projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: Add logic for multisample format descriptions.
[mesa.git] / src / amd / vulkan / radv_pipeline.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 #include "util/mesa-sha1.h"
29 #include "util/u_atomic.h"
30 #include "radv_debug.h"
31 #include "radv_private.h"
32 #include "radv_cs.h"
33 #include "radv_shader.h"
34 #include "nir/nir.h"
35 #include "nir/nir_builder.h"
36 #include "spirv/nir_spirv.h"
37 #include "vk_util.h"
38
39 #include <llvm-c/Core.h>
40 #include <llvm-c/TargetMachine.h>
41
42 #include "sid.h"
43 #include "gfx9d.h"
44 #include "ac_binary.h"
45 #include "ac_llvm_util.h"
46 #include "ac_nir_to_llvm.h"
47 #include "vk_format.h"
48 #include "util/debug.h"
49 #include "ac_exp_param.h"
50 #include "ac_shader_util.h"
51 #include "main/menums.h"
52
53 struct radv_blend_state {
54 uint32_t blend_enable_4bit;
55 uint32_t need_src_alpha;
56
57 uint32_t cb_color_control;
58 uint32_t cb_target_mask;
59 uint32_t cb_target_enabled_4bit;
60 uint32_t sx_mrt_blend_opt[8];
61 uint32_t cb_blend_control[8];
62
63 uint32_t spi_shader_col_format;
64 uint32_t cb_shader_mask;
65 uint32_t db_alpha_to_mask;
66
67 uint32_t commutative_4bit;
68
69 bool single_cb_enable;
70 bool mrt0_is_dual_src;
71 };
72
73 struct radv_dsa_order_invariance {
74 /* Whether the final result in Z/S buffers is guaranteed to be
75 * invariant under changes to the order in which fragments arrive.
76 */
77 bool zs;
78
79 /* Whether the set of fragments that pass the combined Z/S test is
80 * guaranteed to be invariant under changes to the order in which
81 * fragments arrive.
82 */
83 bool pass_set;
84 };
85
86 struct radv_tessellation_state {
87 uint32_t ls_hs_config;
88 unsigned num_patches;
89 unsigned lds_size;
90 uint32_t tf_param;
91 };
92
93 struct radv_gs_state {
94 uint32_t vgt_gs_onchip_cntl;
95 uint32_t vgt_gs_max_prims_per_subgroup;
96 uint32_t vgt_esgs_ring_itemsize;
97 uint32_t lds_size;
98 };
99
100 static void
101 radv_pipeline_destroy(struct radv_device *device,
102 struct radv_pipeline *pipeline,
103 const VkAllocationCallbacks* allocator)
104 {
105 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i)
106 if (pipeline->shaders[i])
107 radv_shader_variant_destroy(device, pipeline->shaders[i]);
108
109 if (pipeline->gs_copy_shader)
110 radv_shader_variant_destroy(device, pipeline->gs_copy_shader);
111
112 if(pipeline->cs.buf)
113 free(pipeline->cs.buf);
114 vk_free2(&device->alloc, allocator, pipeline);
115 }
116
117 void radv_DestroyPipeline(
118 VkDevice _device,
119 VkPipeline _pipeline,
120 const VkAllocationCallbacks* pAllocator)
121 {
122 RADV_FROM_HANDLE(radv_device, device, _device);
123 RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
124
125 if (!_pipeline)
126 return;
127
128 radv_pipeline_destroy(device, pipeline, pAllocator);
129 }
130
131 static uint32_t get_hash_flags(struct radv_device *device)
132 {
133 uint32_t hash_flags = 0;
134
135 if (device->instance->debug_flags & RADV_DEBUG_UNSAFE_MATH)
136 hash_flags |= RADV_HASH_SHADER_UNSAFE_MATH;
137 if (device->instance->perftest_flags & RADV_PERFTEST_SISCHED)
138 hash_flags |= RADV_HASH_SHADER_SISCHED;
139 return hash_flags;
140 }
141
142 static VkResult
143 radv_pipeline_scratch_init(struct radv_device *device,
144 struct radv_pipeline *pipeline)
145 {
146 unsigned scratch_bytes_per_wave = 0;
147 unsigned max_waves = 0;
148 unsigned min_waves = 1;
149
150 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
151 if (pipeline->shaders[i]) {
152 unsigned max_stage_waves = device->scratch_waves;
153
154 scratch_bytes_per_wave = MAX2(scratch_bytes_per_wave,
155 pipeline->shaders[i]->config.scratch_bytes_per_wave);
156
157 max_stage_waves = MIN2(max_stage_waves,
158 4 * device->physical_device->rad_info.num_good_compute_units *
159 (256 / pipeline->shaders[i]->config.num_vgprs));
160 max_waves = MAX2(max_waves, max_stage_waves);
161 }
162 }
163
164 if (pipeline->shaders[MESA_SHADER_COMPUTE]) {
165 unsigned group_size = pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[0] *
166 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[1] *
167 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[2];
168 min_waves = MAX2(min_waves, round_up_u32(group_size, 64));
169 }
170
171 if (scratch_bytes_per_wave)
172 max_waves = MIN2(max_waves, 0xffffffffu / scratch_bytes_per_wave);
173
174 if (scratch_bytes_per_wave && max_waves < min_waves) {
175 /* Not really true at this moment, but will be true on first
176 * execution. Avoid having hanging shaders. */
177 return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
178 }
179 pipeline->scratch_bytes_per_wave = scratch_bytes_per_wave;
180 pipeline->max_waves = max_waves;
181 return VK_SUCCESS;
182 }
183
184 static uint32_t si_translate_blend_logic_op(VkLogicOp op)
185 {
186 switch (op) {
187 case VK_LOGIC_OP_CLEAR:
188 return V_028808_ROP3_CLEAR;
189 case VK_LOGIC_OP_AND:
190 return V_028808_ROP3_AND;
191 case VK_LOGIC_OP_AND_REVERSE:
192 return V_028808_ROP3_AND_REVERSE;
193 case VK_LOGIC_OP_COPY:
194 return V_028808_ROP3_COPY;
195 case VK_LOGIC_OP_AND_INVERTED:
196 return V_028808_ROP3_AND_INVERTED;
197 case VK_LOGIC_OP_NO_OP:
198 return V_028808_ROP3_NO_OP;
199 case VK_LOGIC_OP_XOR:
200 return V_028808_ROP3_XOR;
201 case VK_LOGIC_OP_OR:
202 return V_028808_ROP3_OR;
203 case VK_LOGIC_OP_NOR:
204 return V_028808_ROP3_NOR;
205 case VK_LOGIC_OP_EQUIVALENT:
206 return V_028808_ROP3_EQUIVALENT;
207 case VK_LOGIC_OP_INVERT:
208 return V_028808_ROP3_INVERT;
209 case VK_LOGIC_OP_OR_REVERSE:
210 return V_028808_ROP3_OR_REVERSE;
211 case VK_LOGIC_OP_COPY_INVERTED:
212 return V_028808_ROP3_COPY_INVERTED;
213 case VK_LOGIC_OP_OR_INVERTED:
214 return V_028808_ROP3_OR_INVERTED;
215 case VK_LOGIC_OP_NAND:
216 return V_028808_ROP3_NAND;
217 case VK_LOGIC_OP_SET:
218 return V_028808_ROP3_SET;
219 default:
220 unreachable("Unhandled logic op");
221 }
222 }
223
224
225 static uint32_t si_translate_blend_function(VkBlendOp op)
226 {
227 switch (op) {
228 case VK_BLEND_OP_ADD:
229 return V_028780_COMB_DST_PLUS_SRC;
230 case VK_BLEND_OP_SUBTRACT:
231 return V_028780_COMB_SRC_MINUS_DST;
232 case VK_BLEND_OP_REVERSE_SUBTRACT:
233 return V_028780_COMB_DST_MINUS_SRC;
234 case VK_BLEND_OP_MIN:
235 return V_028780_COMB_MIN_DST_SRC;
236 case VK_BLEND_OP_MAX:
237 return V_028780_COMB_MAX_DST_SRC;
238 default:
239 return 0;
240 }
241 }
242
243 static uint32_t si_translate_blend_factor(VkBlendFactor factor)
244 {
245 switch (factor) {
246 case VK_BLEND_FACTOR_ZERO:
247 return V_028780_BLEND_ZERO;
248 case VK_BLEND_FACTOR_ONE:
249 return V_028780_BLEND_ONE;
250 case VK_BLEND_FACTOR_SRC_COLOR:
251 return V_028780_BLEND_SRC_COLOR;
252 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
253 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
254 case VK_BLEND_FACTOR_DST_COLOR:
255 return V_028780_BLEND_DST_COLOR;
256 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR:
257 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
258 case VK_BLEND_FACTOR_SRC_ALPHA:
259 return V_028780_BLEND_SRC_ALPHA;
260 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
261 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
262 case VK_BLEND_FACTOR_DST_ALPHA:
263 return V_028780_BLEND_DST_ALPHA;
264 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
265 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
266 case VK_BLEND_FACTOR_CONSTANT_COLOR:
267 return V_028780_BLEND_CONSTANT_COLOR;
268 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
269 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
270 case VK_BLEND_FACTOR_CONSTANT_ALPHA:
271 return V_028780_BLEND_CONSTANT_ALPHA;
272 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
273 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
274 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
275 return V_028780_BLEND_SRC_ALPHA_SATURATE;
276 case VK_BLEND_FACTOR_SRC1_COLOR:
277 return V_028780_BLEND_SRC1_COLOR;
278 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
279 return V_028780_BLEND_INV_SRC1_COLOR;
280 case VK_BLEND_FACTOR_SRC1_ALPHA:
281 return V_028780_BLEND_SRC1_ALPHA;
282 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
283 return V_028780_BLEND_INV_SRC1_ALPHA;
284 default:
285 return 0;
286 }
287 }
288
289 static uint32_t si_translate_blend_opt_function(VkBlendOp op)
290 {
291 switch (op) {
292 case VK_BLEND_OP_ADD:
293 return V_028760_OPT_COMB_ADD;
294 case VK_BLEND_OP_SUBTRACT:
295 return V_028760_OPT_COMB_SUBTRACT;
296 case VK_BLEND_OP_REVERSE_SUBTRACT:
297 return V_028760_OPT_COMB_REVSUBTRACT;
298 case VK_BLEND_OP_MIN:
299 return V_028760_OPT_COMB_MIN;
300 case VK_BLEND_OP_MAX:
301 return V_028760_OPT_COMB_MAX;
302 default:
303 return V_028760_OPT_COMB_BLEND_DISABLED;
304 }
305 }
306
307 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor, bool is_alpha)
308 {
309 switch (factor) {
310 case VK_BLEND_FACTOR_ZERO:
311 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
312 case VK_BLEND_FACTOR_ONE:
313 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
314 case VK_BLEND_FACTOR_SRC_COLOR:
315 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
316 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
317 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
318 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
319 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
320 case VK_BLEND_FACTOR_SRC_ALPHA:
321 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
322 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
323 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
324 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
325 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
326 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
327 default:
328 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
329 }
330 }
331
332 /**
333 * Get rid of DST in the blend factors by commuting the operands:
334 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
335 */
336 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
337 unsigned *dst_factor, unsigned expected_dst,
338 unsigned replacement_src)
339 {
340 if (*src_factor == expected_dst &&
341 *dst_factor == VK_BLEND_FACTOR_ZERO) {
342 *src_factor = VK_BLEND_FACTOR_ZERO;
343 *dst_factor = replacement_src;
344
345 /* Commuting the operands requires reversing subtractions. */
346 if (*func == VK_BLEND_OP_SUBTRACT)
347 *func = VK_BLEND_OP_REVERSE_SUBTRACT;
348 else if (*func == VK_BLEND_OP_REVERSE_SUBTRACT)
349 *func = VK_BLEND_OP_SUBTRACT;
350 }
351 }
352
353 static bool si_blend_factor_uses_dst(unsigned factor)
354 {
355 return factor == VK_BLEND_FACTOR_DST_COLOR ||
356 factor == VK_BLEND_FACTOR_DST_ALPHA ||
357 factor == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
358 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA ||
359 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR;
360 }
361
362 static bool is_dual_src(VkBlendFactor factor)
363 {
364 switch (factor) {
365 case VK_BLEND_FACTOR_SRC1_COLOR:
366 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
367 case VK_BLEND_FACTOR_SRC1_ALPHA:
368 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
369 return true;
370 default:
371 return false;
372 }
373 }
374
375 static unsigned si_choose_spi_color_format(VkFormat vk_format,
376 bool blend_enable,
377 bool blend_need_alpha)
378 {
379 const struct vk_format_description *desc = vk_format_description(vk_format);
380 unsigned format, ntype, swap;
381
382 /* Alpha is needed for alpha-to-coverage.
383 * Blending may be with or without alpha.
384 */
385 unsigned normal = 0; /* most optimal, may not support blending or export alpha */
386 unsigned alpha = 0; /* exports alpha, but may not support blending */
387 unsigned blend = 0; /* supports blending, but may not export alpha */
388 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
389
390 format = radv_translate_colorformat(vk_format);
391 ntype = radv_translate_color_numformat(vk_format, desc,
392 vk_format_get_first_non_void_channel(vk_format));
393 swap = radv_translate_colorswap(vk_format, false);
394
395 /* Choose the SPI color formats. These are required values for Stoney/RB+.
396 * Other chips have multiple choices, though they are not necessarily better.
397 */
398 switch (format) {
399 case V_028C70_COLOR_5_6_5:
400 case V_028C70_COLOR_1_5_5_5:
401 case V_028C70_COLOR_5_5_5_1:
402 case V_028C70_COLOR_4_4_4_4:
403 case V_028C70_COLOR_10_11_11:
404 case V_028C70_COLOR_11_11_10:
405 case V_028C70_COLOR_8:
406 case V_028C70_COLOR_8_8:
407 case V_028C70_COLOR_8_8_8_8:
408 case V_028C70_COLOR_10_10_10_2:
409 case V_028C70_COLOR_2_10_10_10:
410 if (ntype == V_028C70_NUMBER_UINT)
411 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
412 else if (ntype == V_028C70_NUMBER_SINT)
413 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
414 else
415 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
416 break;
417
418 case V_028C70_COLOR_16:
419 case V_028C70_COLOR_16_16:
420 case V_028C70_COLOR_16_16_16_16:
421 if (ntype == V_028C70_NUMBER_UNORM ||
422 ntype == V_028C70_NUMBER_SNORM) {
423 /* UNORM16 and SNORM16 don't support blending */
424 if (ntype == V_028C70_NUMBER_UNORM)
425 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
426 else
427 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
428
429 /* Use 32 bits per channel for blending. */
430 if (format == V_028C70_COLOR_16) {
431 if (swap == V_028C70_SWAP_STD) { /* R */
432 blend = V_028714_SPI_SHADER_32_R;
433 blend_alpha = V_028714_SPI_SHADER_32_AR;
434 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
435 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
436 else
437 assert(0);
438 } else if (format == V_028C70_COLOR_16_16) {
439 if (swap == V_028C70_SWAP_STD) { /* RG */
440 blend = V_028714_SPI_SHADER_32_GR;
441 blend_alpha = V_028714_SPI_SHADER_32_ABGR;
442 } else if (swap == V_028C70_SWAP_ALT) /* RA */
443 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
444 else
445 assert(0);
446 } else /* 16_16_16_16 */
447 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
448 } else if (ntype == V_028C70_NUMBER_UINT)
449 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
450 else if (ntype == V_028C70_NUMBER_SINT)
451 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
452 else if (ntype == V_028C70_NUMBER_FLOAT)
453 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
454 else
455 assert(0);
456 break;
457
458 case V_028C70_COLOR_32:
459 if (swap == V_028C70_SWAP_STD) { /* R */
460 blend = normal = V_028714_SPI_SHADER_32_R;
461 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
462 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
463 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
464 else
465 assert(0);
466 break;
467
468 case V_028C70_COLOR_32_32:
469 if (swap == V_028C70_SWAP_STD) { /* RG */
470 blend = normal = V_028714_SPI_SHADER_32_GR;
471 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
472 } else if (swap == V_028C70_SWAP_ALT) /* RA */
473 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
474 else
475 assert(0);
476 break;
477
478 case V_028C70_COLOR_32_32_32_32:
479 case V_028C70_COLOR_8_24:
480 case V_028C70_COLOR_24_8:
481 case V_028C70_COLOR_X24_8_32_FLOAT:
482 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
483 break;
484
485 default:
486 unreachable("unhandled blend format");
487 }
488
489 if (blend_enable && blend_need_alpha)
490 return blend_alpha;
491 else if(blend_need_alpha)
492 return alpha;
493 else if(blend_enable)
494 return blend;
495 else
496 return normal;
497 }
498
499 static void
500 radv_pipeline_compute_spi_color_formats(struct radv_pipeline *pipeline,
501 const VkGraphicsPipelineCreateInfo *pCreateInfo,
502 struct radv_blend_state *blend)
503 {
504 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
505 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
506 unsigned col_format = 0;
507 unsigned num_targets;
508
509 for (unsigned i = 0; i < (blend->single_cb_enable ? 1 : subpass->color_count); ++i) {
510 unsigned cf;
511
512 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) {
513 cf = V_028714_SPI_SHADER_ZERO;
514 } else {
515 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->color_attachments[i].attachment;
516 bool blend_enable =
517 blend->blend_enable_4bit & (0xfu << (i * 4));
518
519 cf = si_choose_spi_color_format(attachment->format,
520 blend_enable,
521 blend->need_src_alpha & (1 << i));
522 }
523
524 col_format |= cf << (4 * i);
525 }
526
527 if (!col_format && blend->need_src_alpha & (1 << 0)) {
528 /* When a subpass doesn't have any color attachments, write the
529 * alpha channel of MRT0 when alpha coverage is enabled because
530 * the depth attachment needs it.
531 */
532 col_format |= V_028714_SPI_SHADER_32_AR;
533 }
534
535 /* If the i-th target format is set, all previous target formats must
536 * be non-zero to avoid hangs.
537 */
538 num_targets = (util_last_bit(col_format) + 3) / 4;
539 for (unsigned i = 0; i < num_targets; i++) {
540 if (!(col_format & (0xf << (i * 4)))) {
541 col_format |= V_028714_SPI_SHADER_32_R << (i * 4);
542 }
543 }
544
545 blend->cb_shader_mask = ac_get_cb_shader_mask(col_format);
546
547 if (blend->mrt0_is_dual_src)
548 col_format |= (col_format & 0xf) << 4;
549 blend->spi_shader_col_format = col_format;
550 }
551
552 static bool
553 format_is_int8(VkFormat format)
554 {
555 const struct vk_format_description *desc = vk_format_description(format);
556 int channel = vk_format_get_first_non_void_channel(format);
557
558 return channel >= 0 && desc->channel[channel].pure_integer &&
559 desc->channel[channel].size == 8;
560 }
561
562 static bool
563 format_is_int10(VkFormat format)
564 {
565 const struct vk_format_description *desc = vk_format_description(format);
566
567 if (desc->nr_channels != 4)
568 return false;
569 for (unsigned i = 0; i < 4; i++) {
570 if (desc->channel[i].pure_integer && desc->channel[i].size == 10)
571 return true;
572 }
573 return false;
574 }
575
576 /*
577 * Ordered so that for each i,
578 * radv_format_meta_fs_key(radv_fs_key_format_exemplars[i]) == i.
