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
= vk_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 vk_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 vk_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
);
147 NIR_PASS(progress
, shader
, nir_opt_conditional_discard
);
152 radv_shader_compile_to_nir(struct radv_device
*device
,
153 struct radv_shader_module
*module
,
154 const char *entrypoint_name
,
155 gl_shader_stage stage
,
156 const VkSpecializationInfo
*spec_info
,
159 if (strcmp(entrypoint_name
, "main") != 0) {
160 radv_finishme("Multiple shaders per module not really supported");
164 nir_function
*entry_point
;
166 /* Some things such as our meta clear/blit code will give us a NIR
167 * shader directly. In that case, we just ignore the SPIR-V entirely
168 * and just use the NIR shader */
170 nir
->options
= &nir_options
;
171 nir_validate_shader(nir
);
173 assert(exec_list_length(&nir
->functions
) == 1);
174 struct exec_node
*node
= exec_list_get_head(&nir
->functions
);
175 entry_point
= exec_node_data(nir_function
, node
, node
);
177 uint32_t *spirv
= (uint32_t *) module
->data
;
178 assert(module
->size
% 4 == 0);
180 uint32_t num_spec_entries
= 0;
181 struct nir_spirv_specialization
*spec_entries
= NULL
;
182 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
183 num_spec_entries
= spec_info
->mapEntryCount
;
184 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
185 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
186 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
187 const void *data
= spec_info
->pData
+ entry
.offset
;
188 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
190 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
191 if (spec_info
->dataSize
== 8)
192 spec_entries
[i
].data64
= *(const uint64_t *)data
;
194 spec_entries
[i
].data32
= *(const uint32_t *)data
;
197 const struct nir_spirv_supported_extensions supported_ext
= {
199 entry_point
= spirv_to_nir(spirv
, module
->size
/ 4,
200 spec_entries
, num_spec_entries
,
201 stage
, entrypoint_name
, &supported_ext
, &nir_options
);
202 nir
= entry_point
->shader
;
203 assert(nir
->stage
== stage
);
204 nir_validate_shader(nir
);
208 /* We have to lower away local constant initializers right before we
209 * inline functions. That way they get properly initialized at the top
210 * of the function and not at the top of its caller.
212 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_local
);
213 NIR_PASS_V(nir
, nir_lower_returns
);
214 NIR_PASS_V(nir
, nir_inline_functions
);
216 /* Pick off the single entrypoint that we want */
217 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
218 if (func
!= entry_point
)
219 exec_node_remove(&func
->node
);
221 assert(exec_list_length(&nir
->functions
) == 1);
222 entry_point
->name
= ralloc_strdup(entry_point
, "main");
224 NIR_PASS_V(nir
, nir_remove_dead_variables
,
225 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
227 /* Now that we've deleted all but the main function, we can go ahead and
228 * lower the rest of the constant initializers.
230 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
231 NIR_PASS_V(nir
, nir_lower_system_values
);
234 /* Vulkan uses the separate-shader linking model */
235 nir
->info
->separate_shader
= true;
237 // nir = brw_preprocess_nir(compiler, nir);
239 nir_shader_gather_info(nir
, entry_point
->impl
);
241 nir_variable_mode indirect_mask
= 0;
242 // if (compiler->glsl_compiler_options[stage].EmitNoIndirectInput)
243 indirect_mask
|= nir_var_shader_in
;
244 // if (compiler->glsl_compiler_options[stage].EmitNoIndirectTemp)
245 indirect_mask
|= nir_var_local
;
247 nir_lower_indirect_derefs(nir
, indirect_mask
);
249 static const nir_lower_tex_options tex_options
= {
253 nir_lower_tex(nir
, &tex_options
);
255 nir_lower_vars_to_ssa(nir
);
256 nir_lower_var_copies(nir
);
257 nir_lower_global_vars_to_local(nir
);
258 nir_remove_dead_variables(nir
, nir_var_local
);
259 radv_optimize_nir(nir
);
262 nir_print_shader(nir
, stderr
);
267 static const char *radv_get_shader_name(struct radv_shader_variant
*var
,
268 gl_shader_stage stage
)
271 case MESA_SHADER_VERTEX
: return "Vertex Shader as VS";
272 case MESA_SHADER_FRAGMENT
: return "Pixel Shader";
273 case MESA_SHADER_COMPUTE
: return "Compute Shader";
275 return "Unknown shader";
279 static void radv_dump_pipeline_stats(struct radv_device
*device
, struct radv_pipeline
*pipeline
)
281 unsigned lds_increment
= device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
? 512 : 256;
282 struct radv_shader_variant
*var
;
283 struct ac_shader_config
*conf
;
286 unsigned max_simd_waves
= 10;
287 unsigned lds_per_wave
= 0;
289 for (i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
290 if (!pipeline
->shaders
[i
])
292 var
= pipeline
->shaders
[i
];
296 if (i
== MESA_SHADER_FRAGMENT
) {
297 lds_per_wave
= conf
->lds_size
* lds_increment
+
298 align(var
->info
.fs
.num_interp
* 48, lds_increment
);
301 if (conf
->num_sgprs
) {
302 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
)
303 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
305 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
309 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
311 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
315 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
317 fprintf(file
, "\n%s:\n",
318 radv_get_shader_name(var
, i
));
319 if (i
== MESA_SHADER_FRAGMENT
) {
320 fprintf(file
, "*** SHADER CONFIG ***\n"
321 "SPI_PS_INPUT_ADDR = 0x%04x\n"
322 "SPI_PS_INPUT_ENA = 0x%04x\n",
323 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
325 fprintf(file
, "*** SHADER STATS ***\n"
328 "Spilled SGPRs: %d\n"
