_ngg;
5892 shader_es.key.mono = shader->key.mono;
5893 shader_es.key.opt = shader->key.opt;
5894 shader_es.is_monolithic = true;
5895 si_llvm_context_set_ir(&ctx, &shader_es);
5896
5897 if (!si_build_main_function(&ctx, nir, free_nir)) {
5898 si_llvm_dispose(&ctx);
5899 return -1;
5900 }
5901 shader->info.uses_instanceid |= es->info.uses_instanceid;
5902 es_main = ctx.main_fn;
5903
5904 /* ES prolog */
5905 if (es->type == PIPE_SHADER_VERTEX &&
5906 si_vs_needs_prolog(es, &shader->key.part.gs.vs_prolog)) {
5907 union si_shader_part_key vs_prolog_key;
5908 si_get_vs_prolog_key(&es->info,
5909 shader_es.info.num_input_sgprs,
5910 &shader->key.part.gs.vs_prolog,
5911 shader, &vs_prolog_key);
5912 vs_prolog_key.vs_prolog.is_monolithic = true;
5913 si_build_vs_prolog_function(&ctx, &vs_prolog_key);
5914 es_prolog = ctx.main_fn;
5915 }
5916
5917 /* Reset the shader context. */
5918 ctx.shader = shader;
5919 ctx.type = PIPE_SHADER_GEOMETRY;
5920
5921 /* Prepare the array of shader parts. */
5922 LLVMValueRef parts[4];
5923 unsigned num_parts = 0, main_part, next_first_part;
5924
5925 if (es_prolog)
5926 parts[num_parts++] = es_prolog;
5927
5928 parts[main_part = num_parts++] = es_main;
5929 parts[next_first_part = num_parts++] = gs_prolog;
5930 parts[num_parts++] = gs_main;
5931
5932 si_build_wrapper_function(&ctx, parts, num_parts,
5933 main_part, next_first_part);
5934 } else {
5935 LLVMValueRef parts[2];
5936 union si_shader_part_key prolog_key;
5937
5938 parts[1] = ctx.main_fn;
5939
5940 memset(&prolog_key, 0, sizeof(prolog_key));
5941 prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
5942 si_build_gs_prolog_function(&ctx, &prolog_key);
5943 parts[0] = ctx.main_fn;
5944
5945 si_build_wrapper_function(&ctx, parts, 2, 1, 0);
5946 }
5947 } else if (shader->is_monolithic && ctx.type == PIPE_SHADER_FRAGMENT) {
5948 LLVMValueRef parts[3];
5949 union si_shader_part_key prolog_key;
5950 union si_shader_part_key epilog_key;
5951 bool need_prolog;
5952
5953 si_get_ps_prolog_key(shader, &prolog_key, false);
5954 need_prolog = si_need_ps_prolog(&prolog_key);
5955
5956 parts[need_prolog ? 1 : 0] = ctx.main_fn;
5957
5958 if (need_prolog) {
5959 si_build_ps_prolog_function(&ctx, &prolog_key);
5960 parts[0] = ctx.main_fn;
5961 }
5962
5963 si_get_ps_epilog_key(shader, &epilog_key);
5964 si_build_ps_epilog_function(&ctx, &epilog_key);
5965 parts[need_prolog ? 2 : 1] = ctx.main_fn;
5966
5967 si_build_wrapper_function(&ctx, parts, need_prolog ? 3 : 2,
5968 need_prolog ? 1 : 0, 0);
5969 }
5970
5971 si_llvm_optimize_module(&ctx);
5972
5973 /* Post-optimization transformations and analysis. */
5974 si_optimize_vs_outputs(&ctx);
5975
5976 if ((debug && debug->debug_message) ||
5977 si_can_dump_shader(sscreen, ctx.type)) {
5978 ctx.shader->info.private_mem_vgprs =
5979 ac_count_scratch_private_memory(ctx.main_fn);
5980 }
5981
5982 /* Make sure the input is a pointer and not integer followed by inttoptr. */
5983 assert(LLVMGetTypeKind(LLVMTypeOf(LLVMGetParam(ctx.main_fn, 0))) ==
5984 LLVMPointerTypeKind);
5985
5986 /* Compile to bytecode. */
5987 r = si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler,
5988 ctx.ac.module, debug, ctx.type, ctx.ac.wave_size,
5989 si_get_shader_name(shader),
5990 si_should_optimize_less(compiler, shader->selector));
5991 si_llvm_dispose(&ctx);
5992 if (r) {
5993 fprintf(stderr, "LLVM failed to compile shader\n");
5994 return r;
5995 }
5996
5997 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
5998 * LLVM 3.9svn has this bug.
5999 */
6000 if (sel->type == PIPE_SHADER_COMPUTE) {
6001 unsigned wave_size = sscreen->compute_wave_size;
6002 unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd *
6003 (wave_size == 32 ? 2 : 1);
6004 unsigned max_sgprs = sscreen->info.num_physical_sgprs_per_simd;
6005 unsigned max_sgprs_per_wave = 128;
6006 unsigned simds_per_tg = 4; /* assuming WGP mode on gfx10 */
6007 unsigned threads_per_tg = si_get_max_workgroup_size(shader);
6008 unsigned waves_per_tg = DIV_ROUND_UP(threads_per_tg, wave_size);
6009 unsigned waves_per_simd = DIV_ROUND_UP(waves_per_tg, simds_per_tg);
6010
6011 max_vgprs = max_vgprs / waves_per_simd;
6012 max_sgprs = MIN2(max_sgprs / waves_per_simd, max_sgprs_per_wave);
6013
6014 if (shader->config.num_sgprs > max_sgprs ||
6015 shader->config.num_vgprs > max_vgprs) {
6016 fprintf(stderr, "LLVM failed to compile a shader correctly: "
6017 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
6018 shader->config.num_sgprs, shader->config.num_vgprs,
6019 max_sgprs, max_vgprs);
6020
6021 /* Just terminate the process, because dependent
6022 * shaders can hang due to bad input data, but use
6023 * the env var to allow shader-db to work.
6024 */
6025 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
6026 abort();
6027 }
6028 }
6029
6030 /* Add the scratch offset to input SGPRs. */
6031 if (shader->config.scratch_bytes_per_wave && !is_merged_shader(&ctx))
6032 shader->info.num_input_sgprs += 1; /* scratch byte offset */
6033
6034 /* Calculate the number of fragment input VGPRs. */
6035 if (ctx.type == PIPE_SHADER_FRAGMENT) {
6036 shader->info.num_input_vgprs = ac_get_fs_input_vgpr_cnt(&shader->config,
6037 &shader->info.face_vgpr_index,
6038 &shader->info.ancillary_vgpr_index);
6039 }
6040
6041 si_calculate_max_simd_waves(shader);
6042 si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
6043 return 0;
6044 }
6045
6046 /**
6047 * Create, compile and return a shader part (prolog or epilog).
