2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30 #include "util/mesa-sha1.h"
31 #include "util/os_time.h"
32 #include "common/gen_l3_config.h"
33 #include "common/gen_disasm.h"
34 #include "anv_private.h"
35 #include "compiler/brw_nir.h"
37 #include "nir/nir_xfb_info.h"
38 #include "spirv/nir_spirv.h"
41 /* Needed for SWIZZLE macros */
42 #include "program/prog_instruction.h"
46 VkResult
anv_CreateShaderModule(
48 const VkShaderModuleCreateInfo
* pCreateInfo
,
49 const VkAllocationCallbacks
* pAllocator
,
50 VkShaderModule
* pShaderModule
)
52 ANV_FROM_HANDLE(anv_device
, device
, _device
);
53 struct anv_shader_module
*module
;
55 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
56 assert(pCreateInfo
->flags
== 0);
58 module
= vk_alloc2(&device
->alloc
, pAllocator
,
59 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
60 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
62 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
64 module
->size
= pCreateInfo
->codeSize
;
65 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
67 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
69 *pShaderModule
= anv_shader_module_to_handle(module
);
74 void anv_DestroyShaderModule(
76 VkShaderModule _module
,
77 const VkAllocationCallbacks
* pAllocator
)
79 ANV_FROM_HANDLE(anv_device
, device
, _device
);
80 ANV_FROM_HANDLE(anv_shader_module
, module
, _module
);
85 vk_free2(&device
->alloc
, pAllocator
, module
);
88 #define SPIR_V_MAGIC_NUMBER 0x07230203
90 static const uint64_t stage_to_debug
[] = {
91 [MESA_SHADER_VERTEX
] = DEBUG_VS
,
92 [MESA_SHADER_TESS_CTRL
] = DEBUG_TCS
,
93 [MESA_SHADER_TESS_EVAL
] = DEBUG_TES
,
94 [MESA_SHADER_GEOMETRY
] = DEBUG_GS
,
95 [MESA_SHADER_FRAGMENT
] = DEBUG_WM
,
96 [MESA_SHADER_COMPUTE
] = DEBUG_CS
,
99 struct anv_spirv_debug_data
{
100 struct anv_device
*device
;
101 const struct anv_shader_module
*module
;
104 static void anv_spirv_nir_debug(void *private_data
,
105 enum nir_spirv_debug_level level
,
109 struct anv_spirv_debug_data
*debug_data
= private_data
;
110 struct anv_instance
*instance
= debug_data
->device
->physical
->instance
;
112 static const VkDebugReportFlagsEXT vk_flags
[] = {
113 [NIR_SPIRV_DEBUG_LEVEL_INFO
] = VK_DEBUG_REPORT_INFORMATION_BIT_EXT
,
114 [NIR_SPIRV_DEBUG_LEVEL_WARNING
] = VK_DEBUG_REPORT_WARNING_BIT_EXT
,
115 [NIR_SPIRV_DEBUG_LEVEL_ERROR
] = VK_DEBUG_REPORT_ERROR_BIT_EXT
,
119 snprintf(buffer
, sizeof(buffer
), "SPIR-V offset %lu: %s", (unsigned long) spirv_offset
, message
);
121 vk_debug_report(&instance
->debug_report_callbacks
,
123 VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT
,
124 (uint64_t) (uintptr_t) debug_data
->module
,
125 0, 0, "anv", buffer
);
128 /* Eventually, this will become part of anv_CreateShader. Unfortunately,
129 * we can't do that yet because we don't have the ability to copy nir.
