2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
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:
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
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
28 #include "util/mesa-sha1.h"
29 #include "radv_private.h"
31 #include "nir/nir_builder.h"
32 #include "spirv/nir_spirv.h"
34 #include <llvm-c/Core.h>
35 #include <llvm-c/TargetMachine.h>
38 #include "r600d_common.h"
39 #include "ac_binary.h"
40 #include "ac_llvm_util.h"
41 #include "ac_nir_to_llvm.h"
42 #include "vk_format.h"
43 #include "util/debug.h"
44 void radv_shader_variant_destroy(struct radv_device
*device
,
45 struct radv_shader_variant
*variant
);
47 static const struct nir_shader_compiler_options nir_options
= {
48 .vertex_id_zero_based
= true,
52 .lower_pack_snorm_2x16
= true,
53 .lower_pack_snorm_4x8
= true,
54 .lower_pack_unorm_2x16
= true,
55 .lower_pack_unorm_4x8
= true,
56 .lower_unpack_snorm_2x16
= true,
57 .lower_unpack_snorm_4x8
= true,
58 .lower_unpack_unorm_2x16
= true,
59 .lower_unpack_unorm_4x8
= true,
60 .lower_extract_byte
= true,
61 .lower_extract_word
= true,
64 VkResult
radv_CreateShaderModule(
66 const VkShaderModuleCreateInfo
* pCreateInfo
,
67 const VkAllocationCallbacks
* pAllocator
,
68 VkShaderModule
* pShaderModule
)
70 RADV_FROM_HANDLE(radv_device
, device
, _device
);
71 struct radv_shader_module
*module
;
73 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
74 assert(pCreateInfo
->flags
== 0);
76 module
= radv_alloc2(&device
->alloc
, pAllocator
,
77 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
78 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
80 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
83 module
->size
= pCreateInfo
->codeSize
;
84 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
86 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
88 *pShaderModule
= radv_shader_module_to_handle(module
);
93 void radv_DestroyShaderModule(
95 VkShaderModule _module
,
96 const VkAllocationCallbacks
* pAllocator
)
98 RADV_FROM_HANDLE(radv_device
, device
, _device
);
99 RADV_FROM_HANDLE(radv_shader_module
, module
, _module
);
104 radv_free2(&device
->alloc
, pAllocator
, module
);
107 void radv_DestroyPipeline(
109 VkPipeline _pipeline
,
110 const VkAllocationCallbacks
* pAllocator
)
112 RADV_FROM_HANDLE(radv_device
, device
, _device
);
113 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
118 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
119 if (pipeline
->shaders
[i
])
120 radv_shader_variant_destroy(device
, pipeline
->shaders
[i
]);
122 radv_free2(&device
->alloc
, pAllocator
, pipeline
);
127 radv_optimize_nir(struct nir_shader
*shader
)
134 NIR_PASS_V(shader
, nir_lower_vars_to_ssa
);
135 NIR_PASS_V(shader
, nir_lower_alu_to_scalar
);
136 NIR_PASS_V(shader
, nir_lower_phis_to_scalar
);
138 NIR_PASS(progress
, shader
, nir_copy_prop
);
139 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
140 NIR_PASS(progress
, shader
, nir_opt_dce
);
141 NIR_PASS(progress
, shader
, nir_opt_dead_cf
);
142 NIR_PASS(progress
, shader
, nir_opt_cse
);
143 NIR_PASS(progress
, shader
, nir_opt_peephole_select
, 8);
144 NIR_PASS(progress
, shader
, nir_opt_algebraic
);
145 NIR_PASS(progress
, shader
, nir_opt_constant_folding
);
146 NIR_PASS(progress
, shader
, nir_opt_undef
);
151 radv_shader_compile_to_nir(struct radv_device
*device
,
152 struct radv_shader_module
*module
,
153 const char *entrypoint_name
,
154 gl_shader_stage stage
,
155 const VkSpecializationInfo
*spec_info
,
158 if (strcmp(entrypoint_name
, "main") != 0) {
159 radv_finishme("Multiple shaders per module not really supported");
163 nir_function
*entry_point
;
165 /* Some things such as our meta clear/blit code will give us a NIR
166 * shader directly. In that case, we just ignore the SPIR-V entirely
167 * and just use the NIR shader */
169 nir
->options
= &nir_options
;
170 nir_validate_shader(nir
);
172 assert(exec_list_length(&nir
->functions
) == 1);
173 struct exec_node
*node
= exec_list_get_head(&nir
->functions
);
174 entry_point
= exec_node_data(nir_function
, node
, node
);
176 uint32_t *spirv
= (uint32_t *) module
->data
;
177 assert(module
->size
% 4 == 0);
179 uint32_t num_spec_entries
= 0;
180 struct nir_spirv_specialization
*spec_entries
= NULL
;
181 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
182 num_spec_entries
= spec_info
->mapEntryCount
;
183 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
184 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
185 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
186 const void *data
= spec_info
->pData
+ entry
.offset
;
187 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
189 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
190 spec_entries
[i
].data
= *(const uint32_t *)data
;
194 entry_point
= spirv_to_nir(spirv
, module
->size
/ 4,
195 spec_entries
, num_spec_entries
,
196 stage
, entrypoint_name
, &nir_options
);
197 nir
= entry_point
->shader
;
198 assert(nir
->stage
== stage
);
199 nir_validate_shader(nir
);
203 nir_lower_returns(nir
);
204 nir_validate_shader(nir
);
206 nir_inline_functions(nir
);
207 nir_validate_shader(nir
);
209 /* Pick off the single entrypoint that we want */
210 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
211 if (func
!= entry_point
)
212 exec_node_remove(&func
->node
);
214 assert(exec_list_length(&nir
->functions
) == 1);
215 entry_point
->name
= ralloc_strdup(entry_point
, "main");
217 nir_remove_dead_variables(nir
, nir_var_shader_in
);
218 nir_remove_dead_variables(nir
, nir_var_shader_out
);
219 nir_remove_dead_variables(nir
, nir_var_system_value
);
220 nir_validate_shader(nir
);
222 nir_lower_system_values(nir
);
223 nir_validate_shader(nir
);
226 /* Vulkan uses the separate-shader linking model */
227 nir
->info
.separate_shader
= true;
229 // nir = brw_preprocess_nir(compiler, nir);
231 nir_shader_gather_info(nir