6048 *
6049 * \param sscreen screen
6050 * \param list list of shader parts of the same category
6051 * \param type shader type
6052 * \param key shader part key
6053 * \param prolog whether the part being requested is a prolog
6054 * \param tm LLVM target machine
6055 * \param debug debug callback
6056 * \param build the callback responsible for building the main function
6057 * \return non-NULL on success
6058 */
6059 static struct si_shader_part *
6060 si_get_shader_part(struct si_screen *sscreen,
6061 struct si_shader_part **list,
6062 enum pipe_shader_type type,
6063 bool prolog,
6064 union si_shader_part_key *key,
6065 struct ac_llvm_compiler *compiler,
6066 struct pipe_debug_callback *debug,
6067 void (*build)(struct si_shader_context *,
6068 union si_shader_part_key *),
6069 const char *name)
6070 {
6071 struct si_shader_part *result;
6072
6073 simple_mtx_lock(&sscreen->shader_parts_mutex);
6074
6075 /* Find existing. */
6076 for (result = *list; result; result = result->next) {
6077 if (memcmp(&result->key, key, sizeof(*key)) == 0) {
6078 simple_mtx_unlock(&sscreen->shader_parts_mutex);
6079 return result;
6080 }
6081 }
6082
6083 /* Compile a new one. */
6084 result = CALLOC_STRUCT(si_shader_part);
6085 result->key = *key;
6086
6087 struct si_shader shader = {};
6088
6089 switch (type) {
6090 case PIPE_SHADER_VERTEX:
6091 shader.key.as_ls = key->vs_prolog.as_ls;
6092 shader.key.as_es = key->vs_prolog.as_es;
6093 shader.key.as_ngg = key->vs_prolog.as_ngg;
6094 break;
6095 case PIPE_SHADER_TESS_CTRL:
6096 assert(!prolog);
6097 shader.key.part.tcs.epilog = key->tcs_epilog.states;
6098 break;
6099 case PIPE_SHADER_GEOMETRY:
6100 assert(prolog);
6101 shader.key.as_ngg = key->gs_prolog.as_ngg;
6102 break;
6103 case PIPE_SHADER_FRAGMENT:
6104 if (prolog)
6105 shader.key.part.ps.prolog = key->ps_prolog.states;
6106 else
6107 shader.key.part.ps.epilog = key->ps_epilog.states;
6108 break;
6109 default:
6110 unreachable("bad shader part");
6111 }
6112
6113 struct si_shader_context ctx;
6114 si_llvm_context_init(&ctx, sscreen, compiler,
6115 si_get_wave_size(sscreen, type, shader.key.as_ngg,
6116 shader.key.as_es),
6117 64);
6118 ctx.shader = &shader;
6119 ctx.type = type;
6120
6121 build(&ctx, key);
6122
6123 /* Compile. */
6124 si_llvm_optimize_module(&ctx);
6125
6126 if (si_compile_llvm(sscreen, &result->binary, &result->config, compiler,
6127 ctx.ac.module, debug, ctx.type, ctx.ac.wave_size,
6128 name, false)) {
6129 FREE(result);
6130 result = NULL;
6131 goto out;
6132 }
6133
6134 result->next = *list;
6135 *list = result;
6136
6137 out:
6138 si_llvm_dispose(&ctx);
6139 simple_mtx_unlock(&sscreen->shader_parts_mutex);
6140 return result;
6141 }
6142
6143 static LLVMValueRef si_prolog_get_rw_buffers(struct si_shader_context *ctx)
6144 {
6145 LLVMValueRef ptr[2], list;
6146 bool merged_shader = is_merged_shader(ctx);
6147
6148 ptr[0] = LLVMGetParam(ctx->main_fn, (merged_shader ? 8 : 0) + SI_SGPR_RW_BUFFERS);
6149 list = LLVMBuildIntToPtr(ctx->ac.builder, ptr[0],
6150 ac_array_in_const32_addr_space(ctx->v4i32), "");
6151 return list;
6152 }
6153
6154 /**
6155 * Build the vertex shader prolog function.
6156 *
6157 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
6158 * All inputs are returned unmodified. The vertex load indices are
6159 * stored after them, which will be used by the API VS for fetching inputs.
6160 *
6161 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
6162 * input_v0,
6163 * input_v1,
6164 * input_v2,
6165 * input_v3,
6166 * (VertexID + BaseVertex),
6167 * (InstanceID + StartInstance),
6168 * (InstanceID / 2 + StartInstance)
6169 */
6170 static void si_build_vs_prolog_function(struct si_shader_context *ctx,
6171 union si_shader_part_key *key)
6172 {
6173 LLVMTypeRef *returns;
6174 LLVMValueRef ret, func;
6175 int num_returns, i;
6176 unsigned first_vs_vgpr = key->vs_prolog.num_merged_next_stage_vgprs;
6177 unsigned num_input_vgprs = key->vs_prolog.num_merged_next_stage_vgprs + 4;
6178 struct ac_arg input_sgpr_param[key->vs_prolog.num_input_sgprs];
6179 struct ac_arg input_vgpr_param[9];
6180 LLVMValueRef input_vgprs[9];
6181 unsigned num_all_input_regs = key->vs_prolog.num_input_sgprs +
6182 num_input_vgprs;
6183 unsigned user_sgpr_base = key->vs_prolog.num_merged_next_stage_vgprs ? 8 : 0;
6184
6185 memset(&ctx->args, 0, sizeof(ctx->args));
6186
6187 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
6188 returns = alloca((num_all_input_regs + key->vs_prolog.num_inputs) *
6189 sizeof(LLVMTypeRef));
6190 num_returns = 0;
6191
6192 /* Declare input and output SGPRs. */
6193 for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
6194 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6195 &input_sgpr_param[i]);
6196 returns[num_returns++] = ctx->i32;
6197 }
6198
6199 struct ac_arg merged_wave_info = input_sgpr_param[3];
6200
6201 /* Preloaded VGPRs (outputs must be floats) */
6202 for (i = 0; i < num_input_vgprs; i++) {
6203 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &input_vgpr_param[i]);
6204 returns[num_returns++] = ctx->f32;
6205 }
6206
6207 /* Vertex load indices. */
6208 for (i = 0; i < key->vs_prolog.num_inputs; i++)
6209 returns[num_returns++] = ctx->f32;
6210
6211 /* Create the function. */
6212 si_create_function(ctx, "vs_prolog", returns, num_returns, 0);
6213 func = ctx->main_fn;
6214
6215 for (i = 0; i < num_input_vgprs; i++) {
6216 input_vgprs[i] = ac_get_arg(&ctx->ac, input_vgpr_param[i]);
6217 }
6218
6219 if (key->vs_prolog.num_merged_next_stage_vgprs) {
6220 if (!key->vs_prolog.is_monolithic)
6221 si_init_exec_from_input(ctx, merged_wave_info, 0);
6222
6223 if (key->vs_prolog.as_ls &&
6224 ctx->screen->info.has_ls_vgpr_init_bug) {
6225 /* If there are no HS threads, SPI loads the LS VGPRs
6226 * starting at VGPR 0. Shift them back to where they
6227 * belong.