132 anv_shader_compile_to_nir(struct anv_device
*device
,
134 const struct anv_shader_module
*module
,
135 const char *entrypoint_name
,
136 gl_shader_stage stage
,
137 const VkSpecializationInfo
*spec_info
)
139 const struct anv_physical_device
*pdevice
= device
->physical
;
140 const struct brw_compiler
*compiler
= pdevice
->compiler
;
141 const nir_shader_compiler_options
*nir_options
=
142 compiler
->glsl_compiler_options
[stage
].NirOptions
;
144 uint32_t *spirv
= (uint32_t *) module
->data
;
145 assert(spirv
[0] == SPIR_V_MAGIC_NUMBER
);
146 assert(module
->size
% 4 == 0);
148 uint32_t num_spec_entries
= 0;
149 struct nir_spirv_specialization
*spec_entries
= NULL
;
150 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
151 num_spec_entries
= spec_info
->mapEntryCount
;
152 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
153 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
154 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
155 const void *data
= spec_info
->pData
+ entry
.offset
;
156 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
158 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
159 if (spec_info
->dataSize
== 8)
160 spec_entries
[i
].data64
= *(const uint64_t *)data
;
162 spec_entries
[i
].data32
= *(const uint32_t *)data
;
166 struct anv_spirv_debug_data spirv_debug_data
= {
170 struct spirv_to_nir_options spirv_options
= {
171 .frag_coord_is_sysval
= true,
172 .use_scoped_memory_barrier
= true,
174 .demote_to_helper_invocation
= true,
175 .derivative_group
= true,
176 .descriptor_array_dynamic_indexing
= true,
177 .descriptor_array_non_uniform_indexing
= true,
178 .descriptor_indexing
= true,
179 .device_group
= true,
180 .draw_parameters
= true,
181 .float16
= pdevice
->info
.gen
>= 8,
182 .float64
= pdevice
->info
.gen
>= 8,
183 .fragment_shader_sample_interlock
= pdevice
->info
.gen
>= 9,
184 .fragment_shader_pixel_interlock
= pdevice
->info
.gen
>= 9,
185 .geometry_streams
= true,
186 .image_write_without_format
= true,
187 .int8
= pdevice
->info
.gen
>= 8,
188 .int16
= pdevice
->info
.gen
>= 8,
189 .int64
= pdevice
->info
.gen
>= 8,
190 .int64_atomics
= pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
,
193 .physical_storage_buffer_address
= pdevice
->has_a64_buffer_access
,
194 .post_depth_coverage
= pdevice
->info
.gen
>= 9,
195 .runtime_descriptor_array
= true,
196 .float_controls
= pdevice
->info
.gen
>= 8,
197 .shader_clock
= true,
198 .shader_viewport_index_layer
= true,
199 .stencil_export
= pdevice
->info
.gen
>= 9,
200 .storage_8bit
= pdevice
->info
.gen
>= 8,
201 .storage_16bit
= pdevice
->info
.gen
>= 8,
202 .subgroup_arithmetic
= true,
203 .subgroup_basic
= true,
204 .subgroup_ballot
= true,
205 .subgroup_quad
= true,
206 .subgroup_shuffle
= true,
207 .subgroup_vote
= true,
208 .tessellation
= true,
209 .transform_feedback
= pdevice
->info
.gen
>= 8,
210 .variable_pointers
= true,
211 .vk_memory_model
= true,
212 .vk_memory_model_device_scope
= true,
214 .ubo_addr_format
= nir_address_format_32bit_index_offset
,
216 anv_nir_ssbo_addr_format(pdevice
, device
->robust_buffer_access
),
217 .phys_ssbo_addr_format
= nir_address_format_64bit_global
,
218 .push_const_addr_format
= nir_address_format_logical
,
220 /* TODO: Consider changing this to an address format that has the NULL
221 * pointer equals to 0. That might be a better format to play nice
222 * with certain code / code generators.
224 .shared_addr_format
= nir_address_format_32bit_offset
,
226 .func
= anv_spirv_nir_debug
,
227 .private_data
= &spirv_debug_data
,
233 spirv_to_nir(spirv
, module
->size
/ 4,
234 spec_entries
, num_spec_entries
,
235 stage
, entrypoint_name
, &spirv_options
, nir_options
);
236 assert(nir
->info
.stage
== stage
);
237 nir_validate_shader(nir
, "after spirv_to_nir");
238 ralloc_steal(mem_ctx
, nir
);
242 if (unlikely(INTEL_DEBUG
& stage_to_debug
[stage
])) {
243 fprintf(stderr
, "NIR (from SPIR-V) for %s shader:\n",
244 gl_shader_stage_name(stage
));
245 nir_print_shader(nir
, stderr
);
248 /* We have to lower away local constant initializers right before we
249 * inline functions. That way they get properly initialized at the top
250 * of the function and not at the top of its caller.
252 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
253 NIR_PASS_V(nir
, nir_lower_returns
);
254 NIR_PASS_V(nir
, nir_inline_functions
);
255 NIR_PASS_V(nir
, nir_opt_deref
);
257 /* Pick off the single entrypoint that we want */
258 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
259 if (!func
->is_entrypoint
)
260 exec_node_remove(&func
->node
);
262 assert(exec_list_length(&nir
->functions
) == 1);
264 /* Now that we've deleted all but the main function, we can go ahead and
265 * lower the rest of the constant initializers. We do this here so that
266 * nir_remove_dead_variables and split_per_member_structs below see the
267 * corresponding stores.
269 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
271 /* Split member structs. We do this before lower_io_to_temporaries so that
272 * it doesn't lower system values to temporaries by accident.