, entry_point
->impl
);
233 nir_variable_mode indirect_mask
= 0;
234 // if (compiler->glsl_compiler_options[stage].EmitNoIndirectInput)
235 indirect_mask
|= nir_var_shader_in
;
236 // if (compiler->glsl_compiler_options[stage].EmitNoIndirectTemp)
237 indirect_mask
|= nir_var_local
;
239 nir_lower_indirect_derefs(nir
, indirect_mask
);
241 static const nir_lower_tex_options tex_options
= {
245 nir_lower_tex(nir
, &tex_options
);
247 nir_lower_vars_to_ssa(nir
);
248 nir_lower_var_copies(nir
);
249 nir_lower_global_vars_to_local(nir
);
250 nir_remove_dead_variables(nir
, nir_var_local
);
251 radv_optimize_nir(nir
);
254 nir_print_shader(nir
, stderr
);
259 void radv_shader_variant_destroy(struct radv_device
*device
,
260 struct radv_shader_variant
*variant
)
262 if (__sync_fetch_and_sub(&variant
->ref_count
, 1) != 1)
265 device
->ws
->buffer_destroy(variant
->bo
);
270 struct radv_shader_variant
*radv_shader_variant_create(struct radv_device
*device
,
271 struct nir_shader
*shader
,
272 struct radv_pipeline_layout
*layout
,
273 const union ac_shader_variant_key
*key
,
275 unsigned *code_size_out
,
278 struct radv_shader_variant
*variant
= calloc(1, sizeof(struct radv_shader_variant
));
279 enum radeon_family chip_family
= device
->instance
->physicalDevice
.rad_info
.family
;
280 LLVMTargetMachineRef tm
;
284 struct ac_nir_compiler_options options
= {0};
285 options
.layout
= layout
;
289 struct ac_shader_binary binary
;
291 options
.unsafe_math
= env_var_as_boolean("RADV_UNSAFE_MATH", false);
292 options
.family
= chip_family
;
293 options
.chip_class
= device
->instance
->physicalDevice
.rad_info
.chip_class
;
294 tm
= ac_create_target_machine(chip_family
);
295 ac_compile_nir_shader(tm
, &binary
, &variant
->config
,
296 &variant
->info
, shader
, &options
, dump
);
297 LLVMDisposeTargetMachine(tm
);
299 bool scratch_enabled
= variant
->config
.scratch_bytes_per_wave
> 0;
300 unsigned vgpr_comp_cnt
= 0;
303 radv_finishme("shader scratch space");
304 switch (shader
->stage
) {
305 case MESA_SHADER_VERTEX
:
306 variant
->rsrc2
= S_00B12C_USER_SGPR(variant
->info
.num_user_sgprs
) |
307 S_00B12C_SCRATCH_EN(scratch_enabled
);
308 vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
310 case MESA_SHADER_FRAGMENT
:
311 variant
->rsrc2
= S_00B12C_USER_SGPR(variant
->info
.num_user_sgprs
) |
312 S_00B12C_SCRATCH_EN(scratch_enabled
);
314 case MESA_SHADER_COMPUTE
:
315 variant
->rsrc2
= S_00B84C_USER_SGPR(variant
->info
.num_user_sgprs
) |
316 S_00B84C_SCRATCH_EN(scratch_enabled
) |
317 S_00B84C_TGID_X_EN(1) | S_00B84C_TGID_Y_EN(1) |
318 S_00B84C_TGID_Z_EN(1) | S_00B84C_TIDIG_COMP_CNT(2) |
319 S_00B84C_TG_SIZE_EN(1) |
320 S_00B84C_LDS_SIZE(variant
->config
.lds_size
);
323 unreachable("unsupported shader type");
327 variant
->rsrc1
= S_00B848_VGPRS((variant
->config
.num_vgprs
- 1) / 4) |
328 S_00B848_SGPRS((variant
->config
.num_sgprs
- 1) / 8) |
329 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
330 S_00B848_DX10_CLAMP(1) |
331 S_00B848_FLOAT_MODE(variant
->config
.float_mode
);
333 variant
->bo
= device
->ws
->buffer_create(device
->ws
, binary
.code_size
, 256,
334 RADEON_DOMAIN_GTT
, RADEON_FLAG_CPU_ACCESS
);
336 void *ptr
= device
->ws
->buffer_map(variant
->bo
);
337 memcpy(ptr
, binary
.code
, binary
.code_size
);
338 device
->ws
->buffer_unmap(variant
->bo
);
341 *code_out
= binary
.code
;
342 *code_size_out
= binary
.code_size
;
347 free(binary
.global_symbol_offsets
);
349 free(binary
.disasm_string
);
350 variant
->ref_count
= 1;
355 static struct radv_shader_variant
*
356 radv_pipeline_compile(struct radv_pipeline
*pipeline
,
357 struct radv_pipeline_cache
*cache
,
358 struct radv_shader_module
*module
,
359 const char *entrypoint
,
360 gl_shader_stage stage
,
361 const VkSpecializationInfo
*spec_info
,
362 struct radv_pipeline_layout
*layout
,
363 const union ac_shader_variant_key
*key
,
366 unsigned char sha1
[20];
367 struct radv_shader_variant
*variant
;
370 unsigned code_size
= 0;
373 _mesa_sha1_compute(module
->nir
->info
.name
,
374 strlen(module
->nir
->info
.name
),
377 radv_hash_shader(sha1
, module
, entrypoint
, spec_info
, layout
, key
);
380 variant
= radv_create_shader_variant_from_pipeline_cache(pipeline
->device
,
387 nir
= radv_shader_compile_to_nir(pipeline
->device
,
388 module
, entrypoint
, stage
,
393 variant
= radv_shader_variant_create(pipeline
->device
, nir
, layout
, key
,
394 &code
, &code_size
, dump
);
398 if (variant
&& cache
)
399 variant
= radv_pipeline_cache_insert_shader(cache
, sha1
, variant
,
407 static uint32_t si_translate_blend_function(VkBlendOp op
)
410 case VK_BLEND_OP_ADD
:
411 return V_028780_COMB_DST_PLUS_SRC
;
412 case VK_BLEND_OP_SUBTRACT
:
413 return V_028780_COMB_SRC_MINUS_DST
;
414 case VK_BLEND_OP_REVERSE_SUBTRACT
:
415 return V_028780_COMB_DST_MINUS_SRC
;
416 case VK_BLEND_OP_MIN
:
417 return V_028780_COMB_MIN_DST_SRC
;
418 case VK_BLEND_OP_MAX
:
419 return V_028780_COMB_MAX_DST_SRC
;
425 static uint32_t si_translate_blend_factor(VkBlendFactor factor
)
428 case VK_BLEND_FACTOR_ZERO
:
429 return V_028780_BLEND_ZERO
;
430 case VK_BLEND_FACTOR_ONE
:
431 return V_028780_BLEND_ONE
;
432 case VK_BLEND_FACTOR_SRC_COLOR
:
433 return V_028780_BLEND_SRC_COLOR
;
434 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
435 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
436 case VK_BLEND_FACTOR_DST_COLOR
:
437 return V_028780_BLEND_DST_COLOR
;
438 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
:
439 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
440 case VK_BLEND_FACTOR_SRC_ALPHA
:
441 return V_028780_BLEND_SRC_ALPHA
;
442 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
443 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
444 case VK_BLEND_FACTOR_DST_ALPHA
:
445 return V_028780_BLEND_DST_ALPHA
;
446 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
:
447 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