6228 */
6229 LLVMValueRef has_hs_threads =
6230 LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
6231 si_unpack_param(ctx, input_sgpr_param[3], 8, 8),
6232 ctx->i32_0, "");
6233
6234 for (i = 4; i > 0; --i) {
6235 input_vgprs[i + 1] =
6236 LLVMBuildSelect(ctx->ac.builder, has_hs_threads,
6237 input_vgprs[i + 1],
6238 input_vgprs[i - 1], "");
6239 }
6240 }
6241 }
6242
6243 unsigned vertex_id_vgpr = first_vs_vgpr;
6244 unsigned instance_id_vgpr =
6245 ctx->screen->info.chip_class >= GFX10 ?
6246 first_vs_vgpr + 3 :
6247 first_vs_vgpr + (key->vs_prolog.as_ls ? 2 : 1);
6248
6249 ctx->abi.vertex_id = input_vgprs[vertex_id_vgpr];
6250 ctx->abi.instance_id = input_vgprs[instance_id_vgpr];
6251
6252 /* InstanceID = VertexID >> 16;
6253 * VertexID = VertexID & 0xffff;
6254 */
6255 if (key->vs_prolog.states.unpack_instance_id_from_vertex_id) {
6256 ctx->abi.instance_id = LLVMBuildLShr(ctx->ac.builder, ctx->abi.vertex_id,
6257 LLVMConstInt(ctx->i32, 16, 0), "");
6258 ctx->abi.vertex_id = LLVMBuildAnd(ctx->ac.builder, ctx->abi.vertex_id,
6259 LLVMConstInt(ctx->i32, 0xffff, 0), "");
6260 }
6261
6262 /* Copy inputs to outputs. This should be no-op, as the registers match,
6263 * but it will prevent the compiler from overwriting them unintentionally.
6264 */
6265 ret = ctx->return_value;
6266 for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
6267 LLVMValueRef p = LLVMGetParam(func, i);
6268 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
6269 }
6270 for (i = 0; i < num_input_vgprs; i++) {
6271 LLVMValueRef p = input_vgprs[i];
6272
6273 if (i == vertex_id_vgpr)
6274 p = ctx->abi.vertex_id;
6275 else if (i == instance_id_vgpr)
6276 p = ctx->abi.instance_id;
6277
6278 p = ac_to_float(&ctx->ac, p);
6279 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p,
6280 key->vs_prolog.num_input_sgprs + i, "");
6281 }
6282
6283 /* Compute vertex load indices from instance divisors. */
6284 LLVMValueRef instance_divisor_constbuf = NULL;
6285
6286 if (key->vs_prolog.states.instance_divisor_is_fetched) {
6287 LLVMValueRef list = si_prolog_get_rw_buffers(ctx);
6288 LLVMValueRef buf_index =
6289 LLVMConstInt(ctx->i32, SI_VS_CONST_INSTANCE_DIVISORS, 0);
6290 instance_divisor_constbuf =
6291 ac_build_load_to_sgpr(&ctx->ac, list, buf_index);
6292 }
6293
6294 for (i = 0; i < key->vs_prolog.num_inputs; i++) {
6295 bool divisor_is_one =
6296 key->vs_prolog.states.instance_divisor_is_one & (1u << i);
6297 bool divisor_is_fetched =
6298 key->vs_prolog.states.instance_divisor_is_fetched & (1u << i);
6299 LLVMValueRef index = NULL;
6300
6301 if (divisor_is_one) {
6302 index = ctx->abi.instance_id;
6303 } else if (divisor_is_fetched) {
6304 LLVMValueRef udiv_factors[4];
6305
6306 for (unsigned j = 0; j < 4; j++) {
6307 udiv_factors[j] =
6308 buffer_load_const(ctx, instance_divisor_constbuf,
6309 LLVMConstInt(ctx->i32, i*16 + j*4, 0));
6310 udiv_factors[j] = ac_to_integer(&ctx->ac, udiv_factors[j]);
6311 }
6312 /* The faster NUW version doesn't work when InstanceID == UINT_MAX.
6313 * Such InstanceID might not be achievable in a reasonable time though.
6314 */
6315 index = ac_build_fast_udiv_nuw(&ctx->ac, ctx->abi.instance_id,
6316 udiv_factors[0], udiv_factors[1],
6317 udiv_factors[2], udiv_factors[3]);
6318 }
6319
6320 if (divisor_is_one || divisor_is_fetched) {
6321 /* Add StartInstance. */
6322 index = LLVMBuildAdd(ctx->ac.builder, index,
6323 LLVMGetParam(ctx->main_fn, user_sgpr_base +
6324 SI_SGPR_START_INSTANCE), "");
6325 } else {
6326 /* VertexID + BaseVertex */
6327 index = LLVMBuildAdd(ctx->ac.builder,
6328 ctx->abi.vertex_id,
6329 LLVMGetParam(func, user_sgpr_base +
6330 SI_SGPR_BASE_VERTEX), "");
6331 }
6332
6333 index = ac_to_float(&ctx->ac, index);
6334 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, index,
6335 ctx->args.arg_count + i, "");
6336 }
6337
6338 si_llvm_build_ret(ctx, ret);
6339 }
6340
6341 static bool si_get_vs_prolog(struct si_screen *sscreen,
6342 struct ac_llvm_compiler *compiler,
6343 struct si_shader *shader,
6344 struct pipe_debug_callback *debug,
6345 struct si_shader *main_part,
6346 const struct si_vs_prolog_bits *key)
6347 {
6348 struct si_shader_selector *vs = main_part->selector;
6349
6350 if (!si_vs_needs_prolog(vs, key))
6351 return true;
6352
6353 /* Get the prolog. */
6354 union si_shader_part_key prolog_key;
6355 si_get_vs_prolog_key(&vs->info, main_part->info.num_input_sgprs,
6356 key, shader, &prolog_key);
6357
6358 shader->prolog =
6359 si_get_shader_part(sscreen, &sscreen->vs_prologs,
6360 PIPE_SHADER_VERTEX, true, &prolog_key, compiler,
6361 debug, si_build_vs_prolog_function,
6362 "Vertex Shader Prolog");
6363 return shader->prolog != NULL;
6364 }
6365
6366 /**
6367 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
6368 */
6369 static bool si_shader_select_vs_parts(struct si_screen *sscreen,
6370 struct ac_llvm_compiler *compiler,
6371 struct si_shader *shader,
6372 struct pipe_debug_callback *debug)
6373 {
6374 return si_get_vs_prolog(sscreen, compiler, shader, debug, shader,
6375 &shader->key.part.vs.prolog);
6376 }
6377
6378 /**
6379 * Compile the TCS epilog function. This writes tesselation factors to memory
6380 * based on the output primitive type of the tesselator (determined by TES).