579 */
580 const VkFormat radv_fs_key_format_exemplars[NUM_META_FS_KEYS] = {
581 VK_FORMAT_R32_SFLOAT,
582 VK_FORMAT_R32G32_SFLOAT,
583 VK_FORMAT_R8G8B8A8_UNORM,
584 VK_FORMAT_R16G16B16A16_UNORM,
585 VK_FORMAT_R16G16B16A16_SNORM,
586 VK_FORMAT_R16G16B16A16_UINT,
587 VK_FORMAT_R16G16B16A16_SINT,
588 VK_FORMAT_R32G32B32A32_SFLOAT,
589 VK_FORMAT_R8G8B8A8_UINT,
590 VK_FORMAT_R8G8B8A8_SINT,
591 VK_FORMAT_A2R10G10B10_UINT_PACK32,
592 VK_FORMAT_A2R10G10B10_SINT_PACK32,
593 };
594
595 unsigned radv_format_meta_fs_key(VkFormat format)
596 {
597 unsigned col_format = si_choose_spi_color_format(format, false, false);
598
599 assert(col_format != V_028714_SPI_SHADER_32_AR);
600 if (col_format >= V_028714_SPI_SHADER_32_AR)
601 --col_format; /* Skip V_028714_SPI_SHADER_32_AR since there is no such VkFormat */
602
603 --col_format; /* Skip V_028714_SPI_SHADER_ZERO */
604 bool is_int8 = format_is_int8(format);
605 bool is_int10 = format_is_int10(format);
606
607 return col_format + (is_int8 ? 3 : is_int10 ? 5 : 0);
608 }
609
610 static void
611 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo *pCreateInfo,
612 unsigned *is_int8, unsigned *is_int10)
613 {
614 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
615 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
616 *is_int8 = 0;
617 *is_int10 = 0;
618
619 for (unsigned i = 0; i < subpass->color_count; ++i) {
620 struct radv_render_pass_attachment *attachment;
621
622 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
623 continue;
624
625 attachment = pass->attachments + subpass->color_attachments[i].attachment;
626
627 if (format_is_int8(attachment->format))
628 *is_int8 |= 1 << i;
629 if (format_is_int10(attachment->format))
630 *is_int10 |= 1 << i;
631 }
632 }
633
634 static void
635 radv_blend_check_commutativity(struct radv_blend_state *blend,
636 VkBlendOp op, VkBlendFactor src,
637 VkBlendFactor dst, unsigned chanmask)
638 {
639 /* Src factor is allowed when it does not depend on Dst. */
640 static const uint32_t src_allowed =
641 (1u << VK_BLEND_FACTOR_ONE) |
642 (1u << VK_BLEND_FACTOR_SRC_COLOR) |
643 (1u << VK_BLEND_FACTOR_SRC_ALPHA) |
644 (1u << VK_BLEND_FACTOR_SRC_ALPHA_SATURATE) |
645 (1u << VK_BLEND_FACTOR_CONSTANT_COLOR) |
646 (1u << VK_BLEND_FACTOR_CONSTANT_ALPHA) |
647 (1u << VK_BLEND_FACTOR_SRC1_COLOR) |
648 (1u << VK_BLEND_FACTOR_SRC1_ALPHA) |
649 (1u << VK_BLEND_FACTOR_ZERO) |
650 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR) |
651 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA) |
652 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR) |
653 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA) |
654 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR) |
655 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA);
656
657 if (dst == VK_BLEND_FACTOR_ONE &&
658 (src_allowed & (1u << src))) {
659 /* Addition is commutative, but floating point addition isn't
660 * associative: subtle changes can be introduced via different
661 * rounding. Be conservative, only enable for min and max.
662 */
663 if (op == VK_BLEND_OP_MAX || op == VK_BLEND_OP_MIN)
664 blend->commutative_4bit |= chanmask;
665 }
666 }
667
668 static struct radv_blend_state
669 radv_pipeline_init_blend_state(struct radv_pipeline *pipeline,
670 const VkGraphicsPipelineCreateInfo *pCreateInfo,
671 const struct radv_graphics_pipeline_create_info *extra)
672 {
673 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
674 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState;
675 struct radv_blend_state blend = {0};
676 unsigned mode = V_028808_CB_NORMAL;
677 int i;
678
679 if (!vkblend)
680 return blend;
681
682 if (extra && extra->custom_blend_mode) {
683 blend.single_cb_enable = true;
684 mode = extra->custom_blend_mode;
685 }
686 blend.cb_color_control = 0;
687 if (vkblend->logicOpEnable)
688 blend.cb_color_control |= S_028808_ROP3(si_translate_blend_logic_op(vkblend->logicOp));
689 else
690 blend.cb_color_control |= S_028808_ROP3(V_028808_ROP3_COPY);
691
692 blend.db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(3) |
693 S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
694 S_028B70_ALPHA_TO_MASK_OFFSET2(0) |
695 S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
696 S_028B70_OFFSET_ROUND(1);
697
698 if (vkms && vkms->alphaToCoverageEnable) {
699 blend.db_alpha_to_mask |= S_028B70_ALPHA_TO_MASK_ENABLE(1);
700 blend.need_src_alpha |= 0x1;
701 }
702
703 blend.cb_target_mask = 0;
704 for (i = 0; i < vkblend->attachmentCount; i++) {
705 const VkPipelineColorBlendAttachmentState *att = &vkblend->pAttachments[i];
706 unsigned blend_cntl = 0;
707 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
708 VkBlendOp eqRGB = att->colorBlendOp;
709 VkBlendFactor srcRGB = att->srcColorBlendFactor;
710 VkBlendFactor dstRGB = att->dstColorBlendFactor;
711 VkBlendOp eqA = att->alphaBlendOp;
712 VkBlendFactor srcA = att->srcAlphaBlendFactor;
713 VkBlendFactor dstA = att->dstAlphaBlendFactor;
714
715 blend.sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
716
717 if (!att->colorWriteMask)
718 continue;
719
720 blend.cb_target_mask |= (unsigned)att->colorWriteMask << (4 * i);
721 blend.cb_target_enabled_4bit |= 0xf << (4 * i);
722 if (!att->blendEnable) {
723 blend.cb_blend_control[i] = blend_cntl;
724 continue;
725 }
726
727 if (is_dual_src(srcRGB) || is_dual_src(dstRGB) || is_dual_src(srcA) || is_dual_src(dstA))
728 if (i == 0)
729 blend.mrt0_is_dual_src = true;
730
731 if (eqRGB == VK_BLEND_OP_MIN || eqRGB == VK_BLEND_OP_MAX) {
732 srcRGB = VK_BLEND_FACTOR_ONE;
733 dstRGB = VK_BLEND_FACTOR_ONE;
734 }
735 if (eqA == VK_BLEND_OP_MIN || eqA == VK_BLEND_OP_MAX) {
736 srcA = VK_BLEND_FACTOR_ONE;
737 dstA = VK_BLEND_FACTOR_ONE;
738 }
739
740 radv_blend_check_commutativity(&blend, eqRGB, srcRGB, dstRGB,
741 0x7 << (4 * i));
742 radv_blend_check_commutativity(&blend, eqA, srcA, dstA,
743 0x8 << (4 * i));
744
745 /* Blending optimizations for RB+.
746 * These transformations don't change the behavior.
747 *
748 * First, get rid of DST in the blend factors:
749 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
750 */
751 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
752 VK_BLEND_FACTOR_DST_COLOR,
753 VK_BLEND_FACTOR_SRC_COLOR);
754
755 si_blend_remove_dst(&eqA, &srcA, &dstA,
756 VK_BLEND_FACTOR_DST_COLOR,
757 VK_BLEND_FACTOR_SRC_COLOR);
758
759 si_blend_remove_dst(&eqA, &srcA, &dstA,
760 VK_BLEND_FACTOR_DST_ALPHA,
761 VK_BLEND_FACTOR_SRC_ALPHA);
762
763 /* Look up the ideal settings from tables. */
764 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
765 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
766 srcA_opt = si_translate_blend_opt_factor(srcA, true);
767 dstA_opt = si_translate_blend_opt_factor(dstA, true);
768
769 /* Handle interdependencies. */
770 if (si_blend_factor_uses_dst(srcRGB))
771 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
772 if (si_blend_factor_uses_dst(srcA))
773 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
774
775 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE &&
776 (dstRGB == VK_BLEND_FACTOR_ZERO ||
777 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
778 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE))
779 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
780
781 /* Set the final value. */
782 blend.sx_mrt_blend_opt[i] =
783 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
784 S_028760_COLOR_DST_OPT(dstRGB_opt) |
785 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
786 S_028760_ALPHA_SRC_OPT(srcA_opt) |
787 S_028760_ALPHA_DST_OPT(dstA_opt) |
788 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
789 blend_cntl |= S_028780_ENABLE(1);
790
791 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
792 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
793 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
794 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
795 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
796 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
797 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
798 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
799 }
800 blend.cb_blend_control[i] = blend_cntl;
801
802 blend.blend_enable_4bit |= 0xfu << (i * 4);
803
804 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
805 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
806 srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
807 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
808 srcRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA ||
809 dstRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA)
810 blend.need_src_alpha |= 1 << i;
811 }
812 for (i = vkblend->attachmentCount; i < 8; i++) {
813 blend.cb_blend_control[i] = 0;
814 blend.sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
815 }
816
817 if (pipeline->device->physical_device->has_rbplus) {
818 /* Disable RB+ blend optimizations for dual source blending. */
819 if (blend.mrt0_is_dual_src) {
820 for (i = 0; i < 8; i++) {
821 blend.sx_mrt_blend_opt[i] =
822 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
823 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
824 }
825 }
826
827 /* RB+ doesn't work with dual source blending, logic op and
828 * RESOLVE.
829 */
830 if (blend.mrt0_is_dual_src || vkblend->logicOpEnable ||
831 mode == V_028808_CB_RESOLVE)
832 blend.cb_color_control |= S_028808_DISABLE_DUAL_QUAD(1);
833 }
834
835 if (blend.cb_target_mask)
836 blend.cb_color_control |= S_028808_MODE(mode);
837 else
838 blend.cb_color_control |= S_028808_MODE(V_028808_CB_DISABLE);
839
840 radv_pipeline_compute_spi_color_formats(pipeline, pCreateInfo, &blend);
841 return blend;
842 }
843
844 static uint32_t si_translate_stencil_op(enum VkStencilOp op)
845 {
846 switch (op) {
847 case VK_STENCIL_OP_KEEP:
848 return V_02842C_STENCIL_KEEP;
849 case VK_STENCIL_OP_ZERO:
850 return V_02842C_STENCIL_ZERO;
851 case VK_STENCIL_OP_REPLACE:
852 return V_02842C_STENCIL_REPLACE_TEST;
853 case VK_STENCIL_OP_INCREMENT_AND_CLAMP:
854 return V_02842C_STENCIL_ADD_CLAMP;
855 case VK_STENCIL_OP_DECREMENT_AND_CLAMP:
856 return V_02842C_STENCIL_SUB_CLAMP;
857 case VK_STENCIL_OP_INVERT:
858 return V_02842C_STENCIL_INVERT;
859 case VK_STENCIL_OP_INCREMENT_AND_WRAP:
860 return V_02842C_STENCIL_ADD_WRAP;
861 case VK_STENCIL_OP_DECREMENT_AND_WRAP:
862 return V_02842C_STENCIL_SUB_WRAP;
863 default:
864 return 0;
865 }
866 }
867
868 static uint32_t si_translate_fill(VkPolygonMode func)
869 {
870 switch(func) {
871 case VK_POLYGON_MODE_FILL:
872 return V_028814_X_DRAW_TRIANGLES;
873 case VK_POLYGON_MODE_LINE:
874 return V_028814_X_DRAW_LINES;
875 case VK_POLYGON_MODE_POINT:
876 return V_028814_X_DRAW_POINTS;
877 default:
878 assert(0);
879 return V_028814_X_DRAW_POINTS;
880 }
881 }
882
883 static uint8_t radv_pipeline_get_ps_iter_samples(const VkPipelineMultisampleStateCreateInfo *vkms)
884 {
885 uint32_t num_samples = vkms->rasterizationSamples;
886 uint32_t ps_iter_samples = 1;
887
888 if (vkms->sampleShadingEnable) {
889 ps_iter_samples = ceil(vkms->minSampleShading * num_samples);
890 ps_iter_samples = util_next_power_of_two(ps_iter_samples);
891 }
892 return ps_iter_samples;
893 }
894
895 static bool
896 radv_is_depth_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
897 {
898 return pCreateInfo->depthTestEnable &&
899 pCreateInfo->depthWriteEnable &&
900 pCreateInfo->depthCompareOp != VK_COMPARE_OP_NEVER;
901 }
902
903 static bool
904 radv_writes_stencil(const VkStencilOpState *state)
905 {
906 return state->writeMask &&
907 (state->failOp != VK_STENCIL_OP_KEEP ||
908 state->passOp != VK_STENCIL_OP_KEEP ||
909 state->depthFailOp != VK_STENCIL_OP_KEEP);
910 }
911
912 static bool
913 radv_is_stencil_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
914 {
915 return pCreateInfo->stencilTestEnable &&
916 (radv_writes_stencil(&pCreateInfo->front) ||
917 radv_writes_stencil(&pCreateInfo->back));
918 }
919
920 static bool
921 radv_is_ds_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
922 {
923 return radv_is_depth_write_enabled(pCreateInfo) ||
924 radv_is_stencil_write_enabled(pCreateInfo);
925 }
926
927 static bool
928 radv_order_invariant_stencil_op(VkStencilOp op)
929 {
930 /* REPLACE is normally order invariant, except when the stencil
931 * reference value is written by the fragment shader. Tracking this
932 * interaction does not seem worth the effort, so be conservative.
933 */
934 return op != VK_STENCIL_OP_INCREMENT_AND_CLAMP &&
935 op != VK_STENCIL_OP_DECREMENT_AND_CLAMP &&
936 op != VK_STENCIL_OP_REPLACE;
937 }
938
939 static bool
940 radv_order_invariant_stencil_state(const VkStencilOpState *state)
941 {
942 /* Compute whether, assuming Z writes are disabled, this stencil state
943 * is order invariant in the sense that the set of passing fragments as
944 * well as the final stencil buffer result does not depend on the order
945 * of fragments.
946 */
947 return !state->writeMask ||
948 /* The following assumes that Z writes are disabled. */
949 (state->compareOp == VK_COMPARE_OP_ALWAYS &&
950 radv_order_invariant_stencil_op(state->passOp) &&
951 radv_order_invariant_stencil_op(state->depthFailOp)) ||
952 (state->compareOp == VK_COMPARE_OP_NEVER &&
953 radv_order_invariant_stencil_op(state->failOp));
954 }
955
956 static bool
957 radv_pipeline_out_of_order_rast(struct radv_pipeline *pipeline,
958 struct radv_blend_state *blend,
959 const VkGraphicsPipelineCreateInfo *pCreateInfo)
960 {
961 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
962 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
963 unsigned colormask = blend->cb_target_enabled_4bit;
964
965 if (!pipeline->device->physical_device->out_of_order_rast_allowed)
966 return false;
967
968 /* Be conservative if a logic operation is enabled with color buffers. */
969 if (colormask && pCreateInfo->pColorBlendState->logicOpEnable)
970 return false;
971
972 /* Default depth/stencil invariance when no attachment is bound. */
973 struct radv_dsa_order_invariance dsa_order_invariant = {
974 .zs = true, .pass_set = true
975 };
976
977 if (pCreateInfo->pDepthStencilState &&
978 subpass->depth_stencil_attachment) {
979 const VkPipelineDepthStencilStateCreateInfo *vkds =
980 pCreateInfo->pDepthStencilState;
981 struct radv_render_pass_attachment *attachment =
982 pass->attachments + subpass->depth_stencil_attachment->attachment;
983 bool has_stencil = vk_format_is_stencil(attachment->format);
984 struct radv_dsa_order_invariance order_invariance[2];
985 struct radv_shader_variant *ps =
986 pipeline->shaders[MESA_SHADER_FRAGMENT];
987
988 /* Compute depth/stencil order invariance in order to know if
989 * it's safe to enable out-of-order.
990 */
991 bool zfunc_is_ordered =
992 vkds->depthCompareOp == VK_COMPARE_OP_NEVER ||
993 vkds->depthCompareOp == VK_COMPARE_OP_LESS ||
994 vkds->depthCompareOp == VK_COMPARE_OP_LESS_OR_EQUAL ||
995 vkds->depthCompareOp == VK_COMPARE_OP_GREATER ||
996 vkds->depthCompareOp == VK_COMPARE_OP_GREATER_OR_EQUAL;
997
998 bool nozwrite_and_order_invariant_stencil =
999 !radv_is_ds_write_enabled(vkds) ||
1000 (!radv_is_depth_write_enabled(vkds) &&
1001 radv_order_invariant_stencil_state(&vkds->front) &&
1002 radv_order_invariant_stencil_state(&vkds->back));
1003
1004 order_invariance[1].zs =
1005 nozwrite_and_order_invariant_stencil ||
1006 (!radv_is_stencil_write_enabled(vkds) &&
1007 zfunc_is_ordered);
1008 order_invariance[0].zs =
1009 !radv_is_depth_write_enabled(vkds) || zfunc_is_ordered;
1010
1011 order_invariance[1].pass_set =
1012 nozwrite_and_order_invariant_stencil ||
1013 (!radv_is_stencil_write_enabled(vkds) &&
1014 (vkds->depthCompareOp == VK_COMPARE_OP_ALWAYS ||
1015 vkds->depthCompareOp == VK_COMPARE_OP_NEVER));
1016 order_invariance[0].pass_set =
1017 !radv_is_depth_write_enabled(vkds) ||
1018 (vkds->depthCompareOp == VK_COMPARE_OP_ALWAYS ||
1019 vkds->depthCompareOp == VK_COMPARE_OP_NEVER);
1020
1021 dsa_order_invariant = order_invariance[has_stencil];
1022 if (!dsa_order_invariant.zs)
1023 return false;
1024
1025 /* The set of PS invocations is always order invariant,
1026 * except when early Z/S tests are requested.
1027 */
1028 if (ps &&
1029 ps->info.info.ps.writes_memory &&
1030 ps->info.fs.early_fragment_test &&
1031 !dsa_order_invariant.pass_set)
1032 return false;
1033
1034 /* Determine if out-of-order rasterization should be disabled
1035 * when occlusion queries are used.
1036 */
1037 pipeline->graphics.disable_out_of_order_rast_for_occlusion =
1038 !dsa_order_invariant.pass_set;
1039 }
1040
1041 /* No color buffers are enabled for writing. */
1042 if (!colormask)
1043 return true;
1044
1045 unsigned blendmask = colormask & blend->blend_enable_4bit;
1046
1047 if (blendmask) {
1048 /* Only commutative blending. */
1049 if (blendmask & ~blend->commutative_4bit)
1050 return false;
1051
1052 if (!dsa_order_invariant.pass_set)
1053 return false;
1054 }
1055
1056 if (colormask & ~blendmask)
1057 return false;
1058
1059 return true;
1060 }
1061
1062 static void
1063 radv_pipeline_init_multisample_state(struct radv_pipeline *pipeline,
1064 struct radv_blend_state *blend,
1065 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1066 {
1067 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState;
1068 struct radv_multisample_state *ms = &pipeline->graphics.ms;
1069 unsigned num_tile_pipes = pipeline->device->physical_device->rad_info.num_tile_pipes;
1070 bool out_of_order_rast = false;
1071 int ps_iter_samples = 1;
1072 uint32_t mask = 0xffff;
1073
1074 if (vkms)
1075 ms->num_samples = vkms->rasterizationSamples;
1076 else
1077 ms->num_samples = 1;
1078
1079 if (vkms)
1080 ps_iter_samples = radv_pipeline_get_ps_iter_samples(vkms);
1081 if (vkms && !vkms->sampleShadingEnable && pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.force_persample) {
1082 ps_iter_samples = ms->num_samples;
1083 }
1084
1085 const struct VkPipelineRasterizationStateRasterizationOrderAMD *raster_order =
1086 vk_find_struct_const(pCreateInfo->pRasterizationState->pNext, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD);
1087 if (raster_order && raster_order->rasterizationOrder == VK_RASTERIZATION_ORDER_RELAXED_AMD) {
1088 /* Out-of-order rasterization is explicitly enabled by the
1089 * application.