329 "Spilled VGPRs: %d\n"
330 "Code Size: %d bytes\n"
332 "Scratch: %d bytes per wave\n"
334 "********************\n\n\n",
335 conf
->num_sgprs
, conf
->num_vgprs
,
336 conf
->spilled_sgprs
, conf
->spilled_vgprs
, var
->code_size
,
337 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
342 void radv_shader_variant_destroy(struct radv_device
*device
,
343 struct radv_shader_variant
*variant
)
345 if (__sync_fetch_and_sub(&variant
->ref_count
, 1) != 1)
348 device
->ws
->buffer_destroy(variant
->bo
);
352 static void radv_fill_shader_variant(struct radv_device
*device
,
353 struct radv_shader_variant
*variant
,
354 struct ac_shader_binary
*binary
,
355 gl_shader_stage stage
)
357 variant
->code_size
= binary
->code_size
;
358 bool scratch_enabled
= variant
->config
.scratch_bytes_per_wave
> 0;
359 unsigned vgpr_comp_cnt
= 0;
362 radv_finishme("shader scratch space");
365 case MESA_SHADER_VERTEX
:
366 variant
->rsrc2
= S_00B12C_USER_SGPR(variant
->info
.num_user_sgprs
) |
367 S_00B12C_SCRATCH_EN(scratch_enabled
);
368 vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
370 case MESA_SHADER_FRAGMENT
:
371 variant
->rsrc2
= S_00B12C_USER_SGPR(variant
->info
.num_user_sgprs
) |
372 S_00B12C_SCRATCH_EN(scratch_enabled
);
374 case MESA_SHADER_COMPUTE
:
375 variant
->rsrc2
= S_00B84C_USER_SGPR(variant
->info
.num_user_sgprs
) |
376 S_00B84C_SCRATCH_EN(scratch_enabled
) |
377 S_00B84C_TGID_X_EN(1) | S_00B84C_TGID_Y_EN(1) |
378 S_00B84C_TGID_Z_EN(1) | S_00B84C_TIDIG_COMP_CNT(2) |
379 S_00B84C_TG_SIZE_EN(1) |
380 S_00B84C_LDS_SIZE(variant
->config
.lds_size
);
383 unreachable("unsupported shader type");
387 variant
->rsrc1
= S_00B848_VGPRS((variant
->config
.num_vgprs
- 1) / 4) |
388 S_00B848_SGPRS((variant
->config
.num_sgprs
- 1) / 8) |
389 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
390 S_00B848_DX10_CLAMP(1) |
391 S_00B848_FLOAT_MODE(variant
->config
.float_mode
);
393 variant
->bo
= device
->ws
->buffer_create(device
->ws
, binary
->code_size
, 256,
394 RADEON_DOMAIN_GTT
, RADEON_FLAG_CPU_ACCESS
);
396 void *ptr
= device
->ws
->buffer_map(variant
->bo
);
397 memcpy(ptr
, binary
->code
, binary
->code_size
);
398 device
->ws
->buffer_unmap(variant
->bo
);
403 static struct radv_shader_variant
*radv_shader_variant_create(struct radv_device
*device
,
404 struct nir_shader
*shader
,
405 struct radv_pipeline_layout
*layout
,
406 const union ac_shader_variant_key
*key
,
408 unsigned *code_size_out
,
411 struct radv_shader_variant
*variant
= calloc(1, sizeof(struct radv_shader_variant
));
412 enum radeon_family chip_family
= device
->instance
->physicalDevice
.rad_info
.family
;
413 LLVMTargetMachineRef tm
;
417 struct ac_nir_compiler_options options
= {0};
418 options
.layout
= layout
;
422 struct ac_shader_binary binary
;
424 options
.unsafe_math
= !!(device
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
);
425 options
.family
= chip_family
;
426 options
.chip_class
= device
->instance
->physicalDevice
.rad_info
.chip_class
;
427 tm
= ac_create_target_machine(chip_family
);
428 ac_compile_nir_shader(tm
, &binary
, &variant
->config
,
429 &variant
->info
, shader
, &options
, dump
);
430 LLVMDisposeTargetMachine(tm
);
432 radv_fill_shader_variant(device
, variant
, &binary
, shader
->stage
);
435 *code_out
= binary
.code
;
436 *code_size_out
= binary
.code_size
;
441 free(binary
.global_symbol_offsets
);
443 free(binary
.disasm_string
);
444 variant
->ref_count
= 1;
449 static struct radv_shader_variant
*
450 radv_pipeline_compile(struct radv_pipeline
*pipeline
,
451 struct radv_pipeline_cache
*cache
,
452 struct radv_shader_module
*module
,
453 const char *entrypoint
,
454 gl_shader_stage stage
,
455 const VkSpecializationInfo
*spec_info
,
456 struct radv_pipeline_layout
*layout
,
457 const union ac_shader_variant_key
*key
)
459 unsigned char sha1
[20];
460 struct radv_shader_variant
*variant
;
463 unsigned code_size
= 0;
464 bool dump
= (pipeline
->device
->debug_flags
& RADV_DEBUG_DUMP_SHADERS
);
467 _mesa_sha1_compute(module
->nir
->info
->name
,
468 strlen(module
->nir
->info
->name
),
471 radv_hash_shader(sha1
, module
, entrypoint
, spec_info
, layout
, key
);
474 variant
= radv_create_shader_variant_from_pipeline_cache(pipeline
->device
,
481 nir
= radv_shader_compile_to_nir(pipeline
->device
,
482 module
, entrypoint
, stage
,
487 variant
= radv_shader_variant_create(pipeline
->device
, nir
, layout
, key
,
488 &code
, &code_size
, dump
);
492 if (variant
&& cache
)
493 variant
= radv_pipeline_cache_insert_shader(cache
, sha1
, variant
,
501 static uint32_t si_translate_blend_function(VkBlendOp op
)
504 case VK_BLEND_OP_ADD
:
505 return V_028780_COMB_DST_PLUS_SRC
;
506 case VK_BLEND_OP_SUBTRACT
:
507 return V_028780_COMB_SRC_MINUS_DST
;
508 case VK_BLEND_OP_REVERSE_SUBTRACT
:
509 return V_028780_COMB_DST_MINUS_SRC
;
510 case VK_BLEND_OP_MIN
:
511 return V_028780_COMB_MIN_DST_SRC
;
512 case VK_BLEND_OP_MAX
:
513 return V_028780_COMB_MAX_DST_SRC
;
519 static uint32_t si_translate_blend_factor(VkBlendFactor factor
)
522 case VK_BLEND_FACTOR_ZERO
:
523 return V_028780_BLEND_ZERO
;
524 case VK_BLEND_FACTOR_ONE
:
525 return V_028780_BLEND_ONE
;
526 case VK_BLEND_FACTOR_SRC_COLOR
:
527 return V_028780_BLEND_SRC_COLOR
;
528 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
529 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
530 case VK_BLEND_FACTOR_DST_COLOR
:
531 return V_028780_BLEND_DST_COLOR
;
532 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
:
533 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
534 case VK_BLEND_FACTOR_SRC_ALPHA
:
535 return V_028780_BLEND_SRC_ALPHA
;
536 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