274 NIR_PASS_V(nir
, nir_split_var_copies
);
275 NIR_PASS_V(nir
, nir_split_per_member_structs
);
277 NIR_PASS_V(nir
, nir_remove_dead_variables
,
278 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
280 NIR_PASS_V(nir
, nir_propagate_invariant
);
281 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
282 nir_shader_get_entrypoint(nir
), true, false);
284 NIR_PASS_V(nir
, nir_lower_frexp
);
286 /* Vulkan uses the separate-shader linking model */
287 nir
->info
.separate_shader
= true;
289 brw_preprocess_nir(compiler
, nir
, NULL
);
294 void anv_DestroyPipeline(
296 VkPipeline _pipeline
,
297 const VkAllocationCallbacks
* pAllocator
)
299 ANV_FROM_HANDLE(anv_device
, device
, _device
);
300 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, _pipeline
);
305 anv_reloc_list_finish(&pipeline
->batch_relocs
,
306 pAllocator
? pAllocator
: &device
->alloc
);
308 ralloc_free(pipeline
->mem_ctx
);
310 if (pipeline
->blend_state
.map
)
311 anv_state_pool_free(&device
->dynamic_state_pool
, pipeline
->blend_state
);
313 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
314 if (pipeline
->shaders
[s
])
315 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
318 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
321 static const uint32_t vk_to_gen_primitive_type
[] = {
322 [VK_PRIMITIVE_TOPOLOGY_POINT_LIST
] = _3DPRIM_POINTLIST
,
323 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST
] = _3DPRIM_LINELIST
,
324 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
325 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
] = _3DPRIM_TRILIST
,
326 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
327 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
328 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
329 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
330 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
331 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
335 populate_sampler_prog_key(const struct gen_device_info
*devinfo
,
336 struct brw_sampler_prog_key_data
*key
)
338 /* Almost all multisampled textures are compressed. The only time when we
339 * don't compress a multisampled texture is for 16x MSAA with a surface
340 * width greater than 8k which is a bit of an edge case. Since the sampler
341 * just ignores the MCS parameter to ld2ms when MCS is disabled, it's safe
342 * to tell the compiler to always assume compression.
344 key
->compressed_multisample_layout_mask
= ~0;
346 /* SkyLake added support for 16x MSAA. With this came a new message for
347 * reading from a 16x MSAA surface with compression. The new message was
348 * needed because now the MCS data is 64 bits instead of 32 or lower as is
349 * the case for 8x, 4x, and 2x. The key->msaa_16 bit-field controls which
350 * message we use. Fortunately, the 16x message works for 8x, 4x, and 2x
351 * so we can just use it unconditionally. This may not be quite as
352 * efficient but it saves us from recompiling.
354 if (devinfo
->gen
>= 9)
357 /* XXX: Handle texture swizzle on HSW- */
358 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
359 /* Assume color sampler, no swizzling. (Works for BDW+) */
360 key
->swizzles
[i
] = SWIZZLE_XYZW
;
365 populate_base_prog_key(const struct gen_device_info
*devinfo
,
366 VkPipelineShaderStageCreateFlags flags
,
367 struct brw_base_prog_key
*key
)
369 if (flags
& VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT_EXT
)
370 key
->subgroup_size_type
= BRW_SUBGROUP_SIZE_VARYING
;
372 key
->subgroup_size_type
= BRW_SUBGROUP_SIZE_API_CONSTANT
;
374 populate_sampler_prog_key(devinfo
, &key
->tex
);
378 populate_vs_prog_key(const struct gen_device_info
*devinfo
,
379 VkPipelineShaderStageCreateFlags flags
,
380 struct brw_vs_prog_key
*key
)
382 memset(key
, 0, sizeof(*key
));
384 populate_base_prog_key(devinfo
, flags
, &key
->base