1090 */
1091 out_of_order_rast = true;
1092 } else {
1093 /* Determine if the driver can enable out-of-order
1094 * rasterization internally.
1095 */
1096 out_of_order_rast =
1097 radv_pipeline_out_of_order_rast(pipeline, blend, pCreateInfo);
1098 }
1099
1100 ms->pa_sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
1101 ms->pa_sc_aa_config = 0;
1102 ms->db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
1103 S_028804_INCOHERENT_EQAA_READS(1) |
1104 S_028804_INTERPOLATE_COMP_Z(1) |
1105 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
1106 ms->pa_sc_mode_cntl_1 =
1107 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
1108 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
1109 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
1110 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
1111 /* always 1: */
1112 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
1113 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
1114 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
1115 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
1116 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
1117 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
1118 ms->pa_sc_mode_cntl_0 = S_028A48_ALTERNATE_RBS_PER_TILE(pipeline->device->physical_device->rad_info.chip_class >= GFX9) |
1119 S_028A48_VPORT_SCISSOR_ENABLE(1);
1120
1121 if (ms->num_samples > 1) {
1122 unsigned log_samples = util_logbase2(ms->num_samples);
1123 unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
1124 ms->pa_sc_mode_cntl_0 |= S_028A48_MSAA_ENABLE(1);
1125 ms->pa_sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
1126 ms->db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_samples) |
1127 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
1128 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
1129 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
1130 ms->pa_sc_aa_config |= S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
1131 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples)) |
1132 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples); /* CM_R_028BE0_PA_SC_AA_CONFIG */
1133 ms->pa_sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
1134 if (ps_iter_samples > 1)
1135 pipeline->graphics.spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1136 }
1137
1138 if (vkms && vkms->pSampleMask) {
1139 mask = vkms->pSampleMask[0] & 0xffff;
1140 }
1141
1142 ms->pa_sc_aa_mask[0] = mask | (mask << 16);
1143 ms->pa_sc_aa_mask[1] = mask | (mask << 16);
1144 }
1145
1146 static bool
1147 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology)
1148 {
1149 switch (topology) {
1150 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1151 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1152 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1153 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1154 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1155 return false;
1156 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1157 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1158 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1159 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1160 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1161 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1162 return true;
1163 default:
1164 unreachable("unhandled primitive type");
1165 }
1166 }
1167
1168 static uint32_t
1169 si_translate_prim(enum VkPrimitiveTopology topology)
1170 {
1171 switch (topology) {
1172 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1173 return V_008958_DI_PT_POINTLIST;
1174 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1175 return V_008958_DI_PT_LINELIST;
1176 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1177 return V_008958_DI_PT_LINESTRIP;
1178 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1179 return V_008958_DI_PT_TRILIST;
1180 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1181 return V_008958_DI_PT_TRISTRIP;
1182 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1183 return V_008958_DI_PT_TRIFAN;
1184 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1185 return V_008958_DI_PT_LINELIST_ADJ;
1186 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1187 return V_008958_DI_PT_LINESTRIP_ADJ;
1188 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1189 return V_008958_DI_PT_TRILIST_ADJ;
1190 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1191 return V_008958_DI_PT_TRISTRIP_ADJ;
1192 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1193 return V_008958_DI_PT_PATCH;
1194 default:
1195 assert(0);
1196 return 0;
1197 }
1198 }
1199
1200 static uint32_t
1201 si_conv_gl_prim_to_gs_out(unsigned gl_prim)
1202 {
1203 switch (gl_prim) {
1204 case 0: /* GL_POINTS */
1205 return V_028A6C_OUTPRIM_TYPE_POINTLIST;
1206 case 1: /* GL_LINES */
1207 case 3: /* GL_LINE_STRIP */
1208 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
1209 case 0x8E7A: /* GL_ISOLINES */
1210 return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
1211
1212 case 4: /* GL_TRIANGLES */
1213 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
1214 case 5: /* GL_TRIANGLE_STRIP */
1215 case 7: /* GL_QUADS */
1216 return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
1217 default:
1218 assert(0);
1219 return 0;
1220 }
1221 }
1222
1223 static uint32_t
1224 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology)
1225 {
1226 switch (topology) {
1227 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1228 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1229 return V_028A6C_OUTPRIM_TYPE_POINTLIST;
1230 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1231 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1232 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1233 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1234 return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
1235 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1236 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1237 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1238 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1239 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1240 return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
1241 default:
1242 assert(0);
1243 return 0;
1244 }
1245 }
1246
1247 static unsigned radv_dynamic_state_mask(VkDynamicState state)
1248 {
1249 switch(state) {
1250 case VK_DYNAMIC_STATE_VIEWPORT:
1251 return RADV_DYNAMIC_VIEWPORT;
1252 case VK_DYNAMIC_STATE_SCISSOR:
1253 return RADV_DYNAMIC_SCISSOR;
1254 case VK_DYNAMIC_STATE_LINE_WIDTH:
1255 return RADV_DYNAMIC_LINE_WIDTH;
1256 case VK_DYNAMIC_STATE_DEPTH_BIAS:
1257 return RADV_DYNAMIC_DEPTH_BIAS;
1258 case VK_DYNAMIC_STATE_BLEND_CONSTANTS:
1259 return RADV_DYNAMIC_BLEND_CONSTANTS;
1260 case VK_DYNAMIC_STATE_DEPTH_BOUNDS:
1261 return RADV_DYNAMIC_DEPTH_BOUNDS;
1262 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK:
1263 return RADV_DYNAMIC_STENCIL_COMPARE_MASK;
1264 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK:
1265 return RADV_DYNAMIC_STENCIL_WRITE_MASK;
1266 case VK_DYNAMIC_STATE_STENCIL_REFERENCE:
1267 return RADV_DYNAMIC_STENCIL_REFERENCE;
1268 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT:
1269 return RADV_DYNAMIC_DISCARD_RECTANGLE;
1270 default:
1271 unreachable("Unhandled dynamic state");
1272 }
1273 }
1274
1275 static uint32_t radv_pipeline_needed_dynamic_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
1276 {
1277 uint32_t states = RADV_DYNAMIC_ALL;
1278
1279 /* If rasterization is disabled we do not care about any of the dynamic states,
1280 * since they are all rasterization related only. */
1281 if (pCreateInfo->pRasterizationState->rasterizerDiscardEnable)
1282 return 0;
1283
1284 if (!pCreateInfo->pRasterizationState->depthBiasEnable)
1285 states &= ~RADV_DYNAMIC_DEPTH_BIAS;
1286
1287 if (!pCreateInfo->pDepthStencilState ||
1288 !pCreateInfo->pDepthStencilState->depthBoundsTestEnable)
1289 states &= ~RADV_DYNAMIC_DEPTH_BOUNDS;
1290
1291 if (!pCreateInfo->pDepthStencilState ||
1292 !pCreateInfo->pDepthStencilState->stencilTestEnable)
1293 states &= ~(RADV_DYNAMIC_STENCIL_COMPARE_MASK |
1294 RADV_DYNAMIC_STENCIL_WRITE_MASK |
1295 RADV_DYNAMIC_STENCIL_REFERENCE);
1296
1297 if (!vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT))
1298 states &= ~RADV_DYNAMIC_DISCARD_RECTANGLE;
1299
1300 /* TODO: blend constants & line width. */
1301
1302 return states;
1303 }
1304
1305
1306 static void
1307 radv_pipeline_init_dynamic_state(struct radv_pipeline *pipeline,
1308 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1309 {
1310 uint32_t needed_states = radv_pipeline_needed_dynamic_state(pCreateInfo);
1311 uint32_t states = needed_states;
1312 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
1313 struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
1314
1315 pipeline->dynamic_state = default_dynamic_state;
1316 pipeline->graphics.needed_dynamic_state = needed_states;
1317
1318 if (pCreateInfo->pDynamicState) {
1319 /* Remove all of the states that are marked as dynamic */
1320 uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
1321 for (uint32_t s = 0; s < count; s++)
1322 states &= ~radv_dynamic_state_mask(pCreateInfo->pDynamicState->pDynamicStates[s]);
1323 }
1324
1325 struct radv_dynamic_state *dynamic = &pipeline->dynamic_state;
1326
1327 if (needed_states & RADV_DYNAMIC_VIEWPORT) {
1328 assert(pCreateInfo->pViewportState);
1329
1330 dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount;
1331 if (states & RADV_DYNAMIC_VIEWPORT) {
1332 typed_memcpy(dynamic->viewport.viewports,
1333 pCreateInfo->pViewportState->pViewports,
1334 pCreateInfo->pViewportState->viewportCount);
1335 }
1336 }
1337
1338 if (needed_states & RADV_DYNAMIC_SCISSOR) {
1339 dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount;
1340 if (states & RADV_DYNAMIC_SCISSOR) {
1341 typed_memcpy(dynamic->scissor.scissors,
1342 pCreateInfo->pViewportState->pScissors,
1343 pCreateInfo->pViewportState->scissorCount);
1344 }
1345 }
1346
1347 if (states & RADV_DYNAMIC_LINE_WIDTH) {
1348 assert(pCreateInfo->pRasterizationState);
1349 dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth;
1350 }
1351
1352 if (states & RADV_DYNAMIC_DEPTH_BIAS) {
1353 assert(pCreateInfo->pRasterizationState);
1354 dynamic->depth_bias.bias =
1355 pCreateInfo->pRasterizationState->depthBiasConstantFactor;
1356 dynamic->depth_bias.clamp =
1357 pCreateInfo->pRasterizationState->depthBiasClamp;
1358 dynamic->depth_bias.slope =
1359 pCreateInfo->pRasterizationState->depthBiasSlopeFactor;
1360 }
1361
1362 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1363 *
1364 * pColorBlendState is [...] NULL if the pipeline has rasterization
1365 * disabled or if the subpass of the render pass the pipeline is
1366 * created against does not use any color attachments.
1367 */
1368 if (subpass->has_color_att && states & RADV_DYNAMIC_BLEND_CONSTANTS) {
1369 assert(pCreateInfo->pColorBlendState);
1370 typed_memcpy(dynamic->blend_constants,
1371 pCreateInfo->pColorBlendState->blendConstants, 4);
1372 }
1373
1374 /* If there is no depthstencil attachment, then don't read
1375 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1376 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1377 * no need to override the depthstencil defaults in
1378 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1379 *
1380 * Section 9.2 of the Vulkan 1.0.15 spec says:
1381 *
1382 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1383 * disabled or if the subpass of the render pass the pipeline is created
1384 * against does not use a depth/stencil attachment.
1385 */
1386 if (needed_states && subpass->depth_stencil_attachment) {
1387 assert(pCreateInfo->pDepthStencilState);
1388
1389 if (states & RADV_DYNAMIC_DEPTH_BOUNDS) {
1390 dynamic->depth_bounds.min =
1391 pCreateInfo->pDepthStencilState->minDepthBounds;
1392 dynamic->depth_bounds.max =
1393 pCreateInfo->pDepthStencilState->maxDepthBounds;
1394 }
1395
1396 if (states & RADV_DYNAMIC_STENCIL_COMPARE_MASK) {
1397 dynamic->stencil_compare_mask.front =
1398 pCreateInfo->pDepthStencilState->front.compareMask;
1399 dynamic->stencil_compare_mask.back =
1400 pCreateInfo->pDepthStencilState->back.compareMask;
1401 }
1402
1403 if (states & RADV_DYNAMIC_STENCIL_WRITE_MASK) {
1404 dynamic->stencil_write_mask.front =
1405 pCreateInfo->pDepthStencilState->front.writeMask;
1406 dynamic->stencil_write_mask.back =
1407 pCreateInfo->pDepthStencilState->back.writeMask;
1408 }
1409
1410 if (states & RADV_DYNAMIC_STENCIL_REFERENCE) {
1411 dynamic->stencil_reference.front =
1412 pCreateInfo->pDepthStencilState->front.reference;
1413 dynamic->stencil_reference.back =
1414 pCreateInfo->pDepthStencilState->back.reference;
1415 }
1416 }
1417
1418 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
1419 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
1420 if (states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1421 dynamic->discard_rectangle.count = discard_rectangle_info->discardRectangleCount;
1422 typed_memcpy(dynamic->discard_rectangle.rectangles,
1423 discard_rectangle_info->pDiscardRectangles,
1424 discard_rectangle_info->discardRectangleCount);
1425 }
1426
1427 pipeline->dynamic_state.mask = states;
1428 }
1429
1430 static struct radv_gs_state
1431 calculate_gs_info(const VkGraphicsPipelineCreateInfo *pCreateInfo,
1432 const struct radv_pipeline *pipeline)
1433 {
1434 struct radv_gs_state gs = {0};
1435 struct radv_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
1436 struct radv_es_output_info *es_info;
1437 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
1438 es_info = radv_pipeline_has_tess(pipeline) ? &gs_info->tes.es_info : &gs_info->vs.es_info;
1439 else
1440 es_info = radv_pipeline_has_tess(pipeline) ?
1441 &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.es_info :
1442 &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.es_info;
1443
1444 unsigned gs_num_invocations = MAX2(gs_info->gs.invocations, 1);
1445 bool uses_adjacency;
1446 switch(pCreateInfo->pInputAssemblyState->topology) {
1447 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1448 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1449 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1450 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1451 uses_adjacency = true;
1452 break;
1453 default:
1454 uses_adjacency = false;
1455 break;
1456 }
1457
1458 /* All these are in dwords: */
1459 /* We can't allow using the whole LDS, because GS waves compete with
1460 * other shader stages for LDS space. */
1461 const unsigned max_lds_size = 8 * 1024;
1462 const unsigned esgs_itemsize = es_info->esgs_itemsize / 4;
1463 unsigned esgs_lds_size;
1464
1465 /* All these are per subgroup: */
1466 const unsigned max_out_prims = 32 * 1024;
1467 const unsigned max_es_verts = 255;
1468 const unsigned ideal_gs_prims = 64;
1469 unsigned max_gs_prims, gs_prims;
1470 unsigned min_es_verts, es_verts, worst_case_es_verts;
1471
1472 if (uses_adjacency || gs_num_invocations > 1)
1473 max_gs_prims = 127 / gs_num_invocations;
1474 else
1475 max_gs_prims = 255;
1476
1477 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1478 * Make sure we don't go over the maximum value.
1479 */
1480 if (gs_info->gs.vertices_out > 0) {
1481 max_gs_prims = MIN2(max_gs_prims,
1482 max_out_prims /
1483 (gs_info->gs.vertices_out * gs_num_invocations));
1484 }
1485 assert(max_gs_prims > 0);
1486
1487 /* If the primitive has adjacency, halve the number of vertices
1488 * that will be reused in multiple primitives.
1489 */
1490 min_es_verts = gs_info->gs.vertices_in / (uses_adjacency ? 2 : 1);
1491
1492 gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
1493 worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
1494
1495 /* Compute ESGS LDS size based on the worst case number of ES vertices
1496 * needed to create the target number of GS prims per subgroup.
1497 */
1498 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1499
1500 /* If total LDS usage is too big, refactor partitions based on ratio
1501 * of ESGS item sizes.
1502 */
1503 if (esgs_lds_size > max_lds_size) {
1504 /* Our target GS Prims Per Subgroup was too large. Calculate
1505 * the maximum number of GS Prims Per Subgroup that will fit
1506 * into LDS, capped by the maximum that the hardware can support.
1507 */
1508 gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
1509 max_gs_prims);
1510 assert(gs_prims > 0);
1511 worst_case_es_verts = MIN2(min_es_verts * gs_prims,
1512 max_es_verts);
1513
1514 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1515 assert(esgs_lds_size <= max_lds_size);
1516 }
1517
1518 /* Now calculate remaining ESGS information. */
1519 if (esgs_lds_size)
1520 es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
1521 else
1522 es_verts = max_es_verts;
1523
1524 /* Vertices for adjacency primitives are not always reused, so restore
1525 * it for ES_VERTS_PER_SUBGRP.
1526 */
1527 min_es_verts = gs_info->gs.vertices_in;
1528
1529 /* For normal primitives, the VGT only checks if they are past the ES
1530 * verts per subgroup after allocating a full GS primitive and if they
1531 * are, kick off a new subgroup. But if those additional ES verts are
1532 * unique (e.g. not reused) we need to make sure there is enough LDS
1533 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1534 */
1535 es_verts -= min_es_verts - 1;
1536
1537 uint32_t es_verts_per_subgroup = es_verts;
1538 uint32_t gs_prims_per_subgroup = gs_prims;
1539 uint32_t gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
1540 uint32_t max_prims_per_subgroup = gs_inst_prims_in_subgroup * gs_info->gs.vertices_out;
1541 gs.lds_size = align(esgs_lds_size, 128) / 128;
1542 gs.vgt_gs_onchip_cntl = S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup) |
1543 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup) |
1544 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup);
1545 gs.vgt_gs_max_prims_per_subgroup = S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup);
1546 gs.vgt_esgs_ring_itemsize = esgs_itemsize;
1547 assert(max_prims_per_subgroup <= max_out_prims);
1548
1549 return gs;
1550 }
1551
1552 static void
1553 calculate_gs_ring_sizes(struct radv_pipeline *pipeline, const struct radv_gs_state *gs)
1554 {
1555 struct radv_device *device = pipeline->device;
1556 unsigned num_se = device->physical_device->rad_info.max_se;
1557 unsigned wave_size = 64;
1558 unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
1559 /* On SI-CI, the value comes from VGT_GS_VERTEX_REUSE = 16.
1560 * On VI+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
1561 */
1562 unsigned gs_vertex_reuse =
1563 (device->physical_device->rad_info.chip_class >= VI ? 32 : 16) * num_se;
1564 unsigned alignment = 256 * num_se;
1565 /* The maximum size is 63.999 MB per SE. */
1566 unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
1567 struct radv_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
1568
1569 /* Calculate the minimum size. */
1570 unsigned min_esgs_ring_size = align(gs->vgt_esgs_ring_itemsize * 4 * gs_vertex_reuse *
1571 wave_size, alignment);
1572 /* These are recommended sizes, not minimum sizes. */
1573 unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
1574 gs->vgt_esgs_ring_itemsize * 4 * gs_info->gs.vertices_in;
1575 unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
1576 gs_info->gs.max_gsvs_emit_size;
1577
1578 min_esgs_ring_size = align(min_esgs_ring_size, alignment);
1579 esgs_ring_size = align(esgs_ring_size, alignment);
1580 gsvs_ring_size = align(gsvs_ring_size, alignment);
1581
1582 if (pipeline->device->physical_device->rad_info.chip_class <= VI)
1583 pipeline->graphics.esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
1584
1585 pipeline->graphics.gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
1586 }
1587
1588 static void si_multiwave_lds_size_workaround(struct radv_device *device,
1589 unsigned *lds_size)
1590 {
1591 /* If tessellation is all offchip and on-chip GS isn't used, this
1592 * workaround is not needed.
1593 */
1594 return;
1595
1596 /* SPI barrier management bug:
1597 * Make sure we have at least 4k of LDS in use to avoid the bug.
1598 * It applies to workgroup sizes of more than one wavefront.
