radv: clean up the sample locations codebase
[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_get_default_max_sample_dist(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 (needed_states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1421 dynamic->discard_rectangle.count = discard_rectangle_info->discardRectangleCount;
1422 if (states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1423 typed_memcpy(dynamic->discard_rectangle.rectangles,
1424 discard_rectangle_info->pDiscardRectangles,
1425 discard_rectangle_info->discardRectangleCount);
1426 }
1427 }
1428
1429 pipeline->dynamic_state.mask = states;
1430 }
1431
1432 static struct radv_gs_state
1433 calculate_gs_info(const VkGraphicsPipelineCreateInfo *pCreateInfo,
1434 const struct radv_pipeline *pipeline)
1435 {
1436 struct radv_gs_state gs = {0};
1437 struct radv_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
1438 struct radv_es_output_info *es_info;
1439 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
1440 es_info = radv_pipeline_has_tess(pipeline) ? &gs_info->tes.es_info : &gs_info->vs.es_info;
1441 else
1442 es_info = radv_pipeline_has_tess(pipeline) ?
1443 &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.es_info :
1444 &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.es_info;
1445
1446 unsigned gs_num_invocations = MAX2(gs_info->gs.invocations, 1);
1447 bool uses_adjacency;
1448 switch(pCreateInfo->pInputAssemblyState->topology) {
1449 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1450 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1451 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1452 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1453 uses_adjacency = true;
1454 break;
1455 default:
1456 uses_adjacency = false;
1457 break;
1458 }
1459
1460 /* All these are in dwords: */
1461 /* We can't allow using the whole LDS, because GS waves compete with
1462 * other shader stages for LDS space. */
1463 const unsigned max_lds_size = 8 * 1024;
1464 const unsigned esgs_itemsize = es_info->esgs_itemsize / 4;
1465 unsigned esgs_lds_size;
1466
1467 /* All these are per subgroup: */
1468 const unsigned max_out_prims = 32 * 1024;
1469 const unsigned max_es_verts = 255;
1470 const unsigned ideal_gs_prims = 64;
1471 unsigned max_gs_prims, gs_prims;
1472 unsigned min_es_verts, es_verts, worst_case_es_verts;
1473
1474 if (uses_adjacency || gs_num_invocations > 1)
1475 max_gs_prims = 127 / gs_num_invocations;
1476 else
1477 max_gs_prims = 255;
1478
1479 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1480 * Make sure we don't go over the maximum value.
1481 */
1482 if (gs_info->gs.vertices_out > 0) {
1483 max_gs_prims = MIN2(max_gs_prims,
1484 max_out_prims /
1485 (gs_info->gs.vertices_out * gs_num_invocations));
1486 }
1487 assert(max_gs_prims > 0);
1488
1489 /* If the primitive has adjacency, halve the number of vertices
1490 * that will be reused in multiple primitives.
1491 */
1492 min_es_verts = gs_info->gs.vertices_in / (uses_adjacency ? 2 : 1);
1493
1494 gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
1495 worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
1496
1497 /* Compute ESGS LDS size based on the worst case number of ES vertices
1498 * needed to create the target number of GS prims per subgroup.
1499 */
1500 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1501
1502 /* If total LDS usage is too big, refactor partitions based on ratio
1503 * of ESGS item sizes.
1504 */
1505 if (esgs_lds_size > max_lds_size) {
1506 /* Our target GS Prims Per Subgroup was too large. Calculate
1507 * the maximum number of GS Prims Per Subgroup that will fit
1508 * into LDS, capped by the maximum that the hardware can support.
1509 */
1510 gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
1511 max_gs_prims);
1512 assert(gs_prims > 0);
1513 worst_case_es_verts = MIN2(min_es_verts * gs_prims,
1514 max_es_verts);
1515
1516 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1517 assert(esgs_lds_size <= max_lds_size);
1518 }
1519
1520 /* Now calculate remaining ESGS information. */
1521 if (esgs_lds_size)
1522 es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
1523 else
1524 es_verts = max_es_verts;
1525
1526 /* Vertices for adjacency primitives are not always reused, so restore
1527 * it for ES_VERTS_PER_SUBGRP.
1528 */
1529 min_es_verts = gs_info->gs.vertices_in;
1530
1531 /* For normal primitives, the VGT only checks if they are past the ES
1532 * verts per subgroup after allocating a full GS primitive and if they
1533 * are, kick off a new subgroup. But if those additional ES verts are
1534 * unique (e.g. not reused) we need to make sure there is enough LDS
1535 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1536 */
1537 es_verts -= min_es_verts - 1;
1538
1539 uint32_t es_verts_per_subgroup = es_verts;
1540 uint32_t gs_prims_per_subgroup = gs_prims;
1541 uint32_t gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
1542 uint32_t max_prims_per_subgroup = gs_inst_prims_in_subgroup * gs_info->gs.vertices_out;
1543 gs.lds_size = align(esgs_lds_size, 128) / 128;
1544 gs.vgt_gs_onchip_cntl = S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup) |
1545 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup) |
1546 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup);
1547 gs.vgt_gs_max_prims_per_subgroup = S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup);
1548 gs.vgt_esgs_ring_itemsize = esgs_itemsize;
1549 assert(max_prims_per_subgroup <= max_out_prims);
1550
1551 return gs;
1552 }
1553
1554 static void
1555 calculate_gs_ring_sizes(struct radv_pipeline *pipeline, const struct radv_gs_state *gs)
1556 {
1557 struct radv_device *device = pipeline->device;
1558 unsigned num_se = device->physical_device->rad_info.max_se;
1559 unsigned wave_size = 64;
1560 unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
1561 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
1562 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
1563 */
1564 unsigned gs_vertex_reuse =
1565 (device->physical_device->rad_info.chip_class >= GFX8 ? 32 : 16) * num_se;
1566 unsigned alignment = 256 * num_se;
1567 /* The maximum size is 63.999 MB per SE. */
1568 unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
1569 struct radv_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
1570
1571 /* Calculate the minimum size. */
1572 unsigned min_esgs_ring_size = align(gs->vgt_esgs_ring_itemsize * 4 * gs_vertex_reuse *
1573 wave_size, alignment);
1574 /* These are recommended sizes, not minimum sizes. */
1575 unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
1576 gs->vgt_esgs_ring_itemsize * 4 * gs_info->gs.vertices_in;
1577 unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
1578 gs_info->gs.max_gsvs_emit_size;
1579
1580 min_esgs_ring_size = align(min_esgs_ring_size, alignment);
1581 esgs_ring_size = align(esgs_ring_size, alignment);
1582 gsvs_ring_size = align(gsvs_ring_size, alignment);
1583
1584 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
1585 pipeline->graphics.esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
1586
1587 pipeline->graphics.gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
1588 }
1589
1590 static void si_multiwave_lds_size_workaround(struct radv_device *device,
1591 unsigned *lds_size)
1592 {
1593 /* If tessellation is all offchip and on-chip GS isn't used, this
1594 * workaround is not needed.
1595 */
1596 return;
1597
1598 /* SPI barrier management bug:
1599 * Make sure we have at least 4k of LDS in use to avoid the bug.
1600 * It applies to workgroup sizes of more than one wavefront.
1601 */
1602 if (device->physical_device->rad_info.family == CHIP_BONAIRE ||
1603 device->physical_device->rad_info.family == CHIP_KABINI ||
1604 device->physical_device->rad_info.family == CHIP_MULLINS)
1605 *lds_size = MAX2(*lds_size, 8);
1606 }
1607
1608 struct radv_shader_variant *
1609 radv_get_shader(struct radv_pipeline *pipeline,
1610 gl_shader_stage stage)
1611 {
1612 if (stage == MESA_SHADER_VERTEX) {
1613 if (pipeline->shaders[MESA_SHADER_VERTEX])
1614 return pipeline->shaders[MESA_SHADER_VERTEX];
1615 if (pipeline->shaders[MESA_SHADER_TESS_CTRL])
1616 return pipeline->shaders[MESA_SHADER_TESS_CTRL];
1617 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
1618 return pipeline->shaders[MESA_SHADER_GEOMETRY];
1619 } else if (stage == MESA_SHADER_TESS_EVAL) {
1620 if (!radv_pipeline_has_tess(pipeline))
1621 return NULL;
1622 if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
1623 return pipeline->shaders[MESA_SHADER_TESS_EVAL];
1624 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
1625 return pipeline->shaders[MESA_SHADER_GEOMETRY];
1626 }
1627 return pipeline->shaders[stage];
1628 }
1629
1630 static struct radv_tessellation_state
1631 calculate_tess_state(struct radv_pipeline *pipeline,
1632 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1633 {
1634 unsigned num_tcs_input_cp;
1635 unsigned num_tcs_output_cp;
1636 unsigned lds_size;
1637 unsigned num_patches;
1638 struct radv_tessellation_state tess = {0};
1639
1640 num_tcs_input_cp = pCreateInfo->pTessellationState->patchControlPoints;
1641 num_tcs_output_cp = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.tcs_vertices_out; //TCS VERTICES OUT
1642 num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
1643
1644 lds_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.lds_size;
1645
1646 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
1647 assert(lds_size <= 65536);
1648 lds_size = align(lds_size, 512) / 512;
1649 } else {
1650 assert(lds_size <= 32768);
1651 lds_size = align(lds_size, 256) / 256;
1652 }
1653 si_multiwave_lds_size_workaround(pipeline->device, &lds_size);
1654
1655 tess.lds_size = lds_size;
1656
1657 tess.ls_hs_config = S_028B58_NUM_PATCHES(num_patches) |
1658 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
1659 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
1660 tess.num_patches = num_patches;
1661
1662 struct radv_shader_variant *tes = radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL);
1663 unsigned type = 0, partitioning = 0, topology = 0, distribution_mode = 0;
1664
1665 switch (tes->info.tes.primitive_mode) {
1666 case GL_TRIANGLES:
1667 type = V_028B6C_TESS_TRIANGLE;
1668 break;
1669 case GL_QUADS:
1670 type = V_028B6C_TESS_QUAD;
1671 break;
1672 case GL_ISOLINES:
1673 type = V_028B6C_TESS_ISOLINE;
1674 break;
1675 }
1676
1677 switch (tes->info.tes.spacing) {
1678 case TESS_SPACING_EQUAL:
1679 partitioning = V_028B6C_PART_INTEGER;
1680 break;
1681 case TESS_SPACING_FRACTIONAL_ODD:
1682 partitioning = V_028B6C_PART_FRAC_ODD;
1683 break;
1684 case TESS_SPACING_FRACTIONAL_EVEN:
1685 partitioning = V_028B6C_PART_FRAC_EVEN;
1686 break;
1687 default:
1688 break;
1689 }
1690
1691 bool ccw = tes->info.tes.ccw;
1692 const VkPipelineTessellationDomainOriginStateCreateInfo *domain_origin_state =
1693 vk_find_struct_const(pCreateInfo->pTessellationState,
1694 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO);
1695
1696 if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT)
1697 ccw = !ccw;
1698
1699 if (tes->info.tes.point_mode)
1700 topology = V_028B6C_OUTPUT_POINT;
1701 else if (tes->info.tes.primitive_mode == GL_ISOLINES)
1702 topology = V_028B6C_OUTPUT_LINE;
1703 else if (ccw)
1704 topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
1705 else
1706 topology = V_028B6C_OUTPUT_TRIANGLE_CW;
1707
1708 if (pipeline->device->has_distributed_tess) {
1709 if (pipeline->device->physical_device->rad_info.family == CHIP_FIJI ||
1710 pipeline->device->physical_device->rad_info.family >= CHIP_POLARIS10)
1711 distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
1712 else
1713 distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
1714 } else
1715 distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
1716
1717 tess.tf_param = S_028B6C_TYPE(type) |
1718 S_028B6C_PARTITIONING(partitioning) |
1719 S_028B6C_TOPOLOGY(topology) |
1720 S_028B6C_DISTRIBUTION_MODE(distribution_mode);
1721
1722 return tess;
1723 }
1724
1725 static const struct radv_prim_vertex_count prim_size_table[] = {