);
386 /* XXX: Handle vertex input work-arounds */
388 /* XXX: Handle sampler_prog_key */
392 populate_tcs_prog_key(const struct gen_device_info
*devinfo
,
393 VkPipelineShaderStageCreateFlags flags
,
394 unsigned input_vertices
,
395 struct brw_tcs_prog_key
*key
)
397 memset(key
, 0, sizeof(*key
));
399 populate_base_prog_key(devinfo
, flags
, &key
->base
);
401 key
->input_vertices
= input_vertices
;
405 populate_tes_prog_key(const struct gen_device_info
*devinfo
,
406 VkPipelineShaderStageCreateFlags flags
,
407 struct brw_tes_prog_key
*key
)
409 memset(key
, 0, sizeof(*key
));
411 populate_base_prog_key(devinfo
, flags
, &key
->base
);
415 populate_gs_prog_key(const struct gen_device_info
*devinfo
,
416 VkPipelineShaderStageCreateFlags flags
,
417 struct brw_gs_prog_key
*key
)
419 memset(key
, 0, sizeof(*key
));
421 populate_base_prog_key(devinfo
, flags
, &key
->base
);
425 populate_wm_prog_key(const struct gen_device_info
*devinfo
,
426 VkPipelineShaderStageCreateFlags flags
,
427 const struct anv_subpass
*subpass
,
428 const VkPipelineMultisampleStateCreateInfo
*ms_info
,
429 struct brw_wm_prog_key
*key
)
431 memset(key
, 0, sizeof(*key
));
433 populate_base_prog_key(devinfo
, flags
, &key
->base
);
435 /* We set this to 0 here and set to the actual value before we call
438 key
->input_slots_valid
= 0;
440 /* Vulkan doesn't specify a default */
441 key
->high_quality_derivatives
= false;
443 /* XXX Vulkan doesn't appear to specify */
444 key
->clamp_fragment_color
= false;
446 assert(subpass
->color_count
<= MAX_RTS
);
447 for (uint32_t i
= 0; i
< subpass
->color_count
; i
++) {
448 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
449 key
->color_outputs_valid
|= (1 << i
);
452 key
->nr_color_regions
= subpass
->color_count
;
454 /* To reduce possible shader recompilations we would need to know if
455 * there is a SampleMask output variable to compute if we should emit
456 * code to workaround the issue that hardware disables alpha to coverage
457 * when there is SampleMask output.
459 key
->alpha_to_coverage
= ms_info
&& ms_info
->alphaToCoverageEnable
;
461 /* Vulkan doesn't support fixed-function alpha test */
462 key
->alpha_test_replicate_alpha
= false;
465 /* We should probably pull this out of the shader, but it's fairly
466 * harmless to compute it and then let dead-code take care of it.
468 if (ms_info
->rasterizationSamples
> 1) {
469 key
->persample_interp
= ms_info
->sampleShadingEnable
&&
470 (ms_info
->minSampleShading
* ms_info
->rasterizationSamples
) > 1;
471 key
->multisample_fbo
= true;
474 key
->frag_coord_adds_sample_pos
= key
->persample_interp
;
479 populate_cs_prog_key(const struct gen_device_info
*devinfo
,
480 VkPipelineShaderStageCreateFlags flags
,
481 const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT
*rss_info
,
482 struct brw_cs_prog_key
*key
)
484 memset(key
, 0, sizeof(*key
));
486 populate_base_prog_key(devinfo
, flags
, &key
->base
);
489 assert(key
->base
.subgroup_size_type
!= BRW_SUBGROUP_SIZE_VARYING
);
491 /* These enum values are expressly chosen to be equal to the subgroup
492 * size that they require.
494 assert(rss_info
->requiredSubgroupSize
== 8 ||
495 rss_info
->requiredSubgroupSize
== 16 ||
496 rss_info
->requiredSubgroupSize
== 32);
497 key
->base
.subgroup_size_type
= rss_info
->requiredSubgroupSize
;
498 } else if (flags
& VK_PIPELINE_SHADER_STAGE_CREATE_REQUIRE_FULL_SUBGROUPS_BIT_EXT
) {
499 /* If the client expressly requests full subgroups and they don't
500 * specify a subgroup size, we need to pick one. If they're requested
501 * varying subgroup sizes, we set it to UNIFORM and let the back-end
502 * compiler pick. Otherwise, we specify the API value of 32.
503 * Performance will likely be terrible in this case but there's nothing
504 * we can do about that. The client should have chosen a size.