448 case VK_BLEND_FACTOR_CONSTANT_COLOR
:
449 return V_028780_BLEND_CONSTANT_COLOR
;
450 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
:
451 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
452 case VK_BLEND_FACTOR_CONSTANT_ALPHA
:
453 return V_028780_BLEND_CONSTANT_ALPHA
;
454 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
:
455 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
456 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
457 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
458 case VK_BLEND_FACTOR_SRC1_COLOR
:
459 return V_028780_BLEND_SRC1_COLOR
;
460 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
461 return V_028780_BLEND_INV_SRC1_COLOR
;
462 case VK_BLEND_FACTOR_SRC1_ALPHA
:
463 return V_028780_BLEND_SRC1_ALPHA
;
464 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
465 return V_028780_BLEND_INV_SRC1_ALPHA
;
471 static bool is_dual_src(VkBlendFactor factor
)
474 case VK_BLEND_FACTOR_SRC1_COLOR
:
475 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
476 case VK_BLEND_FACTOR_SRC1_ALPHA
:
477 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
484 static unsigned si_choose_spi_color_format(VkFormat vk_format
,
486 bool blend_need_alpha
)
488 const struct vk_format_description
*desc
= vk_format_description(vk_format
);
489 unsigned format
, ntype
, swap
;
491 /* Alpha is needed for alpha-to-coverage.
492 * Blending may be with or without alpha.
494 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
495 unsigned alpha
= 0; /* exports alpha, but may not support blending */
496 unsigned blend
= 0; /* supports blending, but may not export alpha */
497 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
499 format
= radv_translate_colorformat(vk_format
);
500 ntype
= radv_translate_color_numformat(vk_format
, desc
,
501 vk_format_get_first_non_void_channel(vk_format
));
502 swap
= radv_translate_colorswap(vk_format
, false);
504 /* Choose the SPI color formats. These are required values for Stoney/RB+.
505 * Other chips have multiple choices, though they are not necessarily better.
508 case V_028C70_COLOR_5_6_5
:
509 case V_028C70_COLOR_1_5_5_5
:
510 case V_028C70_COLOR_5_5_5_1
:
511 case V_028C70_COLOR_4_4_4_4
:
512 case V_028C70_COLOR_10_11_11
:
513 case V_028C70_COLOR_11_11_10
:
514 case V_028C70_COLOR_8
:
515 case V_028C70_COLOR_8_8
:
516 case V_028C70_COLOR_8_8_8_8
:
517 case V_028C70_COLOR_10_10_10_2
:
518 case V_028C70_COLOR_2_10_10_10
:
519 if (ntype
== V_028C70_NUMBER_UINT
)
520 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
521 else if (ntype
== V_028C70_NUMBER_SINT
)
522 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
524 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
527 case V_028C70_COLOR_16
:
528 case V_028C70_COLOR_16_16
:
529 case V_028C70_COLOR_16_16_16_16
:
530 if (ntype
== V_028C70_NUMBER_UNORM
||
531 ntype
== V_028C70_NUMBER_SNORM
) {
532 /* UNORM16 and SNORM16 don't support blending */
533 if (ntype
== V_028C70_NUMBER_UNORM
)
534 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
536 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
538 /* Use 32 bits per channel for blending. */
539 if (format
== V_028C70_COLOR_16
) {
540 if (swap
== V_028C70_SWAP_STD
) { /* R */
541 blend
= V_028714_SPI_SHADER_32_R
;
542 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
543 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
544 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
547 } else if (format
== V_028C70_COLOR_16_16
) {
548 if (swap
== V_028C70_SWAP_STD
) { /* RG */
549 blend
= V_028714_SPI_SHADER_32_GR
;
550 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
551 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
552 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
555 } else /* 16_16_16_16 */
556 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
557 } else if (ntype
== V_028C70_NUMBER_UINT
)
558 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
559 else if (ntype
== V_028C70_NUMBER_SINT
)
560 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
561 else if (ntype
== V_028C70_NUMBER_FLOAT
)
562 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
567 case V_028C70_COLOR_32
:
568 if (swap
== V_028C70_SWAP_STD
) { /* R */
569 blend
= normal
= V_028714_SPI_SHADER_32_R
;
570 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
571 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
572 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
577 case V_028C70_COLOR_32_32
:
578 if (swap
== V_028C70_SWAP_STD
) { /* RG */
579 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
580 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
581 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
582 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
587 case V_028C70_COLOR_32_32_32_32
:
588 case V_028C70_COLOR_8_24
:
589 case V_028C70_COLOR_24_8
:
590 case V_028C70_COLOR_X24_8_32_FLOAT
:
591 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
595 unreachable("unhandled blend format");
598 if (blend_enable
&& blend_need_alpha
)
600 else if(blend_need_alpha
)
602 else if(blend_enable
)
608 static unsigned si_get_cb_shader_mask(unsigned spi_shader_col_format
)
610 unsigned i
, cb_shader_mask
= 0;
612 for (i
= 0; i
< 8; i
++) {
613 switch ((spi_shader_col_format
>> (i
* 4)) & 0xf) {
614 case V_028714_SPI_SHADER_ZERO
:
616 case V_028714_SPI_SHADER_32_R
:
617 cb_shader_mask
|= 0x1 << (i
* 4);
619 case V_028714_SPI_SHADER_32_GR
:
620 cb_shader_mask
|= 0x3 << (i
* 4);
622 case V_028714_SPI_SHADER_32_AR
:
623 cb_shader_mask
|= 0x9 << (i
* 4);
625 case V_028714_SPI_SHADER_FP16_ABGR
:
626 case V_028714_SPI_SHADER_UNORM16_ABGR
:
627 case V_028714_SPI_SHADER_SNORM16_ABGR
:
628 case V_028714_SPI_SHADER_UINT16_ABGR
:
629 case V_028714_SPI_SHADER_SINT16_ABGR
:
630 case V_028714_SPI_SHADER_32_ABGR
:
631 cb_shader_mask
|= 0xf << (i
* 4);
637 return cb_shader_mask