1726 [V_008958_DI_PT_NONE] = {0, 0},
1727 [V_008958_DI_PT_POINTLIST] = {1, 1},
1728 [V_008958_DI_PT_LINELIST] = {2, 2},
1729 [V_008958_DI_PT_LINESTRIP] = {2, 1},
1730 [V_008958_DI_PT_TRILIST] = {3, 3},
1731 [V_008958_DI_PT_TRIFAN] = {3, 1},
1732 [V_008958_DI_PT_TRISTRIP] = {3, 1},
1733 [V_008958_DI_PT_LINELIST_ADJ] = {4, 4},
1734 [V_008958_DI_PT_LINESTRIP_ADJ] = {4, 1},
1735 [V_008958_DI_PT_TRILIST_ADJ] = {6, 6},
1736 [V_008958_DI_PT_TRISTRIP_ADJ] = {6, 2},
1737 [V_008958_DI_PT_RECTLIST] = {3, 3},
1738 [V_008958_DI_PT_LINELOOP] = {2, 1},
1739 [V_008958_DI_PT_POLYGON] = {3, 1},
1740 [V_008958_DI_PT_2D_TRI_STRIP] = {0, 0},
1741 };
1742
1743 static const struct radv_vs_output_info *get_vs_output_info(const struct radv_pipeline *pipeline)
1744 {
1745 if (radv_pipeline_has_gs(pipeline))
1746 return &pipeline->gs_copy_shader->info.vs.outinfo;
1747 else if (radv_pipeline_has_tess(pipeline))
1748 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.outinfo;
1749 else
1750 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.outinfo;
1751 }
1752
1753 static void
1754 radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders)
1755 {
1756 nir_shader* ordered_shaders[MESA_SHADER_STAGES];
1757 int shader_count = 0;
1758
1759 if(shaders[MESA_SHADER_FRAGMENT]) {
1760 ordered_shaders[shader_count++] = shaders[MESA_SHADER_FRAGMENT];
1761 }
1762 if(shaders[MESA_SHADER_GEOMETRY]) {
1763 ordered_shaders[shader_count++] = shaders[MESA_SHADER_GEOMETRY];
1764 }
1765 if(shaders[MESA_SHADER_TESS_EVAL]) {
1766 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_EVAL];
1767 }
1768 if(shaders[MESA_SHADER_TESS_CTRL]) {
1769 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_CTRL];
1770 }
1771 if(shaders[MESA_SHADER_VERTEX]) {
1772 ordered_shaders[shader_count++] = shaders[MESA_SHADER_VERTEX];
1773 }
1774
1775 if (shader_count > 1) {
1776 unsigned first = ordered_shaders[shader_count - 1]->info.stage;
1777 unsigned last = ordered_shaders[0]->info.stage;
1778
1779 if (ordered_shaders[0]->info.stage == MESA_SHADER_FRAGMENT &&
1780 ordered_shaders[1]->info.has_transform_feedback_varyings)
1781 nir_link_xfb_varyings(ordered_shaders[1], ordered_shaders[0]);
1782
1783 for (int i = 0; i < shader_count; ++i) {
1784 nir_variable_mode mask = 0;
1785
1786 if (ordered_shaders[i]->info.stage != first)
1787 mask = mask | nir_var_shader_in;
1788
1789 if (ordered_shaders[i]->info.stage != last)
1790 mask = mask | nir_var_shader_out;
1791
1792 nir_lower_io_to_scalar_early(ordered_shaders[i], mask);
1793 radv_optimize_nir(ordered_shaders[i], false, false);
1794 }
1795 }
1796
1797 for (int i = 1; i < shader_count; ++i) {
1798 nir_lower_io_arrays_to_elements(ordered_shaders[i],
1799 ordered_shaders[i - 1]);
1800
1801 if (nir_link_opt_varyings(ordered_shaders[i],
1802 ordered_shaders[i - 1]))
1803 radv_optimize_nir(ordered_shaders[i - 1], false, false);
1804
1805 nir_remove_dead_variables(ordered_shaders[i],
1806 nir_var_shader_out);
1807 nir_remove_dead_variables(ordered_shaders[i - 1],
1808 nir_var_shader_in);
1809
1810 bool progress = nir_remove_unused_varyings(ordered_shaders[i],
1811 ordered_shaders[i - 1]);
1812
1813 nir_compact_varyings(ordered_shaders[i],
1814 ordered_shaders[i - 1], true);
1815
1816 if (progress) {
1817 if (nir_lower_global_vars_to_local(ordered_shaders[i])) {
1818 ac_lower_indirect_derefs(ordered_shaders[i],
1819 pipeline->device->physical_device->rad_info.chip_class);
1820 }
1821 radv_optimize_nir(ordered_shaders[i], false, false);
1822
1823 if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) {
1824 ac_lower_indirect_derefs(ordered_shaders[i - 1],
1825 pipeline->device->physical_device->rad_info.chip_class);
1826 }
1827 radv_optimize_nir(ordered_shaders[i - 1], false, false);
1828 }
1829 }
1830 }
1831
1832 static uint32_t
1833 radv_get_attrib_stride(const VkPipelineVertexInputStateCreateInfo *input_state,
1834 uint32_t attrib_binding)
1835 {
1836 for (uint32_t i = 0; i < input_state->vertexBindingDescriptionCount; i++) {
1837 const VkVertexInputBindingDescription *input_binding =
1838 &input_state->pVertexBindingDescriptions[i];
1839
1840 if (input_binding->binding == attrib_binding)
1841 return input_binding->stride;
1842 }
1843
1844 return 0;
1845 }
1846
1847 static struct radv_pipeline_key
1848 radv_generate_graphics_pipeline_key(struct radv_pipeline *pipeline,
1849 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1850 const struct radv_blend_state *blend,
1851 bool has_view_index)
1852 {
1853 const VkPipelineVertexInputStateCreateInfo *input_state =
1854 pCreateInfo->pVertexInputState;
1855 const VkPipelineVertexInputDivisorStateCreateInfoEXT *divisor_state =
1856 vk_find_struct_const(input_state->pNext, PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT);
1857
1858 struct radv_pipeline_key key;
1859 memset(&key, 0, sizeof(key));
1860
1861 if (pCreateInfo->flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)
1862 key.optimisations_disabled = 1;
1863
1864 key.has_multiview_view_index = has_view_index;
1865
1866 uint32_t binding_input_rate = 0;
1867 uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
1868 for (unsigned i = 0; i < input_state->vertexBindingDescriptionCount; ++i) {
1869 if (input_state->pVertexBindingDescriptions[i].inputRate) {
1870 unsigned binding = input_state->pVertexBindingDescriptions[i].binding;
1871 binding_input_rate |= 1u << binding;
1872 instance_rate_divisors[binding] = 1;
1873 }
1874 }
1875 if (divisor_state) {
1876 for (unsigned i = 0; i < divisor_state->vertexBindingDivisorCount; ++i) {
1877 instance_rate_divisors[divisor_state->pVertexBindingDivisors[i].binding] =
1878 divisor_state->pVertexBindingDivisors[i].divisor;
1879 }
1880 }
1881
1882 for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) {
1883 const VkVertexInputAttributeDescription *desc =
1884 &input_state->pVertexAttributeDescriptions[i];
1885 const struct vk_format_description *format_desc;
1886 unsigned location = desc->location;
1887 unsigned binding = desc->binding;
1888 unsigned num_format, data_format;
1889 int first_non_void;
1890
1891 if (binding_input_rate & (1u << binding)) {
1892 key.instance_rate_inputs |= 1u << location;
1893 key.instance_rate_divisors[location] = instance_rate_divisors[binding];
1894 }
1895
1896 format_desc = vk_format_description(desc->format);
1897 first_non_void = vk_format_get_first_non_void_channel(desc->format);
1898
1899 num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
1900 data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
1901
1902 key.vertex_attribute_formats[location] = data_format | (num_format << 4);
1903 key.vertex_attribute_bindings[location] = desc->binding;
1904 key.vertex_attribute_offsets[location] = desc->offset;
1905 key.vertex_attribute_strides[location] = radv_get_attrib_stride(input_state, desc->binding);
1906
1907 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8 &&
1908 pipeline->device->physical_device->rad_info.family != CHIP_STONEY) {
1909 VkFormat format = input_state->pVertexAttributeDescriptions[i].format;
1910 uint64_t adjust;
1911 switch(format) {
1912 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
1913 case VK_FORMAT_A2B10G10R10_SNORM_PACK32:
1914 adjust = RADV_ALPHA_ADJUST_SNORM;
1915 break;
1916 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
1917 case VK_FORMAT_A2B10G10R10_SSCALED_PACK32:
1918 adjust = RADV_ALPHA_ADJUST_SSCALED;
1919 break;
1920 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
1921 case VK_FORMAT_A2B10G10R10_SINT_PACK32:
1922 adjust = RADV_ALPHA_ADJUST_SINT;
1923 break;
1924 default:
1925 adjust = 0;
1926 break;
1927 }
1928 key.vertex_alpha_adjust |= adjust << (2 * location);
1929 }
1930
1931 switch (desc->format) {
1932 case VK_FORMAT_B8G8R8A8_UNORM:
1933 case VK_FORMAT_B8G8R8A8_SNORM:
1934 case VK_FORMAT_B8G8R8A8_USCALED:
1935 case VK_FORMAT_B8G8R8A8_SSCALED:
1936 case VK_FORMAT_B8G8R8A8_UINT:
1937 case VK_FORMAT_B8G8R8A8_SINT:
1938 case VK_FORMAT_B8G8R8A8_SRGB:
1939 case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
1940 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
1941 case VK_FORMAT_A2R10G10B10_USCALED_PACK32:
1942 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
1943 case VK_FORMAT_A2R10G10B10_UINT_PACK32:
1944 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
1945 key.vertex_post_shuffle |= 1 << location;
1946 break;
1947 default:
1948 break;
1949 }
1950 }
1951
1952 if (pCreateInfo->pTessellationState)
1953 key.tess_input_vertices = pCreateInfo->pTessellationState->patchControlPoints;
1954
1955
1956 if (pCreateInfo->pMultisampleState &&
1957 pCreateInfo->pMultisampleState->rasterizationSamples > 1) {
1958 uint32_t num_samples = pCreateInfo->pMultisampleState->rasterizationSamples;
1959 uint32_t ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo->pMultisampleState);
1960 key.num_samples = num_samples;
1961 key.log2_ps_iter_samples = util_logbase2(ps_iter_samples);
1962 }
1963
1964 key.col_format = blend->spi_shader_col_format;
1965 if (pipeline->device->physical_device->rad_info.chip_class < GFX8)
1966 radv_pipeline_compute_get_int_clamp(pCreateInfo, &key.is_int8, &key.is_int10);
1967
1968 return key;
1969 }
1970
1971 static void
1972 radv_fill_shader_keys(struct radv_shader_variant_key *keys,
1973 const struct radv_pipeline_key *key,
1974 nir_shader **nir)
1975 {
1976 keys[MESA_SHADER_VERTEX].vs.instance_rate_inputs = key->instance_rate_inputs;
1977 keys[MESA_SHADER_VERTEX].vs.alpha_adjust = key->vertex_alpha_adjust;
1978 keys[MESA_SHADER_VERTEX].vs.post_shuffle = key->vertex_post_shuffle;
1979 for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; ++i) {
1980 keys[MESA_SHADER_VERTEX].vs.instance_rate_divisors[i] = key->instance_rate_divisors[i];
1981 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_formats[i] = key->vertex_attribute_formats[i];
1982 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_bindings[i] = key->vertex_attribute_bindings[i];
1983 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_offsets[i] = key->vertex_attribute_offsets[i];
1984 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_strides[i] = key->vertex_attribute_strides[i];
1985 }
1986
1987 if (nir[MESA_SHADER_TESS_CTRL]) {
1988 keys[MESA_SHADER_VERTEX].vs.as_ls = true;
1989 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = 0;
1990 keys[MESA_SHADER_TESS_CTRL].tcs.input_vertices = key->tess_input_vertices;
1991 keys[MESA_SHADER_TESS_CTRL].tcs.primitive_mode = nir[MESA_SHADER_TESS_EVAL]->info.tess.primitive_mode;
1992
1993 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));
1994 }
1995
1996 if (nir[MESA_SHADER_GEOMETRY]) {
1997 if (nir[MESA_SHADER_TESS_CTRL])
1998 keys[MESA_SHADER_TESS_EVAL].tes.as_es = true;
1999 else
2000 keys[MESA_SHADER_VERTEX].vs.as_es = true;
2001 }
2002
2003 for(int i = 0; i < MESA_SHADER_STAGES; ++i)
2004 keys[i].has_multiview_view_index = key->has_multiview_view_index;
2005
2006 keys[MESA_SHADER_FRAGMENT].fs.col_format = key->col_format;
2007 keys[MESA_SHADER_FRAGMENT].fs.is_int8 = key->is_int8;
2008 keys[MESA_SHADER_FRAGMENT].fs.is_int10 = key->is_int10;
2009 keys[MESA_SHADER_FRAGMENT].fs.log2_ps_iter_samples = key->log2_ps_iter_samples;
2010 keys[MESA_SHADER_FRAGMENT].fs.num_samples = key->num_samples;
2011 }
2012
2013 static void
2014 merge_tess_info(struct shader_info *tes_info,
2015 const struct shader_info *tcs_info)
2016 {
2017 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
2018 *
2019 * "PointMode. Controls generation of points rather than triangles
2020 * or lines. This functionality defaults to disabled, and is
2021 * enabled if either shader stage includes the execution mode.
2022 *
2023 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
2024 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
2025 * and OutputVertices, it says:
2026 *
2027 * "One mode must be set in at least one of the tessellation
2028 * shader stages."
2029 *
2030 * So, the fields can be set in either the TCS or TES, but they must
2031 * agree if set in both. Our backend looks at TES, so bitwise-or in
2032 * the values from the TCS.