537 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
538 case VK_BLEND_FACTOR_DST_ALPHA
:
539 return V_028780_BLEND_DST_ALPHA
;
540 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
:
541 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
542 case VK_BLEND_FACTOR_CONSTANT_COLOR
:
543 return V_028780_BLEND_CONSTANT_COLOR
;
544 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
:
545 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
546 case VK_BLEND_FACTOR_CONSTANT_ALPHA
:
547 return V_028780_BLEND_CONSTANT_ALPHA
;
548 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
:
549 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
550 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
551 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
552 case VK_BLEND_FACTOR_SRC1_COLOR
:
553 return V_028780_BLEND_SRC1_COLOR
;
554 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
555 return V_028780_BLEND_INV_SRC1_COLOR
;
556 case VK_BLEND_FACTOR_SRC1_ALPHA
:
557 return V_028780_BLEND_SRC1_ALPHA
;
558 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
559 return V_028780_BLEND_INV_SRC1_ALPHA
;
565 static bool is_dual_src(VkBlendFactor factor
)
568 case VK_BLEND_FACTOR_SRC1_COLOR
:
569 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
570 case VK_BLEND_FACTOR_SRC1_ALPHA
:
571 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
578 static unsigned si_choose_spi_color_format(VkFormat vk_format
,
580 bool blend_need_alpha
)
582 const struct vk_format_description
*desc
= vk_format_description(vk_format
);
583 unsigned format
, ntype
, swap
;
585 /* Alpha is needed for alpha-to-coverage.
586 * Blending may be with or without alpha.
588 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
589 unsigned alpha
= 0; /* exports alpha, but may not support blending */
590 unsigned blend
= 0; /* supports blending, but may not export alpha */
591 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
593 format
= radv_translate_colorformat(vk_format
);
594 ntype
= radv_translate_color_numformat(vk_format
, desc
,
595 vk_format_get_first_non_void_channel(vk_format
));
596 swap
= radv_translate_colorswap(vk_format
, false);
598 /* Choose the SPI color formats. These are required values for Stoney/RB+.
599 * Other chips have multiple choices, though they are not necessarily better.
602 case V_028C70_COLOR_5_6_5
:
603 case V_028C70_COLOR_1_5_5_5
:
604 case V_028C70_COLOR_5_5_5_1
:
605 case V_028C70_COLOR_4_4_4_4
:
606 case V_028C70_COLOR_10_11_11
:
607 case V_028C70_COLOR_11_11_10
:
608 case V_028C70_COLOR_8
:
609 case V_028C70_COLOR_8_8
:
610 case V_028C70_COLOR_8_8_8_8
:
611 case V_028C70_COLOR_10_10_10_2
:
612 case V_028C70_COLOR_2_10_10_10
:
613 if (ntype
== V_028C70_NUMBER_UINT
)
614 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
615 else if (ntype
== V_028C70_NUMBER_SINT
)
616 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
618 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
621 case V_028C70_COLOR_16
:
622 case V_028C70_COLOR_16_16
:
623 case V_028C70_COLOR_16_16_16_16
:
624 if (ntype
== V_028C70_NUMBER_UNORM
||
625 ntype
== V_028C70_NUMBER_SNORM
) {
626 /* UNORM16 and SNORM16 don't support blending */
627 if (ntype
== V_028C70_NUMBER_UNORM
)
628 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
630 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
632 /* Use 32 bits per channel for blending. */
633 if (format
== V_028C70_COLOR_16
) {
634 if (swap
== V_028C70_SWAP_STD
) { /* R */
635 blend
= V_028714_SPI_SHADER_32_R
;
636 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
637 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
638 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
641 } else if (format
== V_028C70_COLOR_16_16
) {
642 if (swap
== V_028C70_SWAP_STD
) { /* RG */
643 blend
= V_028714_SPI_SHADER_32_GR
;
644 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
645 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
646 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
649 } else /* 16_16_16_16 */
650 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
651 } else if (ntype
== V_028C70_NUMBER_UINT
)
652 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
653 else if (ntype
== V_028C70_NUMBER_SINT
)
654 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
655 else if (ntype
== V_028C70_NUMBER_FLOAT
)
656 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
661 case V_028C70_COLOR_32
:
662 if (swap
== V_028C70_SWAP_STD
) { /* R */
663 blend
= normal
= V_028714_SPI_SHADER_32_R
;
664 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
665 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
666 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
671 case V_028C70_COLOR_32_32
:
672 if (swap
== V_028C70_SWAP_STD
) { /* RG */
673 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
674 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
675 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
676 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
681 case V_028C70_COLOR_32_32_32_32
:
682 case V_028C70_COLOR_8_24
:
683 case V_028C70_COLOR_24_8
:
684 case V_028C70_COLOR_X24_8_32_FLOAT
:
685 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
689 unreachable("unhandled blend format");
692 if (blend_enable
&& blend_need_alpha
)
694 else if(blend_need_alpha
)
696 else if(blend_enable
)
702 static unsigned si_get_cb_shader_mask(unsigned spi_shader_col_format
)
704 unsigned i
, cb_shader_mask
= 0;
706 for (i
= 0; i
< 8; i
++) {
707 switch ((spi_shader_col_format
>> (i
* 4)) & 0xf) {
708 case V_028714_SPI_SHADER_ZERO
:
710 case V_028714_SPI_SHADER_32_R
:
711 cb_shader_mask
|= 0x1 << (i