506 if (flags
& VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT_EXT
)
507 key
->base
.subgroup_size_type
= BRW_SUBGROUP_SIZE_UNIFORM
;
509 key
->base
.subgroup_size_type
= BRW_SUBGROUP_SIZE_REQUIRE_32
;
513 struct anv_pipeline_stage
{
514 gl_shader_stage stage
;
516 const struct anv_shader_module
*module
;
517 const char *entrypoint
;
518 const VkSpecializationInfo
*spec_info
;
520 unsigned char shader_sha1
[20];
522 union brw_any_prog_key key
;
525 gl_shader_stage stage
;
526 unsigned char sha1
[20];
531 struct anv_pipeline_binding surface_to_descriptor
[256];
532 struct anv_pipeline_binding sampler_to_descriptor
[256];
533 struct anv_pipeline_bind_map bind_map
;
535 union brw_any_prog_data prog_data
;
538 struct brw_compile_stats stats
[3];
541 VkPipelineCreationFeedbackEXT feedback
;
543 const unsigned *code
;
547 anv_pipeline_hash_shader(const struct anv_shader_module
*module
,
548 const char *entrypoint
,
549 gl_shader_stage stage
,
550 const VkSpecializationInfo
*spec_info
,
551 unsigned char *sha1_out
)
553 struct mesa_sha1 ctx
;
554 _mesa_sha1_init(&ctx
);
556 _mesa_sha1_update(&ctx
, module
->sha1
, sizeof(module
->sha1
));
557 _mesa_sha1_update(&ctx
, entrypoint
, strlen(entrypoint
));
558 _mesa_sha1_update(&ctx
, &stage
, sizeof(stage
));
560 _mesa_sha1_update(&ctx
, spec_info
->pMapEntries
,
561 spec_info
->mapEntryCount
*
562 sizeof(*spec_info
->pMapEntries
));
563 _mesa_sha1_update(&ctx
, spec_info
->pData
,
564 spec_info
->dataSize
);
567 _mesa_sha1_final(&ctx
, sha1_out
);
571 anv_pipeline_hash_graphics(struct anv_pipeline
*pipeline
,
572 struct anv_pipeline_layout
*layout
,
573 struct anv_pipeline_stage
*stages
,
574 unsigned char *sha1_out
)
576 struct mesa_sha1 ctx
;
577 _mesa_sha1_init(&ctx
);
579 _mesa_sha1_update(&ctx
, &pipeline
->subpass
->view_mask
,
580 sizeof(pipeline
->subpass
->view_mask
));
583 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
585 const bool rba
= pipeline
->device
->robust_buffer_access
;
586 _mesa_sha1_update(&ctx
, &rba
, sizeof(rba
));
588 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
589 if (stages
[s
].entrypoint
) {
590 _mesa_sha1_update(&ctx
, stages
[s
].shader_sha1
,
591 sizeof(stages
[s
].shader_sha1
));
592 _mesa_sha1_update(&ctx
, &stages
[s
].key
, brw_prog_key_size(s
));
596 _mesa_sha1_final(&ctx
, sha1_out
);
600 anv_pipeline_hash_compute(struct anv_pipeline
*pipeline
,
601 struct anv_pipeline_layout
*layout
,
602 struct anv_pipeline_stage
*stage
,
603 unsigned char *sha1_out
)
605 struct mesa_sha1 ctx
;
606 _mesa_sha1_init(&ctx
);
609 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
611 const bool rba
= pipeline
->device
->robust_buffer_access
;
612 _mesa_sha1_update(&ctx
, &rba
, sizeof(rba
));
614 _mesa_sha1_update(&ctx
, stage
->shader_sha1
,
615 sizeof(stage
->shader_sha1
));
616 _mesa_sha1_update(&ctx
, &stage
->key
.cs
, sizeof(stage
->key
.cs
));
618 _mesa_sha1_final(&ctx
, sha1_out
);
622 anv_pipeline_stage_get_nir(struct anv_pipeline
*pipeline
,
623 struct anv_pipeline_cache
*cache
,
625 struct anv_pipeline_stage
*stage
)
627 const struct brw_compiler
*compiler
=
628 pipeline
->device
->physical
->compiler
;
629 const nir_shader_compiler_options
*nir_options
=
630 compiler
->glsl_compiler_options
[stage
->stage
].NirOptions
;
633 nir
= anv_device_search_for_nir(pipeline
->device
, cache
,
638 assert(nir
->info
.stage
== stage
->stage
);
642 nir
= anv_shader_compile_to_nir(pipeline
->device
,
649 anv_device_upload_nir(pipeline
->device
, cache
, nir
, stage
->shader_sha1
);
657 anv_pipeline_lower_nir(struct anv_pipeline
*pipeline
,
659 struct anv_pipeline_stage
*stage
,
660 struct anv_pipeline_layout
*layout
)
662 const struct anv_physical_device
*pdevice
= pipeline
->device
->physical
;
663 const struct brw_compiler
*compiler
= pdevice
->compiler
;
665 struct brw_stage_prog_data
*prog_data
= &stage
->prog_data
.base
;
666 nir_shader
*nir
= stage
->nir
;
668 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
669 NIR_PASS_V(nir
, nir_lower_wpos_center
, pipeline
->sample_shading_enable
);
670 NIR_PASS_V(nir
, nir_lower_input_attachments
, true);
673 NIR_PASS_V(nir
, anv_nir_lower_ycbcr_textures
, layout
);
675 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
676 NIR_PASS_V(nir
, anv_nir_lower_multiview
, pipeline
->subpass
->view_mask
);
678 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
680 if (nir
->info
.num_ssbos
> 0 || nir
->info
.num_images
> 0)
681 pipeline
->needs_data_cache
= true;
683 NIR_PASS_V(nir
, brw_nir_lower_image_load_store
, compiler
->devinfo
);
685 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_global
,
686 nir_address_format_64bit_global
);
688 /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
689 anv_nir_apply_pipeline_layout(pdevice
,
690 pipeline
->device
->robust_buffer_access
,
691 layout
, nir
, &stage
->bind_map
);
693 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_ubo
,
694 nir_address_format_32bit_index_offset
);
695 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_ssbo
,
696 anv_nir_ssbo_addr_format(pdevice
,
697 pipeline
->device
->robust_buffer_access
));
699 NIR_PASS_V(nir
, nir_opt_constant_folding
);
701 /* We don't support non-uniform UBOs and non-uniform SSBO access is
702 * handled naturally by falling back to A64 messages.