2033 */
2034 assert(tcs_info->tess.tcs_vertices_out == 0 ||
2035 tes_info->tess.tcs_vertices_out == 0 ||
2036 tcs_info->tess.tcs_vertices_out == tes_info->tess.tcs_vertices_out);
2037 tes_info->tess.tcs_vertices_out |= tcs_info->tess.tcs_vertices_out;
2038
2039 assert(tcs_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2040 tes_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2041 tcs_info->tess.spacing == tes_info->tess.spacing);
2042 tes_info->tess.spacing |= tcs_info->tess.spacing;
2043
2044 assert(tcs_info->tess.primitive_mode == 0 ||
2045 tes_info->tess.primitive_mode == 0 ||
2046 tcs_info->tess.primitive_mode == tes_info->tess.primitive_mode);
2047 tes_info->tess.primitive_mode |= tcs_info->tess.primitive_mode;
2048 tes_info->tess.ccw |= tcs_info->tess.ccw;
2049 tes_info->tess.point_mode |= tcs_info->tess.point_mode;
2050 }
2051
2052 static
2053 void radv_init_feedback(const VkPipelineCreationFeedbackCreateInfoEXT *ext)
2054 {
2055 if (!ext)
2056 return;
2057
2058 if (ext->pPipelineCreationFeedback) {
2059 ext->pPipelineCreationFeedback->flags = 0;
2060 ext->pPipelineCreationFeedback->duration = 0;
2061 }
2062
2063 for (unsigned i = 0; i < ext->pipelineStageCreationFeedbackCount; ++i) {
2064 ext->pPipelineStageCreationFeedbacks[i].flags = 0;
2065 ext->pPipelineStageCreationFeedbacks[i].duration = 0;
2066 }
2067 }
2068
2069 static
2070 void radv_start_feedback(VkPipelineCreationFeedbackEXT *feedback)
2071 {
2072 if (!feedback)
2073 return;
2074
2075 feedback->duration -= radv_get_current_time();
2076 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT;
2077 }
2078
2079 static
2080 void radv_stop_feedback(VkPipelineCreationFeedbackEXT *feedback, bool cache_hit)
2081 {
2082 if (!feedback)
2083 return;
2084
2085 feedback->duration += radv_get_current_time();
2086 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT |
2087 (cache_hit ? VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT : 0);
2088 }
2089
2090 static
2091 void radv_create_shaders(struct radv_pipeline *pipeline,
2092 struct radv_device *device,
2093 struct radv_pipeline_cache *cache,
2094 const struct radv_pipeline_key *key,
2095 const VkPipelineShaderStageCreateInfo **pStages,
2096 const VkPipelineCreateFlags flags,
2097 VkPipelineCreationFeedbackEXT *pipeline_feedback,
2098 VkPipelineCreationFeedbackEXT **stage_feedbacks)
2099 {
2100 struct radv_shader_module fs_m = {0};
2101 struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
2102 nir_shader *nir[MESA_SHADER_STAGES] = {0};
2103 void *codes[MESA_SHADER_STAGES] = {0};
2104 unsigned code_sizes[MESA_SHADER_STAGES] = {0};
2105 struct radv_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{0}}}};
2106 unsigned char hash[20], gs_copy_hash[20];
2107
2108 radv_start_feedback(pipeline_feedback);
2109
2110 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2111 if (pStages[i]) {
2112 modules[i] = radv_shader_module_from_handle(pStages[i]->module);
2113 if (modules[i]->nir)
2114 _mesa_sha1_compute(modules[i]->nir->info.name,
2115 strlen(modules[i]->nir->info.name),
2116 modules[i]->sha1);
2117
2118 pipeline->active_stages |= mesa_to_vk_shader_stage(i);
2119 }
2120 }
2121
2122 radv_hash_shaders(hash, pStages, pipeline->layout, key, get_hash_flags(device));
2123 memcpy(gs_copy_hash, hash, 20);
2124 gs_copy_hash[0] ^= 1;
2125
2126 bool found_in_application_cache = true;
2127 if (modules[MESA_SHADER_GEOMETRY]) {
2128 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2129 radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants,
2130 &found_in_application_cache);
2131 pipeline->gs_copy_shader = variants[MESA_SHADER_GEOMETRY];
2132 }
2133
2134 if (radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders,
2135 &found_in_application_cache) &&
2136 (!modules[MESA_SHADER_GEOMETRY] || pipeline->gs_copy_shader)) {
2137 radv_stop_feedback(pipeline_feedback, found_in_application_cache);
2138 return;
2139 }
2140
2141 if (!modules[MESA_SHADER_FRAGMENT] && !modules[MESA_SHADER_COMPUTE]) {
2142 nir_builder fs_b;
2143 nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
2144 fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "noop_fs");
2145 fs_m.nir = fs_b.shader;
2146 modules[MESA_SHADER_FRAGMENT] = &fs_m;
2147 }
2148
2149 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2150 const VkPipelineShaderStageCreateInfo *stage = pStages[i];
2151
2152 if (!modules[i])
2153 continue;
2154
2155 radv_start_feedback(stage_feedbacks[i]);
2156
2157 nir[i] = radv_shader_compile_to_nir(device, modules[i],
2158 stage ? stage->pName : "main", i,
2159 stage ? stage->pSpecializationInfo : NULL,
2160 flags, pipeline->layout);
2161
2162 /* We don't want to alter meta shaders IR directly so clone it
2163 * first.
2164 */
2165 if (nir[i]->info.name) {
2166 nir[i] = nir_shader_clone(NULL, nir[i]);
2167 }
2168
2169 radv_stop_feedback(stage_feedbacks[i], false);
2170 }
2171
2172 if (nir[MESA_SHADER_TESS_CTRL]) {
2173 nir_lower_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out, NULL);
2174 merge_tess_info(&nir[MESA_SHADER_TESS_EVAL]->info, &nir[MESA_SHADER_TESS_CTRL]->info);
2175 }
2176
2177 if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT))
2178 radv_link_shaders(pipeline, nir);
2179
2180 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2181 if (nir[i]) {
2182 NIR_PASS_V(nir[i], nir_lower_bool_to_int32);
2183 NIR_PASS_V(nir[i], nir_lower_non_uniform_access,
2184 nir_lower_non_uniform_ubo_access |
2185 nir_lower_non_uniform_ssbo_access |
2186 nir_lower_non_uniform_texture_access |
2187 nir_lower_non_uniform_image_access);
2188 }
2189
2190 if (radv_can_dump_shader(device, modules[i], false))
2191 nir_print_shader(nir[i], stderr);
2192 }
2193
2194 radv_fill_shader_keys(keys, key, nir);
2195
2196 if (nir[MESA_SHADER_FRAGMENT]) {
2197 if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) {
2198 radv_start_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT]);
2199
2200 pipeline->shaders[MESA_SHADER_FRAGMENT] =
2201 radv_shader_variant_create(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1,
2202 pipeline->layout, keys + MESA_SHADER_FRAGMENT,
2203 &codes[MESA_SHADER_FRAGMENT], &code_sizes[MESA_SHADER_FRAGMENT]);
2204
2205 radv_stop_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT], false);
2206 }
2207
2208 /* TODO: These are no longer used as keys we should refactor this */
2209 keys[MESA_SHADER_VERTEX].vs.export_prim_id =
2210 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input;
2211 keys[MESA_SHADER_VERTEX].vs.export_layer_id =
2212 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.layer_input;
2213 keys[MESA_SHADER_TESS_EVAL].tes.export_prim_id =
2214 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input;
2215 keys[MESA_SHADER_TESS_EVAL].tes.export_layer_id =
2216 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.layer_input;
2217 }
2218
2219 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_TESS_CTRL]) {
2220 if (!pipeline->shaders[MESA_SHADER_TESS_CTRL]) {
2221 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
2222 struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
2223 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
2224
2225 radv_start_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL]);
2226
2227 pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_create(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2,
2228 pipeline->layout,
2229 &key, &codes[MESA_SHADER_TESS_CTRL],
2230 &code_sizes[MESA_SHADER_TESS_CTRL]);
2231
2232 radv_stop_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL], false);
2233 }
2234 modules[MESA_SHADER_VERTEX] = NULL;
2235 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
2236 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.tcs.outputs_written);
2237 }
2238
2239 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_GEOMETRY]) {
2240 gl_shader_stage pre_stage = modules[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
2241 if (!pipeline->shaders[MESA_SHADER_GEOMETRY]) {
2242 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
2243
2244 radv_start_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY]);
2245
2246 pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_create(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2,
2247 pipeline->layout,
2248 &keys[pre_stage] , &codes[MESA_SHADER_GEOMETRY],
2249 &code_sizes[MESA_SHADER_GEOMETRY]);
2250
2251 radv_stop_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY], false);
2252 }
2253 modules[pre_stage] = NULL;
2254 }
2255
2256 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2257 if(modules[i] && !pipeline->shaders[i]) {
2258 if (i == MESA_SHADER_TESS_CTRL) {
2259 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = util_last_bit64(pipeline->shaders[MESA_SHADER_VERTEX]->info.info.vs.ls_outputs_written);
2260 }
2261 if (i == MESA_SHADER_TESS_EVAL) {
2262 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
2263 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.tcs.outputs_written);
2264 }
2265
2266 radv_start_feedback(stage_feedbacks[i]);
2267
2268 pipeline->shaders[i] = radv_shader_variant_create(device, modules[i], &nir[i], 1,
2269 pipeline->layout,
2270 keys + i, &codes[i],
2271 &code_sizes[i]);
2272
2273 radv_stop_feedback(stage_feedbacks[i], false);
2274 }
2275 }
2276
2277 if(modules[MESA_SHADER_GEOMETRY]) {
2278 void *gs_copy_code = NULL;
2279 unsigned gs_copy_code_size = 0;
2280 if (!pipeline->gs_copy_shader) {
2281 pipeline->gs_copy_shader = radv_create_gs_copy_shader(
2282 device, nir[MESA_SHADER_GEOMETRY], &gs_copy_code,
2283 &gs_copy_code_size,
2284 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index);
2285 }
2286
2287 if (pipeline->gs_copy_shader) {
2288 void *code[MESA_SHADER_STAGES] = {0};
2289 unsigned code_size[MESA_SHADER_STAGES] = {0};
2290 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2291
2292 code[MESA_SHADER_GEOMETRY] = gs_copy_code;
2293 code_size[MESA_SHADER_GEOMETRY] = gs_copy_code_size;
2294 variants[MESA_SHADER_GEOMETRY] = pipeline->gs_copy_shader;
2295
2296 radv_pipeline_cache_insert_shaders(device, cache,
2297 gs_copy_hash,
2298 variants,
2299 (const void**)code,
2300 code_size);
2301 }
2302 free(gs_copy_code);
2303 }
2304
2305 radv_pipeline_cache_insert_shaders(device, cache, hash, pipeline->shaders,
2306 (const void**)codes, code_sizes);
2307
2308 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2309 free(codes[i]);
2310 if (nir[i]) {
2311 if (!pipeline->device->keep_shader_info)
2312 ralloc_free(nir[i]);
2313
2314 if (radv_can_dump_shader_stats(device, modules[i]))
2315 radv_shader_dump_stats(device,
2316 pipeline->shaders[i],
2317 i, stderr);
2318 }
2319 }
2320
2321 if (fs_m.nir)
2322 ralloc_free(fs_m.nir);
2323
2324 radv_stop_feedback(pipeline_feedback, false);
2325 }
2326
2327 static uint32_t
2328 radv_pipeline_stage_to_user_data_0(struct radv_pipeline *pipeline,
2329 gl_shader_stage stage, enum chip_class chip_class)
2330 {
2331 bool has_gs = radv_pipeline_has_gs(pipeline);
2332 bool has_tess = radv_pipeline_has_tess(pipeline);
2333 switch (stage) {
2334 case MESA_SHADER_FRAGMENT:
2335 return R_00B030_SPI_SHADER_USER_DATA_PS_0;
2336 case MESA_SHADER_VERTEX:
2337 if (chip_class >= GFX9) {
2338 return has_tess ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
2339 has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2340 R_00B130_SPI_SHADER_USER_DATA_VS_0;
2341 }
2342 if (has_tess)
2343 return R_00B530_SPI_SHADER_USER_DATA_LS_0;
2344 else
2345 return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : R_00B130_SPI_SHADER_USER_DATA_VS_0;
2346 case MESA_SHADER_GEOMETRY:
2347 return chip_class >= GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2348 R_00B230_SPI_SHADER_USER_DATA_GS_0;
2349 case MESA_SHADER_COMPUTE:
2350 return R_00B900_COMPUTE_USER_DATA_0;
2351 case MESA_SHADER_TESS_CTRL:
2352 return chip_class >= GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
2353 R_00B430_SPI_SHADER_USER_DATA_HS_0;
2354 case MESA_SHADER_TESS_EVAL:
2355 if (chip_class >= GFX9) {
2356 return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2357 R_00B130_SPI_SHADER_USER_DATA_VS_0;
2358 }
2359 if (has_gs)
2360 return R_00B330_SPI_SHADER_USER_DATA_ES_0;
2361 else
2362 return R_00B130_SPI_SHADER_USER_DATA_VS_0;
2363 default:
2364 unreachable("unknown shader");
2365 }
2366 }
2367