;
641 radv_pipeline_compute_spi_color_formats(struct radv_pipeline
*pipeline
,
642 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
643 uint32_t blend_enable
,
644 uint32_t blend_need_alpha
,
645 bool single_cb_enable
)
647 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
648 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
649 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
650 unsigned col_format
= 0;
652 for (unsigned i
= 0; i
< (single_cb_enable
? 1 : subpass
->color_count
); ++i
) {
653 struct radv_render_pass_attachment
*attachment
;
656 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
658 cf
= si_choose_spi_color_format(attachment
->format
,
659 blend_enable
& (1 << i
),
660 blend_need_alpha
& (1 << i
));
662 col_format
|= cf
<< (4 * i
);
665 blend
->cb_shader_mask
= si_get_cb_shader_mask(col_format
);
668 col_format
|= V_028714_SPI_SHADER_32_R
;
669 blend
->spi_shader_col_format
= col_format
;
673 format_is_int8(VkFormat format
)
675 const struct vk_format_description
*desc
= vk_format_description(format
);
676 int channel
= vk_format_get_first_non_void_channel(format
);
678 return channel
>= 0 && desc
->channel
[channel
].pure_integer
&&
679 desc
->channel
[channel
].size
== 8;
682 unsigned radv_format_meta_fs_key(VkFormat format
)
684 unsigned col_format
= si_choose_spi_color_format(format
, false, false) - 1;
685 bool is_int8
= format_is_int8(format
);
687 return col_format
+ (is_int8
? 3 : 0);
691 radv_pipeline_compute_is_int8(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
693 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
694 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
695 unsigned is_int8
= 0;
697 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
698 struct radv_render_pass_attachment
*attachment
;
700 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
702 if (format_is_int8(attachment
->format
))
710 radv_pipeline_init_blend_state(struct radv_pipeline
*pipeline
,
711 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
712 const struct radv_graphics_pipeline_create_info
*extra
)
714 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
715 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
716 unsigned mode
= V_028808_CB_NORMAL
;
717 uint32_t blend_enable
= 0, blend_need_alpha
= 0;
719 bool single_cb_enable
= false;
720 if (extra
&& extra
->custom_blend_mode
) {
721 single_cb_enable
= true;
722 mode
= extra
->custom_blend_mode
;
724 blend
->cb_color_control
= 0;
725 if (vkblend
->logicOpEnable
)
726 blend
->cb_color_control
|= S_028808_ROP3(vkblend
->logicOp
| (vkblend
->logicOp
<< 4));
728 blend
->cb_color_control
|= S_028808_ROP3(0xcc);
730 blend
->db_alpha_to_mask
= S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
731 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
732 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
733 S_028B70_ALPHA_TO_MASK_OFFSET3(2);
735 blend
->cb_target_mask
= 0;
736 for (i
= 0; i
< vkblend
->attachmentCount
; i
++) {
737 const VkPipelineColorBlendAttachmentState
*att
= &vkblend
->pAttachments
[i
];
738 unsigned blend_cntl
= 0;
739 VkBlendOp eqRGB
= att
->colorBlendOp
;
740 VkBlendFactor srcRGB
= att
->srcColorBlendFactor
;
741 VkBlendFactor dstRGB
= att
->dstColorBlendFactor
;
742 VkBlendOp eqA
= att
->alphaBlendOp
;
743 VkBlendFactor srcA
= att
->srcAlphaBlendFactor
;
744 VkBlendFactor dstA
= att
->dstAlphaBlendFactor
;
746 blend
->sx_mrt0_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
);
748 if (!att
->colorWriteMask
)
751 blend
->cb_target_mask
|= (unsigned)att
->colorWriteMask
<< (4 * i
);
752 if (!att
->blendEnable
) {
753 blend
->cb_blend_control
[i
] = blend_cntl
;
757 if (is_dual_src(srcRGB
) || is_dual_src(dstRGB
) || is_dual_src(srcA
) || is_dual_src(dstA
))
758 radv_finishme("dual source blending");
759 if (eqRGB
== VK_BLEND_OP_MIN
|| eqRGB
== VK_BLEND_OP_MAX
) {
760 srcRGB
= VK_BLEND_FACTOR_ONE
;
761 dstRGB
= VK_BLEND_FACTOR_ONE
;
763 if (eqA
== VK_BLEND_OP_MIN
|| eqA
== VK_BLEND_OP_MAX
) {
764 srcA
= VK_BLEND_FACTOR_ONE
;
765 dstA
= VK_BLEND_FACTOR_ONE
;
768 blend_cntl
|= S_028780_ENABLE(1);
770 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
771 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
772 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
773 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
774 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
775 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
776 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
777 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
779 blend
->cb_blend_control
[i
] = blend_cntl
;
781 blend_enable
|= 1 << i
;
783 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
784 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
785 srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
786 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
787 srcRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
||
788 dstRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
)
789 blend_need_alpha
|= 1 << i
;
791 for (i
= vkblend
->attachmentCount
; i
< 8; i
++)
792 blend
->cb_blend_control
[i
] = 0;
794 if (blend
->cb_target_mask
)
795 blend
->cb_color_control
|= S_028808_MODE(mode
);
797 blend
->cb_color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
799 radv_pipeline_compute_spi_color_formats(pipeline
, pCreateInfo
,
800 blend_enable
, blend_need_alpha
, single_cb_enable
);
803 static uint32_t si_translate_stencil_op(enum VkStencilOp op
)
806 case VK_STENCIL_OP_KEEP
:
807 return V_02842C_STENCIL_KEEP
;
808 case VK_STENCIL_OP_ZERO
:
809 return V_02842C_STENCIL_ZERO
;
810 case VK_STENCIL_OP_REPLACE
:
811 return V_02842C_STENCIL_REPLACE_TEST
;
812 case VK_STENCIL_OP_INCREMENT_AND_CLAMP
:
813 return V_02842C_STENCIL_ADD_CLAMP
;
814 case VK_STENCIL_OP_DECREMENT_AND_CLAMP
:
815 return V_02842C_STENCIL_SUB_CLAMP
;
816 case VK_STENCIL_OP_INVERT