704 NIR_PASS_V(nir
, nir_lower_non_uniform_access
,
705 nir_lower_non_uniform_texture_access
|
706 nir_lower_non_uniform_image_access
);
708 anv_nir_compute_push_layout(pdevice
, nir
, prog_data
,
709 &stage
->bind_map
, mem_ctx
);
715 anv_pipeline_link_vs(const struct brw_compiler
*compiler
,
716 struct anv_pipeline_stage
*vs_stage
,
717 struct anv_pipeline_stage
*next_stage
)
720 brw_nir_link_shaders(compiler
, vs_stage
->nir
, next_stage
->nir
);
724 anv_pipeline_compile_vs(const struct brw_compiler
*compiler
,
726 struct anv_device
*device
,
727 struct anv_pipeline_stage
*vs_stage
)
729 brw_compute_vue_map(compiler
->devinfo
,
730 &vs_stage
->prog_data
.vs
.base
.vue_map
,
731 vs_stage
->nir
->info
.outputs_written
,
732 vs_stage
->nir
->info
.separate_shader
);
734 vs_stage
->num_stats
= 1;
735 vs_stage
->code
= brw_compile_vs(compiler
, device
, mem_ctx
,
737 &vs_stage
->prog_data
.vs
,
739 vs_stage
->stats
, NULL
);
743 merge_tess_info(struct shader_info
*tes_info
,
744 const struct shader_info
*tcs_info
)
746 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
748 * "PointMode. Controls generation of points rather than triangles
749 * or lines. This functionality defaults to disabled, and is
750 * enabled if either shader stage includes the execution mode.
752 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
753 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
754 * and OutputVertices, it says:
756 * "One mode must be set in at least one of the tessellation
759 * So, the fields can be set in either the TCS or TES, but they must
760 * agree if set in both. Our backend looks at TES, so bitwise-or in
761 * the values from the TCS.
763 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
764 tes_info
->tess
.tcs_vertices_out
== 0 ||
765 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
766 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
768 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
769 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
770 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
771 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
773 assert(tcs_info
->tess
.primitive_mode
== 0 ||
774 tes_info
->tess
.primitive_mode
== 0 ||
775 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
776 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
777 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
778 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
782 anv_pipeline_link_tcs(const struct brw_compiler
*compiler
,
783 struct anv_pipeline_stage
*tcs_stage
,
784 struct anv_pipeline_stage
*tes_stage
)
786 assert(tes_stage
&& tes_stage
->stage
== MESA_SHADER_TESS_EVAL
);
788 brw_nir_link_shaders(compiler
, tcs_stage
->nir
, tes_stage
->nir
);
790 nir_lower_patch_vertices(tes_stage
->nir
,
791 tcs_stage
->nir
->info
.tess
.tcs_vertices_out
,
794 /* Copy TCS info into the TES info */
795 merge_tess_info(&tes_stage
->nir
->info
, &tcs_stage
->nir
->info
);
797 /* Whacking the key after cache lookup is a bit sketchy, but all of
798 * this comes from the SPIR-V, which is part of the hash used for the
799 * pipeline cache. So it should be safe.