2368 struct radv_bin_size_entry {
2369 unsigned bpp;
2370 VkExtent2D extent;
2371 };
2372
2373 static VkExtent2D
2374 radv_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo)
2375 {
2376 static const struct radv_bin_size_entry color_size_table[][3][9] = {
2377 {
2378 /* One RB / SE */
2379 {
2380 /* One shader engine */
2381 { 0, {128, 128}},
2382 { 1, { 64, 128}},
2383 { 2, { 32, 128}},
2384 { 3, { 16, 128}},
2385 { 17, { 0, 0}},
2386 { UINT_MAX, { 0, 0}},
2387 },
2388 {
2389 /* Two shader engines */
2390 { 0, {128, 128}},
2391 { 2, { 64, 128}},
2392 { 3, { 32, 128}},
2393 { 5, { 16, 128}},
2394 { 17, { 0, 0}},
2395 { UINT_MAX, { 0, 0}},
2396 },
2397 {
2398 /* Four shader engines */
2399 { 0, {128, 128}},
2400 { 3, { 64, 128}},
2401 { 5, { 16, 128}},
2402 { 17, { 0, 0}},
2403 { UINT_MAX, { 0, 0}},
2404 },
2405 },
2406 {
2407 /* Two RB / SE */
2408 {
2409 /* One shader engine */
2410 { 0, {128, 128}},
2411 { 2, { 64, 128}},
2412 { 3, { 32, 128}},
2413 { 5, { 16, 128}},
2414 { 33, { 0, 0}},
2415 { UINT_MAX, { 0, 0}},
2416 },
2417 {
2418 /* Two shader engines */
2419 { 0, {128, 128}},
2420 { 3, { 64, 128}},
2421 { 5, { 32, 128}},
2422 { 9, { 16, 128}},
2423 { 33, { 0, 0}},
2424 { UINT_MAX, { 0, 0}},
2425 },
2426 {
2427 /* Four shader engines */
2428 { 0, {256, 256}},
2429 { 2, {128, 256}},
2430 { 3, {128, 128}},
2431 { 5, { 64, 128}},
2432 { 9, { 16, 128}},
2433 { 33, { 0, 0}},
2434 { UINT_MAX, { 0, 0}},
2435 },
2436 },
2437 {
2438 /* Four RB / SE */
2439 {
2440 /* One shader engine */
2441 { 0, {128, 256}},
2442 { 2, {128, 128}},
2443 { 3, { 64, 128}},
2444 { 5, { 32, 128}},
2445 { 9, { 16, 128}},
2446 { 33, { 0, 0}},
2447 { UINT_MAX, { 0, 0}},
2448 },
2449 {
2450 /* Two shader engines */
2451 { 0, {256, 256}},
2452 { 2, {128, 256}},
2453 { 3, {128, 128}},
2454 { 5, { 64, 128}},
2455 { 9, { 32, 128}},
2456 { 17, { 16, 128}},
2457 { 33, { 0, 0}},
2458 { UINT_MAX, { 0, 0}},
2459 },
2460 {
2461 /* Four shader engines */
2462 { 0, {256, 512}},
2463 { 2, {256, 256}},
2464 { 3, {128, 256}},
2465 { 5, {128, 128}},
2466 { 9, { 64, 128}},
2467 { 17, { 16, 128}},
2468 { 33, { 0, 0}},
2469 { UINT_MAX, { 0, 0}},
2470 },
2471 },
2472 };
2473 static const struct radv_bin_size_entry ds_size_table[][3][9] = {
2474 {
2475 // One RB / SE
2476 {
2477 // One shader engine
2478 { 0, {128, 256}},
2479 { 2, {128, 128}},
2480 { 4, { 64, 128}},
2481 { 7, { 32, 128}},
2482 { 13, { 16, 128}},
2483 { 49, { 0, 0}},
2484 { UINT_MAX, { 0, 0}},
2485 },
2486 {
2487 // Two shader engines
2488 { 0, {256, 256}},
2489 { 2, {128, 256}},
2490 { 4, {128, 128}},
2491 { 7, { 64, 128}},
2492 { 13, { 32, 128}},
2493 { 25, { 16, 128}},
2494 { 49, { 0, 0}},
2495 { UINT_MAX, { 0, 0}},
2496 },
2497 {
2498 // Four shader engines
2499 { 0, {256, 512}},
2500 { 2, {256, 256}},
2501 { 4, {128, 256}},
2502 { 7, {128, 128}},
2503 { 13, { 64, 128}},
2504 { 25, { 16, 128}},
2505 { 49, { 0, 0}},
2506 { UINT_MAX, { 0, 0}},
2507 },
2508 },
2509 {
2510 // Two RB / SE
2511 {
2512 // One shader engine
2513 { 0, {256, 256}},
2514 { 2, {128, 256}},
2515 { 4, {128, 128}},
2516 { 7, { 64, 128}},
2517 { 13, { 32, 128}},
2518 { 25, { 16, 128}},
2519 { 97, { 0, 0}},
2520 { UINT_MAX, { 0, 0}},
2521 },
2522 {
2523 // Two shader engines
2524 { 0, {256, 512}},
2525 { 2, {256, 256}},
2526 { 4, {128, 256}},
2527 { 7, {128, 128}},
2528 { 13, { 64, 128}},
2529 { 25, { 32, 128}},
2530 { 49, { 16, 128}},
2531 { 97, { 0, 0}},
2532 { UINT_MAX, { 0, 0}},
2533 },
2534 {
2535 // Four shader engines
2536 { 0, {512, 512}},
2537 { 2, {256, 512}},
2538 { 4, {256, 256}},
2539 { 7, {128, 256}},
2540 { 13, {128, 128}},
2541 { 25, { 64, 128}},
2542 { 49, { 16, 128}},
2543 { 97, { 0, 0}},
2544 { UINT_MAX, { 0, 0}},
2545 },
2546 },
2547 {
2548 // Four RB / SE
2549 {
2550 // One shader engine
2551 { 0, {256, 512}},
2552 { 2, {256, 256}},
2553 { 4, {128, 256}},
2554 { 7, {128, 128}},
2555 { 13, { 64, 128}},
2556 { 25, { 32, 128}},
2557 { 49, { 16, 128}},
2558 { UINT_MAX, { 0, 0}},
2559 },
2560 {
2561 // Two shader engines
2562 { 0, {512, 512}},
2563 { 2, {256, 512}},
2564 { 4, {256, 256}},
2565 { 7, {128, 256}},
2566 { 13, {128, 128}},
2567 { 25, { 64, 128}},
2568 { 49, { 32, 128}},
2569 { 97, { 16, 128}},
2570 { UINT_MAX, { 0, 0}},
2571 },
2572 {
2573 // Four shader engines
2574 { 0, {512, 512}},
2575 { 4, {256, 512}},
2576 { 7, {256, 256}},
2577 { 13, {128, 256}},
2578 { 25, {128, 128}},
2579 { 49, { 64, 128}},
2580 { 97, { 16, 128}},
2581 { UINT_MAX, { 0, 0}},
2582 },
2583 },
2584 };
2585
2586 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
2587 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
2588 VkExtent2D extent = {512, 512};
2589
2590 unsigned log_num_rb_per_se =
2591 util_logbase2_ceil(pipeline->device->physical_device->rad_info.num_render_backends /
2592 pipeline->device->physical_device->rad_info.max_se);
2593 unsigned log_num_se = util_logbase2_ceil(pipeline->device->physical_device->rad_info.max_se);
2594
2595 unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
2596 unsigned ps_iter_samples = 1u << G_028804_PS_ITER_SAMPLES(pipeline->graphics.ms.db_eqaa);
2597 unsigned effective_samples = total_samples;
2598 unsigned color_bytes_per_pixel = 0;
2599
2600 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
2601 if (vkblend) {
2602 for (unsigned i = 0; i < subpass->color_count; i++) {
2603 if (!vkblend->pAttachments[i].colorWriteMask)
2604 continue;
2605
2606 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
2607 continue;
2608
2609 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
2610 color_bytes_per_pixel += vk_format_get_blocksize(format);
2611 }
2612
2613 /* MSAA images typically don't use all samples all the time. */
2614 if (effective_samples >= 2 && ps_iter_samples <= 1)
2615 effective_samples = 2;
2616 color_bytes_per_pixel *= effective_samples;
2617 }
2618
2619 const struct radv_bin_size_entry *color_entry = color_size_table[log_num_rb_per_se][log_num_se];
2620 while(color_entry[1].bpp <= color_bytes_per_pixel)
2621 ++color_entry;
2622
2623 extent = color_entry->extent;
2624
2625 if (subpass->depth_stencil_attachment) {
2626 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
2627
2628 /* Coefficients taken from AMDVLK */
2629 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
2630 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0;
2631 unsigned ds_bytes_per_pixel = 4 * (depth_coeff + stencil_coeff) * total_samples;
2632
2633 const struct radv_bin_size_entry *ds_entry = ds_size_table[log_num_rb_per_se][log_num_se];
2634 while(ds_entry[1].bpp <= ds_bytes_per_pixel)
2635 ++ds_entry;
2636
2637 extent.width = MIN2(extent.width, ds_entry->extent.width);
2638 extent.height = MIN2(extent.height, ds_entry->extent.height);
2639 }
2640
2641 return extent;
2642 }
2643
2644 static void
2645 radv_pipeline_generate_binning_state(struct radeon_cmdbuf *ctx_cs,
2646 struct radv_pipeline *pipeline,
2647 const VkGraphicsPipelineCreateInfo *pCreateInfo)
2648 {
2649 if (pipeline->device->physical_device->rad_info.chip_class < GFX9)
2650 return;
2651
2652 uint32_t pa_sc_binner_cntl_0 =
2653 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
2654 S_028C44_DISABLE_START_OF_PRIM(1);
2655 uint32_t db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF);
2656
2657 VkExtent2D bin_size = radv_compute_bin_size(pipeline, pCreateInfo);
2658
2659 unsigned context_states_per_bin; /* allowed range: [1, 6] */
2660 unsigned persistent_states_per_bin; /* allowed range: [1, 32] */
2661 unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */
2662
2663 switch (pipeline->device->physical_device->rad_info.family) {
2664 case CHIP_VEGA10:
2665 case CHIP_VEGA12:
2666 case CHIP_VEGA20:
2667 context_states_per_bin = 1;
2668 persistent_states_per_bin = 1;
2669 fpovs_per_batch = 63;
2670 break;
2671 case CHIP_RAVEN:
2672 case CHIP_RAVEN2:
2673 context_states_per_bin = 6;
2674 persistent_states_per_bin = 32;
2675 fpovs_per_batch = 63;
2676 break;
2677 default:
2678 unreachable("unhandled family while determining binning state.");
2679 }
2680
2681 if (pipeline->device->pbb_allowed && bin_size.width && bin_size.height) {
2682 pa_sc_binner_cntl_0 =
2683 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) |
2684 S_028C44_BIN_SIZE_X(bin_size.width == 16) |
2685 S_028C44_BIN_SIZE_Y(bin_size.height == 16) |
2686 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size.width, 32)) - 5) |
2687 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size.height, 32)) - 5) |
2688 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin - 1) |
2689 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin - 1) |
2690 S_028C44_DISABLE_START_OF_PRIM(1) |
2691 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch) |
2692 S_028C44_OPTIMAL_BIN_SELECTION(1);
2693 }
2694
2695 radeon_set_context_reg(ctx_cs, R_028C44_PA_SC_BINNER_CNTL_0,
2696 pa_sc_binner_cntl_0);
2697 radeon_set_context_reg(ctx_cs, R_028060_DB_DFSM_CONTROL,
2698 db_dfsm_control);
2699 }
2700
2701
2702 static void
2703 radv_pipeline_generate_depth_stencil_state(struct radeon_cmdbuf *ctx_cs,
2704 struct radv_pipeline *pipeline,
2705 const VkGraphicsPipelineCreateInfo *pCreateInfo,
2706 const struct radv_graphics_pipeline_create_info *extra)
2707 {
2708 const VkPipelineDepthStencilStateCreateInfo *vkds = pCreateInfo->pDepthStencilState;
2709 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
2710 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
2711 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
2712 struct radv_render_pass_attachment *attachment = NULL;
2713 uint32_t db_depth_control = 0, db_stencil_control = 0;
2714 uint32_t db_render_control = 0, db_render_override2 = 0;
2715 uint32_t db_render_override = 0;
2716
2717 if (subpass->depth_stencil_attachment)
2718 attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
2719
2720 bool has_depth_attachment = attachment && vk_format_is_depth(attachment->format);
2721 bool has_stencil_attachment = attachment && vk_format_is_stencil(attachment->format);
2722
2723 if (vkds && has_depth_attachment) {
2724 db_depth_control = S_028800_Z_ENABLE(vkds->depthTestEnable ? 1 : 0) |
2725 S_028800_Z_WRITE_ENABLE(vkds->depthWriteEnable ? 1 : 0) |
2726 S_028800_ZFUNC(vkds->depthCompareOp) |
2727 S_028800_DEPTH_BOUNDS_ENABLE(vkds->depthBoundsTestEnable ? 1 : 0);
2728
2729 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
2730 db_render_override2 |= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment->samples > 2);
2731 }
2732
2733 if (has_stencil_attachment && vkds && vkds->stencilTestEnable) {
2734 db_depth_control |= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
2735 db_depth_control |= S_028800_STENCILFUNC(vkds->front.compareOp);
2736 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds->front.failOp));
2737 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds->front.passOp));
2738 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds->front.depthFailOp));
2739
2740 db_depth_control |= S_028800_STENCILFUNC_BF(vkds->back.compareOp);
2741 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds->back.failOp));
2742 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds->back.passOp));
2743 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds->back.depthFailOp));
2744 }
2745
2746 if (attachment && extra) {
2747 db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(extra->db_depth_clear);
2748 db_render_control |= S_028000_STENCIL_CLEAR_ENABLE(extra->db_stencil_clear);
2749
2750 db_render_control |= S_028000_RESUMMARIZE_ENABLE(extra->db_resummarize);
2751 db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(extra->db_flush_depth_inplace);
2752 db_render_control |= S_028000_STENCIL_COMPRESS_DISABLE(extra->db_flush_stencil_inplace);
2753 db_render_override2 |= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra->db_depth_disable_expclear);
2754 db_render_override2 |= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra->db_stencil_disable_expclear);
2755 }
2756
2757 db_render_override |= S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
2758 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE);
2759
2760 if (!pCreateInfo->pRasterizationState->depthClampEnable &&
2761 ps->info.info.ps.writes_z) {
2762 /* From VK_EXT_depth_range_unrestricted spec:
2763 *
2764 * "The behavior described in Primitive Clipping still applies.