:
817 return V_02842C_STENCIL_INVERT
;
818 case VK_STENCIL_OP_INCREMENT_AND_WRAP
:
819 return V_02842C_STENCIL_ADD_WRAP
;
820 case VK_STENCIL_OP_DECREMENT_AND_WRAP
:
821 return V_02842C_STENCIL_SUB_WRAP
;
827 radv_pipeline_init_depth_stencil_state(struct radv_pipeline
*pipeline
,
828 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
829 const struct radv_graphics_pipeline_create_info
*extra
)
831 const VkPipelineDepthStencilStateCreateInfo
*vkds
= pCreateInfo
->pDepthStencilState
;
832 struct radv_depth_stencil_state
*ds
= &pipeline
->graphics
.ds
;
834 memset(ds
, 0, sizeof(*ds
));
837 ds
->db_depth_control
= S_028800_Z_ENABLE(vkds
->depthTestEnable
? 1 : 0) |
838 S_028800_Z_WRITE_ENABLE(vkds
->depthWriteEnable
? 1 : 0) |
839 S_028800_ZFUNC(vkds
->depthCompareOp
) |
840 S_028800_DEPTH_BOUNDS_ENABLE(vkds
->depthBoundsTestEnable
? 1 : 0);
842 if (vkds
->stencilTestEnable
) {
843 ds
->db_depth_control
|= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
844 ds
->db_depth_control
|= S_028800_STENCILFUNC(vkds
->front
.compareOp
);
845 ds
->db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds
->front
.failOp
));
846 ds
->db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds
->front
.passOp
));
847 ds
->db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds
->front
.depthFailOp
));
849 ds
->db_depth_control
|= S_028800_STENCILFUNC_BF(vkds
->back
.compareOp
);
850 ds
->db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds
->back
.failOp
));
851 ds
->db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds
->back
.passOp
));
852 ds
->db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds
->back
.depthFailOp
));
857 ds
->db_render_control
|= S_028000_DEPTH_CLEAR_ENABLE(extra
->db_depth_clear
);
858 ds
->db_render_control
|= S_028000_STENCIL_CLEAR_ENABLE(extra
->db_stencil_clear
);
860 ds
->db_render_control
|= S_028000_RESUMMARIZE_ENABLE(extra
->db_resummarize
);
861 ds
->db_render_control
|= S_028000_DEPTH_COMPRESS_DISABLE(extra
->db_flush_depth_inplace
);
862 ds
->db_render_control
|= S_028000_STENCIL_COMPRESS_DISABLE(extra
->db_flush_stencil_inplace
);
863 ds
->db_render_override2
|= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra
->db_depth_disable_expclear
);
864 ds
->db_render_override2
|= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra
->db_stencil_disable_expclear
);
868 static uint32_t si_translate_fill(VkPolygonMode 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
;
879 return V_028814_X_DRAW_POINTS
;
883 radv_pipeline_init_raster_state(struct radv_pipeline
*pipeline
,
884 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
886 const VkPipelineRasterizationStateCreateInfo
*vkraster
= pCreateInfo
->pRasterizationState
;
887 struct radv_raster_state
*raster
= &pipeline
->graphics
.raster
;
889 memset(raster
, 0, sizeof(*raster
));
891 raster
->spi_interp_control
=
892 S_0286D4_FLAT_SHADE_ENA(1) |
893 S_0286D4_PNT_SPRITE_ENA(1) |
894 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
895 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
896 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
897 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
898 S_0286D4_PNT_SPRITE_TOP_1(0); // vulkan is top to bottom - 1.0 at bottom
900 raster
->pa_cl_vs_out_cntl
= S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(1);
901 raster
->pa_cl_clip_cntl
= S_028810_PS_UCP_MODE(3) |
902 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
903 S_028810_ZCLIP_NEAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
904 S_028810_ZCLIP_FAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
905 S_028810_DX_RASTERIZATION_KILL(vkraster
->rasterizerDiscardEnable
? 1 : 0) |
906 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
908 raster
->pa_su_vtx_cntl
=
909 S_028BE4_PIX_CENTER(1) | // TODO verify
910 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
911 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
);
913 raster
->pa_su_sc_mode_cntl
=
914 S_028814_FACE(vkraster
->frontFace
) |
915 S_028814_CULL_FRONT(!!(vkraster
->cullMode
& VK_CULL_MODE_FRONT_BIT
)) |
916 S_028814_CULL_BACK(!!(vkraster
->cullMode
& VK_CULL_MODE_BACK_BIT
)) |
917 S_028814_POLY_MODE(vkraster
->polygonMode
!= VK_POLYGON_MODE_FILL
) |
918 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
919 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
920 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
921 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
922 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster
->depthBiasEnable
? 1 : 0);
927 radv_pipeline_init_multisample_state(struct radv_pipeline
*pipeline
,
928 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
930 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
931 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
932 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
933 unsigned num_tile_pipes
= pipeline
->device
->instance
->physicalDevice
.rad_info
.num_tile_pipes
;
934 int ps_iter_samples
= 1;
935 uint32_t mask
= 0xffff;
937 ms
->num_samples
= vkms
->rasterizationSamples
;
938 ms
->pa_sc_line_cntl
= S_028BDC_DX10_DIAMOND_TEST_ENA(1);
939 ms
->pa_sc_aa_config
= 0;
940 ms
->db_eqaa
= S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
941 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
942 ms
->pa_sc_mode_cntl_1
=
943 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
944 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
946 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
947 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
948 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
949 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
950 EG_S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
951 EG_S_028A4C_FORCE_EOV_REZ_ENABLE(1);
953 if (vkms
->rasterizationSamples
> 1) {
954 unsigned log_samples
= util_logbase2(vkms
->rasterizationSamples