6381 */
6382 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
6383 union si_shader_part_key *key)
6384 {
6385 memset(&ctx->args, 0, sizeof(ctx->args));
6386
6387 if (ctx->screen->info.chip_class >= GFX9) {
6388 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6389 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6390 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6391 &ctx->tcs_offchip_offset);
6392 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* wave info */
6393 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6394 &ctx->tcs_factor_offset);
6395 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6396 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6397 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6398 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6399 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6400 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6401 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6402 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6403 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6404 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6405 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6406 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6407 &ctx->tcs_offchip_layout);
6408 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6409 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6410 &ctx->tcs_out_lds_layout);
6411 } else {
6412 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6413 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6414 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6415 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6416 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6417 &ctx->tcs_offchip_layout);
6418 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6419 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6420 &ctx->tcs_out_lds_layout);
6421 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6422 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6423 &ctx->tcs_offchip_offset);
6424 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6425 &ctx->tcs_factor_offset);
6426 }
6427
6428 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* VGPR gap */
6429 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* VGPR gap */
6430 struct ac_arg rel_patch_id; /* patch index within the wave (REL_PATCH_ID) */
6431 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &rel_patch_id);
6432 struct ac_arg invocation_id; /* invocation ID within the patch */
6433 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &invocation_id);
6434 struct ac_arg tcs_out_current_patch_data_offset; /* LDS offset where tess factors should be loaded from */
6435 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
6436 &tcs_out_current_patch_data_offset);
6437
6438 struct ac_arg tess_factors[6];
6439 for (unsigned i = 0; i < 6; i++)
6440 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &tess_factors[i]);
6441
6442 /* Create the function. */
6443 si_create_function(ctx, "tcs_epilog", NULL, 0,
6444 ctx->screen->info.chip_class >= GFX7 ? 128 : 0);
6445 ac_declare_lds_as_pointer(&ctx->ac);
6446
6447 LLVMValueRef invoc0_tess_factors[6];
6448 for (unsigned i = 0; i < 6; i++)
6449 invoc0_tess_factors[i] = ac_get_arg(&ctx->ac, tess_factors[i]);
6450
6451 si_write_tess_factors(ctx,
6452 ac_get_arg(&ctx->ac, rel_patch_id),
6453 ac_get_arg(&ctx->ac, invocation_id),
6454 ac_get_arg(&ctx->ac, tcs_out_current_patch_data_offset),
6455 invoc0_tess_factors, invoc0_tess_factors + 4);
6456
6457 LLVMBuildRetVoid(ctx->ac.builder);
6458 }
6459
6460 /**
6461 * Select and compile (or reuse) TCS parts (epilog).
6462 */
6463 static bool si_shader_select_tcs_parts(struct si_screen *sscreen,
6464 struct ac_llvm_compiler *compiler,
6465 struct si_shader *shader,
6466 struct pipe_debug_callback *debug)
6467 {
6468 if (sscreen->info.chip_class >= GFX9) {
6469 struct si_shader *ls_main_part =
6470 shader->key.part.tcs.ls->main_shader_part_ls;
6471
6472 if (!si_get_vs_prolog(sscreen, compiler, shader, debug, ls_main_part,
6473 &shader->key.part.tcs.ls_prolog))
6474 return false;
6475
6476 shader->previous_stage = ls_main_part;
6477 }
6478
6479 /* Get the epilog. */
6480 union si_shader_part_key epilog_key;
6481 memset(&epilog_key, 0, sizeof(epilog_key));
6482 epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
6483
6484 shader->epilog = si_get_shader_part(sscreen, &sscreen->tcs_epilogs,
6485 PIPE_SHADER_TESS_CTRL, false,
6486 &epilog_key, compiler, debug,
6487 si_build_tcs_epilog_function,
6488 "Tessellation Control Shader Epilog");
6489 return shader->epilog != NULL;
6490 }
6491
6492 /**
6493 * Select and compile (or reuse) GS parts (prolog).
6494 */
6495 static bool si_shader_select_gs_parts(struct si_screen *sscreen,
6496 struct ac_llvm_compiler *compiler,
6497 struct si_shader *shader,
6498 struct pipe_debug_callback *debug)
6499 {
6500 if (sscreen->info.chip_class >= GFX9) {
6501 struct si_shader *es_main_part;
6502 enum pipe_shader_type es_type = shader->key.part.gs.es->type;
6503
6504 if (shader->key.as_ngg)
6505 es_main_part = shader->key.part.gs.es->main_shader_part_ngg_es;
6506 else
6507 es_main_part = shader->key.part.gs.es->main_shader_part_es;
6508
6509 if (es_type == PIPE_SHADER_VERTEX &&
6510 !si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
6511 &shader->key.part.gs.vs_prolog))
6512 return false;
6513
6514 shader->previous_stage = es_main_part;
6515 }
6516
6517 if (!shader->key.part.gs.prolog.tri_strip_adj_fix)
6518 return true;
6519
6520 union si_shader_part_key prolog_key;
6521 memset(&prolog_key, 0, sizeof(prolog_key));
6522 prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
6523 prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
6524
6525 shader->prolog2 = si_get_shader_part(sscreen, &sscreen->gs_prologs,
6526 PIPE_SHADER_GEOMETRY, true,
6527 &prolog_key, compiler, debug,
6528 si_build_gs_prolog_function,
6529 "Geometry Shader Prolog");
6530 return shader->prolog2 != NULL;
6531 }
6532
6533 /**
6534 * Build the pixel shader prolog function. This handles:
6535 * - two-side color selection and interpolation
6536 * - overriding interpolation parameters for the API PS
6537 * - polygon stippling
6538 *
6539 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
6540 * overriden by other states. (e.g. per-sample interpolation)
6541 * Interpolated colors are stored after the preloaded VGPRs.