2765 * If depth clamping is disabled the depth values are still
2766 * clipped to 0 ≤ zc ≤ wc before the viewport transform. If
2767 * depth clamping is enabled the above equation is ignored and
2768 * the depth values are instead clamped to the VkViewport
2769 * minDepth and maxDepth values, which in the case of this
2770 * extension can be outside of the 0.0 to 1.0 range."
2771 */
2772 db_render_override |= S_02800C_DISABLE_VIEWPORT_CLAMP(1);
2773 }
2774
2775 radeon_set_context_reg(ctx_cs, R_028800_DB_DEPTH_CONTROL, db_depth_control);
2776 radeon_set_context_reg(ctx_cs, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
2777
2778 radeon_set_context_reg(ctx_cs, R_028000_DB_RENDER_CONTROL, db_render_control);
2779 radeon_set_context_reg(ctx_cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
2780 radeon_set_context_reg(ctx_cs, R_028010_DB_RENDER_OVERRIDE2, db_render_override2);
2781 }
2782
2783 static void
2784 radv_pipeline_generate_blend_state(struct radeon_cmdbuf *ctx_cs,
2785 struct radv_pipeline *pipeline,
2786 const struct radv_blend_state *blend)
2787 {
2788 radeon_set_context_reg_seq(ctx_cs, R_028780_CB_BLEND0_CONTROL, 8);
2789 radeon_emit_array(ctx_cs, blend->cb_blend_control,
2790 8);
2791 radeon_set_context_reg(ctx_cs, R_028808_CB_COLOR_CONTROL, blend->cb_color_control);
2792 radeon_set_context_reg(ctx_cs, R_028B70_DB_ALPHA_TO_MASK, blend->db_alpha_to_mask);
2793
2794 if (pipeline->device->physical_device->has_rbplus) {
2795
2796 radeon_set_context_reg_seq(ctx_cs, R_028760_SX_MRT0_BLEND_OPT, 8);
2797 radeon_emit_array(ctx_cs, blend->sx_mrt_blend_opt, 8);
2798 }
2799
2800 radeon_set_context_reg(ctx_cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
2801
2802 radeon_set_context_reg(ctx_cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
2803 radeon_set_context_reg(ctx_cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
2804
2805 pipeline->graphics.col_format = blend->spi_shader_col_format;
2806 pipeline->graphics.cb_target_mask = blend->cb_target_mask;
2807 }
2808
2809 static const VkConservativeRasterizationModeEXT
2810 radv_get_conservative_raster_mode(const VkPipelineRasterizationStateCreateInfo *pCreateInfo)
2811 {
2812 const VkPipelineRasterizationConservativeStateCreateInfoEXT *conservative_raster =
2813 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_RASTERIZATION_CONSERVATIVE_STATE_CREATE_INFO_EXT);
2814
2815 if (!conservative_raster)
2816 return VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT;
2817 return conservative_raster->conservativeRasterizationMode;
2818 }
2819
2820 static void
2821 radv_pipeline_generate_raster_state(struct radeon_cmdbuf *ctx_cs,
2822 struct radv_pipeline *pipeline,
2823 const VkGraphicsPipelineCreateInfo *pCreateInfo)
2824 {
2825 const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState;
2826 const VkConservativeRasterizationModeEXT mode =
2827 radv_get_conservative_raster_mode(vkraster);
2828 uint32_t pa_sc_conservative_rast = S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1);
2829 bool depth_clip_disable = vkraster->depthClampEnable;
2830
2831 const VkPipelineRasterizationDepthClipStateCreateInfoEXT *depth_clip_state =
2832 vk_find_struct_const(vkraster->pNext, PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT);
2833 if (depth_clip_state) {
2834 depth_clip_disable = !depth_clip_state->depthClipEnable;
2835 }
2836
2837 radeon_set_context_reg(ctx_cs, R_028810_PA_CL_CLIP_CNTL,
2838 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
2839 S_028810_ZCLIP_NEAR_DISABLE(depth_clip_disable ? 1 : 0) |
2840 S_028810_ZCLIP_FAR_DISABLE(depth_clip_disable ? 1 : 0) |
2841 S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) |
2842 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
2843
2844 radeon_set_context_reg(ctx_cs, R_0286D4_SPI_INTERP_CONTROL_0,
2845 S_0286D4_FLAT_SHADE_ENA(1) |
2846 S_0286D4_PNT_SPRITE_ENA(1) |
2847 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
2848 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
2849 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
2850 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
2851 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
2852
2853 radeon_set_context_reg(ctx_cs, R_028BE4_PA_SU_VTX_CNTL,
2854 S_028BE4_PIX_CENTER(1) | // TODO verify
2855 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) |
2856 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
2857
2858 radeon_set_context_reg(ctx_cs, R_028814_PA_SU_SC_MODE_CNTL,
2859 S_028814_FACE(vkraster->frontFace) |
2860 S_028814_CULL_FRONT(!!(vkraster->cullMode & VK_CULL_MODE_FRONT_BIT)) |
2861 S_028814_CULL_BACK(!!(vkraster->cullMode & VK_CULL_MODE_BACK_BIT)) |
2862 S_028814_POLY_MODE(vkraster->polygonMode != VK_POLYGON_MODE_FILL) |
2863 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster->polygonMode)) |
2864 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster->polygonMode)) |
2865 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
2866 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
2867 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster->depthBiasEnable ? 1 : 0));
2868
2869 /* Conservative rasterization. */
2870 if (mode != VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
2871 struct radv_multisample_state *ms = &pipeline->graphics.ms;
2872
2873 ms->pa_sc_aa_config |= S_028BE0_AA_MASK_CENTROID_DTMN(1);
2874 ms->db_eqaa |= S_028804_ENABLE_POSTZ_OVERRASTERIZATION(1) |
2875 S_028804_OVERRASTERIZATION_AMOUNT(4);
2876
2877 pa_sc_conservative_rast = S_028C4C_PREZ_AA_MASK_ENABLE(1) |
2878 S_028C4C_POSTZ_AA_MASK_ENABLE(1) |
2879 S_028C4C_CENTROID_SAMPLE_OVERRIDE(1);
2880
2881 if (mode == VK_CONSERVATIVE_RASTERIZATION_MODE_OVERESTIMATE_EXT) {
2882 pa_sc_conservative_rast |=
2883 S_028C4C_OVER_RAST_ENABLE(1) |
2884 S_028C4C_OVER_RAST_SAMPLE_SELECT(0) |
2885 S_028C4C_UNDER_RAST_ENABLE(0) |
2886 S_028C4C_UNDER_RAST_SAMPLE_SELECT(1) |
2887 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(1);
2888 } else {
2889 assert(mode == VK_CONSERVATIVE_RASTERIZATION_MODE_UNDERESTIMATE_EXT);
2890 pa_sc_conservative_rast |=
2891 S_028C4C_OVER_RAST_ENABLE(0) |
2892 S_028C4C_OVER_RAST_SAMPLE_SELECT(1) |
2893 S_028C4C_UNDER_RAST_ENABLE(1) |
2894 S_028C4C_UNDER_RAST_SAMPLE_SELECT(0) |
2895 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(0);
2896 }
2897 }
2898
2899 radeon_set_context_reg(ctx_cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
2900 pa_sc_conservative_rast);
2901 }
2902
2903
2904 static void
2905 radv_pipeline_generate_multisample_state(struct radeon_cmdbuf *ctx_cs,
2906 struct radv_pipeline *pipeline)
2907 {
2908 struct radv_multisample_state *ms = &pipeline->graphics.ms;
2909
2910 radeon_set_context_reg_seq(ctx_cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
2911 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[0]);
2912 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[1]);
2913
2914 radeon_set_context_reg(ctx_cs, R_028804_DB_EQAA, ms->db_eqaa);
2915 radeon_set_context_reg(ctx_cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
2916
2917 /* The exclusion bits can be set to improve rasterization efficiency
2918 * if no sample lies on the pixel boundary (-8 sample offset). It's
2919 * currently always TRUE because the driver doesn't support 16 samples.