);
955 unsigned log_ps_iter_samples
= util_logbase2(util_next_power_of_two(ps_iter_samples
));
956 ms
->pa_sc_mode_cntl_0
= S_028A48_MSAA_ENABLE(1);
957 ms
->pa_sc_line_cntl
|= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
958 ms
->db_eqaa
|= S_028804_MAX_ANCHOR_SAMPLES(log_samples
) |
959 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples
) |
960 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples
) |
961 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples
);
962 ms
->pa_sc_aa_config
|= S_028BE0_MSAA_NUM_SAMPLES(log_samples
) |
963 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples
)) |
964 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples
); /* CM_R_028BE0_PA_SC_AA_CONFIG */
965 ms
->pa_sc_mode_cntl_1
|= EG_S_028A4C_PS_ITER_SAMPLE(ps_iter_samples
> 1);
968 if (vkms
->alphaToCoverageEnable
)
969 blend
->db_alpha_to_mask
|= S_028B70_ALPHA_TO_MASK_ENABLE(1);
971 if (vkms
->pSampleMask
) {
972 mask
= vkms
->pSampleMask
[0] & 0xffff;
975 ms
->pa_sc_aa_mask
[0] = mask
| (mask
<< 16);
976 ms
->pa_sc_aa_mask
[1] = mask
| (mask
<< 16);
980 si_translate_prim(enum VkPrimitiveTopology topology
)
983 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
984 return V_008958_DI_PT_POINTLIST
;
985 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
986 return V_008958_DI_PT_LINELIST
;
987 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
988 return V_008958_DI_PT_LINESTRIP
;
989 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
990 return V_008958_DI_PT_TRILIST
;
991 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
992 return V_008958_DI_PT_TRISTRIP
;
993 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
994 return V_008958_DI_PT_TRIFAN
;
995 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
996 return V_008958_DI_PT_LINELIST_ADJ
;
997 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
998 return V_008958_DI_PT_LINESTRIP_ADJ
;
999 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1000 return V_008958_DI_PT_TRILIST_ADJ
;
1001 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1002 return V_008958_DI_PT_TRISTRIP_ADJ
;
1003 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1004 return V_008958_DI_PT_PATCH
;
1012 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology
)
1015 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1016 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1017 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
1018 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1019 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1020 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1021 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1022 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
1023 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1024 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1025 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1026 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1027 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1028 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1035 static unsigned si_map_swizzle(unsigned swizzle
)
1039 return V_008F0C_SQ_SEL_Y
;
1041 return V_008F0C_SQ_SEL_Z
;
1043 return V_008F0C_SQ_SEL_W
;
1045 return V_008F0C_SQ_SEL_0
;
1047 return V_008F0C_SQ_SEL_1
;
1048 default: /* VK_SWIZZLE_X */
1049 return V_008F0C_SQ_SEL_X
;
1054 radv_pipeline_init_dynamic_state(struct radv_pipeline
*pipeline
,
1055 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1057 radv_cmd_dirty_mask_t states
= RADV_CMD_DIRTY_DYNAMIC_ALL
;
1058 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
1059 struct radv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
1061 pipeline
->dynamic_state
= default_dynamic_state
;
1063 if (pCreateInfo
->pDynamicState
) {
1064 /* Remove all of the states that are marked as dynamic */
1065 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1066 for (uint32_t s
= 0; s
< count
; s
++)
1067 states
&= ~(1 << pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1070 struct radv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1072 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1073 if (states
& (1 << VK_DYNAMIC_STATE_VIEWPORT
)) {
1074 typed_memcpy(dynamic
->viewport
.viewports
,
1075 pCreateInfo
->pViewportState
->pViewports
,
1076 pCreateInfo
->pViewportState
->viewportCount
);
1079 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1080 if (states
& (1 << VK_DYNAMIC_STATE_SCISSOR
)) {
1081 typed_memcpy(dynamic
->scissor
.scissors
,
1082 pCreateInfo
->pViewportState
->pScissors
,
1083 pCreateInfo
->pViewportState
->scissorCount
);
1086 if (states
& (1 << VK_DYNAMIC_STATE_LINE_WIDTH
)) {
1087 assert(pCreateInfo
->pRasterizationState
);
1088 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1091 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BIAS
)) {
1092 assert(pCreateInfo
->pRasterizationState
);
1093 dynamic
->depth_bias
.bias
=
1094 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1095 dynamic
->depth_bias
.clamp
=
1096 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1097 dynamic
->depth_bias
.slope
=
1098 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1101 if (states
& (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS
)) {
1102 assert(pCreateInfo
->pColorBlendState
);
1103 typed_memcpy(dynamic
->blend_constants
,
1104 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1107 /* If there is no depthstencil attachment, then don't read
1108 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1109 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1110 * no need to override the depthstencil defaults in
1111 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1113 * From the Vulkan spec (20 Oct 2015, git-aa308cb):
1115 * pDepthStencilState [...] may only be NULL if renderPass and subpass
1116 * specify a subpass that has no depth/stencil attachment.