1599 */
1600 if (device->physical_device->rad_info.family == CHIP_BONAIRE ||
1601 device->physical_device->rad_info.family == CHIP_KABINI ||
1602 device->physical_device->rad_info.family == CHIP_MULLINS)
1603 *lds_size = MAX2(*lds_size, 8);
1604 }
1605
1606 struct radv_shader_variant *
1607 radv_get_shader(struct radv_pipeline *pipeline,
1608 gl_shader_stage stage)
1609 {
1610 if (stage == MESA_SHADER_VERTEX) {
1611 if (pipeline->shaders[MESA_SHADER_VERTEX])
1612 return pipeline->shaders[MESA_SHADER_VERTEX];
1613 if (pipeline->shaders[MESA_SHADER_TESS_CTRL])
1614 return pipeline->shaders[MESA_SHADER_TESS_CTRL];
1615 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
1616 return pipeline->shaders[MESA_SHADER_GEOMETRY];
1617 } else if (stage == MESA_SHADER_TESS_EVAL) {
1618 if (!radv_pipeline_has_tess(pipeline))
1619 return NULL;
1620 if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
1621 return pipeline->shaders[MESA_SHADER_TESS_EVAL];
1622 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
1623 return pipeline->shaders[MESA_SHADER_GEOMETRY];
1624 }
1625 return pipeline->shaders[stage];
1626 }
1627
1628 static struct radv_tessellation_state
1629 calculate_tess_state(struct radv_pipeline *pipeline,
1630 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1631 {
1632 unsigned num_tcs_input_cp;
1633 unsigned num_tcs_output_cp;
1634 unsigned lds_size;
1635 unsigned num_patches;
1636 struct radv_tessellation_state tess = {0};
1637
1638 num_tcs_input_cp = pCreateInfo->pTessellationState->patchControlPoints;
1639 num_tcs_output_cp = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.tcs_vertices_out; //TCS VERTICES OUT
1640 num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
1641
1642 lds_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.lds_size;
1643
1644 if (pipeline->device->physical_device->rad_info.chip_class >= CIK) {
1645 assert(lds_size <= 65536);
1646 lds_size = align(lds_size, 512) / 512;
1647 } else {
1648 assert(lds_size <= 32768);
1649 lds_size = align(lds_size, 256) / 256;
1650 }
1651 si_multiwave_lds_size_workaround(pipeline->device, &lds_size);
1652
1653 tess.lds_size = lds_size;
1654
1655 tess.ls_hs_config = S_028B58_NUM_PATCHES(num_patches) |
1656 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
1657 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
1658 tess.num_patches = num_patches;
1659
1660 struct radv_shader_variant *tes = radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL);
1661 unsigned type = 0, partitioning = 0, topology = 0, distribution_mode = 0;
1662
1663 switch (tes->info.tes.primitive_mode) {
1664 case GL_TRIANGLES:
1665 type = V_028B6C_TESS_TRIANGLE;
1666 break;
1667 case GL_QUADS:
1668 type = V_028B6C_TESS_QUAD;
1669 break;
1670 case GL_ISOLINES:
1671 type = V_028B6C_TESS_ISOLINE;
1672 break;
1673 }
1674
1675 switch (tes->info.tes.spacing) {
1676 case TESS_SPACING_EQUAL:
1677 partitioning = V_028B6C_PART_INTEGER;
1678 break;
1679 case TESS_SPACING_FRACTIONAL_ODD:
1680 partitioning = V_028B6C_PART_FRAC_ODD;
1681 break;
1682 case TESS_SPACING_FRACTIONAL_EVEN:
1683 partitioning = V_028B6C_PART_FRAC_EVEN;
1684 break;
1685 default:
1686 break;
1687 }
1688
1689 bool ccw = tes->info.tes.ccw;
1690 const VkPipelineTessellationDomainOriginStateCreateInfo *domain_origin_state =
1691 vk_find_struct_const(pCreateInfo->pTessellationState,
1692 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO);
1693
1694 if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT)
1695 ccw = !ccw;
1696
1697 if (tes->info.tes.point_mode)
1698 topology = V_028B6C_OUTPUT_POINT;
1699 else if (tes->info.tes.primitive_mode == GL_ISOLINES)
1700 topology = V_028B6C_OUTPUT_LINE;
1701 else if (ccw)
1702 topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
1703 else
1704 topology = V_028B6C_OUTPUT_TRIANGLE_CW;
1705
1706 if (pipeline->device->has_distributed_tess) {
1707 if (pipeline->device->physical_device->rad_info.family == CHIP_FIJI ||
1708 pipeline->device->physical_device->rad_info.family >= CHIP_POLARIS10)
1709 distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
1710 else
1711 distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
1712 } else
1713 distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
1714
1715 tess.tf_param = S_028B6C_TYPE(type) |
1716 S_028B6C_PARTITIONING(partitioning) |
1717 S_028B6C_TOPOLOGY(topology) |
1718 S_028B6C_DISTRIBUTION_MODE(distribution_mode);
1719
1720 return tess;
1721 }
1722
1723 static const struct radv_prim_vertex_count prim_size_table[] = {
1724 [V_008958_DI_PT_NONE] = {0, 0},
1725 [V_008958_DI_PT_POINTLIST] = {1, 1},
1726 [V_008958_DI_PT_LINELIST] = {2, 2},
1727 [V_008958_DI_PT_LINESTRIP] = {2, 1},
1728 [V_008958_DI_PT_TRILIST] = {3, 3},
1729 [V_008958_DI_PT_TRIFAN] = {3, 1},
1730 [V_008958_DI_PT_TRISTRIP] = {3, 1},
1731 [V_008958_DI_PT_LINELIST_ADJ] = {4, 4},
1732 [V_008958_DI_PT_LINESTRIP_ADJ] = {4, 1},
1733 [V_008958_DI_PT_TRILIST_ADJ] = {6, 6},
1734 [V_008958_DI_PT_TRISTRIP_ADJ] = {6, 2},
1735 [V_008958_DI_PT_RECTLIST] = {3, 3},
1736 [V_008958_DI_PT_LINELOOP] = {2, 1},
1737 [V_008958_DI_PT_POLYGON] = {3, 1},
1738 [V_008958_DI_PT_2D_TRI_STRIP] = {0, 0},
1739 };
1740
1741 static const struct radv_vs_output_info *get_vs_output_info(const struct radv_pipeline *pipeline)
1742 {
1743 if (radv_pipeline_has_gs(pipeline))
1744 return &pipeline->gs_copy_shader->info.vs.outinfo;
1745 else if (radv_pipeline_has_tess(pipeline))
1746 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.outinfo;
1747 else
1748 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.outinfo;
1749 }
1750
1751 static void
1752 radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders)
1753 {
1754 nir_shader* ordered_shaders[MESA_SHADER_STAGES];
1755 int shader_count = 0;
1756
1757 if(shaders[MESA_SHADER_FRAGMENT]) {
1758 ordered_shaders[shader_count++] = shaders[MESA_SHADER_FRAGMENT];
1759 }
1760 if(shaders[MESA_SHADER_GEOMETRY]) {
1761 ordered_shaders[shader_count++] = shaders[MESA_SHADER_GEOMETRY];
1762 }
1763 if(shaders[MESA_SHADER_TESS_EVAL]) {
1764 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_EVAL];
1765 }
1766 if(shaders[MESA_SHADER_TESS_CTRL]) {
1767 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_CTRL];
1768 }
1769 if(shaders[MESA_SHADER_VERTEX]) {
1770 ordered_shaders[shader_count++] = shaders[MESA_SHADER_VERTEX];
1771 }
1772
1773 if (shader_count > 1) {
1774 unsigned first = ordered_shaders[shader_count - 1]->info.stage;
1775 unsigned last = ordered_shaders[0]->info.stage;
1776
1777 if (ordered_shaders[0]->info.stage == MESA_SHADER_FRAGMENT &&
1778 ordered_shaders[1]->info.has_transform_feedback_varyings)
1779 nir_link_xfb_varyings(ordered_shaders[1], ordered_shaders[0]);
1780
1781 for (int i = 0; i < shader_count; ++i) {
1782 nir_variable_mode mask = 0;
1783
1784 if (ordered_shaders[i]->info.stage != first)
1785 mask = mask | nir_var_shader_in;
1786
1787 if (ordered_shaders[i]->info.stage != last)
1788 mask = mask | nir_var_shader_out;
1789
1790 nir_lower_io_to_scalar_early(ordered_shaders[i], mask);
1791 radv_optimize_nir(ordered_shaders[i], false, false);
1792 }
1793 }
1794
1795 for (int i = 1; i < shader_count; ++i) {
1796 nir_lower_io_arrays_to_elements(ordered_shaders[i],
1797 ordered_shaders[i - 1]);
1798
1799 if (nir_link_opt_varyings(ordered_shaders[i],
1800 ordered_shaders[i - 1]))
1801 radv_optimize_nir(ordered_shaders[i - 1], false, false);
1802
1803 nir_remove_dead_variables(ordered_shaders[i],
1804 nir_var_shader_out);
1805 nir_remove_dead_variables(ordered_shaders[i - 1],
1806 nir_var_shader_in);
1807
1808 bool progress = nir_remove_unused_varyings(ordered_shaders[i],
1809 ordered_shaders[i - 1]);
1810
1811 nir_compact_varyings(ordered_shaders[i],
1812 ordered_shaders[i - 1], true);
1813
1814 if (progress) {
1815 if (nir_lower_global_vars_to_local(ordered_shaders[i])) {
1816 ac_lower_indirect_derefs(ordered_shaders[i],
1817 pipeline->device->physical_device->rad_info.chip_class);
1818 }
1819 radv_optimize_nir(ordered_shaders[i], false, false);
1820
1821 if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) {
1822 ac_lower_indirect_derefs(ordered_shaders[i - 1],
1823 pipeline->device->physical_device->rad_info.chip_class);
1824 }
1825 radv_optimize_nir(ordered_shaders[i - 1], false, false);
1826 }
1827 }
1828 }
1829
1830 static uint32_t
1831 radv_get_attrib_stride(const VkPipelineVertexInputStateCreateInfo *input_state,
1832 uint32_t attrib_binding)
1833 {
1834 for (uint32_t i = 0; i < input_state->vertexBindingDescriptionCount; i++) {
1835 const VkVertexInputBindingDescription *input_binding =
1836 &input_state->pVertexBindingDescriptions[i];
1837
1838 if (input_binding->binding == attrib_binding)
1839 return input_binding->stride;
1840 }
1841
1842 return 0;
1843 }
1844
1845 static struct radv_pipeline_key
1846 radv_generate_graphics_pipeline_key(struct radv_pipeline *pipeline,
1847 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1848 const struct radv_blend_state *blend,
1849 bool has_view_index)
1850 {
1851 const VkPipelineVertexInputStateCreateInfo *input_state =
1852 pCreateInfo->pVertexInputState;
1853 const VkPipelineVertexInputDivisorStateCreateInfoEXT *divisor_state =
1854 vk_find_struct_const(input_state->pNext, PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT);
1855
1856 struct radv_pipeline_key key;
1857 memset(&key, 0, sizeof(key));
1858
1859 if (pCreateInfo->flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)
1860 key.optimisations_disabled = 1;
1861
1862 key.has_multiview_view_index = has_view_index;
1863
1864 uint32_t binding_input_rate = 0;
1865 uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
1866 for (unsigned i = 0; i < input_state->vertexBindingDescriptionCount; ++i) {
1867 if (input_state->pVertexBindingDescriptions[i].inputRate) {
1868 unsigned binding = input_state->pVertexBindingDescriptions[i].binding;
1869 binding_input_rate |= 1u << binding;
1870 instance_rate_divisors[binding] = 1;
1871 }
1872 }
1873 if (divisor_state) {
1874 for (unsigned i = 0; i < divisor_state->vertexBindingDivisorCount; ++i) {
1875 instance_rate_divisors[divisor_state->pVertexBindingDivisors[i].binding] =
1876 divisor_state->pVertexBindingDivisors[i].divisor;
1877 }
1878 }
1879
1880 for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) {
1881 const VkVertexInputAttributeDescription *desc =
1882 &input_state->pVertexAttributeDescriptions[i];
1883 const struct vk_format_description *format_desc;
1884 unsigned location = desc->location;
1885 unsigned binding = desc->binding;
1886 unsigned num_format, data_format;
1887 int first_non_void;
1888
1889 if (binding_input_rate & (1u << binding)) {
1890 key.instance_rate_inputs |= 1u << location;
1891 key.instance_rate_divisors[location] = instance_rate_divisors[binding];
1892 }
1893
1894 format_desc = vk_format_description(desc->format);
1895 first_non_void = vk_format_get_first_non_void_channel(desc->format);
1896
1897 num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
1898 data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
1899
1900 key.vertex_attribute_formats[location] = data_format | (num_format << 4);
1901 key.vertex_attribute_bindings[location] = desc->binding;
1902 key.vertex_attribute_offsets[location] = desc->offset;
1903 key.vertex_attribute_strides[location] = radv_get_attrib_stride(input_state, desc->binding);
1904
1905 if (pipeline->device->physical_device->rad_info.chip_class <= VI &&
1906 pipeline->device->physical_device->rad_info.family != CHIP_STONEY) {
1907 VkFormat format = input_state->pVertexAttributeDescriptions[i].format;
1908 uint64_t adjust;
1909 switch(format) {
1910 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
1911 case VK_FORMAT_A2B10G10R10_SNORM_PACK32:
1912 adjust = RADV_ALPHA_ADJUST_SNORM;
1913 break;
1914 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
1915 case VK_FORMAT_A2B10G10R10_SSCALED_PACK32:
1916 adjust = RADV_ALPHA_ADJUST_SSCALED;
1917 break;
1918 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
1919 case VK_FORMAT_A2B10G10R10_SINT_PACK32:
1920 adjust = RADV_ALPHA_ADJUST_SINT;
1921 break;
1922 default:
1923 adjust = 0;
1924 break;
1925 }
1926 key.vertex_alpha_adjust |= adjust << (2 * location);
1927 }
1928
1929 switch (desc->format) {
1930 case VK_FORMAT_B8G8R8A8_UNORM:
1931 case VK_FORMAT_B8G8R8A8_SNORM:
1932 case VK_FORMAT_B8G8R8A8_USCALED:
1933 case VK_FORMAT_B8G8R8A8_SSCALED:
1934 case VK_FORMAT_B8G8R8A8_UINT:
1935 case VK_FORMAT_B8G8R8A8_SINT:
1936 case VK_FORMAT_B8G8R8A8_SRGB:
1937 case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
1938 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
1939 case VK_FORMAT_A2R10G10B10_USCALED_PACK32:
1940 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
1941 case VK_FORMAT_A2R10G10B10_UINT_PACK32:
1942 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
1943 key.vertex_post_shuffle |= 1 << location;
1944 break;
1945 default:
1946 break;
1947 }
1948 }
1949
1950 if (pCreateInfo->pTessellationState)
1951 key.tess_input_vertices = pCreateInfo->pTessellationState->patchControlPoints;
1952
1953
1954 if (pCreateInfo->pMultisampleState &&
1955 pCreateInfo->pMultisampleState->rasterizationSamples > 1) {
1956 uint32_t num_samples = pCreateInfo->pMultisampleState->rasterizationSamples;
1957 uint32_t ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo->pMultisampleState);
1958 key.num_samples = num_samples;
1959 key.log2_ps_iter_samples = util_logbase2(ps_iter_samples);
1960 }
1961
1962 key.col_format = blend->spi_shader_col_format;
1963 if (pipeline->device->physical_device->rad_info.chip_class < VI)
1964 radv_pipeline_compute_get_int_clamp(pCreateInfo, &key.is_int8, &key.is_int10);
1965
1966 return key;
1967 }
1968
1969 static void
1970 radv_fill_shader_keys(struct radv_shader_variant_key *keys,
1971 const struct radv_pipeline_key *key,
1972 nir_shader **nir)
1973 {
1974 keys[MESA_SHADER_VERTEX].vs.instance_rate_inputs = key->instance_rate_inputs;
1975 keys[MESA_SHADER_VERTEX].vs.alpha_adjust = key->vertex_alpha_adjust;
1976 keys[MESA_SHADER_VERTEX].vs.post_shuffle = key->vertex_post_shuffle;
1977 for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; ++i) {
1978 keys[MESA_SHADER_VERTEX].vs.instance_rate_divisors[i] = key->instance_rate_divisors[i];
1979 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_formats[i] = key->vertex_attribute_formats[i];
1980 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_bindings[i] = key->vertex_attribute_bindings[i];
1981 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_offsets[i] = key->vertex_attribute_offsets[i];
1982 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_strides[i] = key->vertex_attribute_strides[i];
1983 }
1984
1985 if (nir[MESA_SHADER_TESS_CTRL]) {
1986 keys[MESA_SHADER_VERTEX].vs.as_ls = true;
1987 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = 0;
1988 keys[MESA_SHADER_TESS_CTRL].tcs.input_vertices = key->tess_input_vertices;
1989 keys[MESA_SHADER_TESS_CTRL].tcs.primitive_mode = nir[MESA_SHADER_TESS_EVAL]->info.tess.primitive_mode;
1990
1991 keys[MESA_SHADER_TESS_CTRL].tcs.tes_reads_tess_factors = !!(nir[MESA_SHADER_TESS_EVAL]->info.inputs_read & (VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER));
1992 }
1993
1994 if (nir[MESA_SHADER_GEOMETRY]) {
1995 if (nir[MESA_SHADER_TESS_CTRL])
1996 keys[MESA_SHADER_TESS_EVAL].tes.as_es = true;
1997 else
1998 keys[MESA_SHADER_VERTEX].vs.as_es = true;
1999 }
2000
2001 for(int i = 0; i < MESA_SHADER_STAGES; ++i)
2002 keys[i].has_multiview_view_index = key->has_multiview_view_index;
2003
2004 keys[MESA_SHADER_FRAGMENT].fs.col_format = key->col_format;
2005 keys[MESA_SHADER_FRAGMENT].fs.is_int8 = key->is_int8;
2006 keys[MESA_SHADER_FRAGMENT].fs.is_int10 = key->is_int10;
2007 keys[MESA_SHADER_FRAGMENT].fs.log2_ps_iter_samples = key->log2_ps_iter_samples;
2008 keys[MESA_SHADER_FRAGMENT].fs.num_samples = key->num_samples;
2009 }
2010
2011 static void
2012 merge_tess_info(struct shader_info *tes_info,
2013 const struct shader_info *tcs_info)
2014 {
2015 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
2016 *
2017 * "PointMode. Controls generation of points rather than triangles
2018 * or lines. This functionality defaults to disabled, and is
2019 * enabled if either shader stage includes the execution mode.
2020 *
2021 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
2022 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
2023 * and OutputVertices, it says:
2024 *
2025 * "One mode must be set in at least one of the tessellation
2026 * shader stages."
2027 *
2028 * So, the fields can be set in either the TCS or TES, but they must
2029 * agree if set in both. Our backend looks at TES, so bitwise-or in
2030 * the values from the TCS.