2920 */
2921 bool exclusion = pipeline->device->physical_device->rad_info.chip_class >= GFX7;
2922 radeon_set_context_reg(ctx_cs, R_02882C_PA_SU_PRIM_FILTER_CNTL,
2923 S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
2924 S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
2925 }
2926
2927 static void
2928 radv_pipeline_generate_vgt_gs_mode(struct radeon_cmdbuf *ctx_cs,
2929 struct radv_pipeline *pipeline)
2930 {
2931 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
2932
2933 uint32_t vgt_primitiveid_en = false;
2934 uint32_t vgt_gs_mode = 0;
2935
2936 if (radv_pipeline_has_gs(pipeline)) {
2937 const struct radv_shader_variant *gs =
2938 pipeline->shaders[MESA_SHADER_GEOMETRY];
2939
2940 vgt_gs_mode = ac_vgt_gs_mode(gs->info.gs.vertices_out,
2941 pipeline->device->physical_device->rad_info.chip_class);
2942 } else if (outinfo->export_prim_id) {
2943 vgt_gs_mode = S_028A40_MODE(V_028A40_GS_SCENARIO_A);
2944 vgt_primitiveid_en = true;
2945 }
2946
2947 radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN, vgt_primitiveid_en);
2948 radeon_set_context_reg(ctx_cs, R_028A40_VGT_GS_MODE, vgt_gs_mode);
2949 }
2950
2951 static void
2952 radv_pipeline_generate_hw_vs(struct radeon_cmdbuf *ctx_cs,
2953 struct radeon_cmdbuf *cs,
2954 struct radv_pipeline *pipeline,
2955 struct radv_shader_variant *shader)
2956 {
2957 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
2958
2959 radeon_set_sh_reg_seq(cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
2960 radeon_emit(cs, va >> 8);
2961 radeon_emit(cs, S_00B124_MEM_BASE(va >> 40));
2962 radeon_emit(cs, shader->rsrc1);
2963 radeon_emit(cs, shader->rsrc2);
2964
2965 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
2966 unsigned clip_dist_mask, cull_dist_mask, total_mask;
2967 clip_dist_mask = outinfo->clip_dist_mask;
2968 cull_dist_mask = outinfo->cull_dist_mask;
2969 total_mask = clip_dist_mask | cull_dist_mask;
2970 bool misc_vec_ena = outinfo->writes_pointsize ||
2971 outinfo->writes_layer ||
2972 outinfo->writes_viewport_index;
2973
2974 radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG,
2975 S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo->param_exports) - 1));
2976
2977 radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
2978 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
2979 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
2980 V_02870C_SPI_SHADER_4COMP :
2981 V_02870C_SPI_SHADER_NONE) |
2982 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
2983 V_02870C_SPI_SHADER_4COMP :
2984 V_02870C_SPI_SHADER_NONE) |
2985 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
2986 V_02870C_SPI_SHADER_4COMP :
2987 V_02870C_SPI_SHADER_NONE));
2988
2989 radeon_set_context_reg(ctx_cs, R_028818_PA_CL_VTE_CNTL,
2990 S_028818_VTX_W0_FMT(1) |
2991 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
2992 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
2993 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
2994
2995 radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
2996 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
2997 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
2998 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
2999 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
3000 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
3001 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
3002 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
3003 cull_dist_mask << 8 |
3004 clip_dist_mask);
3005
3006 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
3007 radeon_set_context_reg(ctx_cs, R_028AB4_VGT_REUSE_OFF,
3008 outinfo->writes_viewport_index);
3009 }
3010
3011 static void
3012 radv_pipeline_generate_hw_es(struct radeon_cmdbuf *cs,
3013 struct radv_pipeline *pipeline,
3014 struct radv_shader_variant *shader)
3015 {
3016 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3017
3018 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 4);
3019 radeon_emit(cs, va >> 8);
3020 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
3021 radeon_emit(cs, shader->rsrc1);
3022 radeon_emit(cs, shader->rsrc2);
3023 }
3024
3025 static void
3026 radv_pipeline_generate_hw_ls(struct radeon_cmdbuf *cs,
3027 struct radv_pipeline *pipeline,
3028 struct radv_shader_variant *shader,
3029 const struct radv_tessellation_state *tess)
3030 {
3031 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3032 uint32_t rsrc2 = shader->rsrc2;
3033
3034 radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
3035 radeon_emit(cs, va >> 8);
3036 radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
3037
3038 rsrc2 |= S_00B52C_LDS_SIZE(tess->lds_size);
3039 if (pipeline->device->physical_device->rad_info.chip_class == GFX7 &&
3040 pipeline->device->physical_device->rad_info.family != CHIP_HAWAII)
3041 radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, rsrc2);
3042
3043 radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
3044 radeon_emit(cs, shader->rsrc1);
3045 radeon_emit(cs, rsrc2);
3046 }
3047
3048 static void
3049 radv_pipeline_generate_hw_hs(struct radeon_cmdbuf *cs,
3050 struct radv_pipeline *pipeline,
3051 struct radv_shader_variant *shader,
3052 const struct radv_tessellation_state *tess)
3053 {
3054 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3055
3056 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
3057 radeon_set_sh_reg_seq(cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
3058 radeon_emit(cs, va >> 8);
3059 radeon_emit(cs, S_00B414_MEM_BASE(va >> 40));
3060
3061 radeon_set_sh_reg_seq(cs, R_00B428_SPI_SHADER_PGM_RSRC1_HS, 2);
3062 radeon_emit(cs, shader->rsrc1);
3063 radeon_emit(cs, shader->rsrc2 |
3064 S_00B42C_LDS_SIZE(tess->lds_size));
3065 } else {
3066 radeon_set_sh_reg_seq(cs, R_00B420_SPI_SHADER_PGM_LO_HS, 4);
3067 radeon_emit(cs, va >> 8);
3068 radeon_emit(cs, S_00B424_MEM_BASE(va >> 40));
3069 radeon_emit(cs, shader->rsrc1);
3070 radeon_emit(cs, shader->rsrc2);
3071 }
3072 }
3073
3074 static void
3075 radv_pipeline_generate_vertex_shader(struct radeon_cmdbuf *ctx_cs,
3076 struct radeon_cmdbuf *cs,
3077 struct radv_pipeline *pipeline,
3078 const struct radv_tessellation_state *tess)
3079 {
3080 struct radv_shader_variant *vs;
3081
3082 /* Skip shaders merged into HS/GS */
3083 vs = pipeline->shaders[MESA_SHADER_VERTEX];
3084 if (!vs)
3085 return;
3086
3087 if (vs->info.vs.as_ls)
3088 radv_pipeline_generate_hw_ls(cs, pipeline, vs, tess);
3089 else if (vs->info.vs.as_es)
3090 radv_pipeline_generate_hw_es(cs, pipeline, vs);
3091 else
3092 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, vs);
3093 }
3094
3095 static void
3096 radv_pipeline_generate_tess_shaders(struct radeon_cmdbuf *ctx_cs,
3097 struct radeon_cmdbuf *cs,
3098 struct radv_pipeline *pipeline,
3099 const struct radv_tessellation_state *tess)
3100 {
3101 if (!radv_pipeline_has_tess(pipeline))
3102 return;
3103
3104 struct radv_shader_variant *tes, *tcs;
3105
3106 tcs = pipeline->shaders[MESA_SHADER_TESS_CTRL];
3107 tes = pipeline->shaders[MESA_SHADER_TESS_EVAL];
3108
3109 if (tes) {
3110 if (tes->info.tes.as_es)
3111 radv_pipeline_generate_hw_es(cs, pipeline, tes);
3112 else
3113 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, tes);
3114 }
3115
3116 radv_pipeline_generate_hw_hs(cs, pipeline, tcs, tess);
3117
3118 radeon_set_context_reg(ctx_cs, R_028B6C_VGT_TF_PARAM,
3119 tess->tf_param);
3120
3121 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7)
3122 radeon_set_context_reg_idx(ctx_cs, R_028B58_VGT_LS_HS_CONFIG, 2,
3123 tess->ls_hs_config);
3124 else
3125 radeon_set_context_reg(ctx_cs, R_028B58_VGT_LS_HS_CONFIG,
3126 tess->ls_hs_config);
3127 }
3128
3129 static void
3130 radv_pipeline_generate_geometry_shader(struct radeon_cmdbuf *ctx_cs,
3131 struct radeon_cmdbuf *cs,
3132 struct radv_pipeline *pipeline,
3133 const struct radv_gs_state *gs_state)
3134 {
3135 struct radv_shader_variant *gs;
3136 unsigned gs_max_out_vertices;
3137 uint8_t *num_components;
3138 uint8_t max_stream;
3139 unsigned offset;
3140 uint64_t va;
3141
3142 gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
3143 if (!gs)
3144 return;
3145
3146 gs_max_out_vertices = gs->info.gs.vertices_out;
3147 max_stream = gs->info.info.gs.max_stream;
3148 num_components = gs->info.info.gs.num_stream_output_components;
3149
3150 offset = num_components[0] * gs_max_out_vertices;
3151
3152 radeon_set_context_reg_seq(ctx_cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
3153 radeon_emit(ctx_cs, offset);
3154 if (max_stream >= 1)
3155 offset += num_components[1] * gs_max_out_vertices;
3156 radeon_emit(ctx_cs, offset);
3157 if (max_stream >= 2)
3158 offset += num_components[2] * gs_max_out_vertices;
3159 radeon_emit(ctx_cs, offset);
3160 if (max_stream >= 3)
3161 offset += num_components[3] * gs_max_out_vertices;
3162 radeon_set_context_reg(ctx_cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
3163
3164 radeon_set_context_reg(ctx_cs, R_028B38_VGT_GS_MAX_VERT_OUT, gs->info.gs.vertices_out);
3165
3166 radeon_set_context_reg_seq(ctx_cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
3167 radeon_emit(ctx_cs, num_components[0]);
3168 radeon_emit(ctx_cs, (max_stream >= 1) ? num_components[1] : 0);
3169 radeon_emit(ctx_cs, (max_stream >= 2) ? num_components[2] : 0);
3170 radeon_emit(ctx_cs, (max_stream >= 3) ? num_components[3] : 0);
3171
3172 uint32_t gs_num_invocations = gs->info.gs.invocations;
3173 radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
3174 S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
3175 S_028B90_ENABLE(gs_num_invocations > 0));
3176
3177 radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
3178 gs_state->vgt_esgs_ring_itemsize);
3179
3180 va = radv_buffer_get_va(gs->bo) + gs->bo_offset;
3181
3182 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
3183 radeon_set_sh_reg_seq(cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
3184 radeon_emit(cs, va >> 8);
3185 radeon_emit(cs, S_00B214_MEM_BASE(va >> 40));
3186
3187 radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
3188 radeon_emit(cs, gs->rsrc1);
3189 radeon_emit(cs, gs->rsrc2 | S_00B22C_LDS_SIZE(gs_state->lds_size));
3190
3191 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL, gs_state->vgt_gs_onchip_cntl);
3192 radeon_set_context_reg(ctx_cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, gs_state->vgt_gs_max_prims_per_subgroup);
3193 } else {
3194 radeon_set_sh_reg_seq(cs, R_00B220_SPI_SHADER_PGM_LO_GS, 4);
3195 radeon_emit(cs, va >> 8);
3196 radeon_emit(cs, S_00B224_MEM_BASE(va >> 40));
3197 radeon_emit(cs, gs->rsrc1);
3198 radeon_emit(cs, gs->rsrc2);
3199 }
3200
3201 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, pipeline->gs_copy_shader);
3202 }
3203
3204 static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade, bool float16)
3205 {
3206 uint32_t ps_input_cntl;
3207 if (offset <= AC_EXP_PARAM_OFFSET_31) {
3208 ps_input_cntl = S_028644_OFFSET(offset);
3209 if (flat_shade)
3210 ps_input_cntl |= S_028644_FLAT_SHADE(1);
3211 if (float16) {
3212 ps_input_cntl |= S_028644_FP16_INTERP_MODE(1) |
3213 S_028644_ATTR0_VALID(1);
3214 }
3215 } else {
3216 /* The input is a DEFAULT_VAL constant. */
3217 assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
3218 offset <= AC_EXP_PARAM_DEFAULT_VAL_1111);
3219 offset -= AC_EXP_PARAM_DEFAULT_VAL_0000;
3220 ps_input_cntl = S_028644_OFFSET(0x20) |
3221 S_028644_DEFAULT_VAL(offset);
3222 }
3223 return ps_input_cntl;
3224 }
3225
3226 static void
3227 radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
3228 struct radv_pipeline *pipeline)
3229 {
3230 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3231 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3232 uint32_t ps_input_cntl[32];
3233
3234 unsigned ps_offset = 0;
3235
3236 if (ps->info.info.ps.prim_id_input) {
3237 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
3238 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
3239 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
3240 ++ps_offset;
3241 }
3242 }
3243
3244 if (ps->info.info.ps.layer_input ||
3245 ps->info.info.ps.uses_input_attachments ||
3246 ps->info.info.needs_multiview_view_index) {
3247 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
3248 if (vs_offset != AC_EXP_PARAM_UNDEFINED)
3249 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
3250 else
3251 ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false);
3252 ++ps_offset;
3253 }
3254
3255 if (ps->info.info.ps.has_pcoord) {
3256 unsigned val;
3257 val = S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
3258 ps_input_cntl[ps_offset] = val;
3259 ps_offset++;
3260 }
3261
3262 if (ps->info.info.ps.num_input_clips_culls) {
3263 unsigned vs_offset;
3264
3265 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0];
3266 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
3267 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
3268 ++ps_offset;