6542 */
6543 static void si_build_ps_prolog_function(struct si_shader_context *ctx,
6544 union si_shader_part_key *key)
6545 {
6546 LLVMValueRef ret, func;
6547 int num_returns, i, num_color_channels;
6548
6549 assert(si_need_ps_prolog(key));
6550
6551 memset(&ctx->args, 0, sizeof(ctx->args));
6552
6553 /* Declare inputs. */
6554 LLVMTypeRef return_types[AC_MAX_ARGS];
6555 num_returns = 0;
6556 num_color_channels = util_bitcount(key->ps_prolog.colors_read);
6557 assert(key->ps_prolog.num_input_sgprs +
6558 key->ps_prolog.num_input_vgprs +
6559 num_color_channels <= AC_MAX_ARGS);
6560 for (i = 0; i < key->ps_prolog.num_input_sgprs; i++) {
6561 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
6562 return_types[num_returns++] = ctx->i32;
6563
6564 }
6565
6566 struct ac_arg pos_fixed_pt;
6567 struct ac_arg ancillary;
6568 struct ac_arg param_sample_mask;
6569 for (i = 0; i < key->ps_prolog.num_input_vgprs; i++) {
6570 struct ac_arg *arg = NULL;
6571 if (i == key->ps_prolog.ancillary_vgpr_index) {
6572 arg = &ancillary;
6573 } else if (i == key->ps_prolog.ancillary_vgpr_index + 1) {
6574 arg = &param_sample_mask;
6575 } else if (i == key->ps_prolog.num_input_vgprs - 1) {
6576 /* POS_FIXED_PT is always last. */
6577 arg = &pos_fixed_pt;
6578 }
6579 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, arg);
6580 return_types[num_returns++] = ctx->f32;
6581 }
6582
6583 /* Declare outputs (same as inputs + add colors if needed) */
6584 for (i = 0; i < num_color_channels; i++)
6585 return_types[num_returns++] = ctx->f32;
6586
6587 /* Create the function. */
6588 si_create_function(ctx, "ps_prolog", return_types, num_returns, 0);
6589 func = ctx->main_fn;
6590
6591 /* Copy inputs to outputs. This should be no-op, as the registers match,
6592 * but it will prevent the compiler from overwriting them unintentionally.
6593 */
6594 ret = ctx->return_value;
6595 for (i = 0; i < ctx->args.arg_count; i++) {
6596 LLVMValueRef p = LLVMGetParam(func, i);
6597 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
6598 }
6599
6600 /* Polygon stippling. */
6601 if (key->ps_prolog.states.poly_stipple) {
6602 LLVMValueRef list = si_prolog_get_rw_buffers(ctx);
6603
6604 si_llvm_emit_polygon_stipple(ctx, list, pos_fixed_pt);
6605 }
6606
6607 if (key->ps_prolog.states.bc_optimize_for_persp ||
6608 key->ps_prolog.states.bc_optimize_for_linear) {
6609 unsigned i, base = key->ps_prolog.num_input_sgprs;
6610 LLVMValueRef center[2], centroid[2], tmp, bc_optimize;
6611
6612 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
6613 * The hw doesn't compute CENTROID if the whole wave only
6614 * contains fully-covered quads.
6615 *
6616 * PRIM_MASK is after user SGPRs.
6617 */
6618 bc_optimize = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
6619 bc_optimize = LLVMBuildLShr(ctx->ac.builder, bc_optimize,
6620 LLVMConstInt(ctx->i32, 31, 0), "");
6621 bc_optimize = LLVMBuildTrunc(ctx->ac.builder, bc_optimize,
6622 ctx->i1, "");
6623
6624 if (key->ps_prolog.states.bc_optimize_for_persp) {
6625 /* Read PERSP_CENTER. */
6626 for (i = 0; i < 2; i++)
6627 center[i] = LLVMGetParam(func, base + 2 + i);
6628 /* Read PERSP_CENTROID. */
6629 for (i = 0; i < 2; i++)
6630 centroid[i] = LLVMGetParam(func, base + 4 + i);
6631 /* Select PERSP_CENTROID. */
6632 for (i = 0; i < 2; i++) {
6633 tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize,
6634 center[i], centroid[i], "");
6635 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6636 tmp, base + 4 + i, "");
6637 }
6638 }
6639 if (key->ps_prolog.states.bc_optimize_for_linear) {
6640 /* Read LINEAR_CENTER. */
6641 for (i = 0; i < 2; i++)
6642 center[i] = LLVMGetParam(func, base + 8 + i);
6643 /* Read LINEAR_CENTROID. */
6644 for (i = 0; i < 2; i++)
6645 centroid[i] = LLVMGetParam(func, base + 10 + i);
6646 /* Select LINEAR_CENTROID. */
6647 for (i = 0; i < 2; i++) {
6648 tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize,
6649 center[i], centroid[i], "");
6650 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6651 tmp, base + 10 + i, "");
6652 }
6653 }
6654 }
6655
6656 /* Force per-sample interpolation. */
6657 if (key->ps_prolog.states.force_persp_sample_interp) {
6658 unsigned i, base = key->ps_prolog.num_input_sgprs;
6659 LLVMValueRef persp_sample[2];
6660
6661 /* Read PERSP_SAMPLE. */
6662 for (i = 0; i < 2; i++)
6663 persp_sample[i] = LLVMGetParam(func, base + i);
6664 /* Overwrite PERSP_CENTER. */
6665 for (i = 0; i < 2; i++)
6666 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6667 persp_sample[i], base + 2 + i, "");
6668 /* Overwrite PERSP_CENTROID. */
6669 for (i = 0; i < 2; i++)
6670 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6671 persp_sample[i], base + 4 + i, "");
6672 }
6673 if (key->ps_prolog.states.force_linear_sample_interp) {
6674 unsigned i, base = key->ps_prolog.num_input_sgprs;
6675 LLVMValueRef linear_sample[2];
6676
6677 /* Read LINEAR_SAMPLE. */
6678 for (i = 0; i < 2; i++)
6679 linear_sample[i] = LLVMGetParam(func, base + 6 + i);
6680 /* Overwrite LINEAR_CENTER. */
6681 for (i = 0; i < 2; i++)
6682 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6683 linear_sample[i], base + 8 + i, "");
6684 /* Overwrite LINEAR_CENTROID. */
6685 for (i = 0; i < 2; i++)
6686 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6687 linear_sample[i], base + 10 + i, "");
6688 }
6689
6690 /* Force center interpolation. */
6691 if (key->ps_prolog.states.force_persp_center_interp) {
6692 unsigned i, base = key->ps_prolog.num_input_sgprs;
6693 LLVMValueRef persp_center[2];
6694
6695 /* Read PERSP_CENTER. */
6696 for (i = 0; i < 2; i++)
6697 persp_center[i] = LLVMGetParam(func, base + 2 + i);
6698 /* Overwrite PERSP_SAMPLE. */
6699 for (i = 0; i < 2; i++)
6700 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6701 persp_center[i], base + i, "");
6702 /* Overwrite PERSP_CENTROID. */
6703 for (i = 0; i < 2; i++)
6704 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6705 persp_center[i], base + 4 + i, "");
6706 }
6707 if (key->ps_prolog.states.force_linear_center_interp) {
6708 unsigned i, base = key->ps_prolog.num_input_sgprs;
6709 LLVMValueRef linear_center[2];
6710