* 4);
713 case V_028714_SPI_SHADER_32_GR
:
714 cb_shader_mask
|= 0x3 << (i
* 4);
716 case V_028714_SPI_SHADER_32_AR
:
717 cb_shader_mask
|= 0x9 << (i
* 4);
719 case V_028714_SPI_SHADER_FP16_ABGR
:
720 case V_028714_SPI_SHADER_UNORM16_ABGR
:
721 case V_028714_SPI_SHADER_SNORM16_ABGR
:
722 case V_028714_SPI_SHADER_UINT16_ABGR
:
723 case V_028714_SPI_SHADER_SINT16_ABGR
:
724 case V_028714_SPI_SHADER_32_ABGR
:
725 cb_shader_mask
|= 0xf << (i
* 4);
731 return cb_shader_mask
;
735 radv_pipeline_compute_spi_color_formats(struct radv_pipeline
*pipeline
,
736 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
737 uint32_t blend_enable
,
738 uint32_t blend_need_alpha
,
739 bool single_cb_enable
,
740 bool blend_mrt0_is_dual_src
)
742 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
743 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
744 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
745 unsigned col_format
= 0;
747 for (unsigned i
= 0; i
< (single_cb_enable
? 1 : subpass
->color_count
); ++i
) {
748 struct radv_render_pass_attachment
*attachment
;
751 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
753 cf
= si_choose_spi_color_format(attachment
->format
,
754 blend_enable
& (1 << i
),
755 blend_need_alpha
& (1 << i
));
757 col_format
|= cf
<< (4 * i
);
760 blend
->cb_shader_mask
= si_get_cb_shader_mask(col_format
);
762 if (blend_mrt0_is_dual_src
)
763 col_format
|= (col_format
& 0xf) << 4;
765 col_format
|= V_028714_SPI_SHADER_32_R
;
766 blend
->spi_shader_col_format
= col_format
;
770 format_is_int8(VkFormat format
)
772 const struct vk_format_description
*desc
= vk_format_description(format
);
773 int channel
= vk_format_get_first_non_void_channel(format
);
775 return channel
>= 0 && desc
->channel
[channel
].pure_integer
&&
776 desc
->channel
[channel
].size
== 8;
779 unsigned radv_format_meta_fs_key(VkFormat format
)
781 unsigned col_format
= si_choose_spi_color_format(format
, false, false) - 1;
782 bool is_int8
= format_is_int8(format
);
784 return col_format
+ (is_int8
? 3 : 0);
788 radv_pipeline_compute_is_int8(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
790 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
791 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
792 unsigned is_int8
= 0;
794 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
795 struct radv_render_pass_attachment
*attachment
;
797 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
799 if (format_is_int8(attachment
->format
))
807 radv_pipeline_init_blend_state(struct radv_pipeline
*pipeline
,
808 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
809 const struct radv_graphics_pipeline_create_info
*extra
)
811 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
812 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
813 unsigned mode
= V_028808_CB_NORMAL
;
814 uint32_t blend_enable
= 0, blend_need_alpha
= 0;
815 bool blend_mrt0_is_dual_src
= false;
817 bool single_cb_enable
= false;
822 if (extra
&& extra
->custom_blend_mode
) {
823 single_cb_enable
= true;
824 mode
= extra
->custom_blend_mode
;
826 blend
->cb_color_control
= 0;
827 if (vkblend
->logicOpEnable
)
828 blend
->cb_color_control
|= S_028808_ROP3(vkblend
->logicOp
| (vkblend
->logicOp
<< 4));
830 blend
->cb_color_control
|= S_028808_ROP3(0xcc);
832 blend
->db_alpha_to_mask
= S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
833 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
834 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
835 S_028B70_ALPHA_TO_MASK_OFFSET3(2);
837 blend
->cb_target_mask
= 0;
838 for (i
= 0; i
< vkblend
->attachmentCount
; i
++) {
839 const VkPipelineColorBlendAttachmentState
*att
= &vkblend
->pAttachments
[i
];
840 unsigned blend_cntl
= 0;
841 VkBlendOp eqRGB
= att
->colorBlendOp
;
842 VkBlendFactor srcRGB
= att
->srcColorBlendFactor
;
843 VkBlendFactor dstRGB
= att
->dstColorBlendFactor
;
844 VkBlendOp eqA
= att
->alphaBlendOp
;
845 VkBlendFactor srcA
= att
->srcAlphaBlendFactor
;
846 VkBlendFactor dstA
= att
->dstAlphaBlendFactor
;
848 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
);
850 if (!att
->colorWriteMask
)
853 blend
->cb_target_mask
|= (unsigned)att
->colorWriteMask
<< (4 * i
);
854 if (!att
->blendEnable
) {
855 blend
->cb_blend_control
[i
] = blend_cntl
;
859 if (is_dual_src(srcRGB
) || is_dual_src(dstRGB
) || is_dual_src(srcA
) || is_dual_src(dstA
))
861 blend_mrt0_is_dual_src
= true;
863 if (eqRGB
== VK_BLEND_OP_MIN
|| eqRGB
== VK_BLEND_OP_MAX
) {
864 srcRGB
= VK_BLEND_FACTOR_ONE
;
865 dstRGB
= VK_BLEND_FACTOR_ONE
;
867 if (eqA
== VK_BLEND_OP_MIN
|| eqA
== VK_BLEND_OP_MAX
) {
868 srcA
= VK_BLEND_FACTOR_ONE
;
869 dstA
= VK_BLEND_FACTOR_ONE
;
872 blend_cntl
|= S_028780_ENABLE(1);
874 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
875 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
876 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
877 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
878 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
879 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
880 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
881 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
883 blend
->cb_blend_control
[i
] = blend_cntl
;
885 blend_enable
|= 1 << i
;
887 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
888 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
889 srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
890 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
891 srcRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
||
892 dstRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
)
893 blend_need_alpha
|= 1 << i
;
895 for (i
= vkblend