3269 }
3270
3271 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1];
3272 if (vs_offset != AC_EXP_PARAM_UNDEFINED &&
3273 ps->info.info.ps.num_input_clips_culls > 4) {
3274 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
3275 ++ps_offset;
3276 }
3277 }
3278
3279 for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.fs.input_mask; ++i) {
3280 unsigned vs_offset;
3281 bool flat_shade;
3282 bool float16;
3283 if (!(ps->info.fs.input_mask & (1u << i)))
3284 continue;
3285
3286 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VAR0 + i];
3287 if (vs_offset == AC_EXP_PARAM_UNDEFINED) {
3288 ps_input_cntl[ps_offset] = S_028644_OFFSET(0x20);
3289 ++ps_offset;
3290 continue;
3291 }
3292
3293 flat_shade = !!(ps->info.fs.flat_shaded_mask & (1u << ps_offset));
3294 float16 = !!(ps->info.fs.float16_shaded_mask & (1u << ps_offset));
3295
3296 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade, float16);
3297 ++ps_offset;
3298 }
3299
3300 if (ps_offset) {
3301 radeon_set_context_reg_seq(ctx_cs, R_028644_SPI_PS_INPUT_CNTL_0, ps_offset);
3302 for (unsigned i = 0; i < ps_offset; i++) {
3303 radeon_emit(ctx_cs, ps_input_cntl[i]);
3304 }
3305 }
3306 }
3307
3308 static uint32_t
3309 radv_compute_db_shader_control(const struct radv_device *device,
3310 const struct radv_pipeline *pipeline,
3311 const struct radv_shader_variant *ps)
3312 {
3313 unsigned z_order;
3314 if (ps->info.fs.early_fragment_test || !ps->info.info.ps.writes_memory)
3315 z_order = V_02880C_EARLY_Z_THEN_LATE_Z;
3316 else
3317 z_order = V_02880C_LATE_Z;
3318
3319 bool disable_rbplus = device->physical_device->has_rbplus &&
3320 !device->physical_device->rbplus_allowed;
3321
3322 /* It shouldn't be needed to export gl_SampleMask when MSAA is disabled
3323 * but this appears to break Project Cars (DXVK). See
3324 * https://bugs.freedesktop.org/show_bug.cgi?id=109401
3325 */
3326 bool mask_export_enable = ps->info.info.ps.writes_sample_mask;
3327
3328 return S_02880C_Z_EXPORT_ENABLE(ps->info.info.ps.writes_z) |
3329 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.info.ps.writes_stencil) |
3330 S_02880C_KILL_ENABLE(!!ps->info.fs.can_discard) |
3331 S_02880C_MASK_EXPORT_ENABLE(mask_export_enable) |
3332 S_02880C_Z_ORDER(z_order) |
3333 S_02880C_DEPTH_BEFORE_SHADER(ps->info.fs.early_fragment_test) |
3334 S_02880C_EXEC_ON_HIER_FAIL(ps->info.info.ps.writes_memory) |
3335 S_02880C_EXEC_ON_NOOP(ps->info.info.ps.writes_memory) |
3336 S_02880C_DUAL_QUAD_DISABLE(disable_rbplus);
3337 }
3338
3339 static void
3340 radv_pipeline_generate_fragment_shader(struct radeon_cmdbuf *ctx_cs,
3341 struct radeon_cmdbuf *cs,
3342 struct radv_pipeline *pipeline)
3343 {
3344 struct radv_shader_variant *ps;
3345 uint64_t va;
3346 assert (pipeline->shaders[MESA_SHADER_FRAGMENT]);
3347
3348 ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3349 va = radv_buffer_get_va(ps->bo) + ps->bo_offset;
3350
3351 radeon_set_sh_reg_seq(cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
3352 radeon_emit(cs, va >> 8);
3353 radeon_emit(cs, S_00B024_MEM_BASE(va >> 40));
3354 radeon_emit(cs, ps->rsrc1);
3355 radeon_emit(cs, ps->rsrc2);
3356
3357 radeon_set_context_reg(ctx_cs, R_02880C_DB_SHADER_CONTROL,
3358 radv_compute_db_shader_control(pipeline->device,
3359 pipeline, ps));
3360
3361 radeon_set_context_reg(ctx_cs, R_0286CC_SPI_PS_INPUT_ENA,
3362 ps->config.spi_ps_input_ena);
3363
3364 radeon_set_context_reg(ctx_cs, R_0286D0_SPI_PS_INPUT_ADDR,
3365 ps->config.spi_ps_input_addr);
3366
3367 radeon_set_context_reg(ctx_cs, R_0286D8_SPI_PS_IN_CONTROL,
3368 S_0286D8_NUM_INTERP(ps->info.fs.num_interp));
3369
3370 radeon_set_context_reg(ctx_cs, R_0286E0_SPI_BARYC_CNTL, pipeline->graphics.spi_baryc_cntl);
3371
3372 radeon_set_context_reg(ctx_cs, R_028710_SPI_SHADER_Z_FORMAT,
3373 ac_get_spi_shader_z_format(ps->info.info.ps.writes_z,
3374 ps->info.info.ps.writes_stencil,
3375 ps->info.info.ps.writes_sample_mask));
3376
3377 if (pipeline->device->dfsm_allowed) {
3378 /* optimise this? */
3379 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3380 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3381 }
3382 }
3383
3384 static void
3385 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_cmdbuf *ctx_cs,
3386 struct radv_pipeline *pipeline)
3387 {
3388 if (pipeline->device->physical_device->rad_info.family < CHIP_POLARIS10)
3389 return;
3390
3391 unsigned vtx_reuse_depth = 30;
3392 if (radv_pipeline_has_tess(pipeline) &&
3393 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.tes.spacing == TESS_SPACING_FRACTIONAL_ODD) {
3394 vtx_reuse_depth = 14;
3395 }
3396 radeon_set_context_reg(ctx_cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
3397 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth));
3398 }
3399
3400 static uint32_t
3401 radv_compute_vgt_shader_stages_en(const struct radv_pipeline *pipeline)
3402 {
3403 uint32_t stages = 0;
3404 if (radv_pipeline_has_tess(pipeline)) {
3405 stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
3406 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3407
3408 if (radv_pipeline_has_gs(pipeline))
3409 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
3410 S_028B54_GS_EN(1) |
3411 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
3412 else
3413 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
3414
3415 } else if (radv_pipeline_has_gs(pipeline))
3416 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
3417 S_028B54_GS_EN(1) |
3418 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
3419
3420 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
3421 stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3422
3423 return stages;
3424 }
3425
3426 static uint32_t
3427 radv_compute_cliprect_rule(const VkGraphicsPipelineCreateInfo *pCreateInfo)
3428 {
3429 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
3430 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
3431
3432 if (!discard_rectangle_info)
3433 return 0xffff;
3434
3435 unsigned mask = 0;
3436
3437 for (unsigned i = 0; i < (1u << MAX_DISCARD_RECTANGLES); ++i) {
3438 /* Interpret i as a bitmask, and then set the bit in the mask if
3439 * that combination of rectangles in which the pixel is contained
3440 * should pass the cliprect test. */
3441 unsigned relevant_subset = i & ((1u << discard_rectangle_info->discardRectangleCount) - 1);
3442
3443 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT &&
3444 !relevant_subset)
3445 continue;
3446
3447 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT &&
3448 relevant_subset)
3449 continue;
3450
3451 mask |= 1u << i;
3452 }
3453
3454 return mask;
3455 }
3456
3457 static void
3458 radv_pipeline_generate_pm4(struct radv_pipeline *pipeline,
3459 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3460 const struct radv_graphics_pipeline_create_info *extra,
3461 const struct radv_blend_state *blend,
3462 const struct radv_tessellation_state *tess,
3463 const struct radv_gs_state *gs,
3464 unsigned prim, unsigned gs_out)
3465 {
3466 struct radeon_cmdbuf *ctx_cs = &pipeline->ctx_cs;
3467 struct radeon_cmdbuf *cs = &pipeline->cs;
3468
3469 cs->max_dw = 64;
3470 ctx_cs->max_dw = 256;
3471 cs->buf = malloc(4 * (cs->max_dw + ctx_cs->max_dw));
3472 ctx_cs->buf = cs->buf + cs->max_dw;
3473
3474 radv_pipeline_generate_depth_stencil_state(ctx_cs, pipeline, pCreateInfo, extra);
3475 radv_pipeline_generate_blend_state(ctx_cs, pipeline, blend);
3476 radv_pipeline_generate_raster_state(ctx_cs, pipeline, pCreateInfo);
3477 radv_pipeline_generate_multisample_state(ctx_cs, pipeline);
3478 radv_pipeline_generate_vgt_gs_mode(ctx_cs, pipeline);
3479 radv_pipeline_generate_vertex_shader(ctx_cs, cs, pipeline, tess);
3480 radv_pipeline_generate_tess_shaders(ctx_cs, cs, pipeline, tess);
3481 radv_pipeline_generate_geometry_shader(ctx_cs, cs, pipeline, gs);
3482 radv_pipeline_generate_fragment_shader(ctx_cs, cs, pipeline);
3483 radv_pipeline_generate_ps_inputs(ctx_cs, pipeline);
3484 radv_pipeline_generate_vgt_vertex_reuse(ctx_cs, pipeline);
3485 radv_pipeline_generate_binning_state(ctx_cs, pipeline, pCreateInfo);
3486
3487 radeon_set_context_reg(ctx_cs, R_0286E8_SPI_TMPRING_SIZE,
3488 S_0286E8_WAVES(pipeline->max_waves) |
3489 S_0286E8_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
3490
3491 radeon_set_context_reg(ctx_cs, R_028B54_VGT_SHADER_STAGES_EN, radv_compute_vgt_shader_stages_en(pipeline));
3492
3493 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
3494 radeon_set_uconfig_reg_idx(cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
3495 } else {
3496 radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
3497 }
3498 radeon_set_context_reg(ctx_cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out);
3499
3500 radeon_set_context_reg(ctx_cs, R_02820C_PA_SC_CLIPRECT_RULE, radv_compute_cliprect_rule(pCreateInfo));
3501
3502 pipeline->ctx_cs_hash = _mesa_hash_data(ctx_cs->buf, ctx_cs->cdw * 4);
3503
3504 assert(ctx_cs->cdw <= ctx_cs->max_dw);
3505 assert(cs->cdw <= cs->max_dw);
3506 }
3507
3508 static struct radv_ia_multi_vgt_param_helpers
3509 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline *pipeline,
3510 const struct radv_tessellation_state *tess,
3511 uint32_t prim)
3512 {
3513 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param = {0};
3514 const struct radv_device *device = pipeline->device;
3515
3516 if (radv_pipeline_has_tess(pipeline))
3517 ia_multi_vgt_param.primgroup_size = tess->num_patches;
3518 else if (radv_pipeline_has_gs(pipeline))
3519 ia_multi_vgt_param.primgroup_size = 64;
3520 else
3521 ia_multi_vgt_param.primgroup_size = 128; /* recommended without a GS */
3522
3523 /* GS requirement. */
3524 ia_multi_vgt_param.partial_es_wave = false;
3525 if (radv_pipeline_has_gs(pipeline) && device->physical_device->rad_info.chip_class <= GFX8)
3526 if (SI_GS_PER_ES / ia_multi_vgt_param.primgroup_size >= pipeline->device->gs_table_depth - 3)
3527 ia_multi_vgt_param.partial_es_wave = true;
3528
3529 ia_multi_vgt_param.wd_switch_on_eop = false;
3530 if (device->physical_device->rad_info.chip_class >= GFX7) {
3531 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
3532 * 4 shader engines. Set 1 to pass the assertion below.
3533 * The other cases are hardware requirements. */
3534 if (device->physical_device->rad_info.max_se < 4 ||
3535 prim == V_008958_DI_PT_POLYGON ||
3536 prim == V_008958_DI_PT_LINELOOP ||
3537 prim == V_008958_DI_PT_TRIFAN ||
3538 prim == V_008958_DI_PT_TRISTRIP_ADJ ||
3539 (pipeline->graphics.prim_restart_enable &&
3540 (device->physical_device->rad_info.family < CHIP_POLARIS10 ||
3541 (prim != V_008958_DI_PT_POINTLIST &&
3542 prim != V_008958_DI_PT_LINESTRIP))))
3543 ia_multi_vgt_param.wd_switch_on_eop = true;
3544 }
3545
3546 ia_multi_vgt_param.ia_switch_on_eoi = false;
3547 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input)
3548 ia_multi_vgt_param.ia_switch_on_eoi = true;
3549 if (radv_pipeline_has_gs(pipeline) &&
3550 pipeline->shaders[MESA_SHADER_GEOMETRY]->info.info.uses_prim_id)
3551 ia_multi_vgt_param.ia_switch_on_eoi = true;
3552 if (radv_pipeline_has_tess(pipeline)) {
3553 /* SWITCH_ON_EOI must be set if PrimID is used. */
3554 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.uses_prim_id ||
3555 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.info.uses_prim_id)
3556 ia_multi_vgt_param.ia_switch_on_eoi = true;
3557 }
3558
3559 ia_multi_vgt_param.partial_vs_wave = false;
3560 if (radv_pipeline_has_tess(pipeline)) {
3561 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
3562 if ((device->physical_device->rad_info.family == CHIP_TAHITI ||
3563 device->physical_device->rad_info.family == CHIP_PITCAIRN ||
3564 device->physical_device->rad_info.family == CHIP_BONAIRE) &&
3565 radv_pipeline_has_gs(pipeline))
3566 ia_multi_vgt_param.partial_vs_wave = true;
3567 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
3568 if (device->has_distributed_tess) {
3569 if (radv_pipeline_has_gs(pipeline)) {
3570 if (device->physical_device->rad_info.chip_class <= GFX8)
3571 ia_multi_vgt_param.partial_es_wave = true;
3572 } else {
3573 ia_multi_vgt_param.partial_vs_wave = true;
3574 }
3575 }
3576 }
3577
3578 /* Workaround for a VGT hang when strip primitive types are used with
3579 * primitive restart.
3580 */
3581 if (pipeline->graphics.prim_restart_enable &&
3582 (prim == V_008958_DI_PT_LINESTRIP ||
3583 prim == V_008958_DI_PT_TRISTRIP ||
3584 prim == V_008958_DI_PT_LINESTRIP_ADJ ||
3585 prim == V_008958_DI_PT_TRISTRIP_ADJ)) {
3586 ia_multi_vgt_param.partial_vs_wave = true;
3587 }
3588
3589 if (radv_pipeline_has_gs(pipeline)) {
3590 /* On these chips there is the possibility of a hang if the
3591 * pipeline uses a GS and partial_vs_wave is not set.
3592 *
3593 * This mostly does not hit 4-SE chips, as those typically set
3594 * ia_switch_on_eoi and then partial_vs_wave is set for pipelines
3595 * with GS due to another workaround.