1118 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1119 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS
)) {
1120 assert(pCreateInfo
->pDepthStencilState
);
1121 dynamic
->depth_bounds
.min
=
1122 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1123 dynamic
->depth_bounds
.max
=
1124 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1127 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
)) {
1128 assert(pCreateInfo
->pDepthStencilState
);
1129 dynamic
->stencil_compare_mask
.front
=
1130 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1131 dynamic
->stencil_compare_mask
.back
=
1132 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1135 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
)) {
1136 assert(pCreateInfo
->pDepthStencilState
);
1137 dynamic
->stencil_write_mask
.front
=
1138 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1139 dynamic
->stencil_write_mask
.back
=
1140 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1143 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE
)) {
1144 assert(pCreateInfo
->pDepthStencilState
);
1145 dynamic
->stencil_reference
.front
=
1146 pCreateInfo
->pDepthStencilState
->front
.reference
;
1147 dynamic
->stencil_reference
.back
=
1148 pCreateInfo
->pDepthStencilState
->back
.reference
;
1152 pipeline
->dynamic_state_mask
= states
;
1155 static union ac_shader_variant_key
1156 radv_compute_vs_key(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1158 union ac_shader_variant_key key
;
1159 const VkPipelineVertexInputStateCreateInfo
*input_state
=
1160 pCreateInfo
->pVertexInputState
;
1162 memset(&key
, 0, sizeof(key
));
1163 key
.vs
.instance_rate_inputs
= 0;
1165 for (unsigned i
= 0; i
< input_state
->vertexAttributeDescriptionCount
; ++i
) {
1167 binding
= input_state
->pVertexAttributeDescriptions
[i
].binding
;
1168 if (input_state
->pVertexBindingDescriptions
[binding
].inputRate
)
1169 key
.vs
.instance_rate_inputs
|= 1u << input_state
->pVertexAttributeDescriptions
[i
].location
;
1175 radv_pipeline_init(struct radv_pipeline
*pipeline
,
1176 struct radv_device
*device
,
1177 struct radv_pipeline_cache
*cache
,
1178 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1179 const struct radv_graphics_pipeline_create_info
*extra
,
1180 const VkAllocationCallbacks
*alloc
)
1182 struct radv_shader_module fs_m
= {0};
1184 bool dump
= getenv("RADV_DUMP_SHADERS");
1186 alloc
= &device
->alloc
;
1188 pipeline
->device
= device
;
1189 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
1191 radv_pipeline_init_dynamic_state(pipeline
, pCreateInfo
);
1192 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
1193 struct radv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1194 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
1195 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
1196 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
1197 modules
[stage
] = radv_shader_module_from_handle(pStages
[stage
]->module
);
1200 radv_pipeline_init_blend_state(pipeline
, pCreateInfo
, extra
);
1203 if (modules
[MESA_SHADER_VERTEX
]) {
1204 union ac_shader_variant_key key
= radv_compute_vs_key(pCreateInfo
);
1206 pipeline
->shaders
[MESA_SHADER_VERTEX
] =
1207 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_VERTEX
],
1208 pStages
[MESA_SHADER_VERTEX
]->pName
,
1210 pStages
[MESA_SHADER_VERTEX
]->pSpecializationInfo
,
1211 pipeline
->layout
, &key
, dump
);
1213 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_VERTEX
);
1216 if (!modules
[MESA_SHADER_FRAGMENT
]) {
1218 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
1219 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "noop_fs");
1220 fs_m
.nir
= fs_b
.shader
;
1221 modules
[MESA_SHADER_FRAGMENT
] = &fs_m
;
1224 if (modules
[MESA_SHADER_FRAGMENT
]) {
1225 union ac_shader_variant_key key
;
1226 key
.fs
.col_format
= pipeline
->graphics
.blend
.spi_shader_col_format
;
1227 key
.fs
.is_int8
= radv_pipeline_compute_is_int8(pCreateInfo
);
1229 const VkPipelineShaderStageCreateInfo
*stage
= pStages
[MESA_SHADER_FRAGMENT
];
1231 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] =
1232 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_FRAGMENT
],
1233 stage
? stage
->pName
: "main",
1234 MESA_SHADER_FRAGMENT
,
1235 stage
? stage
->pSpecializationInfo
: NULL
,
1236 pipeline
->layout
, &key
, dump
);
1237 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_FRAGMENT
);
1241 ralloc_free(fs_m
.nir
);
1243 radv_pipeline_init_depth_stencil_state(pipeline
, pCreateInfo
, extra
);
1244 radv_pipeline_init_raster_state(pipeline
, pCreateInfo
);
1245 radv_pipeline_init_multisample_state(pipeline
, pCreateInfo
);
1246 pipeline
->graphics
.prim
= si_translate_prim(pCreateInfo
->pInputAssemblyState
->topology
);
1247 pipeline
->graphics
.gs_out
= si_conv_prim_to_gs_out(pCreateInfo
->pInputAssemblyState
->topology
);
1248 if (extra
&& extra
->use_rectlist
) {
1249 pipeline
->graphics
.prim
= V_008958_DI_PT_RECTLIST