->attachmentCount
; i
< 8; i
++)
896 blend
->cb_blend_control
[i
] = 0;
898 if (blend
->cb_target_mask
)
899 blend
->cb_color_control
|= S_028808_MODE(mode
);
901 blend
->cb_color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
903 radv_pipeline_compute_spi_color_formats(pipeline
, pCreateInfo
,
904 blend_enable
, blend_need_alpha
, single_cb_enable
, blend_mrt0_is_dual_src
);
907 static uint32_t si_translate_stencil_op(enum VkStencilOp op
)
910 case VK_STENCIL_OP_KEEP
:
911 return V_02842C_STENCIL_KEEP
;
912 case VK_STENCIL_OP_ZERO
:
913 return V_02842C_STENCIL_ZERO
;
914 case VK_STENCIL_OP_REPLACE
:
915 return V_02842C_STENCIL_REPLACE_TEST
;
916 case VK_STENCIL_OP_INCREMENT_AND_CLAMP
:
917 return V_02842C_STENCIL_ADD_CLAMP
;
918 case VK_STENCIL_OP_DECREMENT_AND_CLAMP
:
919 return V_02842C_STENCIL_SUB_CLAMP
;
920 case VK_STENCIL_OP_INVERT
:
921 return V_02842C_STENCIL_INVERT
;
922 case VK_STENCIL_OP_INCREMENT_AND_WRAP
:
923 return V_02842C_STENCIL_ADD_WRAP
;
924 case VK_STENCIL_OP_DECREMENT_AND_WRAP
:
925 return V_02842C_STENCIL_SUB_WRAP
;
931 radv_pipeline_init_depth_stencil_state(struct radv_pipeline
*pipeline
,
932 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
933 const struct radv_graphics_pipeline_create_info
*extra
)
935 const VkPipelineDepthStencilStateCreateInfo
*vkds
= pCreateInfo
->pDepthStencilState
;
936 struct radv_depth_stencil_state
*ds
= &pipeline
->graphics
.ds
;
938 memset(ds
, 0, sizeof(*ds
));
941 ds
->db_depth_control
= S_028800_Z_ENABLE(vkds
->depthTestEnable
? 1 : 0) |
942 S_028800_Z_WRITE_ENABLE(vkds
->depthWriteEnable
? 1 : 0) |
943 S_028800_ZFUNC(vkds
->depthCompareOp
) |
944 S_028800_DEPTH_BOUNDS_ENABLE(vkds
->depthBoundsTestEnable
? 1 : 0);
946 if (vkds
->stencilTestEnable
) {
947 ds
->db_depth_control
|= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
948 ds
->db_depth_control
|= S_028800_STENCILFUNC(vkds
->front
.compareOp
);
949 ds
->db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds
->front
.failOp
));
950 ds
->db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds
->front
.passOp
));
951 ds
->db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds
->front
.depthFailOp
));
953 ds
->db_depth_control
|= S_028800_STENCILFUNC_BF(vkds
->back
.compareOp
);
954 ds
->db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds
->back
.failOp
));
955 ds
->db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds
->back
.passOp
));
956 ds
->db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds
->back
.depthFailOp
));
961 ds
->db_render_control
|= S_028000_DEPTH_CLEAR_ENABLE(extra
->db_depth_clear
);
962 ds
->db_render_control
|= S_028000_STENCIL_CLEAR_ENABLE(extra
->db_stencil_clear
);
964 ds
->db_render_control
|= S_028000_RESUMMARIZE_ENABLE(extra
->db_resummarize
);
965 ds
->db_render_control
|= S_028000_DEPTH_COMPRESS_DISABLE(extra
->db_flush_depth_inplace
);
966 ds
->db_render_control
|= S_028000_STENCIL_COMPRESS_DISABLE(extra
->db_flush_stencil_inplace
);
967 ds
->db_render_override2
|= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra
->db_depth_disable_expclear
);
968 ds
->db_render_override2
|= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra
->db_stencil_disable_expclear
);
972 static uint32_t si_translate_fill(VkPolygonMode func
)
975 case VK_POLYGON_MODE_FILL
:
976 return V_028814_X_DRAW_TRIANGLES
;
977 case VK_POLYGON_MODE_LINE
:
978 return V_028814_X_DRAW_LINES
;
979 case VK_POLYGON_MODE_POINT
:
980 return V_028814_X_DRAW_POINTS
;
983 return V_028814_X_DRAW_POINTS
;
987 radv_pipeline_init_raster_state(struct radv_pipeline
*pipeline
,
988 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
990 const VkPipelineRasterizationStateCreateInfo
*vkraster
= pCreateInfo
->pRasterizationState
;
991 struct radv_raster_state
*raster
= &pipeline
->graphics
.raster
;
993 memset(raster
, 0, sizeof(*raster
));
995 raster
->spi_interp_control
=
996 S_0286D4_FLAT_SHADE_ENA(1) |
997 S_0286D4_PNT_SPRITE_ENA(1) |
998 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
999 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
1000 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
1001 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
1002 S_0286D4_PNT_SPRITE_TOP_1(0); // vulkan is top to bottom - 1.0 at bottom
1004 raster
->pa_cl_vs_out_cntl
= S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(1);
1005 raster
->pa_cl_clip_cntl
= S_028810_PS_UCP_MODE(3) |
1006 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
1007 S_028810_ZCLIP_NEAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
1008 S_028810_ZCLIP_FAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
1009 S_028810_DX_RASTERIZATION_KILL(vkraster
->rasterizerDiscardEnable
? 1 : 0) |
1010 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
1012 raster
->pa_su_vtx_cntl
=
1013 S_028BE4_PIX_CENTER(1) | // TODO verify
1014 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
1015 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
);
1017 raster
->pa_su_sc_mode_cntl
=
1018 S_028814_FACE(vkraster
->frontFace
) |
1019 S_028814_CULL_FRONT(!!(vkraster
->cullMode
& VK_CULL_MODE_FRONT_BIT
)) |
1020 S_028814_CULL_BACK(!!(vkraster
->cullMode
& VK_CULL_MODE_BACK_BIT
)) |
1021 S_028814_POLY_MODE(vkraster
->polygonMode
!= VK_POLYGON_MODE_FILL
) |
1022 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
1023 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
1024 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
1025 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
1026 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster
->depthBiasEnable
? 1 : 0);
1031 radv_pipeline_init_multisample_state(struct radv_pipeline
*pipeline
,