2031 */
2032 assert(tcs_info->tess.tcs_vertices_out == 0 ||
2033 tes_info->tess.tcs_vertices_out == 0 ||
2034 tcs_info->tess.tcs_vertices_out == tes_info->tess.tcs_vertices_out);
2035 tes_info->tess.tcs_vertices_out |= tcs_info->tess.tcs_vertices_out;
2036
2037 assert(tcs_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2038 tes_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2039 tcs_info->tess.spacing == tes_info->tess.spacing);
2040 tes_info->tess.spacing |= tcs_info->tess.spacing;
2041
2042 assert(tcs_info->tess.primitive_mode == 0 ||
2043 tes_info->tess.primitive_mode == 0 ||
2044 tcs_info->tess.primitive_mode == tes_info->tess.primitive_mode);
2045 tes_info->tess.primitive_mode |= tcs_info->tess.primitive_mode;
2046 tes_info->tess.ccw |= tcs_info->tess.ccw;
2047 tes_info->tess.point_mode |= tcs_info->tess.point_mode;
2048 }
2049
2050 static
2051 void radv_init_feedback(const VkPipelineCreationFeedbackCreateInfoEXT *ext)
2052 {
2053 if (!ext)
2054 return;
2055
2056 if (ext->pPipelineCreationFeedback) {
2057 ext->pPipelineCreationFeedback->flags = 0;
2058 ext->pPipelineCreationFeedback->duration = 0;
2059 }
2060
2061 for (unsigned i = 0; i < ext->pipelineStageCreationFeedbackCount; ++i) {
2062 ext->pPipelineStageCreationFeedbacks[i].flags = 0;
2063 ext->pPipelineStageCreationFeedbacks[i].duration = 0;
2064 }
2065 }
2066
2067 static
2068 void radv_start_feedback(VkPipelineCreationFeedbackEXT *feedback)
2069 {
2070 if (!feedback)
2071 return;
2072
2073 feedback->duration -= radv_get_current_time();
2074 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT;
2075 }
2076
2077 static
2078 void radv_stop_feedback(VkPipelineCreationFeedbackEXT *feedback, bool cache_hit)
2079 {
2080 if (!feedback)
2081 return;
2082
2083 feedback->duration += radv_get_current_time();
2084 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT |
2085 (cache_hit ? VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT : 0);
2086 }
2087
2088 static
2089 void radv_create_shaders(struct radv_pipeline *pipeline,
2090 struct radv_device *device,
2091 struct radv_pipeline_cache *cache,
2092 const struct radv_pipeline_key *key,
2093 const VkPipelineShaderStageCreateInfo **pStages,
2094 const VkPipelineCreateFlags flags,
2095 VkPipelineCreationFeedbackEXT *pipeline_feedback,
2096 VkPipelineCreationFeedbackEXT **stage_feedbacks)
2097 {
2098 struct radv_shader_module fs_m = {0};
2099 struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
2100 nir_shader *nir[MESA_SHADER_STAGES] = {0};
2101 void *codes[MESA_SHADER_STAGES] = {0};
2102 unsigned code_sizes[MESA_SHADER_STAGES] = {0};
2103 struct radv_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{0}}}};
2104 unsigned char hash[20], gs_copy_hash[20];
2105
2106 radv_start_feedback(pipeline_feedback);
2107
2108 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2109 if (pStages[i]) {
2110 modules[i] = radv_shader_module_from_handle(pStages[i]->module);
2111 if (modules[i]->nir)
2112 _mesa_sha1_compute(modules[i]->nir->info.name,
2113 strlen(modules[i]->nir->info.name),
2114 modules[i]->sha1);
2115
2116 pipeline->active_stages |= mesa_to_vk_shader_stage(i);
2117 }
2118 }
2119
2120 radv_hash_shaders(hash, pStages, pipeline->layout, key, get_hash_flags(device));
2121 memcpy(gs_copy_hash, hash, 20);
2122 gs_copy_hash[0] ^= 1;
2123
2124 bool found_in_application_cache = true;
2125 if (modules[MESA_SHADER_GEOMETRY]) {
2126 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2127 radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants,
2128 &found_in_application_cache);
2129 pipeline->gs_copy_shader = variants[MESA_SHADER_GEOMETRY];
2130 }
2131
2132 if (radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders,
2133 &found_in_application_cache) &&
2134 (!modules[MESA_SHADER_GEOMETRY] || pipeline->gs_copy_shader)) {
2135 radv_stop_feedback(pipeline_feedback, found_in_application_cache);
2136 return;
2137 }
2138
2139 if (!modules[MESA_SHADER_FRAGMENT] && !modules[MESA_SHADER_COMPUTE]) {
2140 nir_builder fs_b;
2141 nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
2142 fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "noop_fs");
2143 fs_m.nir = fs_b.shader;
2144 modules[MESA_SHADER_FRAGMENT] = &fs_m;
2145 }
2146
2147 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2148 const VkPipelineShaderStageCreateInfo *stage = pStages[i];
2149
2150 if (!modules[i])
2151 continue;
2152
2153 radv_start_feedback(stage_feedbacks[i]);
2154
2155 nir[i] = radv_shader_compile_to_nir(device, modules[i],
2156 stage ? stage->pName : "main", i,
2157 stage ? stage->pSpecializationInfo : NULL,
2158 flags);
2159
2160 /* We don't want to alter meta shaders IR directly so clone it
2161 * first.
2162 */
2163 if (nir[i]->info.name) {
2164 nir[i] = nir_shader_clone(NULL, nir[i]);
2165 }
2166
2167 radv_stop_feedback(stage_feedbacks[i], false);
2168 }
2169
2170 if (nir[MESA_SHADER_TESS_CTRL]) {
2171 nir_lower_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out, NULL);
2172 merge_tess_info(&nir[MESA_SHADER_TESS_EVAL]->info, &nir[MESA_SHADER_TESS_CTRL]->info);
2173 }
2174
2175 if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT))
2176 radv_link_shaders(pipeline, nir);
2177
2178 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2179 if (nir[i]) {
2180 NIR_PASS_V(nir[i], nir_lower_bool_to_int32);
2181 NIR_PASS_V(nir[i], nir_lower_non_uniform_access,
2182 nir_lower_non_uniform_ubo_access |
2183 nir_lower_non_uniform_ssbo_access |
2184 nir_lower_non_uniform_texture_access |
2185 nir_lower_non_uniform_image_access);
2186 }
2187
2188 if (radv_can_dump_shader(device, modules[i], false))
2189 nir_print_shader(nir[i], stderr);
2190 }
2191
2192 radv_fill_shader_keys(keys, key, nir);
2193
2194 if (nir[MESA_SHADER_FRAGMENT]) {
2195 if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) {
2196 radv_start_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT]);
2197
2198 pipeline->shaders[MESA_SHADER_FRAGMENT] =
2199 radv_shader_variant_create(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1,
2200 pipeline->layout, keys + MESA_SHADER_FRAGMENT,
2201 &codes[MESA_SHADER_FRAGMENT], &code_sizes[MESA_SHADER_FRAGMENT]);
2202
2203 radv_stop_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT], false);
2204 }
2205
2206 /* TODO: These are no longer used as keys we should refactor this */
2207 keys[MESA_SHADER_VERTEX].vs.export_prim_id =
2208 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input;
2209 keys[MESA_SHADER_VERTEX].vs.export_layer_id =
2210 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.layer_input;
2211 keys[MESA_SHADER_TESS_EVAL].tes.export_prim_id =
2212 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input;
2213 keys[MESA_SHADER_TESS_EVAL].tes.export_layer_id =
2214 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.layer_input;
2215 }
2216
2217 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_TESS_CTRL]) {
2218 if (!pipeline->shaders[MESA_SHADER_TESS_CTRL]) {
2219 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
2220 struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
2221 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
2222
2223 radv_start_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL]);
2224
2225 pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_create(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2,
2226 pipeline->layout,
2227 &key, &codes[MESA_SHADER_TESS_CTRL],
2228 &code_sizes[MESA_SHADER_TESS_CTRL]);
2229
2230 radv_stop_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL], false);
2231 }
2232 modules[MESA_SHADER_VERTEX] = NULL;
2233 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
2234 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.tcs.outputs_written);
2235 }
2236
2237 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_GEOMETRY]) {
2238 gl_shader_stage pre_stage = modules[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
2239 if (!pipeline->shaders[MESA_SHADER_GEOMETRY]) {
2240 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
2241
2242 radv_start_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY]);
2243
2244 pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_create(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2,
2245 pipeline->layout,
2246 &keys[pre_stage] , &codes[MESA_SHADER_GEOMETRY],
2247 &code_sizes[MESA_SHADER_GEOMETRY]);
2248
2249 radv_stop_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY], false);
2250 }
2251 modules[pre_stage] = NULL;
2252 }
2253
2254 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2255 if(modules[i] && !pipeline->shaders[i]) {
2256 if (i == MESA_SHADER_TESS_CTRL) {
2257 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = util_last_bit64(pipeline->shaders[MESA_SHADER_VERTEX]->info.info.vs.ls_outputs_written);
2258 }
2259 if (i == MESA_SHADER_TESS_EVAL) {
2260 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
2261 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.tcs.outputs_written);
2262 }
2263
2264 radv_start_feedback(stage_feedbacks[i]);
2265
2266 pipeline->shaders[i] = radv_shader_variant_create(device, modules[i], &nir[i], 1,
2267 pipeline->layout,
2268 keys + i, &codes[i],
2269 &code_sizes[i]);
2270
2271 radv_stop_feedback(stage_feedbacks[i], false);
2272 }
2273 }
2274
2275 if(modules[MESA_SHADER_GEOMETRY]) {
2276 void *gs_copy_code = NULL;
2277 unsigned gs_copy_code_size = 0;
2278 if (!pipeline->gs_copy_shader) {
2279 pipeline->gs_copy_shader = radv_create_gs_copy_shader(
2280 device, nir[MESA_SHADER_GEOMETRY], &gs_copy_code,
2281 &gs_copy_code_size,
2282 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index);
2283 }
2284
2285 if (pipeline->gs_copy_shader) {
2286 void *code[MESA_SHADER_STAGES] = {0};
2287 unsigned code_size[MESA_SHADER_STAGES] = {0};
2288 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2289
2290 code[MESA_SHADER_GEOMETRY] = gs_copy_code;
2291 code_size[MESA_SHADER_GEOMETRY] = gs_copy_code_size;
2292 variants[MESA_SHADER_GEOMETRY] = pipeline->gs_copy_shader;
2293
2294 radv_pipeline_cache_insert_shaders(device, cache,
2295 gs_copy_hash,
2296 variants,
2297 (const void**)code,
2298 code_size);
2299 }
2300 free(gs_copy_code);
2301 }
2302
2303 radv_pipeline_cache_insert_shaders(device, cache, hash, pipeline->shaders,
2304 (const void**)codes, code_sizes);
2305
2306 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2307 free(codes[i]);
2308 if (nir[i]) {
2309 if (!pipeline->device->keep_shader_info)
2310 ralloc_free(nir[i]);
2311
2312 if (radv_can_dump_shader_stats(device, modules[i]))
2313 radv_shader_dump_stats(device,
2314 pipeline->shaders[i],
2315 i, stderr);
2316 }
2317 }
2318
2319 if (fs_m.nir)
2320 ralloc_free(fs_m.nir);
2321
2322 radv_stop_feedback(pipeline_feedback, false);
2323 }
2324
2325 static uint32_t
2326 radv_pipeline_stage_to_user_data_0(struct radv_pipeline *pipeline,
2327 gl_shader_stage stage, enum chip_class chip_class)
2328 {
2329 bool has_gs = radv_pipeline_has_gs(pipeline);
2330 bool has_tess = radv_pipeline_has_tess(pipeline);
2331 switch (stage) {
2332 case MESA_SHADER_FRAGMENT:
2333 return R_00B030_SPI_SHADER_USER_DATA_PS_0;
2334 case MESA_SHADER_VERTEX:
2335 if (chip_class >= GFX9) {
2336 return has_tess ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
2337 has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2338 R_00B130_SPI_SHADER_USER_DATA_VS_0;
2339 }
2340 if (has_tess)
2341 return R_00B530_SPI_SHADER_USER_DATA_LS_0;
2342 else
2343 return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : R_00B130_SPI_SHADER_USER_DATA_VS_0;
2344 case MESA_SHADER_GEOMETRY:
2345 return chip_class >= GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2346 R_00B230_SPI_SHADER_USER_DATA_GS_0;
2347 case MESA_SHADER_COMPUTE:
2348 return R_00B900_COMPUTE_USER_DATA_0;
2349 case MESA_SHADER_TESS_CTRL:
2350 return chip_class >= GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
2351 R_00B430_SPI_SHADER_USER_DATA_HS_0;
2352 case MESA_SHADER_TESS_EVAL:
2353 if (chip_class >= GFX9) {
2354 return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2355 R_00B130_SPI_SHADER_USER_DATA_VS_0;
2356 }
2357 if (has_gs)
2358 return R_00B330_SPI_SHADER_USER_DATA_ES_0;
2359 else
2360 return R_00B130_SPI_SHADER_USER_DATA_VS_0;
2361 default:
2362 unreachable("unknown shader");
2363 }
2364 }
2365
2366 struct radv_bin_size_entry {
2367 unsigned bpp;
2368 VkExtent2D extent;
2369 };
2370
2371 static VkExtent2D
2372 radv_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo)
2373 {
2374 static const struct radv_bin_size_entry color_size_table[][3][9] = {
2375 {
2376 /* One RB / SE */
2377 {
2378 /* One shader engine */
2379 { 0, {128, 128}},
2380 { 1, { 64, 128}},
2381 { 2, { 32, 128}},
2382 { 3, { 16, 128}},
2383 { 17, { 0, 0}},
2384 { UINT_MAX, { 0, 0}},
2385 },
2386 {
2387 /* Two shader engines */
2388 { 0, {128, 128}},
2389 { 2, { 64, 128}},
2390 { 3, { 32, 128}},
2391 { 5, { 16, 128}},
2392 { 17, { 0, 0}},
2393 { UINT_MAX, { 0, 0}},
2394 },
2395 {
2396 /* Four shader engines */
2397 { 0, {128, 128}},
2398 { 3, { 64, 128}},
2399 { 5, { 16, 128}},
2400 { 17, { 0, 0}},
2401 { UINT_MAX, { 0, 0}},
2402 },
2403 },
2404 {
2405 /* Two RB / SE */
2406 {
2407 /* One shader engine */
2408 { 0, {128, 128}},
2409 { 2, { 64, 128}},
2410 { 3, { 32, 128}},
2411 { 5, { 16, 128}},
2412 { 33, { 0, 0}},
2413 { UINT_MAX, { 0, 0}},
2414 },
2415 {
2416 /* Two shader engines */
2417 { 0, {128, 128}},
2418 { 3, { 64, 128}},
2419 { 5, { 32, 128}},
2420 { 9, { 16, 128}},
2421 { 33, { 0, 0}},
2422 { UINT_MAX, { 0, 0}},
2423 },
2424 {
2425 /* Four shader engines */
2426 { 0, {256, 256}},
2427 { 2, {128, 256}},
2428 { 3, {128, 128}},
2429 { 5, { 64, 128}},
2430 { 9, { 16, 128}},
2431 { 33, { 0, 0}},
2432 { UINT_MAX, { 0, 0}},
2433 },
2434 },
2435 {
2436 /* Four RB / SE */
2437 {
2438 /* One shader engine */
2439 { 0, {128, 256}},
2440 { 2, {128, 128}},
2441 { 3, { 64, 128}},
2442 { 5, { 32, 128}},
2443 { 9, { 16, 128}},
2444 { 33, { 0, 0}},
2445 { UINT_MAX, { 0, 0}},
2446 },
2447 {
2448 /* Two shader engines */
2449 { 0, {256, 256}},
2450 { 2, {128, 256}},
2451 { 3, {128, 128}},
2452 { 5, { 64, 128}},
2453 { 9, { 32, 128}},
2454 { 17, { 16, 128}},
2455 { 33, { 0, 0}},
2456 { UINT_MAX, { 0, 0}},
2457 },
2458 {
2459 /* Four shader engines */
2460 { 0, {256, 512}},
2461 { 2, {256, 256}},
2462 { 3, {128, 256}},
2463 { 5, {128, 128}},
2464 { 9, { 64, 128}},
2465 { 17, { 16, 128}},
2466 { 33, { 0, 0}},
2467 { UINT_MAX, { 0, 0}},
2468 },
2469 },
2470 };
2471 static const struct radv_bin_size_entry ds_size_table[][3][9] = {
2472 {
2473 // One RB / SE
2474 {
2475 // One shader engine
2476 { 0, {128, 256}},
2477 { 2, {128, 128}},
2478 { 4, { 64, 128}},
2479 { 7, { 32, 128}},
2480 { 13, { 16, 128}},
2481 { 49, { 0, 0}},
2482 { UINT_MAX, { 0, 0}},
2483 },
2484 {
2485 // Two shader engines
2486 { 0, {256, 256}},
2487 { 2, {128, 256}},
2488 { 4, {128, 128}},
2489 { 7, { 64, 128}},
2490 { 13, { 32, 128}},
2491 { 25, { 16, 128}},
2492 { 49, { 0, 0}},
2493 { UINT_MAX, { 0, 0}},
2494 },
2495 {
2496 // Four shader engines
2497 { 0, {256, 512}},
2498 { 2, {256, 256}},
2499 { 4, {128, 256}},
2500 { 7, {128, 128}},
2501 { 13, { 64, 128}},
2502 { 25, { 16, 128}},
2503 { 49, { 0, 0}},
2504 { UINT_MAX, { 0, 0}},
2505 },
2506 },
2507 {
2508 // Two RB / SE
2509 {
2510 // One shader engine
2511 { 0, {256, 256}},
2512 { 2, {128, 256}},
2513 { 4, {128, 128}},
2514 { 7, { 64, 128}},
2515 { 13, { 32, 128}},
2516 { 25, { 16, 128}},
2517 { 97, { 0, 0}},
2518 { UINT_MAX, { 0, 0}},
2519 },
2520 {
2521 // Two shader engines
2522 { 0, {256, 512}},
2523 { 2, {256, 256}},
2524 { 4, {128, 256}},
2525 { 7, {128, 128}},
2526 { 13, { 64, 128}},
2527 { 25, { 32, 128}},
2528 { 49, { 16, 128}},
2529 { 97, { 0, 0}},
2530 { UINT_MAX, { 0, 0}},
2531 },
2532 {
2533 // Four shader engines
2534 { 0, {512, 512}},
2535 { 2, {256, 512}},
2536 { 4, {256, 256}},
2537 { 7, {128, 256}},
2538 { 13, {128, 128}},
2539 { 25, { 64, 128}},
2540 { 49, { 16, 128}},