;
1250 pipeline
->graphics
.gs_out
= V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1252 pipeline
->graphics
.prim_restart_enable
= !!pCreateInfo
->pInputAssemblyState
->primitiveRestartEnable
;
1254 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1255 pCreateInfo
->pVertexInputState
;
1256 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1257 const VkVertexInputAttributeDescription
*desc
=
1258 &vi_info
->pVertexAttributeDescriptions
[i
];
1259 unsigned loc
= desc
->location
;
1260 const struct vk_format_description
*format_desc
;
1262 uint32_t num_format
, data_format
;
1263 format_desc
= vk_format_description(desc
->format
);
1264 first_non_void
= vk_format_get_first_non_void_channel(desc
->format
);
1266 num_format
= radv_translate_buffer_numformat(format_desc
, first_non_void
);
1267 data_format
= radv_translate_buffer_dataformat(format_desc
, first_non_void
);
1269 pipeline
->va_rsrc_word3
[loc
] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc
->swizzle
[0])) |
1270 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc
->swizzle
[1])) |
1271 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc
->swizzle
[2])) |
1272 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc
->swizzle
[3])) |
1273 S_008F0C_NUM_FORMAT(num_format
) |
1274 S_008F0C_DATA_FORMAT(data_format
);
1275 pipeline
->va_format_size
[loc
] = format_desc
->block
.bits
/ 8;
1276 pipeline
->va_offset
[loc
] = desc
->offset
;
1277 pipeline
->va_binding
[loc
] = desc
->binding
;
1278 pipeline
->num_vertex_attribs
= MAX2(pipeline
->num_vertex_attribs
, loc
+ 1);
1281 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1282 const VkVertexInputBindingDescription
*desc
=
1283 &vi_info
->pVertexBindingDescriptions
[i
];
1285 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
1292 radv_graphics_pipeline_create(
1294 VkPipelineCache _cache
,
1295 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1296 const struct radv_graphics_pipeline_create_info
*extra
,
1297 const VkAllocationCallbacks
*pAllocator
,
1298 VkPipeline
*pPipeline
)
1300 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1301 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
1302 struct radv_pipeline
*pipeline
;
1305 pipeline
= radv_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
1306 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1307 if (pipeline
== NULL
)
1308 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1310 memset(pipeline
, 0, sizeof(*pipeline
));
1311 result
= radv_pipeline_init(pipeline
, device
, cache
,
1312 pCreateInfo
, extra
, pAllocator
);
1313 if (result
!= VK_SUCCESS
) {
1314 radv_free2(&device
->alloc
, pAllocator
, pipeline
);
1318 *pPipeline
= radv_pipeline_to_handle(pipeline
);
1323 VkResult
radv_CreateGraphicsPipelines(
1325 VkPipelineCache pipelineCache
,
1327 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
1328 const VkAllocationCallbacks
* pAllocator
,
1329 VkPipeline
* pPipelines
)
1331 VkResult result
= VK_SUCCESS
;
1334 for (; i
< count
; i
++) {
1335 result
= radv_graphics_pipeline_create(_device
,
1338 NULL
, pAllocator
, &pPipelines
[i
]);
1339 if (result
!= VK_SUCCESS
) {
1340 for (unsigned j
= 0; j
< i
; j
++) {
1341 radv_DestroyPipeline(_device
, pPipelines
[j
], pAllocator
);
1351 static VkResult
radv_compute_pipeline_create(
1353 VkPipelineCache _cache
,
1354 const VkComputePipelineCreateInfo
* pCreateInfo
,
1355 const VkAllocationCallbacks
* pAllocator
,
1356 VkPipeline
* pPipeline
)
1358 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1359 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
1360 RADV_FROM_HANDLE(radv_shader_module
, module
, pCreateInfo
->stage
.module
);
1361 struct radv_pipeline
*pipeline
;
1362 bool dump
= getenv("RADV_DUMP_SHADERS");
1364 pipeline
= radv_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
1365 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1366 if (pipeline
== NULL
)
1367 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1369 memset(pipeline
, 0, sizeof(*pipeline
));
1370 pipeline
->device
= device
;
1371 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
1373 pipeline
->shaders
[MESA_SHADER_COMPUTE
] =
1374 radv_pipeline_compile(pipeline
, cache
, module
,
1375 pCreateInfo
->stage
.pName
,
1376 MESA_SHADER_COMPUTE
,
1377 pCreateInfo
->stage
.pSpecializationInfo
,
1378 pipeline
->layout
, NULL
, dump
);
1380 *pPipeline
= radv_pipeline_to_handle(pipeline
);
1383 VkResult
radv_CreateComputePipelines(
1385 VkPipelineCache pipelineCache
,
1387 const VkComputePipelineCreateInfo
* pCreateInfos
,
1388 const VkAllocationCallbacks
* pAllocator
,
1389 VkPipeline
* pPipelines
)
1391 VkResult result
= VK_SUCCESS
;
1394 for (; i
< count
; i
++) {
1395 result
= radv_compute_pipeline_create(_device
, pipelineCache
,
1397 pAllocator
, &pPipelines
[i
]);
1398 if (result
!= VK_SUCCESS
) {
1399 for (unsigned j
= 0; j
< i
; j
++) {
1400 radv_DestroyPipeline(_device
, pPipelines
[j
], pAllocator
);