801 tcs_stage
->key
.tcs
.tes_primitive_mode
=
802 tes_stage
->nir
->info
.tess
.primitive_mode
;
803 tcs_stage
->key
.tcs
.quads_workaround
=
804 compiler
->devinfo
->gen
< 9 &&
805 tes_stage
->nir
->info
.tess
.primitive_mode
== 7 /* GL_QUADS */ &&
806 tes_stage
->nir
->info
.tess
.spacing
== TESS_SPACING_EQUAL
;
810 anv_pipeline_compile_tcs(const struct brw_compiler
*compiler
,
812 struct anv_device
*device
,
813 struct anv_pipeline_stage
*tcs_stage
,
814 struct anv_pipeline_stage
*prev_stage
)
816 tcs_stage
->key
.tcs
.outputs_written
=
817 tcs_stage
->nir
->info
.outputs_written
;
818 tcs_stage
->key
.tcs
.patch_outputs_written
=
819 tcs_stage
->nir
->info
.patch_outputs_written
;
821 tcs_stage
->num_stats
= 1;
822 tcs_stage
->code
= brw_compile_tcs(compiler
, device
, mem_ctx
,
824 &tcs_stage
->prog_data
.tcs
,
826 tcs_stage
->stats
, NULL
);
830 anv_pipeline_link_tes(const struct brw_compiler
*compiler
,
831 struct anv_pipeline_stage
*tes_stage
,
832 struct anv_pipeline_stage
*next_stage
)
835 brw_nir_link_shaders(compiler
, tes_stage
->nir
, next_stage
->nir
);
839 anv_pipeline_compile_tes(const struct brw_compiler
*compiler
,
841 struct anv_device
*device
,
842 struct anv_pipeline_stage
*tes_stage
,
843 struct anv_pipeline_stage
*tcs_stage
)
845 tes_stage
->key
.tes
.inputs_read
=
846 tcs_stage
->nir
->info
.outputs_written
;
847 tes_stage
->key
.tes
.patch_inputs_read
=
848 tcs_stage
->nir
->info
.patch_outputs_written
;
850 tes_stage
->num_stats
= 1;
851 tes_stage
->code
= brw_compile_tes(compiler
, device
, mem_ctx
,
853 &tcs_stage
->prog_data
.tcs
.base
.vue_map
,
854 &tes_stage
->prog_data
.tes
,
856 tes_stage
->stats
, NULL
);
860 anv_pipeline_link_gs(const struct brw_compiler
*compiler
,
861 struct anv_pipeline_stage
*gs_stage
,
862 struct anv_pipeline_stage
*next_stage
)
865 brw_nir_link_shaders(compiler
, gs_stage
->nir
, next_stage
->nir
);
869 anv_pipeline_compile_gs(const struct brw_compiler
*compiler
,
871 struct anv_device
*device
,
872 struct anv_pipeline_stage
*gs_stage
,
873 struct anv_pipeline_stage
*prev_stage
)
875 brw_compute_vue_map(compiler
->devinfo
,
876 &gs_stage
->prog_data
.gs
.base
.vue_map
,
877 gs_stage
->nir
->info
.outputs_written
,
878 gs_stage
->nir
->info
.separate_shader
);
880 gs_stage
->num_stats
= 1;
881 gs_stage
->code
= brw_compile_gs(compiler
, device
, mem_ctx
,
883 &gs_stage
->prog_data
.gs
,
884 gs_stage
->nir
, NULL
, -1,
885 gs_stage
->stats
, NULL
);
889 anv_pipeline_link_fs(const struct brw_compiler
*compiler
,
890 struct anv_pipeline_stage
*stage
)
892 unsigned num_rt_bindings
;
893 struct anv_pipeline_binding rt_bindings
[MAX_RTS
];
894 if (stage
->key
.wm
.nr_color_regions
> 0) {
895 assert(stage
->key
.wm
.nr_color_regions
<= MAX_RTS
);
896 for (unsigned rt
= 0; rt
< stage
->key
.wm
.nr_color_regions
; rt
++) {
897 if (stage
->key
.wm
.color_outputs_valid
& BITFIELD_BIT(rt
)) {
898 rt_bindings
[rt
] = (struct anv_pipeline_binding
) {
899 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
903 /* Setup a null render target */
904 rt_bindings
[rt
] = (struct anv_pipeline_binding
) {
905 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
910 num_rt_bindings
= stage
->key
.wm
.nr_color_regions
;
912 /* Setup a null render target */
913 rt_bindings
[0] = (struct anv_pipeline_binding
) {
914 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
920 assert(num_rt_bindings
<= MAX_RTS
);
921 assert(stage
->bind_map
.surface_count
== 0);
922 typed_memcpy(stage
->bind_map
.surface_to_descriptor
,
923 rt_bindings
, num_rt_bindings
);
924 stage
->bind_map
.surface_count
+= num_rt_bindings
;
926 /* Now that we've set up the color attachments, we can go through and
927 * eliminate any shader outputs that map to VK_ATTACHMENT_UNUSED in the
928 * hopes that dead code can clean them up in this and any earlier shader
931 nir_function_impl
*impl
= nir_shader_get_entrypoint(stage
->nir
);
932 bool deleted_output
= false;
933 nir_foreach_variable_safe(var
, &stage
->nir
->outputs
) {
934 /* TODO: We don't delete depth/stencil writes. We probably could if the
935 * subpass doesn't have a depth/stencil attachment.