2541 { 97, { 0, 0}},
2542 { UINT_MAX, { 0, 0}},
2543 },
2544 },
2545 {
2546 // Four RB / SE
2547 {
2548 // One shader engine
2549 { 0, {256, 512}},
2550 { 2, {256, 256}},
2551 { 4, {128, 256}},
2552 { 7, {128, 128}},
2553 { 13, { 64, 128}},
2554 { 25, { 32, 128}},
2555 { 49, { 16, 128}},
2556 { UINT_MAX, { 0, 0}},
2557 },
2558 {
2559 // Two shader engines
2560 { 0, {512, 512}},
2561 { 2, {256, 512}},
2562 { 4, {256, 256}},
2563 { 7, {128, 256}},
2564 { 13, {128, 128}},
2565 { 25, { 64, 128}},
2566 { 49, { 32, 128}},
2567 { 97, { 16, 128}},
2568 { UINT_MAX, { 0, 0}},
2569 },
2570 {
2571 // Four shader engines
2572 { 0, {512, 512}},
2573 { 4, {256, 512}},
2574 { 7, {256, 256}},
2575 { 13, {128, 256}},
2576 { 25, {128, 128}},
2577 { 49, { 64, 128}},
2578 { 97, { 16, 128}},
2579 { UINT_MAX, { 0, 0}},
2580 },
2581 },
2582 };
2583
2584 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
2585 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
2586 VkExtent2D extent = {512, 512};
2587
2588 unsigned log_num_rb_per_se =
2589 util_logbase2_ceil(pipeline->device->physical_device->rad_info.num_render_backends /
2590 pipeline->device->physical_device->rad_info.max_se);
2591 unsigned log_num_se = util_logbase2_ceil(pipeline->device->physical_device->rad_info.max_se);
2592
2593 unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
2594 unsigned ps_iter_samples = 1u << G_028804_PS_ITER_SAMPLES(pipeline->graphics.ms.db_eqaa);
2595 unsigned effective_samples = total_samples;
2596 unsigned color_bytes_per_pixel = 0;
2597
2598 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
2599 if (vkblend) {
2600 for (unsigned i = 0; i < subpass->color_count; i++) {
2601 if (!vkblend->pAttachments[i].colorWriteMask)
2602 continue;
2603
2604 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
2605 continue;
2606
2607 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
2608 color_bytes_per_pixel += vk_format_get_blocksize(format);
2609 }
2610
2611 /* MSAA images typically don't use all samples all the time. */
2612 if (effective_samples >= 2 && ps_iter_samples <= 1)
2613 effective_samples = 2;
2614 color_bytes_per_pixel *= effective_samples;
2615 }
2616
2617 const struct radv_bin_size_entry *color_entry = color_size_table[log_num_rb_per_se][log_num_se];
2618 while(color_entry[1].bpp <= color_bytes_per_pixel)
2619 ++color_entry;
2620
2621 extent = color_entry->extent;
2622
2623 if (subpass->depth_stencil_attachment) {
2624 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
2625
2626 /* Coefficients taken from AMDVLK */
2627 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
2628 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0;
2629 unsigned ds_bytes_per_pixel = 4 * (depth_coeff + stencil_coeff) * total_samples;
2630
2631 const struct radv_bin_size_entry *ds_entry = ds_size_table[log_num_rb_per_se][log_num_se];
2632 while(ds_entry[1].bpp <= ds_bytes_per_pixel)
2633 ++ds_entry;
2634
2635 extent.width = MIN2(extent.width, ds_entry->extent.width);
2636 extent.height = MIN2(extent.height, ds_entry->extent.height);
2637 }
2638
2639 return extent;
2640 }
2641
2642 static void
2643 radv_pipeline_generate_binning_state(struct radeon_cmdbuf *ctx_cs,
2644 struct radv_pipeline *pipeline,
2645 const VkGraphicsPipelineCreateInfo *pCreateInfo)
2646 {
2647 if (pipeline->device->physical_device->rad_info.chip_class < GFX9)
2648 return;
2649
2650 uint32_t pa_sc_binner_cntl_0 =
2651 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
2652 S_028C44_DISABLE_START_OF_PRIM(1);
2653 uint32_t db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF);
2654
2655 VkExtent2D bin_size = radv_compute_bin_size(pipeline, pCreateInfo);
2656
2657 unsigned context_states_per_bin; /* allowed range: [1, 6] */
2658 unsigned persistent_states_per_bin; /* allowed range: [1, 32] */
2659 unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */
2660
2661 switch (pipeline->device->physical_device->rad_info.family) {
2662 case CHIP_VEGA10:
2663 case CHIP_VEGA12:
2664 case CHIP_VEGA20:
2665 context_states_per_bin = 1;
2666 persistent_states_per_bin = 1;
2667 fpovs_per_batch = 63;
2668 break;
2669 case CHIP_RAVEN:
2670 case CHIP_RAVEN2:
2671 context_states_per_bin = 6;
2672 persistent_states_per_bin = 32;
2673 fpovs_per_batch = 63;
2674 break;
2675 default:
2676 unreachable("unhandled family while determining binning state.");
2677 }
2678
2679 if (pipeline->device->pbb_allowed && bin_size.width && bin_size.height) {
2680 pa_sc_binner_cntl_0 =
2681 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) |
2682 S_028C44_BIN_SIZE_X(bin_size.width == 16) |
2683 S_028C44_BIN_SIZE_Y(bin_size.height == 16) |
2684 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size.width, 32)) - 5) |
2685 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size.height, 32)) - 5) |
2686 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin - 1) |
2687 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin - 1) |
2688 S_028C44_DISABLE_START_OF_PRIM(1) |
2689 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch) |
2690 S_028C44_OPTIMAL_BIN_SELECTION(1);
2691 }
2692
2693 radeon_set_context_reg(ctx_cs, R_028C44_PA_SC_BINNER_CNTL_0,
2694 pa_sc_binner_cntl_0);
2695 radeon_set_context_reg(ctx_cs, R_028060_DB_DFSM_CONTROL,
2696 db_dfsm_control);
2697 }
2698
2699
2700 static void
2701 radv_pipeline_generate_depth_stencil_state(struct radeon_cmdbuf *ctx_cs,
2702 struct radv_pipeline *pipeline,
2703 const VkGraphicsPipelineCreateInfo *pCreateInfo,
2704 const struct radv_graphics_pipeline_create_info *extra)
2705 {
2706 const VkPipelineDepthStencilStateCreateInfo *vkds = pCreateInfo->pDepthStencilState;
2707 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
2708 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
2709 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
2710 struct radv_render_pass_attachment *attachment = NULL;
2711 uint32_t db_depth_control = 0, db_stencil_control = 0;
2712 uint32_t db_render_control = 0, db_render_override2 = 0;
2713 uint32_t db_render_override = 0;
2714
2715 if (subpass->depth_stencil_attachment)
2716 attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
2717
2718 bool has_depth_attachment = attachment && vk_format_is_depth(attachment->format);
2719 bool has_stencil_attachment = attachment && vk_format_is_stencil(attachment->format);
2720
2721 if (vkds && has_depth_attachment) {
2722 db_depth_control = S_028800_Z_ENABLE(vkds->depthTestEnable ? 1 : 0) |
2723 S_028800_Z_WRITE_ENABLE(vkds->depthWriteEnable ? 1 : 0) |
2724 S_028800_ZFUNC(vkds->depthCompareOp) |
2725 S_028800_DEPTH_BOUNDS_ENABLE(vkds->depthBoundsTestEnable ? 1 : 0);
2726
2727 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
2728 db_render_override2 |= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment->samples > 2);
2729 }
2730
2731 if (has_stencil_attachment && vkds && vkds->stencilTestEnable) {
2732 db_depth_control |= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
2733 db_depth_control |= S_028800_STENCILFUNC(vkds->front.compareOp);
2734 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds->front.failOp));
2735 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds->front.passOp));
2736 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds->front.depthFailOp));
2737
2738 db_depth_control |= S_028800_STENCILFUNC_BF(vkds->back.compareOp);
2739 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds->back.failOp));
2740 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds->back.passOp));
2741 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds->back.depthFailOp));
2742 }
2743
2744 if (attachment && extra) {
2745 db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(extra->db_depth_clear);
2746 db_render_control |= S_028000_STENCIL_CLEAR_ENABLE(extra->db_stencil_clear);
2747
2748 db_render_control |= S_028000_RESUMMARIZE_ENABLE(extra->db_resummarize);
2749 db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(extra->db_flush_depth_inplace);
2750 db_render_control |= S_028000_STENCIL_COMPRESS_DISABLE(extra->db_flush_stencil_inplace);
2751 db_render_override2 |= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra->db_depth_disable_expclear);
2752 db_render_override2 |= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra->db_stencil_disable_expclear);
2753 }
2754
2755 db_render_override |= S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
2756 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE);
2757
2758 if (!pCreateInfo->pRasterizationState->depthClampEnable &&
2759 ps->info.info.ps.writes_z) {
2760 /* From VK_EXT_depth_range_unrestricted spec:
2761 *
2762 * "The behavior described in Primitive Clipping still applies.
2763 * If depth clamping is disabled the depth values are still
2764 * clipped to 0 ≤ zc ≤ wc before the viewport transform. If
2765 * depth clamping is enabled the above equation is ignored and
2766 * the depth values are instead clamped to the VkViewport
2767 * minDepth and maxDepth values, which in the case of this
2768 * extension can be outside of the 0.0 to 1.0 range."
2769 */
2770 db_render_override |= S_02800C_DISABLE_VIEWPORT_CLAMP(1);
2771 }
2772
2773 radeon_set_context_reg(ctx_cs, R_028800_DB_DEPTH_CONTROL, db_depth_control);
2774 radeon_set_context_reg(ctx_cs, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
2775
2776 radeon_set_context_reg(ctx_cs, R_028000_DB_RENDER_CONTROL, db_render_control);
2777 radeon_set_context_reg(ctx_cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
2778 radeon_set_context_reg(ctx_cs, R_028010_DB_RENDER_OVERRIDE2, db_render_override2);
2779 }
2780
2781 static void
2782 radv_pipeline_generate_blend_state(struct radeon_cmdbuf *ctx_cs,
2783 struct radv_pipeline *pipeline,
2784 const struct radv_blend_state *blend)
2785 {
2786 radeon_set_context_reg_seq(ctx_cs, R_028780_CB_BLEND0_CONTROL, 8);
2787 radeon_emit_array(ctx_cs, blend->cb_blend_control,
2788 8);
2789 radeon_set_context_reg(ctx_cs, R_028808_CB_COLOR_CONTROL, blend->cb_color_control);
2790 radeon_set_context_reg(ctx_cs, R_028B70_DB_ALPHA_TO_MASK, blend->db_alpha_to_mask);
2791
2792 if (pipeline->device->physical_device->has_rbplus) {
2793
2794 radeon_set_context_reg_seq(ctx_cs, R_028760_SX_MRT0_BLEND_OPT, 8);
2795 radeon_emit_array(ctx_cs, blend->sx_mrt_blend_opt, 8);
2796 }
2797
2798 radeon_set_context_reg(ctx_cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
2799
2800 radeon_set_context_reg(ctx_cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
2801 radeon_set_context_reg(ctx_cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
2802
2803 pipeline->graphics.col_format = blend->spi_shader_col_format;
2804 pipeline->graphics.cb_target_mask = blend->cb_target_mask;
2805 }
2806
2807 static const VkConservativeRasterizationModeEXT
2808 radv_get_conservative_raster_mode(const VkPipelineRasterizationStateCreateInfo *pCreateInfo)
2809 {
2810 const VkPipelineRasterizationConservativeStateCreateInfoEXT *conservative_raster =
2811 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_RASTERIZATION_CONSERVATIVE_STATE_CREATE_INFO_EXT);
2812
2813 if (!conservative_raster)
2814 return VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT;
2815 return conservative_raster->conservativeRasterizationMode;
2816 }
2817
2818 static void
2819 radv_pipeline_generate_raster_state(struct radeon_cmdbuf *ctx_cs,
2820 struct radv_pipeline *pipeline,
2821 const VkGraphicsPipelineCreateInfo *pCreateInfo)
2822 {
2823 const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState;
2824 const VkConservativeRasterizationModeEXT mode =
2825 radv_get_conservative_raster_mode(vkraster);
2826 uint32_t pa_sc_conservative_rast = S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1);
2827 bool depth_clip_disable = vkraster->depthClampEnable;
2828
2829 const VkPipelineRasterizationDepthClipStateCreateInfoEXT *depth_clip_state =
2830 vk_find_struct_const(vkraster->pNext, PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT);
2831 if (depth_clip_state) {
2832 depth_clip_disable = !depth_clip_state->depthClipEnable;
2833 }
2834
2835 radeon_set_context_reg(ctx_cs, R_028810_PA_CL_CLIP_CNTL,
2836 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
2837 S_028810_ZCLIP_NEAR_DISABLE(depth_clip_disable ? 1 : 0) |
2838 S_028810_ZCLIP_FAR_DISABLE(depth_clip_disable ? 1 : 0) |
2839 S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) |
2840 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
2841
2842 radeon_set_context_reg(ctx_cs, R_0286D4_SPI_INTERP_CONTROL_0,
2843 S_0286D4_FLAT_SHADE_ENA(1) |
2844 S_0286D4_PNT_SPRITE_ENA(1) |
2845 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
2846 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
2847 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
2848 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
2849 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
2850
2851 radeon_set_context_reg(ctx_cs, R_028BE4_PA_SU_VTX_CNTL,
2852 S_028BE4_PIX_CENTER(1) | // TODO verify
2853 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) |
2854 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
2855
2856 radeon_set_context_reg(ctx_cs, R_028814_PA_SU_SC_MODE_CNTL,
2857 S_028814_FACE(vkraster->frontFace) |
2858 S_028814_CULL_FRONT(!!(vkraster->cullMode & VK_CULL_MODE_FRONT_BIT)) |
2859 S_028814_CULL_BACK(!!(vkraster->cullMode & VK_CULL_MODE_BACK_BIT)) |
2860 S_028814_POLY_MODE(vkraster->polygonMode != VK_POLYGON_MODE_FILL) |
2861 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster->polygonMode)) |
2862 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster->polygonMode)) |
2863 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
2864 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
2865 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster->depthBiasEnable ? 1 : 0));
2866
2867 /* Conservative rasterization. */
2868 if (mode != VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
2869 struct radv_multisample_state *ms = &pipeline->graphics.ms;
2870
2871 ms->pa_sc_aa_config |= S_028BE0_AA_MASK_CENTROID_DTMN(1);
2872 ms->db_eqaa |= S_028804_ENABLE_POSTZ_OVERRASTERIZATION(1) |
2873 S_028804_OVERRASTERIZATION_AMOUNT(4);
2874
2875 pa_sc_conservative_rast = S_028C4C_PREZ_AA_MASK_ENABLE(1) |
2876 S_028C4C_POSTZ_AA_MASK_ENABLE(1) |
2877 S_028C4C_CENTROID_SAMPLE_OVERRIDE(1);
2878
2879 if (mode == VK_CONSERVATIVE_RASTERIZATION_MODE_OVERESTIMATE_EXT) {
2880 pa_sc_conservative_rast |=
2881 S_028C4C_OVER_RAST_ENABLE(1) |
2882 S_028C4C_OVER_RAST_SAMPLE_SELECT(0) |
2883 S_028C4C_UNDER_RAST_ENABLE(0) |
2884 S_028C4C_UNDER_RAST_SAMPLE_SELECT(1) |
2885 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(1);
2886 } else {
2887 assert(mode == VK_CONSERVATIVE_RASTERIZATION_MODE_UNDERESTIMATE_EXT);
2888 pa_sc_conservative_rast |=
2889 S_028C4C_OVER_RAST_ENABLE(0) |
2890 S_028C4C_OVER_RAST_SAMPLE_SELECT(1) |
2891 S_028C4C_UNDER_RAST_ENABLE(1) |
2892 S_028C4C_UNDER_RAST_SAMPLE_SELECT(0) |
2893 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(0);
2894 }
2895 }
2896
2897 radeon_set_context_reg(ctx_cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
2898 pa_sc_conservative_rast);
2899 }
2900
2901
2902 static void
2903 radv_pipeline_generate_multisample_state(struct radeon_cmdbuf *ctx_cs,
2904 struct radv_pipeline *pipeline)
2905 {
2906 struct radv_multisample_state *ms = &pipeline->graphics.ms;
2907
2908 radeon_set_context_reg_seq(ctx_cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
2909 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[0]);
2910 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[1]);
2911
2912 radeon_set_context_reg(ctx_cs, R_028804_DB_EQAA, ms->db_eqaa);
2913 radeon_set_context_reg(ctx_cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
2914
2915 /* The exclusion bits can be set to improve rasterization efficiency
2916 * if no sample lies on the pixel boundary (-8 sample offset). It's
2917 * currently always TRUE because the driver doesn't support 16 samples.