1032 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1034 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
1035 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
1036 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
1037 unsigned num_tile_pipes
= pipeline
->device
->instance
->physicalDevice
.rad_info
.num_tile_pipes
;
1038 int ps_iter_samples
= 1;
1039 uint32_t mask
= 0xffff;
1041 ms
->num_samples
= vkms
->rasterizationSamples
;
1043 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.fs
.force_persample
) {
1044 ps_iter_samples
= vkms
->rasterizationSamples
;
1047 ms
->pa_sc_line_cntl
= S_028BDC_DX10_DIAMOND_TEST_ENA(1);
1048 ms
->pa_sc_aa_config
= 0;
1049 ms
->db_eqaa
= S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
1050 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
1051 ms
->pa_sc_mode_cntl_1
=
1052 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
1053 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
1055 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
1056 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
1057 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
1058 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
1059 EG_S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
1060 EG_S_028A4C_FORCE_EOV_REZ_ENABLE(1);
1062 if (vkms
->rasterizationSamples
> 1) {
1063 unsigned log_samples
= util_logbase2(vkms
->rasterizationSamples
);
1064 unsigned log_ps_iter_samples
= util_logbase2(util_next_power_of_two(ps_iter_samples
));
1065 ms
->pa_sc_mode_cntl_0
= S_028A48_MSAA_ENABLE(1);
1066 ms
->pa_sc_line_cntl
|= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
1067 ms
->db_eqaa
|= S_028804_MAX_ANCHOR_SAMPLES(log_samples
) |
1068 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples
) |
1069 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples
) |
1070 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples
);
1071 ms
->pa_sc_aa_config
|= S_028BE0_MSAA_NUM_SAMPLES(log_samples
) |
1072 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples
)) |
1073 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples
); /* CM_R_028BE0_PA_SC_AA_CONFIG */
1074 ms
->pa_sc_mode_cntl_1
|= EG_S_028A4C_PS_ITER_SAMPLE(ps_iter_samples
> 1);
1077 if (vkms
->alphaToCoverageEnable
)
1078 blend
->db_alpha_to_mask
|= S_028B70_ALPHA_TO_MASK_ENABLE(1);
1080 if (vkms
->pSampleMask
) {
1081 mask
= vkms
->pSampleMask
[0] & 0xffff;
1084 ms
->pa_sc_aa_mask
[0] = mask
| (mask
<< 16);
1085 ms
->pa_sc_aa_mask
[1] = mask
| (mask
<< 16);
1089 si_translate_prim(enum VkPrimitiveTopology topology
)
1092 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1093 return V_008958_DI_PT_POINTLIST
;
1094 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1095 return V_008958_DI_PT_LINELIST
;
1096 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1097 return V_008958_DI_PT_LINESTRIP
;
1098 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1099 return V_008958_DI_PT_TRILIST
;
1100 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1101 return V_008958_DI_PT_TRISTRIP
;
1102 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1103 return V_008958_DI_PT_TRIFAN
;
1104 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1105 return V_008958_DI_PT_LINELIST_ADJ
;
1106 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1107 return V_008958_DI_PT_LINESTRIP_ADJ
;
1108 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1109 return V_008958_DI_PT_TRILIST_ADJ
;
1110 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1111 return V_008958_DI_PT_TRISTRIP_ADJ
;
1112 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1113 return V_008958_DI_PT_PATCH
;
1121 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology
)
1124 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1125 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1126 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
1127 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1128 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1129 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1130 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1131 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
1132 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1133 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1134 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1135 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1136 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1137 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1144 static unsigned si_map_swizzle(unsigned swizzle
)
1148 return V_008F0C_SQ_SEL_Y
;
1150 return V_008F0C_SQ_SEL_Z
;
1152 return V_008F0C_SQ_SEL_W
;
1154 return V_008F0C_SQ_SEL_0
;
1156 return V_008F0C_SQ_SEL_1
;
1157 default: /* VK_SWIZZLE_X */
1158 return V_008F0C_SQ_SEL_X
;
1163 radv_pipeline_init_dynamic_state(struct radv_pipeline
*pipeline
,
1164 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1166 radv_cmd_dirty_mask_t states
= RADV_CMD_DIRTY_DYNAMIC_ALL
;
1167 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
1168 struct radv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
1170 pipeline
->dynamic_state
= default_dynamic_state
;
1172 if (pCreateInfo
->pDynamicState
) {
1173 /* Remove all of the states that are marked as dynamic */
1174 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1175 for (uint32_t s
= 0; s
< count
; s
++)
1176 states
&= ~(1 << pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1179 struct radv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1181 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1183 * pViewportState is [...] NULL if the pipeline
1184 * has rasterization disabled.