6711 /* Read LINEAR_CENTER. */
6712 for (i = 0; i < 2; i++)
6713 linear_center[i] = LLVMGetParam(func, base + 8 + i);
6714 /* Overwrite LINEAR_SAMPLE. */
6715 for (i = 0; i < 2; i++)
6716 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6717 linear_center[i], base + 6 + i, "");
6718 /* Overwrite LINEAR_CENTROID. */
6719 for (i = 0; i < 2; i++)
6720 ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
6721 linear_center[i], base + 10 + i, "");
6722 }
6723
6724 /* Interpolate colors. */
6725 unsigned color_out_idx = 0;
6726 for (i = 0; i < 2; i++) {
6727 unsigned writemask = (key->ps_prolog.colors_read >> (i * 4)) & 0xf;
6728 unsigned face_vgpr = key->ps_prolog.num_input_sgprs +
6729 key->ps_prolog.face_vgpr_index;
6730 LLVMValueRef interp[2], color[4];
6731 LLVMValueRef interp_ij = NULL, prim_mask = NULL, face = NULL;
6732
6733 if (!writemask)
6734 continue;
6735
6736 /* If the interpolation qualifier is not CONSTANT (-1). */
6737 if (key->ps_prolog.color_interp_vgpr_index[i] != -1) {
6738 unsigned interp_vgpr = key->ps_prolog.num_input_sgprs +
6739 key->ps_prolog.color_interp_vgpr_index[i];
6740
6741 /* Get the (i,j) updated by bc_optimize handling. */
6742 interp[0] = LLVMBuildExtractValue(ctx->ac.builder, ret,
6743 interp_vgpr, "");
6744 interp[1] = LLVMBuildExtractValue(ctx->ac.builder, ret,
6745 interp_vgpr + 1, "");
6746 interp_ij = ac_build_gather_values(&ctx->ac, interp, 2);
6747 }
6748
6749 /* Use the absolute location of the input. */
6750 prim_mask = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
6751
6752 if (key->ps_prolog.states.color_two_side) {
6753 face = LLVMGetParam(func, face_vgpr);
6754 face = ac_to_integer(&ctx->ac, face);
6755 }
6756
6757 interp_fs_color(ctx,
6758 key->ps_prolog.color_attr_index[i], i,
6759 key->ps_prolog.num_interp_inputs,
6760 key->ps_prolog.colors_read, interp_ij,
6761 prim_mask, face, color);
6762
6763 while (writemask) {
6764 unsigned chan = u_bit_scan(&writemask);
6765 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, color[chan],
6766 ctx->args.arg_count + color_out_idx++, "");
6767 }
6768 }
6769
6770 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
6771 * says:
6772 *
6773 * "When per-sample shading is active due to the use of a fragment
6774 * input qualified by sample or due to the use of the gl_SampleID
6775 * or gl_SamplePosition variables, only the bit for the current
6776 * sample is set in gl_SampleMaskIn. When state specifies multiple
6777 * fragment shader invocations for a given fragment, the sample
6778 * mask for any single fragment shader invocation may specify a
6779 * subset of the covered samples for the fragment. In this case,
6780 * the bit corresponding to each covered sample will be set in
6781 * exactly one fragment shader invocation."
6782 *
6783 * The samplemask loaded by hardware is always the coverage of the
6784 * entire pixel/fragment, so mask bits out based on the sample ID.
6785 */
6786 if (key->ps_prolog.states.samplemask_log_ps_iter) {
6787 /* The bit pattern matches that used by fixed function fragment
6788 * processing. */
6789 static const uint16_t ps_iter_masks[] = {
6790 0xffff, /* not used */
6791 0x5555,
6792 0x1111,
6793 0x0101,
6794 0x0001,
6795 };
6796 assert(key->ps_prolog.states.samplemask_log_ps_iter < ARRAY_SIZE(ps_iter_masks));
6797
6798 uint32_t ps_iter_mask = ps_iter_masks[key->ps_prolog.states.samplemask_log_ps_iter];
6799 LLVMValueRef sampleid = si_unpack_param(ctx, ancillary, 8, 4);
6800 LLVMValueRef samplemask = ac_get_arg(&ctx->ac, param_sample_mask);
6801
6802 samplemask = ac_to_integer(&ctx->ac, samplemask);
6803 samplemask = LLVMBuildAnd(
6804 ctx->ac.builder,
6805 samplemask,
6806 LLVMBuildShl(ctx->ac.builder,
6807 LLVMConstInt(ctx->i32, ps_iter_mask, false),
6808 sampleid, ""),
6809 "");
6810 samplemask = ac_to_float(&ctx->ac, samplemask);
6811
6812 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, samplemask,
6813 param_sample_mask.arg_index, "");
6814 }
6815
6816 /* Tell LLVM to insert WQM instruction sequence when needed. */
6817 if (key->ps_prolog.wqm) {
6818 LLVMAddTargetDependentFunctionAttr(func,
6819 "amdgpu-ps-wqm-outputs", "");
6820 }
6821
6822 si_llvm_build_ret(ctx, ret);
6823 }
6824
6825 /**
6826 * Build the pixel shader epilog function. This handles everything that must be
6827 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
6828 */
6829 static void si_build_ps_epilog_function(struct si_shader_context *ctx,
6830 union si_shader_part_key *key)
6831 {
6832 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
6833 int i;
6834 struct si_ps_exports exp = {};
6835
6836 memset(&ctx->args, 0, sizeof(ctx->args));
6837
6838 /* Declare input SGPRs. */
6839 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->rw_buffers);
6840 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6841 &ctx->bindless_samplers_and_images);
6842 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6843 &ctx->const_and_shader_buffers);
6844 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
6845 &ctx->samplers_and_images);
6846 add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT,
6847 NULL, SI_PARAM_ALPHA_REF);
6848
6849 /* Declare input VGPRs. */
6850 unsigned required_num_params =
6851 ctx->args.num_sgprs_used +
6852 util_bitcount(key->ps_epilog.colors_written) * 4 +
6853 key->ps_epilog.writes_z +
6854 key->ps_epilog.writes_stencil +
6855 key->ps_epilog.writes_samplemask;
6856
6857 required_num_params = MAX2(required_num_params,
6858 ctx->args.num_sgprs_used + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
6859
6860 while (ctx->args.arg_count < required_num_params)
6861 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
6862
6863 /* Create the function. */
6864 si_create_function(ctx, "ps_epilog", NULL, 0, 0);
6865 /* Disable elimination of unused inputs. */
6866 ac_llvm_add_target_dep_function_attr(ctx->main_fn,
6867 "InitialPSInputAddr", 0xffffff);
6868
6869 /* Process colors. */
6870 unsigned vgpr = ctx->args.num_sgprs_used;
6871 unsigned colors_written = key->ps_epilog.colors_written;
6872 int last_color_export = -1;
6873
6874 /* Find the last color export. */
6875 if (!key->ps_epilog.writes_z &&