2918 */
2919 bool exclusion = pipeline->device->physical_device->rad_info.chip_class >= CIK;
2920 radeon_set_context_reg(ctx_cs, R_02882C_PA_SU_PRIM_FILTER_CNTL,
2921 S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
2922 S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
2923 }
2924
2925 static void
2926 radv_pipeline_generate_vgt_gs_mode(struct radeon_cmdbuf *ctx_cs,
2927 struct radv_pipeline *pipeline)
2928 {
2929 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
2930
2931 uint32_t vgt_primitiveid_en = false;
2932 uint32_t vgt_gs_mode = 0;
2933
2934 if (radv_pipeline_has_gs(pipeline)) {
2935 const struct radv_shader_variant *gs =
2936 pipeline->shaders[MESA_SHADER_GEOMETRY];
2937
2938 vgt_gs_mode = ac_vgt_gs_mode(gs->info.gs.vertices_out,
2939 pipeline->device->physical_device->rad_info.chip_class);
2940 } else if (outinfo->export_prim_id) {
2941 vgt_gs_mode = S_028A40_MODE(V_028A40_GS_SCENARIO_A);
2942 vgt_primitiveid_en = true;
2943 }
2944
2945 radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN, vgt_primitiveid_en);
2946 radeon_set_context_reg(ctx_cs, R_028A40_VGT_GS_MODE, vgt_gs_mode);
2947 }
2948
2949 static void
2950 radv_pipeline_generate_hw_vs(struct radeon_cmdbuf *ctx_cs,
2951 struct radeon_cmdbuf *cs,
2952 struct radv_pipeline *pipeline,
2953 struct radv_shader_variant *shader)
2954 {
2955 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
2956
2957 radeon_set_sh_reg_seq(cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
2958 radeon_emit(cs, va >> 8);
2959 radeon_emit(cs, S_00B124_MEM_BASE(va >> 40));
2960 radeon_emit(cs, shader->rsrc1);
2961 radeon_emit(cs, shader->rsrc2);
2962
2963 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
2964 unsigned clip_dist_mask, cull_dist_mask, total_mask;
2965 clip_dist_mask = outinfo->clip_dist_mask;
2966 cull_dist_mask = outinfo->cull_dist_mask;
2967 total_mask = clip_dist_mask | cull_dist_mask;
2968 bool misc_vec_ena = outinfo->writes_pointsize ||
2969 outinfo->writes_layer ||
2970 outinfo->writes_viewport_index;
2971
2972 radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG,
2973 S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo->param_exports) - 1));
2974
2975 radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
2976 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
2977 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
2978 V_02870C_SPI_SHADER_4COMP :
2979 V_02870C_SPI_SHADER_NONE) |
2980 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
2981 V_02870C_SPI_SHADER_4COMP :
2982 V_02870C_SPI_SHADER_NONE) |
2983 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
2984 V_02870C_SPI_SHADER_4COMP :
2985 V_02870C_SPI_SHADER_NONE));
2986
2987 radeon_set_context_reg(ctx_cs, R_028818_PA_CL_VTE_CNTL,
2988 S_028818_VTX_W0_FMT(1) |
2989 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
2990 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
2991 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
2992
2993 radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
2994 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
2995 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
2996 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
2997 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
2998 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
2999 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
3000 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
3001 cull_dist_mask << 8 |
3002 clip_dist_mask);
3003
3004 if (pipeline->device->physical_device->rad_info.chip_class <= VI)
3005 radeon_set_context_reg(ctx_cs, R_028AB4_VGT_REUSE_OFF,
3006 outinfo->writes_viewport_index);
3007 }
3008
3009 static void
3010 radv_pipeline_generate_hw_es(struct radeon_cmdbuf *cs,
3011 struct radv_pipeline *pipeline,
3012 struct radv_shader_variant *shader)
3013 {
3014 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3015
3016 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 4);
3017 radeon_emit(cs, va >> 8);
3018 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
3019 radeon_emit(cs, shader->rsrc1);
3020 radeon_emit(cs, shader->rsrc2);
3021 }
3022
3023 static void
3024 radv_pipeline_generate_hw_ls(struct radeon_cmdbuf *cs,
3025 struct radv_pipeline *pipeline,
3026 struct radv_shader_variant *shader,
3027 const struct radv_tessellation_state *tess)
3028 {
3029 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3030 uint32_t rsrc2 = shader->rsrc2;
3031
3032 radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
3033 radeon_emit(cs, va >> 8);
3034 radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
3035
3036 rsrc2 |= S_00B52C_LDS_SIZE(tess->lds_size);
3037 if (pipeline->device->physical_device->rad_info.chip_class == CIK &&
3038 pipeline->device->physical_device->rad_info.family != CHIP_HAWAII)
3039 radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, rsrc2);
3040
3041 radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
3042 radeon_emit(cs, shader->rsrc1);
3043 radeon_emit(cs, rsrc2);
3044 }
3045
3046 static void
3047 radv_pipeline_generate_hw_hs(struct radeon_cmdbuf *cs,
3048 struct radv_pipeline *pipeline,
3049 struct radv_shader_variant *shader,
3050 const struct radv_tessellation_state *tess)
3051 {
3052 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3053
3054 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
3055 radeon_set_sh_reg_seq(cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
3056 radeon_emit(cs, va >> 8);
3057 radeon_emit(cs, S_00B414_MEM_BASE(va >> 40));
3058
3059 radeon_set_sh_reg_seq(cs, R_00B428_SPI_SHADER_PGM_RSRC1_HS, 2);
3060 radeon_emit(cs, shader->rsrc1);
3061 radeon_emit(cs, shader->rsrc2 |
3062 S_00B42C_LDS_SIZE(tess->lds_size));
3063 } else {
3064 radeon_set_sh_reg_seq(cs, R_00B420_SPI_SHADER_PGM_LO_HS, 4);
3065 radeon_emit(cs, va >> 8);
3066 radeon_emit(cs, S_00B424_MEM_BASE(va >> 40));
3067 radeon_emit(cs, shader->rsrc1);
3068 radeon_emit(cs, shader->rsrc2);
3069 }
3070 }
3071
3072 static void
3073 radv_pipeline_generate_vertex_shader(struct radeon_cmdbuf *ctx_cs,
3074 struct radeon_cmdbuf *cs,
3075 struct radv_pipeline *pipeline,
3076 const struct radv_tessellation_state *tess)
3077 {
3078 struct radv_shader_variant *vs;
3079
3080 /* Skip shaders merged into HS/GS */
3081 vs = pipeline->shaders[MESA_SHADER_VERTEX];
3082 if (!vs)
3083 return;
3084
3085 if (vs->info.vs.as_ls)
3086 radv_pipeline_generate_hw_ls(cs, pipeline, vs, tess);
3087 else if (vs->info.vs.as_es)
3088 radv_pipeline_generate_hw_es(cs, pipeline, vs);
3089 else
3090 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, vs);
3091 }
3092
3093 static void
3094 radv_pipeline_generate_tess_shaders(struct radeon_cmdbuf *ctx_cs,
3095 struct radeon_cmdbuf *cs,
3096 struct radv_pipeline *pipeline,
3097 const struct radv_tessellation_state *tess)
3098 {
3099 if (!radv_pipeline_has_tess(pipeline))
3100 return;
3101
3102 struct radv_shader_variant *tes, *tcs;
3103
3104 tcs = pipeline->shaders[MESA_SHADER_TESS_CTRL];
3105 tes = pipeline->shaders[MESA_SHADER_TESS_EVAL];
3106
3107 if (tes) {
3108 if (tes->info.tes.as_es)
3109 radv_pipeline_generate_hw_es(cs, pipeline, tes);
3110 else
3111 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, tes);
3112 }
3113
3114 radv_pipeline_generate_hw_hs(cs, pipeline, tcs, tess);
3115
3116 radeon_set_context_reg(ctx_cs, R_028B6C_VGT_TF_PARAM,
3117 tess->tf_param);
3118
3119 if (pipeline->device->physical_device->rad_info.chip_class >= CIK)
3120 radeon_set_context_reg_idx(ctx_cs, R_028B58_VGT_LS_HS_CONFIG, 2,
3121 tess->ls_hs_config);
3122 else
3123 radeon_set_context_reg(ctx_cs, R_028B58_VGT_LS_HS_CONFIG,
3124 tess->ls_hs_config);
3125 }
3126
3127 static void
3128 radv_pipeline_generate_geometry_shader(struct radeon_cmdbuf *ctx_cs,
3129 struct radeon_cmdbuf *cs,
3130 struct radv_pipeline *pipeline,
3131 const struct radv_gs_state *gs_state)
3132 {
3133 struct radv_shader_variant *gs;
3134 unsigned gs_max_out_vertices;
3135 uint8_t *num_components;
3136 uint8_t max_stream;
3137 unsigned offset;
3138 uint64_t va;
3139
3140 gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
3141 if (!gs)
3142 return;
3143
3144 gs_max_out_vertices = gs->info.gs.vertices_out;
3145 max_stream = gs->info.info.gs.max_stream;
3146 num_components = gs->info.info.gs.num_stream_output_components;
3147
3148 offset = num_components[0] * gs_max_out_vertices;
3149
3150 radeon_set_context_reg_seq(ctx_cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
3151 radeon_emit(ctx_cs, offset);
3152 if (max_stream >= 1)
3153 offset += num_components[1] * gs_max_out_vertices;
3154 radeon_emit(ctx_cs, offset);
3155 if (max_stream >= 2)
3156 offset += num_components[2] * gs_max_out_vertices;
3157 radeon_emit(ctx_cs, offset);
3158 if (max_stream >= 3)
3159 offset += num_components[3] * gs_max_out_vertices;
3160 radeon_set_context_reg(ctx_cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
3161
3162 radeon_set_context_reg(ctx_cs, R_028B38_VGT_GS_MAX_VERT_OUT, gs->info.gs.vertices_out);
3163
3164 radeon_set_context_reg_seq(ctx_cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
3165 radeon_emit(ctx_cs, num_components[0]);
3166 radeon_emit(ctx_cs, (max_stream >= 1) ? num_components[1] : 0);
3167 radeon_emit(ctx_cs, (max_stream >= 2) ? num_components[2] : 0);
3168 radeon_emit(ctx_cs, (max_stream >= 3) ? num_components[3] : 0);
3169
3170 uint32_t gs_num_invocations = gs->info.gs.invocations;
3171 radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
3172 S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
3173 S_028B90_ENABLE(gs_num_invocations > 0));
3174
3175 radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
3176 gs_state->vgt_esgs_ring_itemsize);
3177
3178 va = radv_buffer_get_va(gs->bo) + gs->bo_offset;
3179
3180 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
3181 radeon_set_sh_reg_seq(cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
3182 radeon_emit(cs, va >> 8);
3183 radeon_emit(cs, S_00B214_MEM_BASE(va >> 40));
3184
3185 radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
3186 radeon_emit(cs, gs->rsrc1);
3187 radeon_emit(cs, gs->rsrc2 | S_00B22C_LDS_SIZE(gs_state->lds_size));
3188
3189 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL, gs_state->vgt_gs_onchip_cntl);
3190 radeon_set_context_reg(ctx_cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, gs_state->vgt_gs_max_prims_per_subgroup);
3191 } else {
3192 radeon_set_sh_reg_seq(cs, R_00B220_SPI_SHADER_PGM_LO_GS, 4);
3193 radeon_emit(cs, va >> 8);
3194 radeon_emit(cs, S_00B224_MEM_BASE(va >> 40));
3195 radeon_emit(cs, gs->rsrc1);
3196 radeon_emit(cs, gs->rsrc2);
3197 }
3198
3199 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, pipeline->gs_copy_shader);
3200 }
3201
3202 static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade, bool float16)
3203 {
3204 uint32_t ps_input_cntl;
3205 if (offset <= AC_EXP_PARAM_OFFSET_31) {
3206 ps_input_cntl = S_028644_OFFSET(offset);
3207 if (flat_shade)
3208 ps_input_cntl |= S_028644_FLAT_SHADE(1);
3209 if (float16) {
3210 ps_input_cntl |= S_028644_FP16_INTERP_MODE(1) |
3211 S_028644_ATTR0_VALID(1);
3212 }
3213 } else {
3214 /* The input is a DEFAULT_VAL constant. */
3215 assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
3216 offset <= AC_EXP_PARAM_DEFAULT_VAL_1111);
3217 offset -= AC_EXP_PARAM_DEFAULT_VAL_0000;
3218 ps_input_cntl = S_028644_OFFSET(0x20) |
3219 S_028644_DEFAULT_VAL(offset);
3220 }
3221 return ps_input_cntl;
3222 }
3223
3224 static void
3225 radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
3226 struct radv_pipeline *pipeline)
3227 {
3228 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3229 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3230 uint32_t ps_input_cntl[32];
3231
3232 unsigned ps_offset = 0;
3233
3234 if (ps->info.info.ps.prim_id_input) {
3235 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
3236 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
3237 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
3238 ++ps_offset;
3239 }
3240 }
3241
3242 if (ps->info.info.ps.layer_input ||
3243 ps->info.info.ps.uses_input_attachments ||
3244 ps->info.info.needs_multiview_view_index) {
3245 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
3246 if (vs_offset != AC_EXP_PARAM_UNDEFINED)
3247 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
3248 else
3249 ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false);
3250 ++ps_offset;
3251 }
3252
3253 if (ps->info.info.ps.has_pcoord) {
3254 unsigned val;
3255 val = S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
3256 ps_input_cntl[ps_offset] = val;
3257 ps_offset++;
3258 }
3259
3260 if (ps->info.info.ps.num_input_clips_culls) {
3261 unsigned vs_offset;
3262
3263 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0];
3264 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
3265 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
3266 ++ps_offset;
3267 }
3268
3269 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1];
3270 if (vs_offset != AC_EXP_PARAM_UNDEFINED &&
3271 ps->info.info.ps.num_input_clips_culls > 4) {
3272 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
3273 ++ps_offset;
3274 }
3275 }
3276
3277 for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.fs.input_mask; ++i) {
3278 unsigned vs_offset;
3279 bool flat_shade;
3280 bool float16;
3281 if (!(ps->info.fs.input_mask & (1u << i)))
3282 continue;
3283
3284 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VAR0 + i];
3285 if (vs_offset == AC_EXP_PARAM_UNDEFINED) {
3286 ps_input_cntl[ps_offset] = S_028644_OFFSET(0x20);
3287 ++ps_offset;
3288 continue;
3289 }
3290
3291 flat_shade = !!(ps->info.fs.flat_shaded_mask & (1u << ps_offset));
3292 float16 = !!(ps->info.fs.float16_shaded_mask & (1u << ps_offset));
3293
3294 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade, float16);
3295 ++ps_offset;
3296 }
3297
3298 if (ps_offset) {
3299 radeon_set_context_reg_seq(ctx_cs, R_028644_SPI_PS_INPUT_CNTL_0, ps_offset);
3300 for (unsigned i = 0; i < ps_offset; i++) {
3301 radeon_emit(ctx_cs, ps_input_cntl[i]);
3302 }
3303 }
3304 }
3305
3306 static uint32_t
3307 radv_compute_db_shader_control(const struct radv_device *device,
3308 const struct radv_pipeline *pipeline,
3309 const struct radv_shader_variant *ps)
3310 {
3311 unsigned z_order;
3312 if (ps->info.fs.early_fragment_test || !ps->info.info.ps.writes_memory)
3313 z_order = V_02880C_EARLY_Z_THEN_LATE_Z;
3314 else
3315 z_order = V_02880C_LATE_Z;
3316
3317 bool disable_rbplus = device->physical_device->has_rbplus &&
3318 !device->physical_device->rbplus_allowed;
3319
3320 /* It shouldn't be needed to export gl_SampleMask when MSAA is disabled
3321 * but this appears to break Project Cars (DXVK). See
3322 * https://bugs.freedesktop.org/show_bug.cgi?id=109401
3323 */
3324 bool mask_export_enable = ps->info.info.ps.writes_sample_mask;
3325
3326 return S_02880C_Z_EXPORT_ENABLE(ps->info.info.ps.writes_z) |
3327 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.info.ps.writes_stencil) |
3328 S_02880C_KILL_ENABLE(!!ps->info.fs.can_discard) |
3329 S_02880C_MASK_EXPORT_ENABLE(mask_export_enable) |
3330 S_02880C_Z_ORDER(z_order) |
3331 S_02880C_DEPTH_BEFORE_SHADER(ps->info.fs.early_fragment_test) |
3332 S_02880C_EXEC_ON_HIER_FAIL(ps->info.info.ps.writes_memory) |
3333 S_02880C_EXEC_ON_NOOP(ps->info.info.ps.writes_memory) |
3334 S_02880C_DUAL_QUAD_DISABLE(disable_rbplus);
3335 }
3336
3337 static void
3338 radv_pipeline_generate_fragment_shader(struct radeon_cmdbuf *ctx_cs,
3339 struct radeon_cmdbuf *cs,
3340 struct radv_pipeline *pipeline)
3341 {
3342 struct radv_shader_variant *ps;
3343 uint64_t va;
3344 assert (pipeline->shaders[MESA_SHADER_FRAGMENT]);
3345
3346 ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3347 va = radv_buffer_get_va(ps->bo) + ps->bo_offset;
3348
3349 radeon_set_sh_reg_seq(cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
3350 radeon_emit(cs, va >> 8);
3351 radeon_emit(cs, S_00B024_MEM_BASE(va >> 40));
3352 radeon_emit(cs, ps->rsrc1);
3353 radeon_emit(cs, ps->rsrc2);
3354
3355 radeon_set_context_reg(ctx_cs, R_02880C_DB_SHADER_CONTROL,
3356 radv_compute_db_shader_control(pipeline->device,
3357 pipeline, ps));
3358
3359 radeon_set_context_reg(ctx_cs, R_0286CC_SPI_PS_INPUT_ENA,
3360 ps->config.spi_ps_input_ena);
3361
3362 radeon_set_context_reg(ctx_cs, R_0286D0_SPI_PS_INPUT_ADDR,
3363 ps->config.spi_ps_input_addr);
3364
3365 radeon_set_context_reg(ctx_cs, R_0286D8_SPI_PS_IN_CONTROL,
3366 S_0286D8_NUM_INTERP(ps->info.fs.num_interp));
3367
3368 radeon_set_context_reg(ctx_cs, R_0286E0_SPI_BARYC_CNTL, pipeline->graphics.spi_baryc_cntl);
3369
3370 radeon_set_context_reg(ctx_cs, R_028710_SPI_SHADER_Z_FORMAT,
3371 ac_get_spi_shader_z_format(ps->info.info.ps.writes_z,
3372 ps->info.info.ps.writes_stencil,
3373 ps->info.info.ps.writes_sample_mask));
3374
3375 if (pipeline->device->dfsm_allowed) {
3376 /* optimise this? */
3377 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3378 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3379 }
3380 }
3381
3382 static void
3383 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_cmdbuf *ctx_cs,
3384 struct radv_pipeline *pipeline)
3385 {
3386 if (pipeline->device->physical_device->rad_info.family < CHIP_POLARIS10)
3387 return;
3388
3389 unsigned vtx_reuse_depth = 30;
3390 if (radv_pipeline_has_tess(pipeline) &&
3391 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.tes.spacing == TESS_SPACING_FRACTIONAL_ODD) {
3392 vtx_reuse_depth = 14;
3393 }
3394 radeon_set_context_reg(ctx_cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
3395 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth));
3396 }
3397
3398 static uint32_t
3399 radv_compute_vgt_shader_stages_en(const struct radv_pipeline *pipeline)
3400 {
3401 uint32_t stages = 0;
3402 if (radv_pipeline_has_tess(pipeline)) {
3403 stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
3404 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3405
3406 if (radv_pipeline_has_gs(pipeline))
3407 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
3408 S_028B54_GS_EN(1) |
3409 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
3410 else
3411 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
3412
3413 } else if (radv_pipeline_has_gs(pipeline))
3414 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
3415 S_028B54_GS_EN(1) |
3416 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
3417
3418 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
3419 stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3420
3421 return stages;
3422 }
3423
3424 static uint32_t
3425 radv_compute_cliprect_rule(const VkGraphicsPipelineCreateInfo *pCreateInfo)
3426 {
3427 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
3428 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
3429
3430 if (!discard_rectangle_info)
3431 return 0xffff;
3432
3433 unsigned mask = 0;
3434
3435 for (unsigned i = 0; i < (1u << MAX_DISCARD_RECTANGLES); ++i) {
3436 /* Interpret i as a bitmask, and then set the bit in the mask if
3437 * that combination of rectangles in which the pixel is contained
3438 * should pass the cliprect test. */
3439 unsigned relevant_subset = i & ((1u << discard_rectangle_info->discardRectangleCount) - 1);
3440
3441 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT &&
3442 !relevant_subset)
3443 continue;
3444
3445 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT &&
3446 relevant_subset)
3447 continue;
3448
3449 mask |= 1u << i;
3450 }
3451
3452 return mask;
3453 }
3454
3455 static void
3456 radv_pipeline_generate_pm4(struct radv_pipeline *pipeline,
3457 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3458 const struct radv_graphics_pipeline_create_info *extra,
3459 const struct radv_blend_state *blend,
3460 const struct radv_tessellation_state *tess,
3461 const struct radv_gs_state *gs,
3462 unsigned prim, unsigned gs_out)
3463 {
3464 struct radeon_cmdbuf *ctx_cs = &pipeline->ctx_cs;
3465 struct radeon_cmdbuf *cs = &pipeline->cs;
3466
3467 cs->max_dw = 64;
3468 ctx_cs->max_dw = 256;
3469 cs->buf = malloc(4 * (cs->max_dw + ctx_cs->max_dw));
3470 ctx_cs->buf = cs->buf + cs->max_dw;
3471
3472 radv_pipeline_generate_depth_stencil_state(ctx_cs, pipeline, pCreateInfo, extra);
3473 radv_pipeline_generate_blend_state(ctx_cs, pipeline, blend);
3474 radv_pipeline_generate_raster_state(ctx_cs, pipeline, pCreateInfo);
3475 radv_pipeline_generate_multisample_state(ctx_cs, pipeline);
3476 radv_pipeline_generate_vgt_gs_mode(ctx_cs, pipeline);
3477 radv_pipeline_generate_vertex_shader(ctx_cs, cs, pipeline, tess);
3478 radv_pipeline_generate_tess_shaders(ctx_cs, cs, pipeline, tess);
3479 radv_pipeline_generate_geometry_shader(ctx_cs, cs, pipeline, gs);
3480 radv_pipeline_generate_fragment_shader(ctx_cs, cs, pipeline);
3481 radv_pipeline_generate_ps_inputs(ctx_cs, pipeline);
3482 radv_pipeline_generate_vgt_vertex_reuse(ctx_cs, pipeline);
3483 radv_pipeline_generate_binning_state(ctx_cs, pipeline, pCreateInfo);
3484
3485 radeon_set_context_reg(ctx_cs, R_0286E8_SPI_TMPRING_SIZE,
3486 S_0286E8_WAVES(pipeline->max_waves) |
3487 S_0286E8_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
3488
3489 radeon_set_context_reg(ctx_cs, R_028B54_VGT_SHADER_STAGES_EN, radv_compute_vgt_shader_stages_en(pipeline));
3490
3491 if (pipeline->device->physical_device->rad_info.chip_class >= CIK) {
3492 radeon_set_uconfig_reg_idx(cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
3493 } else {
3494 radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
3495 }
3496 radeon_set_context_reg(ctx_cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out);
3497
3498 radeon_set_context_reg(ctx_cs, R_02820C_PA_SC_CLIPRECT_RULE, radv_compute_cliprect_rule(pCreateInfo));
3499
3500 pipeline->ctx_cs_hash = _mesa_hash_data(ctx_cs->buf, ctx_cs->cdw * 4);
3501
3502 assert(ctx_cs->cdw <= ctx_cs->max_dw);
3503 assert(cs->cdw <= cs->max_dw);
3504 }
3505
3506 static struct radv_ia_multi_vgt_param_helpers
3507 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline *pipeline,
3508 const struct radv_tessellation_state *tess,
3509 uint32_t prim)
3510 {
3511 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param = {0};
3512 const struct radv_device *device = pipeline->device;
3513
3514 if (radv_pipeline_has_tess(pipeline))
3515 ia_multi_vgt_param.primgroup_size = tess->num_patches;
3516 else if (radv_pipeline_has_gs(pipeline))
3517 ia_multi_vgt_param.primgroup_size = 64;
3518 else
3519 ia_multi_vgt_param.primgroup_size = 128; /* recommended without a GS */
3520
3521 /* GS requirement. */
3522 ia_multi_vgt_param.partial_es_wave = false;
3523 if (radv_pipeline_has_gs(pipeline) && device->physical_device->rad_info.chip_class <= VI)
3524 if (SI_GS_PER_ES / ia_multi_vgt_param.primgroup_size >= pipeline->device->gs_table_depth - 3)
3525 ia_multi_vgt_param.partial_es_wave = true;
3526
3527 ia_multi_vgt_param.wd_switch_on_eop = false;
3528 if (device->physical_device->rad_info.chip_class >= CIK) {
3529 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
3530 * 4 shader engines. Set 1 to pass the assertion below.