1186 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1187 assert(pCreateInfo
->pViewportState
);
1189 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1190 if (states
& (1 << VK_DYNAMIC_STATE_VIEWPORT
)) {
1191 typed_memcpy(dynamic
->viewport
.viewports
,
1192 pCreateInfo
->pViewportState
->pViewports
,
1193 pCreateInfo
->pViewportState
->viewportCount
);
1196 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1197 if (states
& (1 << VK_DYNAMIC_STATE_SCISSOR
)) {
1198 typed_memcpy(dynamic
->scissor
.scissors
,
1199 pCreateInfo
->pViewportState
->pScissors
,
1200 pCreateInfo
->pViewportState
->scissorCount
);
1204 if (states
& (1 << VK_DYNAMIC_STATE_LINE_WIDTH
)) {
1205 assert(pCreateInfo
->pRasterizationState
);
1206 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1209 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BIAS
)) {
1210 assert(pCreateInfo
->pRasterizationState
);
1211 dynamic
->depth_bias
.bias
=
1212 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1213 dynamic
->depth_bias
.clamp
=
1214 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1215 dynamic
->depth_bias
.slope
=
1216 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1219 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1221 * pColorBlendState is [...] NULL if the pipeline has rasterization
1222 * disabled or if the subpass of the render pass the pipeline is
1223 * created against does not use any color attachments.
1225 bool uses_color_att
= false;
1226 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1227 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1228 uses_color_att
= true;
1233 if (uses_color_att
&& states
& (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS
)) {
1234 assert(pCreateInfo
->pColorBlendState
);
1235 typed_memcpy(dynamic
->blend_constants
,
1236 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1239 /* If there is no depthstencil attachment, then don't read
1240 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1241 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1242 * no need to override the depthstencil defaults in
1243 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1245 * Section 9.2 of the Vulkan 1.0.15 spec says:
1247 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1248 * disabled or if the subpass of the render pass the pipeline is created
1249 * against does not use a depth/stencil attachment.
1251 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1252 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1253 assert(pCreateInfo
->pDepthStencilState
);
1255 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS
)) {
1256 dynamic
->depth_bounds
.min
=
1257 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1258 dynamic
->depth_bounds
.max
=
1259 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1262 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
)) {
1263 dynamic
->stencil_compare_mask
.front
=
1264 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1265 dynamic
->stencil_compare_mask
.back
=
1266 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1269 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
)) {
1270 dynamic
->stencil_write_mask
.front
=
1271 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1272 dynamic
->stencil_write_mask
.back
=
1273 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1276 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE
)) {
1277 dynamic
->stencil_reference
.front
=
1278 pCreateInfo
->pDepthStencilState
->front
.reference
;
1279 dynamic
->stencil_reference
.back
=
1280 pCreateInfo
->pDepthStencilState
->back
.reference
;
1284 pipeline
->dynamic_state_mask
= states
;
1287 static union ac_shader_variant_key
1288 radv_compute_vs_key(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1290 union ac_shader_variant_key key
;
1291 const VkPipelineVertexInputStateCreateInfo
*input_state
=
1292 pCreateInfo
->pVertexInputState
;
1294 memset(&key
, 0, sizeof(key
));
1295 key
.vs
.instance_rate_inputs
= 0;
1297 for (unsigned i
= 0; i
< input_state
->vertexAttributeDescriptionCount
; ++i
) {
1299 binding
= input_state
->pVertexAttributeDescriptions
[i
].binding
;
1300 if (input_state
->pVertexBindingDescriptions
[binding
].inputRate
)
1301 key
.vs
.instance_rate_inputs
|= 1u << input_state
->pVertexAttributeDescriptions
[i
].location
;
1307 radv_pipeline_init(struct radv_pipeline
*pipeline
,
1308 struct radv_device
*device
,
1309 struct radv_pipeline_cache
*cache
,
1310 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1311 const struct radv_graphics_pipeline_create_info
*extra
,
1312 const VkAllocationCallbacks
*alloc
)
1314 struct radv_shader_module fs_m
= {0};
1317 alloc
= &device
->alloc
;
1319 pipeline
->device
= device
;
1320 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
1322 radv_pipeline_init_dynamic_state(pipeline
, pCreateInfo
);
1323 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
1324 struct radv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1325 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
1326 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
1327 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
1328 modules
[stage
] = radv_shader_module_from_handle(pStages
[stage
]->module
);
1331 radv_pipeline_init_blend_state(pipeline
, pCreateInfo
, extra
);
1334 if (modules
[MESA_SHADER_VERTEX
]) {
1335 union ac_shader_variant_key key
= radv_compute_vs_key(pCreateInfo
);
1337 pipeline
->shaders
[MESA_SHADER_VERTEX
] =
1338 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_VERTEX
],
1339 pStages
[MESA_SHADER_VERTEX
]->pName
,
1341 pStages
[MESA_SHADER_VERTEX
]->pSpecializationInfo
,
1342 pipeline
->layout
, &key
);
1344 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_VERTEX
);
1347 if (!modules
[MESA_SHADER_FRAGMENT
]) {
1349 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
1350 fs_b
.shader
->info
->name
= ralloc_strdup(fs_b
.shader
, "noop_fs");
1351 fs_m
.nir
= fs_b
.shader
;
1352 modules
[MESA_SHADER_FRAGMENT
] = &fs_m
;
1355 if (modules
[MESA_SHADER_FRAGMENT
]) {
1356 union ac_shader_variant_key key