6876 !key->ps_epilog.writes_stencil &&
6877 !key->ps_epilog.writes_samplemask) {
6878 unsigned spi_format = key->ps_epilog.states.spi_shader_col_format;
6879
6880 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
6881 if (colors_written == 0x1 && key->ps_epilog.states.last_cbuf > 0) {
6882 /* Just set this if any of the colorbuffers are enabled. */
6883 if (spi_format &
6884 ((1ull << (4 * (key->ps_epilog.states.last_cbuf + 1))) - 1))
6885 last_color_export = 0;
6886 } else {
6887 for (i = 0; i < 8; i++)
6888 if (colors_written & (1 << i) &&
6889 (spi_format >> (i * 4)) & 0xf)
6890 last_color_export = i;
6891 }
6892 }
6893
6894 while (colors_written) {
6895 LLVMValueRef color[4];
6896 int mrt = u_bit_scan(&colors_written);
6897
6898 for (i = 0; i < 4; i++)
6899 color[i] = LLVMGetParam(ctx->main_fn, vgpr++);
6900
6901 si_export_mrt_color(ctx, color, mrt,
6902 ctx->args.arg_count - 1,
6903 mrt == last_color_export, &exp);
6904 }
6905
6906 /* Process depth, stencil, samplemask. */
6907 if (key->ps_epilog.writes_z)
6908 depth = LLVMGetParam(ctx->main_fn, vgpr++);
6909 if (key->ps_epilog.writes_stencil)
6910 stencil = LLVMGetParam(ctx->main_fn, vgpr++);
6911 if (key->ps_epilog.writes_samplemask)
6912 samplemask = LLVMGetParam(ctx->main_fn, vgpr++);
6913
6914 if (depth || stencil || samplemask)
6915 si_export_mrt_z(ctx, depth, stencil, samplemask, &exp);
6916 else if (last_color_export == -1)
6917 ac_build_export_null(&ctx->ac);
6918
6919 if (exp.num)
6920 si_emit_ps_exports(ctx, &exp);
6921
6922 /* Compile. */
6923 LLVMBuildRetVoid(ctx->ac.builder);
6924 }
6925
6926 /**
6927 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
6928 */
6929 static bool si_shader_select_ps_parts(struct si_screen *sscreen,
6930 struct ac_llvm_compiler *compiler,
6931 struct si_shader *shader,
6932 struct pipe_debug_callback *debug)
6933 {
6934 union si_shader_part_key prolog_key;
6935 union si_shader_part_key epilog_key;
6936
6937 /* Get the prolog. */
6938 si_get_ps_prolog_key(shader, &prolog_key, true);
6939
6940 /* The prolog is a no-op if these aren't set. */
6941 if (si_need_ps_prolog(&prolog_key)) {
6942 shader->prolog =
6943 si_get_shader_part(sscreen, &sscreen->ps_prologs,
6944 PIPE_SHADER_FRAGMENT, true,
6945 &prolog_key, compiler, debug,
6946 si_build_ps_prolog_function,
6947 "Fragment Shader Prolog");
6948 if (!shader->prolog)
6949 return false;
6950 }
6951
6952 /* Get the epilog. */
6953 si_get_ps_epilog_key(shader, &epilog_key);
6954
6955 shader->epilog =
6956 si_get_shader_part(sscreen, &sscreen->ps_epilogs,
6957 PIPE_SHADER_FRAGMENT, false,
6958 &epilog_key, compiler, debug,
6959 si_build_ps_epilog_function,
6960 "Fragment Shader Epilog");
6961 if (!shader->epilog)
6962 return false;
6963
6964 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
6965 if (shader->key.part.ps.prolog.poly_stipple) {
6966 shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
6967 assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
6968 }
6969
6970 /* Set up the enable bits for per-sample shading if needed. */
6971 if (shader->key.part.ps.prolog.force_persp_sample_interp &&
6972 (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
6973 G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
6974 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
6975 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
6976 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
6977 }
6978 if (shader->key.part.ps.prolog.force_linear_sample_interp &&
6979 (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
6980 G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
6981 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
6982 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
6983 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
6984 }
6985 if (shader->key.part.ps.prolog.force_persp_center_interp &&
6986 (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
6987 G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
6988 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
6989 shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
6990 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
6991 }
6992 if (shader->key.part.ps.prolog.force_linear_center_interp &&
6993 (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
6994 G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
6995 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
6996 shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
6997 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
6998 }
6999
7000 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7001 if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_ena) &&
7002 !(shader->config.spi_ps_input_ena & 0xf)) {
7003 shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
7004 assert(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr));
7005 }
7006
7007 /* At least one pair of interpolation weights must be enabled. */
7008 if (!(shader->config.spi_ps_input_ena & 0x7f)) {
7009 shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
7010 assert(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr));
7011 }
7012
7013 /* Samplemask fixup requires the sample ID. */
7014 if (shader->key.part.ps.prolog.samplemask_log_ps_iter) {
7015 shader->config.spi_ps_input_ena |= S_0286CC_ANCILLARY_ENA(1);
7016 assert(G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr));
7017 }
7018
7019 /* The sample mask input is always enabled, because the API shader always
7020 * passes it through to the epilog. Disable it here if it's unused.
7021 */
7022 if (!shader->key.part.ps.epilog.poly_line_smoothing &&
7023 !shader->selector->info.reads_samplemask)
7024 shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;
7025
7026 return true;
7027 }
7028
7029 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
7030 unsigned *lds_size)
7031 {
7032 /* If tessellation is all offchip and on-chip GS isn't used, this
7033 * workaround is not needed.
7034 */
7035 return;
7036
7037 /* SPI barrier management bug:
7038 * Make sure we have at least 4k of LDS in use to avoid the bug.
7039 * It applies to workgroup sizes of more than one wavefront.