3596 *
3597 * Reproducer: https://bugs.freedesktop.org/show_bug.cgi?id=109242
3598 */
3599 if (device->physical_device->rad_info.family == CHIP_TONGA ||
3600 device->physical_device->rad_info.family == CHIP_FIJI ||
3601 device->physical_device->rad_info.family == CHIP_POLARIS10 ||
3602 device->physical_device->rad_info.family == CHIP_POLARIS11 ||
3603 device->physical_device->rad_info.family == CHIP_POLARIS12 ||
3604 device->physical_device->rad_info.family == CHIP_VEGAM) {
3605 ia_multi_vgt_param.partial_vs_wave = true;
3606 }
3607 }
3608
3609 ia_multi_vgt_param.base =
3610 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param.primgroup_size - 1) |
3611 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
3612 S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == GFX8 ? 2 : 0) |
3613 S_030960_EN_INST_OPT_BASIC(device->physical_device->rad_info.chip_class >= GFX9) |
3614 S_030960_EN_INST_OPT_ADV(device->physical_device->rad_info.chip_class >= GFX9);
3615
3616 return ia_multi_vgt_param;
3617 }
3618
3619
3620 static void
3621 radv_compute_vertex_input_state(struct radv_pipeline *pipeline,
3622 const VkGraphicsPipelineCreateInfo *pCreateInfo)
3623 {
3624 const VkPipelineVertexInputStateCreateInfo *vi_info =
3625 pCreateInfo->pVertexInputState;
3626 struct radv_vertex_elements_info *velems = &pipeline->vertex_elements;
3627
3628 for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
3629 const VkVertexInputAttributeDescription *desc =
3630 &vi_info->pVertexAttributeDescriptions[i];
3631 unsigned loc = desc->location;
3632 const struct vk_format_description *format_desc;
3633
3634 format_desc = vk_format_description(desc->format);
3635
3636 velems->format_size[loc] = format_desc->block.bits / 8;
3637 }
3638
3639 for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
3640 const VkVertexInputBindingDescription *desc =
3641 &vi_info->pVertexBindingDescriptions[i];
3642
3643 pipeline->binding_stride[desc->binding] = desc->stride;
3644 pipeline->num_vertex_bindings =
3645 MAX2(pipeline->num_vertex_bindings, desc->binding + 1);
3646 }
3647 }
3648
3649 static struct radv_shader_variant *
3650 radv_pipeline_get_streamout_shader(struct radv_pipeline *pipeline)
3651 {
3652 int i;
3653
3654 for (i = MESA_SHADER_GEOMETRY; i >= MESA_SHADER_VERTEX; i--) {
3655 struct radv_shader_variant *shader =
3656 radv_get_shader(pipeline, i);
3657
3658 if (shader && shader->info.info.so.num_outputs > 0)
3659 return shader;
3660 }
3661
3662 return NULL;
3663 }
3664
3665 static VkResult
3666 radv_pipeline_init(struct radv_pipeline *pipeline,
3667 struct radv_device *device,
3668 struct radv_pipeline_cache *cache,
3669 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3670 const struct radv_graphics_pipeline_create_info *extra)
3671 {
3672 VkResult result;
3673 bool has_view_index = false;
3674
3675 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3676 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3677 if (subpass->view_mask)
3678 has_view_index = true;
3679
3680 pipeline->device = device;
3681 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
3682 assert(pipeline->layout);
3683
3684 struct radv_blend_state blend = radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra);
3685
3686 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
3687 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
3688 radv_init_feedback(creation_feedback);
3689
3690 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
3691
3692 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
3693 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
3694 for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
3695 gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
3696 pStages[stage] = &pCreateInfo->pStages[i];
3697 if(creation_feedback)
3698 stage_feedbacks[stage] = &creation_feedback->pPipelineStageCreationFeedbacks[i];
3699 }
3700
3701 struct radv_pipeline_key key = radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend, has_view_index);
3702 radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
3703
3704 pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
3705 radv_pipeline_init_multisample_state(pipeline, &blend, pCreateInfo);
3706 uint32_t gs_out;
3707 uint32_t prim = si_translate_prim(pCreateInfo->pInputAssemblyState->topology);
3708
3709 pipeline->graphics.can_use_guardband = radv_prim_can_use_guardband(pCreateInfo->pInputAssemblyState->topology);
3710
3711 if (radv_pipeline_has_gs(pipeline)) {
3712 gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.output_prim);
3713 pipeline->graphics.can_use_guardband = gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP;
3714 } else {
3715 gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology);
3716 }
3717 if (extra && extra->use_rectlist) {
3718 prim = V_008958_DI_PT_RECTLIST;
3719 gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP;
3720 pipeline->graphics.can_use_guardband = true;
3721 }
3722 pipeline->graphics.prim_restart_enable = !!pCreateInfo->pInputAssemblyState->primitiveRestartEnable;
3723 /* prim vertex count will need TESS changes */
3724 pipeline->graphics.prim_vertex_count = prim_size_table[prim];
3725
3726 radv_pipeline_init_dynamic_state(pipeline, pCreateInfo);
3727
3728 /* Ensure that some export memory is always allocated, for two reasons:
3729 *
3730 * 1) Correctness: The hardware ignores the EXEC mask if no export
3731 * memory is allocated, so KILL and alpha test do not work correctly
3732 * without this.
3733 * 2) Performance: Every shader needs at least a NULL export, even when
3734 * it writes no color/depth output. The NULL export instruction
3735 * stalls without this setting.
3736 *
3737 * Don't add this to CB_SHADER_MASK.
3738 */
3739 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3740 if (!blend.spi_shader_col_format) {
3741 if (!ps->info.info.ps.writes_z &&
3742 !ps->info.info.ps.writes_stencil &&
3743 !ps->info.info.ps.writes_sample_mask)
3744 blend.spi_shader_col_format = V_028714_SPI_SHADER_32_R;
3745 }
3746
3747 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
3748 if (pipeline->shaders[i]) {
3749 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[i]->info.need_indirect_descriptor_sets;
3750 }
3751 }
3752
3753 struct radv_gs_state gs = {0};
3754 if (radv_pipeline_has_gs(pipeline)) {
3755 gs = calculate_gs_info(pCreateInfo, pipeline);
3756 calculate_gs_ring_sizes(pipeline, &gs);
3757 }
3758
3759 struct radv_tessellation_state tess = {0};
3760 if (radv_pipeline_has_tess(pipeline)) {
3761 if (prim == V_008958_DI_PT_PATCH) {
3762 pipeline->graphics.prim_vertex_count.min = pCreateInfo->pTessellationState->patchControlPoints;
3763 pipeline->graphics.prim_vertex_count.incr = 1;
3764 }
3765 tess = calculate_tess_state(pipeline, pCreateInfo);
3766 }
3767
3768 pipeline->graphics.ia_multi_vgt_param = radv_compute_ia_multi_vgt_param_helpers(pipeline, &tess, prim);
3769
3770 radv_compute_vertex_input_state(pipeline, pCreateInfo);
3771
3772 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++)
3773 pipeline->user_data_0[i] = radv_pipeline_stage_to_user_data_0(pipeline, i, device->physical_device->rad_info.chip_class);
3774
3775 struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_VERTEX,
3776 AC_UD_VS_BASE_VERTEX_START_INSTANCE);
3777 if (loc->sgpr_idx != -1) {
3778 pipeline->graphics.vtx_base_sgpr = pipeline->user_data_0[MESA_SHADER_VERTEX];
3779 pipeline->graphics.vtx_base_sgpr += loc->sgpr_idx * 4;
3780 if (radv_get_shader(pipeline, MESA_SHADER_VERTEX)->info.info.vs.needs_draw_id)
3781 pipeline->graphics.vtx_emit_num = 3;
3782 else
3783 pipeline->graphics.vtx_emit_num = 2;
3784 }
3785
3786 /* Find the last vertex shader stage that eventually uses streamout. */
3787 pipeline->streamout_shader = radv_pipeline_get_streamout_shader(pipeline);
3788
3789 result = radv_pipeline_scratch_init(device, pipeline);
3790 radv_pipeline_generate_pm4(pipeline, pCreateInfo, extra, &blend, &tess, &gs, prim, gs_out);
3791
3792 return result;
3793 }
3794
3795 VkResult
3796 radv_graphics_pipeline_create(
3797 VkDevice _device,
3798 VkPipelineCache _cache,
3799 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3800 const struct radv_graphics_pipeline_create_info *extra,
3801 const VkAllocationCallbacks *pAllocator,
3802 VkPipeline *pPipeline)
3803 {
3804 RADV_FROM_HANDLE(radv_device, device, _device);
3805 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
3806 struct radv_pipeline *pipeline;
3807 VkResult result;
3808
3809 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
3810 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3811 if (pipeline == NULL)
3812 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
3813
3814 result = radv_pipeline_init(pipeline, device, cache,
3815 pCreateInfo, extra);
3816 if (result != VK_SUCCESS) {
3817 radv_pipeline_destroy(device, pipeline, pAllocator);
3818 return result;
3819 }
3820
3821 *pPipeline = radv_pipeline_to_handle(pipeline);
3822
3823 return VK_SUCCESS;
3824 }
3825
3826 VkResult radv_CreateGraphicsPipelines(
3827 VkDevice _device,
3828 VkPipelineCache pipelineCache,
3829 uint32_t count,
3830 const VkGraphicsPipelineCreateInfo* pCreateInfos,
3831 const VkAllocationCallbacks* pAllocator,
3832 VkPipeline* pPipelines)
3833 {
3834 VkResult result = VK_SUCCESS;
3835 unsigned i = 0;
3836
3837 for (; i < count; i++) {
3838 VkResult r;
3839 r = radv_graphics_pipeline_create(_device,
3840 pipelineCache,
3841 &pCreateInfos[i],
3842 NULL, pAllocator, &pPipelines[i]);
3843 if (r != VK_SUCCESS) {
3844 result = r;
3845 pPipelines[i] = VK_NULL_HANDLE;
3846 }
3847 }
3848
3849 return result;
3850 }
3851
3852
3853 static void
3854 radv_compute_generate_pm4(struct radv_pipeline *pipeline)
3855 {
3856 struct radv_shader_variant *compute_shader;
3857 struct radv_device *device = pipeline->device;
3858 unsigned compute_resource_limits;
3859 unsigned waves_per_threadgroup;
3860 uint64_t va;
3861
3862 pipeline->cs.buf = malloc(20 * 4);
3863 pipeline->cs.max_dw = 20;
3864
3865 compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
3866 va = radv_buffer_get_va(compute_shader->bo) + compute_shader->bo_offset;
3867
3868 radeon_set_sh_reg_seq(&pipeline->cs, R_00B830_COMPUTE_PGM_LO, 2);
3869 radeon_emit(&pipeline->cs, va >> 8);
3870 radeon_emit(&pipeline->cs, S_00B834_DATA(va >> 40));
3871
3872 radeon_set_sh_reg_seq(&pipeline->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
3873 radeon_emit(&pipeline->cs, compute_shader->rsrc1);
3874 radeon_emit(&pipeline->cs, compute_shader->rsrc2);
3875
3876 radeon_set_sh_reg(&pipeline->cs, R_00B860_COMPUTE_TMPRING_SIZE,
3877 S_00B860_WAVES(pipeline->max_waves) |
3878 S_00B860_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
3879
3880 /* Calculate best compute resource limits. */
3881 waves_per_threadgroup =
3882 DIV_ROUND_UP(compute_shader->info.cs.block_size[0] *
3883 compute_shader->info.cs.block_size[1] *
3884 compute_shader->info.cs.block_size[2], 64);
3885 compute_resource_limits =
3886 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
3887
3888 if (device->physical_device->rad_info.chip_class >= GFX7) {
3889 unsigned num_cu_per_se =
3890 device->physical_device->rad_info.num_good_compute_units /
3891 device->physical_device->rad_info.max_se;
3892
3893 /* Force even distribution on all SIMDs in CU if the workgroup
3894 * size is 64. This has shown some good improvements if # of
3895 * CUs per SE is not a multiple of 4.
3896 */
3897 if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
3898 compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
3899 }
3900
3901 radeon_set_sh_reg(&pipeline->cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
3902 compute_resource_limits);
3903
3904 radeon_set_sh_reg_seq(&pipeline->cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
3905 radeon_emit(&pipeline->cs,
3906 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[0]));
3907 radeon_emit(&pipeline->cs,
3908 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[1]));
3909 radeon_emit(&pipeline->cs,
3910 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[2]));
3911
3912 assert(pipeline->cs.cdw <= pipeline->cs.max_dw);
3913 }
3914
3915 static VkResult radv_compute_pipeline_create(
3916 VkDevice _device,
3917 VkPipelineCache _cache,
3918 const VkComputePipelineCreateInfo* pCreateInfo,
3919 const VkAllocationCallbacks* pAllocator,
3920 VkPipeline* pPipeline)
3921 {
3922 RADV_FROM_HANDLE(radv_device, device, _device);
3923 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
3924 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
3925 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
3926 struct radv_pipeline *pipeline;
3927 VkResult result;
3928
3929 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
3930 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3931 if (pipeline == NULL)
3932 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
3933
3934 pipeline->device = device;
3935 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
3936 assert(pipeline->layout);
3937
3938 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
3939 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
3940 radv_init_feedback(creation_feedback);
3941
3942 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
3943 if (creation_feedback)
3944 stage_feedbacks[MESA_SHADER_COMPUTE] = &creation_feedback->pPipelineStageCreationFeedbacks[0];
3945
3946 pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage;
3947 radv_create_shaders(pipeline, device, cache, &(struct radv_pipeline_key) {0}, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
3948
3949 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);
3950 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets;
3951 result = radv_pipeline_scratch_init(device, pipeline);
3952 if (result != VK_SUCCESS) {
3953 radv_pipeline_destroy(device, pipeline, pAllocator);
3954 return result;
3955 }
3956
3957 radv_compute_generate_pm4(pipeline);
3958
3959 *pPipeline = radv_pipeline_to_handle(pipeline);
3960
3961 return VK_SUCCESS;
3962 }
3963
3964 VkResult radv_CreateComputePipelines(
3965 VkDevice _device,
3966 VkPipelineCache pipelineCache,
3967 uint32_t count,
3968 const VkComputePipelineCreateInfo* pCreateInfos,
3969 const VkAllocationCallbacks* pAllocator,
3970 VkPipeline* pPipelines)
3971 {
3972 VkResult result = VK_SUCCESS;
3973
3974 unsigned i = 0;
3975 for (; i < count; i++) {
3976 VkResult r;
3977 r = radv_compute_pipeline_create(_device, pipelineCache,
3978 &pCreateInfos[i],
3979 pAllocator, &pPipelines[i]);
3980 if (r != VK_SUCCESS) {
3981 result = r;
3982 pPipelines[i] = VK_NULL_HANDLE;
3983 }
3984 }
3985
3986 return result;
3987 }