;
1357 key
.fs
.col_format
= pipeline
->graphics
.blend
.spi_shader_col_format
;
1358 key
.fs
.is_int8
= radv_pipeline_compute_is_int8(pCreateInfo
);
1360 const VkPipelineShaderStageCreateInfo
*stage
= pStages
[MESA_SHADER_FRAGMENT
];
1362 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] =
1363 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_FRAGMENT
],
1364 stage
? stage
->pName
: "main",
1365 MESA_SHADER_FRAGMENT
,
1366 stage
? stage
->pSpecializationInfo
: NULL
,
1367 pipeline
->layout
, &key
);
1368 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_FRAGMENT
);
1372 ralloc_free(fs_m
.nir
);
1374 radv_pipeline_init_depth_stencil_state(pipeline
, pCreateInfo
, extra
);
1375 radv_pipeline_init_raster_state(pipeline
, pCreateInfo
);
1376 radv_pipeline_init_multisample_state(pipeline
, pCreateInfo
);
1377 pipeline
->graphics
.prim
= si_translate_prim(pCreateInfo
->pInputAssemblyState
->topology
);
1378 pipeline
->graphics
.gs_out
= si_conv_prim_to_gs_out(pCreateInfo
->pInputAssemblyState
->topology
);
1379 if (extra
&& extra
->use_rectlist
) {
1380 pipeline
->graphics
.prim
= V_008958_DI_PT_RECTLIST
;
1381 pipeline
->graphics
.gs_out
= V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1383 pipeline
->graphics
.prim_restart_enable
= !!pCreateInfo
->pInputAssemblyState
->primitiveRestartEnable
;
1385 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1386 pCreateInfo
->pVertexInputState
;
1387 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1388 const VkVertexInputAttributeDescription
*desc
=
1389 &vi_info
->pVertexAttributeDescriptions
[i
];
1390 unsigned loc
= desc
->location
;
1391 const struct vk_format_description
*format_desc
;
1393 uint32_t num_format
, data_format
;
1394 format_desc
= vk_format_description(desc
->format
);
1395 first_non_void
= vk_format_get_first_non_void_channel(desc
->format
);
1397 num_format
= radv_translate_buffer_numformat(format_desc
, first_non_void
);
1398 data_format
= radv_translate_buffer_dataformat(format_desc
, first_non_void
);
1400 pipeline
->va_rsrc_word3
[loc
] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc
->swizzle
[0])) |
1401 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc
->swizzle
[1])) |
1402 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc
->swizzle
[2])) |
1403 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc
->swizzle
[3])) |
1404 S_008F0C_NUM_FORMAT(num_format
) |
1405 S_008F0C_DATA_FORMAT(data_format
);
1406 pipeline
->va_format_size
[loc
] = format_desc
->block
.bits
/ 8;
1407 pipeline
->va_offset
[loc
] = desc
->offset
;
1408 pipeline
->va_binding
[loc
] = desc
->binding
;
1409 pipeline
->num_vertex_attribs
= MAX2(pipeline
->num_vertex_attribs
, loc
+ 1);
1412 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1413 const VkVertexInputBindingDescription
*desc
=
1414 &vi_info
->pVertexBindingDescriptions
[i
];
1416 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
1419 if (device
->debug_flags
& RADV_DEBUG_DUMP_SHADER_STATS
) {
1420 radv_dump_pipeline_stats(device
, pipeline
);
1427 radv_graphics_pipeline_create(
1429 VkPipelineCache _cache
,
1430 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1431 const struct radv_graphics_pipeline_create_info
*extra
,
1432 const VkAllocationCallbacks
*pAllocator
,
1433 VkPipeline
*pPipeline
)
1435 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1436 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
1437 struct radv_pipeline
*pipeline
;
1440 pipeline
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
1441 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1442 if (pipeline
== NULL
)
1443 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1445 memset(pipeline
, 0, sizeof(*pipeline
));
1446 result
= radv_pipeline_init(pipeline
, device
, cache
,
1447 pCreateInfo
, extra
, pAllocator
);
1448 if (result
!= VK_SUCCESS
) {
1449 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
1453 *pPipeline
= radv_pipeline_to_handle(pipeline
);
1458 VkResult
radv_CreateGraphicsPipelines(
1460 VkPipelineCache pipelineCache
,
1462 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
1463 const VkAllocationCallbacks
* pAllocator
,
1464 VkPipeline
* pPipelines
)
1466 VkResult result
= VK_SUCCESS
;
1469 for (; i
< count
; i
++) {
1471 r
= radv_graphics_pipeline_create(_device
,
1474 NULL
, pAllocator
, &pPipelines
[i
]);
1475 if (r
!= VK_SUCCESS
) {
1477 pPipelines
[i
] = VK_NULL_HANDLE
;
1484 static VkResult
radv_compute_pipeline_create(
1486 VkPipelineCache _cache
,
1487 const VkComputePipelineCreateInfo
* pCreateInfo
,
1488 const VkAllocationCallbacks
* pAllocator
,
1489 VkPipeline
* pPipeline
)
1491 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1492 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
1493 RADV_FROM_HANDLE(radv_shader_module
, module
, pCreateInfo
->stage
.module
);
1494 struct radv_pipeline
*pipeline
;
1496 pipeline
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
1497 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1498 if (pipeline
== NULL
)
1499 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1501 memset(pipeline
, 0, sizeof(*pipeline
));
1502 pipeline
->device
= device
;
1503 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
1505 pipeline
->shaders
[MESA_SHADER_COMPUTE
] =
1506 radv_pipeline_compile(pipeline
, cache
, module
,
1507 pCreateInfo
->stage
.pName
,
1508 MESA_SHADER_COMPUTE
,
1509 pCreateInfo
->stage
.pSpecializationInfo
,
1510 pipeline
->layout
, NULL
);
1512 *pPipeline
= radv_pipeline_to_handle(pipeline
);
1514 if (device
->debug_flags
& RADV_DEBUG_DUMP_SHADER_STATS
) {
1515 radv_dump_pipeline_stats(device
, pipeline
);
1519 VkResult
radv_CreateComputePipelines(
1521 VkPipelineCache pipelineCache
,
1523 const VkComputePipelineCreateInfo
* pCreateInfos
,
1524 const VkAllocationCallbacks
* pAllocator
,
1525 VkPipeline
* pPipelines
)
1527 VkResult result
= VK_SUCCESS
;
1530 for (; i
< count
; i
++) {
1532 r
= radv_compute_pipeline_create(_device
, pipelineCache
,
1534 pAllocator
, &pPipelines
[i
]);
1535 if (r
!= VK_SUCCESS
) {
1537 pPipelines
[i
] = VK_NULL_HANDLE
;