7040 */
7041 if (sscreen->info.family == CHIP_BONAIRE ||
7042 sscreen->info.family == CHIP_KABINI)
7043 *lds_size = MAX2(*lds_size, 8);
7044 }
7045
7046 static void si_fix_resource_usage(struct si_screen *sscreen,
7047 struct si_shader *shader)
7048 {
7049 unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */
7050
7051 shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
7052
7053 if (shader->selector->type == PIPE_SHADER_COMPUTE &&
7054 si_get_max_workgroup_size(shader) > sscreen->compute_wave_size) {
7055 si_multiwave_lds_size_workaround(sscreen,
7056 &shader->config.lds_size);
7057 }
7058 }
7059
7060 bool si_create_shader_variant(struct si_screen *sscreen,
7061 struct ac_llvm_compiler *compiler,
7062 struct si_shader *shader,
7063 struct pipe_debug_callback *debug)
7064 {
7065 struct si_shader_selector *sel = shader->selector;
7066 struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key);
7067 int r;
7068
7069 /* LS, ES, VS are compiled on demand if the main part hasn't been
7070 * compiled for that stage.
7071 *
7072 * GS are compiled on demand if the main part hasn't been compiled
7073 * for the chosen NGG-ness.
7074 *
7075 * Vertex shaders are compiled on demand when a vertex fetch
7076 * workaround must be applied.
7077 */
7078 if (shader->is_monolithic) {
7079 /* Monolithic shader (compiled as a whole, has many variants,
7080 * may take a long time to compile).
7081 */
7082 r = si_compile_shader(sscreen, compiler, shader, debug);
7083 if (r)
7084 return false;
7085 } else {
7086 /* The shader consists of several parts:
7087 *
7088 * - the middle part is the user shader, it has 1 variant only
7089 * and it was compiled during the creation of the shader
7090 * selector
7091 * - the prolog part is inserted at the beginning
7092 * - the epilog part is inserted at the end
7093 *
7094 * The prolog and epilog have many (but simple) variants.
7095 *
7096 * Starting with gfx9, geometry and tessellation control
7097 * shaders also contain the prolog and user shader parts of
7098 * the previous shader stage.
7099 */
7100
7101 if (!mainp)
7102 return false;
7103
7104 /* Copy the compiled shader data over. */
7105 shader->is_binary_shared = true;
7106 shader->binary = mainp->binary;
7107 shader->config = mainp->config;
7108 shader->info.num_input_sgprs = mainp->info.num_input_sgprs;
7109 shader->info.num_input_vgprs = mainp->info.num_input_vgprs;
7110 shader->info.face_vgpr_index = mainp->info.face_vgpr_index;
7111 shader->info.ancillary_vgpr_index = mainp->info.ancillary_vgpr_index;
7112 memcpy(shader->info.vs_output_param_offset,
7113 mainp->info.vs_output_param_offset,
7114 sizeof(mainp->info.vs_output_param_offset));
7115 shader->info.uses_instanceid = mainp->info.uses_instanceid;
7116 shader->info.nr_pos_exports = mainp->info.nr_pos_exports;
7117 shader->info.nr_param_exports = mainp->info.nr_param_exports;
7118
7119 /* Select prologs and/or epilogs. */
7120 switch (sel->type) {
7121 case PIPE_SHADER_VERTEX:
7122 if (!si_shader_select_vs_parts(sscreen, compiler, shader, debug))
7123 return false;
7124 break;
7125 case PIPE_SHADER_TESS_CTRL:
7126 if (!si_shader_select_tcs_parts(sscreen, compiler, shader, debug))
7127 return false;
7128 break;
7129 case PIPE_SHADER_TESS_EVAL:
7130 break;
7131 case PIPE_SHADER_GEOMETRY:
7132 if (!si_shader_select_gs_parts(sscreen, compiler, shader, debug))
7133 return false;
7134 break;
7135 case PIPE_SHADER_FRAGMENT:
7136 if (!si_shader_select_ps_parts(sscreen, compiler, shader, debug))
7137 return false;
7138
7139 /* Make sure we have at least as many VGPRs as there
7140 * are allocated inputs.
7141 */
7142 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
7143 shader->info.num_input_vgprs);
7144 break;
7145 default:;
7146 }
7147
7148 /* Update SGPR and VGPR counts. */
7149 if (shader->prolog) {
7150 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
7151 shader->prolog->config.num_sgprs);
7152 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
7153 shader->prolog->config.num_vgprs);
7154 }
7155 if (shader->previous_stage) {
7156 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
7157 shader->previous_stage->config.num_sgprs);
7158 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
7159 shader->previous_stage->config.num_vgprs);
7160 shader->config.spilled_sgprs =
7161 MAX2(shader->config.spilled_sgprs,
7162 shader->previous_stage->config.spilled_sgprs);
7163 shader->config.spilled_vgprs =
7164 MAX2(shader->config.spilled_vgprs,
7165 shader->previous_stage->config.spilled_vgprs);
7166 shader->info.private_mem_vgprs =
7167 MAX2(shader->info.private_mem_vgprs,
7168 shader->previous_stage->info.private_mem_vgprs);
7169 shader->config.scratch_bytes_per_wave =
7170 MAX2(shader->config.scratch_bytes_per_wave,
7171 shader->previous_stage->config.scratch_bytes_per_wave);
7172 shader->info.uses_instanceid |=
7173 shader->previous_stage->info.uses_instanceid;
7174 }
7175 if (shader->prolog2) {
7176 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
7177 shader->prolog2->config.num_sgprs);
7178 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
7179 shader->prolog2->config.num_vgprs);
7180 }
7181 if (shader->epilog) {
7182 shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
7183 shader->epilog->config.num_sgprs);
7184 shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
7185 shader->epilog->config.num_vgprs);
7186 }
7187 si_calculate_max_simd_waves(shader);
7188 }
7189
7190 if (shader->key.as_ngg) {
7191 assert(!shader->key.as_es && !shader->key.as_ls);
7192 gfx10_ngg_calculate_subgroup_info(shader);
7193 } else if (sscreen->info.chip_class >= GFX9 && sel->type == PIPE_SHADER_GEOMETRY) {
7194 gfx9_get_gs_info(shader->previous_stage_sel, sel, &shader->gs_info);
7195 }
7196
7197 si_fix_resource_usage(sscreen, shader);
7198 si_shader_dump(sscreen, shader, debug, stderr, true);
7199
7200 /* Upload. */
7201 if (!si_shader_binary_upload(sscreen, shader, 0)) {
7202 fprintf(stderr, "LLVM failed to upload shader\n");
7203 return false;
7204 }
7205
7206 return true;
7207 }
7208
7209 void si_shader_destroy(struct si_shader *shader)
7210 {
7211 if (shader->scratch_bo)
7212 si_resource_reference(&shader->scratch_bo, NULL);
7213
7214 si_resource_reference(&shader->bo, NULL);
7215
7216 if (!shader->is_binary_shared)
7217 si_shader_binary_clean(&shader->binary);
7218
7219 free(shader->shader_log);
7220 }