3531 * The other cases are hardware requirements. */
3532 if (device->physical_device->rad_info.max_se < 4 ||
3533 prim == V_008958_DI_PT_POLYGON ||
3534 prim == V_008958_DI_PT_LINELOOP ||
3535 prim == V_008958_DI_PT_TRIFAN ||
3536 prim == V_008958_DI_PT_TRISTRIP_ADJ ||
3537 (pipeline->graphics.prim_restart_enable &&
3538 (device->physical_device->rad_info.family < CHIP_POLARIS10 ||
3539 (prim != V_008958_DI_PT_POINTLIST &&
3540 prim != V_008958_DI_PT_LINESTRIP))))
3541 ia_multi_vgt_param.wd_switch_on_eop = true;
3542 }
3543
3544 ia_multi_vgt_param.ia_switch_on_eoi = false;
3545 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input)
3546 ia_multi_vgt_param.ia_switch_on_eoi = true;
3547 if (radv_pipeline_has_gs(pipeline) &&
3548 pipeline->shaders[MESA_SHADER_GEOMETRY]->info.info.uses_prim_id)
3549 ia_multi_vgt_param.ia_switch_on_eoi = true;
3550 if (radv_pipeline_has_tess(pipeline)) {
3551 /* SWITCH_ON_EOI must be set if PrimID is used. */
3552 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.uses_prim_id ||
3553 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.info.uses_prim_id)
3554 ia_multi_vgt_param.ia_switch_on_eoi = true;
3555 }
3556
3557 ia_multi_vgt_param.partial_vs_wave = false;
3558 if (radv_pipeline_has_tess(pipeline)) {
3559 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
3560 if ((device->physical_device->rad_info.family == CHIP_TAHITI ||
3561 device->physical_device->rad_info.family == CHIP_PITCAIRN ||
3562 device->physical_device->rad_info.family == CHIP_BONAIRE) &&
3563 radv_pipeline_has_gs(pipeline))
3564 ia_multi_vgt_param.partial_vs_wave = true;
3565 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
3566 if (device->has_distributed_tess) {
3567 if (radv_pipeline_has_gs(pipeline)) {
3568 if (device->physical_device->rad_info.chip_class <= VI)
3569 ia_multi_vgt_param.partial_es_wave = true;
3570 } else {
3571 ia_multi_vgt_param.partial_vs_wave = true;
3572 }
3573 }
3574 }
3575
3576 /* Workaround for a VGT hang when strip primitive types are used with
3577 * primitive restart.
3578 */
3579 if (pipeline->graphics.prim_restart_enable &&
3580 (prim == V_008958_DI_PT_LINESTRIP ||
3581 prim == V_008958_DI_PT_TRISTRIP ||
3582 prim == V_008958_DI_PT_LINESTRIP_ADJ ||
3583 prim == V_008958_DI_PT_TRISTRIP_ADJ)) {
3584 ia_multi_vgt_param.partial_vs_wave = true;
3585 }
3586
3587 if (radv_pipeline_has_gs(pipeline)) {
3588 /* On these chips there is the possibility of a hang if the
3589 * pipeline uses a GS and partial_vs_wave is not set.
3590 *
3591 * This mostly does not hit 4-SE chips, as those typically set
3592 * ia_switch_on_eoi and then partial_vs_wave is set for pipelines
3593 * with GS due to another workaround.
3594 *
3595 * Reproducer: https://bugs.freedesktop.org/show_bug.cgi?id=109242
3596 */
3597 if (device->physical_device->rad_info.family == CHIP_TONGA ||
3598 device->physical_device->rad_info.family == CHIP_FIJI ||
3599 device->physical_device->rad_info.family == CHIP_POLARIS10 ||
3600 device->physical_device->rad_info.family == CHIP_POLARIS11 ||
3601 device->physical_device->rad_info.family == CHIP_POLARIS12 ||
3602 device->physical_device->rad_info.family == CHIP_VEGAM) {
3603 ia_multi_vgt_param.partial_vs_wave = true;
3604 }
3605 }
3606
3607 ia_multi_vgt_param.base =
3608 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param.primgroup_size - 1) |
3609 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
3610 S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == VI ? 2 : 0) |
3611 S_030960_EN_INST_OPT_BASIC(device->physical_device->rad_info.chip_class >= GFX9) |
3612 S_030960_EN_INST_OPT_ADV(device->physical_device->rad_info.chip_class >= GFX9);
3613
3614 return ia_multi_vgt_param;
3615 }
3616
3617
3618 static void
3619 radv_compute_vertex_input_state(struct radv_pipeline *pipeline,
3620 const VkGraphicsPipelineCreateInfo *pCreateInfo)
3621 {
3622 const VkPipelineVertexInputStateCreateInfo *vi_info =
3623 pCreateInfo->pVertexInputState;
3624 struct radv_vertex_elements_info *velems = &pipeline->vertex_elements;
3625
3626 for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
3627 const VkVertexInputAttributeDescription *desc =
3628 &vi_info->pVertexAttributeDescriptions[i];
3629 unsigned loc = desc->location;
3630 const struct vk_format_description *format_desc;
3631
3632 format_desc = vk_format_description(desc->format);
3633
3634 velems->format_size[loc] = format_desc->block.bits / 8;
3635 }
3636
3637 for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
3638 const VkVertexInputBindingDescription *desc =
3639 &vi_info->pVertexBindingDescriptions[i];
3640
3641 pipeline->binding_stride[desc->binding] = desc->stride;
3642 pipeline->num_vertex_bindings =
3643 MAX2(pipeline->num_vertex_bindings, desc->binding + 1);
3644 }
3645 }
3646
3647 static struct radv_shader_variant *
3648 radv_pipeline_get_streamout_shader(struct radv_pipeline *pipeline)
3649 {
3650 int i;
3651
3652 for (i = MESA_SHADER_GEOMETRY; i >= MESA_SHADER_VERTEX; i--) {
3653 struct radv_shader_variant *shader =
3654 radv_get_shader(pipeline, i);
3655
3656 if (shader && shader->info.info.so.num_outputs > 0)
3657 return shader;
3658 }
3659
3660 return NULL;
3661 }
3662
3663 static VkResult
3664 radv_pipeline_init(struct radv_pipeline *pipeline,
3665 struct radv_device *device,
3666 struct radv_pipeline_cache *cache,
3667 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3668 const struct radv_graphics_pipeline_create_info *extra)
3669 {
3670 VkResult result;
3671 bool has_view_index = false;
3672
3673 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3674 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3675 if (subpass->view_mask)
3676 has_view_index = true;
3677
3678 pipeline->device = device;
3679 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
3680 assert(pipeline->layout);
3681
3682 struct radv_blend_state blend = radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra);
3683
3684 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
3685 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
3686 radv_init_feedback(creation_feedback);
3687
3688 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
3689
3690 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
3691 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
3692 for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
3693 gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
3694 pStages[stage] = &pCreateInfo->pStages[i];
3695 if(creation_feedback)
3696 stage_feedbacks[stage] = &creation_feedback->pPipelineStageCreationFeedbacks[i];
3697 }
3698
3699 struct radv_pipeline_key key = radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend, has_view_index);
3700 radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
3701
3702 pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
3703 radv_pipeline_init_multisample_state(pipeline, &blend, pCreateInfo);
3704 uint32_t gs_out;
3705 uint32_t prim = si_translate_prim(pCreateInfo->pInputAssemblyState->topology);
3706
3707 pipeline->graphics.can_use_guardband = radv_prim_can_use_guardband(pCreateInfo->pInputAssemblyState->topology);
3708
3709 if (radv_pipeline_has_gs(pipeline)) {
3710 gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.output_prim);
3711 pipeline->graphics.can_use_guardband = gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP;
3712 } else {
3713 gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology);
3714 }
3715 if (extra && extra->use_rectlist) {
3716 prim = V_008958_DI_PT_RECTLIST;
3717 gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP;
3718 pipeline->graphics.can_use_guardband = true;
3719 }
3720 pipeline->graphics.prim_restart_enable = !!pCreateInfo->pInputAssemblyState->primitiveRestartEnable;
3721 /* prim vertex count will need TESS changes */
3722 pipeline->graphics.prim_vertex_count = prim_size_table[prim];
3723
3724 radv_pipeline_init_dynamic_state(pipeline, pCreateInfo);
3725
3726 /* Ensure that some export memory is always allocated, for two reasons:
3727 *
3728 * 1) Correctness: The hardware ignores the EXEC mask if no export
3729 * memory is allocated, so KILL and alpha test do not work correctly
3730 * without this.
3731 * 2) Performance: Every shader needs at least a NULL export, even when
3732 * it writes no color/depth output. The NULL export instruction
3733 * stalls without this setting.
3734 *
3735 * Don't add this to CB_SHADER_MASK.
3736 */
3737 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3738 if (!blend.spi_shader_col_format) {
3739 if (!ps->info.info.ps.writes_z &&
3740 !ps->info.info.ps.writes_stencil &&
3741 !ps->info.info.ps.writes_sample_mask)
3742 blend.spi_shader_col_format = V_028714_SPI_SHADER_32_R;
3743 }
3744
3745 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
3746 if (pipeline->shaders[i]) {
3747 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[i]->info.need_indirect_descriptor_sets;
3748 }
3749 }
3750
3751 struct radv_gs_state gs = {0};
3752 if (radv_pipeline_has_gs(pipeline)) {
3753 gs = calculate_gs_info(pCreateInfo, pipeline);
3754 calculate_gs_ring_sizes(pipeline, &gs);
3755 }
3756
3757 struct radv_tessellation_state tess = {0};
3758 if (radv_pipeline_has_tess(pipeline)) {
3759 if (prim == V_008958_DI_PT_PATCH) {
3760 pipeline->graphics.prim_vertex_count.min = pCreateInfo->pTessellationState->patchControlPoints;
3761 pipeline->graphics.prim_vertex_count.incr = 1;
3762 }
3763 tess = calculate_tess_state(pipeline, pCreateInfo);
3764 }
3765
3766 pipeline->graphics.ia_multi_vgt_param = radv_compute_ia_multi_vgt_param_helpers(pipeline, &tess, prim);
3767
3768 radv_compute_vertex_input_state(pipeline, pCreateInfo);
3769
3770 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++)
3771 pipeline->user_data_0[i] = radv_pipeline_stage_to_user_data_0(pipeline, i, device->physical_device->rad_info.chip_class);
3772
3773 struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_VERTEX,
3774 AC_UD_VS_BASE_VERTEX_START_INSTANCE);
3775 if (loc->sgpr_idx != -1) {
3776 pipeline->graphics.vtx_base_sgpr = pipeline->user_data_0[MESA_SHADER_VERTEX];
3777 pipeline->graphics.vtx_base_sgpr += loc->sgpr_idx * 4;
3778 if (radv_get_shader(pipeline, MESA_SHADER_VERTEX)->info.info.vs.needs_draw_id)
3779 pipeline->graphics.vtx_emit_num = 3;
3780 else
3781 pipeline->graphics.vtx_emit_num = 2;
3782 }
3783
3784 /* Find the last vertex shader stage that eventually uses streamout. */
3785 pipeline->streamout_shader = radv_pipeline_get_streamout_shader(pipeline);
3786
3787 result = radv_pipeline_scratch_init(device, pipeline);
3788 radv_pipeline_generate_pm4(pipeline, pCreateInfo, extra, &blend, &tess, &gs, prim, gs_out);
3789
3790 return result;
3791 }
3792
3793 VkResult
3794 radv_graphics_pipeline_create(
3795 VkDevice _device,
3796 VkPipelineCache _cache,
3797 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3798 const struct radv_graphics_pipeline_create_info *extra,
3799 const VkAllocationCallbacks *pAllocator,
3800 VkPipeline *pPipeline)
3801 {
3802 RADV_FROM_HANDLE(radv_device, device, _device);
3803 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
3804 struct radv_pipeline *pipeline;
3805 VkResult result;
3806
3807 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
3808 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3809 if (pipeline == NULL)
3810 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
3811
3812 result = radv_pipeline_init(pipeline, device, cache,
3813 pCreateInfo, extra);
3814 if (result != VK_SUCCESS) {
3815 radv_pipeline_destroy(device, pipeline, pAllocator);
3816 return result;
3817 }
3818
3819 *pPipeline = radv_pipeline_to_handle(pipeline);
3820
3821 return VK_SUCCESS;
3822 }
3823
3824 VkResult radv_CreateGraphicsPipelines(
3825 VkDevice _device,
3826 VkPipelineCache pipelineCache,
3827 uint32_t count,
3828 const VkGraphicsPipelineCreateInfo* pCreateInfos,
3829 const VkAllocationCallbacks* pAllocator,
3830 VkPipeline* pPipelines)
3831 {
3832 VkResult result = VK_SUCCESS;
3833 unsigned i = 0;
3834
3835 for (; i < count; i++) {
3836 VkResult r;
3837 r = radv_graphics_pipeline_create(_device,
3838 pipelineCache,
3839 &pCreateInfos[i],
3840 NULL, pAllocator, &pPipelines[i]);
3841 if (r != VK_SUCCESS) {
3842 result = r;
3843 pPipelines[i] = VK_NULL_HANDLE;
3844 }
3845 }
3846
3847 return result;
3848 }
3849
3850
3851 static void
3852 radv_compute_generate_pm4(struct radv_pipeline *pipeline)
3853 {
3854 struct radv_shader_variant *compute_shader;
3855 struct radv_device *device = pipeline->device;
3856 unsigned compute_resource_limits;
3857 unsigned waves_per_threadgroup;
3858 uint64_t va;
3859
3860 pipeline->cs.buf = malloc(20 * 4);
3861 pipeline->cs.max_dw = 20;
3862
3863 compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
3864 va = radv_buffer_get_va(compute_shader->bo) + compute_shader->bo_offset;
3865
3866 radeon_set_sh_reg_seq(&pipeline->cs, R_00B830_COMPUTE_PGM_LO, 2);
3867 radeon_emit(&pipeline->cs, va >> 8);
3868 radeon_emit(&pipeline->cs, S_00B834_DATA(va >> 40));
3869
3870 radeon_set_sh_reg_seq(&pipeline->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
3871 radeon_emit(&pipeline->cs, compute_shader->rsrc1);
3872 radeon_emit(&pipeline->cs, compute_shader->rsrc2);
3873
3874 radeon_set_sh_reg(&pipeline->cs, R_00B860_COMPUTE_TMPRING_SIZE,
3875 S_00B860_WAVES(pipeline->max_waves) |
3876 S_00B860_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
3877
3878 /* Calculate best compute resource limits. */
3879 waves_per_threadgroup =
3880 DIV_ROUND_UP(compute_shader->info.cs.block_size[0] *
3881 compute_shader->info.cs.block_size[1] *
3882 compute_shader->info.cs.block_size[2], 64);
3883 compute_resource_limits =
3884 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
3885
3886 if (device->physical_device->rad_info.chip_class >= CIK) {
3887 unsigned num_cu_per_se =
3888 device->physical_device->rad_info.num_good_compute_units /
3889 device->physical_device->rad_info.max_se;
3890
3891 /* Force even distribution on all SIMDs in CU if the workgroup
3892 * size is 64. This has shown some good improvements if # of
3893 * CUs per SE is not a multiple of 4.
3894 */
3895 if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
3896 compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
3897 }
3898
3899 radeon_set_sh_reg(&pipeline->cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
3900 compute_resource_limits);
3901
3902 radeon_set_sh_reg_seq(&pipeline->cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
3903 radeon_emit(&pipeline->cs,
3904 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[0]));
3905 radeon_emit(&pipeline->cs,
3906 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[1]));
3907 radeon_emit(&pipeline->cs,
3908 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[2]));
3909
3910 assert(pipeline->cs.cdw <= pipeline->cs.max_dw);
3911 }
3912
3913 static VkResult radv_compute_pipeline_create(
3914 VkDevice _device,
3915 VkPipelineCache _cache,
3916 const VkComputePipelineCreateInfo* pCreateInfo,
3917 const VkAllocationCallbacks* pAllocator,
3918 VkPipeline* pPipeline)
3919 {
3920 RADV_FROM_HANDLE(radv_device, device, _device);
3921 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
3922 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
3923 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
3924 struct radv_pipeline *pipeline;
3925 VkResult result;
3926
3927 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
3928 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3929 if (pipeline == NULL)
3930 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
3931
3932 pipeline->device = device;
3933 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
3934 assert(pipeline->layout);
3935
3936 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
3937 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
3938 radv_init_feedback(creation_feedback);
3939
3940 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
3941 if (creation_feedback)
3942 stage_feedbacks[MESA_SHADER_COMPUTE] = &creation_feedback->pPipelineStageCreationFeedbacks[0];
3943
3944 pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage;
3945 radv_create_shaders(pipeline, device, cache, &(struct radv_pipeline_key) {0}, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
3946
3947 pipeline->user_data_0[MESA_SHADER_COMPUTE] = radv_pipeline_stage_to_user_data_0(pipeline, MESA_SHADER_COMPUTE, device->physical_device->rad_info.chip_class);
3948 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets;
3949 result = radv_pipeline_scratch_init(device, pipeline);
3950 if (result != VK_SUCCESS) {
3951 radv_pipeline_destroy(device, pipeline, pAllocator);
3952 return result;
3953 }
3954
3955 radv_compute_generate_pm4(pipeline);
3956
3957 *pPipeline = radv_pipeline_to_handle(pipeline);
3958
3959 return VK_SUCCESS;
3960 }
3961
3962 VkResult radv_CreateComputePipelines(
3963 VkDevice _device,
3964 VkPipelineCache pipelineCache,
3965 uint32_t count,
3966 const VkComputePipelineCreateInfo* pCreateInfos,
3967 const VkAllocationCallbacks* pAllocator,
3968 VkPipeline* pPipelines)
3969 {
3970 VkResult result = VK_SUCCESS;
3971
3972 unsigned i = 0;
3973 for (; i < count; i++) {
3974 VkResult r;
3975 r = radv_compute_pipeline_create(_device, pipelineCache,
3976 &pCreateInfos[i],
3977 pAllocator, &pPipelines[i]);
3978 if (r != VK_SUCCESS) {
3979 result = r;
3980 pPipelines[i] = VK_NULL_HANDLE;
3981 }
3982 }
3983
3984 return result;
3985 }