937 if (var
->data
.location
< FRAG_RESULT_DATA0
)
940 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
942 /* If this is the RT at location 0 and we have alpha to coverage
943 * enabled we still need that write because it will affect the coverage
944 * mask even if it's never written to a color target.
946 if (rt
== 0 && stage
->key
.wm
.alpha_to_coverage
)
949 const unsigned array_len
=
950 glsl_type_is_array(var
->type
) ? glsl_get_length(var
->type
) : 1;
951 assert(rt
+ array_len
<= MAX_RTS
);
953 if (rt
>= MAX_RTS
|| !(stage
->key
.wm
.color_outputs_valid
&
954 BITFIELD_RANGE(rt
, array_len
))) {
955 deleted_output
= true;
956 var
->data
.mode
= nir_var_function_temp
;
957 exec_node_remove(&var
->node
);
958 exec_list_push_tail(&impl
->locals
, &var
->node
);
963 nir_fixup_deref_modes(stage
->nir
);
965 /* We stored the number of subpass color attachments in nr_color_regions
966 * when calculating the key for caching. Now that we've computed the bind
967 * map, we can reduce this to the actual max before we go into the back-end
970 stage
->key
.wm
.nr_color_regions
=
971 util_last_bit(stage
->key
.wm
.color_outputs_valid
);
975 anv_pipeline_compile_fs(const struct brw_compiler
*compiler
,
977 struct anv_device
*device
,
978 struct anv_pipeline_stage
*fs_stage
,
979 struct anv_pipeline_stage
*prev_stage
)
981 /* TODO: we could set this to 0 based on the information in nir_shader, but
982 * we need this before we call spirv_to_nir.
985 fs_stage
->key
.wm
.input_slots_valid
=
986 prev_stage
->prog_data
.vue
.vue_map
.slots_valid
;
988 fs_stage
->code
= brw_compile_fs(compiler
, device
, mem_ctx
,
990 &fs_stage
->prog_data
.wm
,
991 fs_stage
->nir
, -1, -1, -1,
993 fs_stage
->stats
, NULL
);
995 fs_stage
->num_stats
= (uint32_t)fs_stage
->prog_data
.wm
.dispatch_8
+
996 (uint32_t)fs_stage
->prog_data
.wm
.dispatch_16
+
997 (uint32_t)fs_stage
->prog_data
.wm
.dispatch_32
;
999 if (fs_stage
->key
.wm
.color_outputs_valid
== 0 &&
1000 !fs_stage
->prog_data
.wm
.has_side_effects
&&
1001 !fs_stage
->prog_data
.wm
.uses_omask
&&
1002 !fs_stage
->key
.wm
.alpha_to_coverage
&&
1003 !fs_stage
->prog_data
.wm
.uses_kill
&&
1004 fs_stage
->prog_data
.wm
.computed_depth_mode
== BRW_PSCDEPTH_OFF
&&
1005 !fs_stage
->prog_data
.wm
.computed_stencil
) {
1006 /* This fragment shader has no outputs and no side effects. Go ahead
1007 * and return the code pointer so we don't accidentally think the
1008 * compile failed but zero out prog_data which will set program_size to
1009 * zero and disable the stage.
1011 memset(&fs_stage
->prog_data
, 0, sizeof(fs_stage
->prog_data
));
1016 anv_pipeline_add_executable(struct anv_pipeline
*pipeline
,
1017 struct anv_pipeline_stage
*stage
,
1018 struct brw_compile_stats
*stats
,
1019 uint32_t code_offset
)
1024 VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
)) {
1025 char *stream_data
= NULL
;
1026 size_t stream_size
= 0;
1027 FILE *stream
= open_memstream(&stream_data
, &stream_size
);
1029 nir_print_shader(stage
->nir
, stream
);
1033 /* Copy it to a ralloc'd thing */
1034 nir
= ralloc_size(pipeline
->mem_ctx
, stream_size
+ 1);
1035 memcpy(nir
, stream_data
, stream_size
);
1036 nir
[stream_size
] = 0;
1041 char *disasm
= NULL
;
1044 VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
)) {
1045 char *stream_data
= NULL
;
1046 size_t stream_size
= 0;
1047 FILE *stream
= open_memstream(&stream_data
, &stream_size
);
1049 /* Creating this is far cheaper than it looks. It's perfectly fine to
1050 * do it for every binary.
1052 struct gen_disasm
*d
= gen_disasm_create(&pipeline
->device
->info
);
1053 gen_disasm_disassemble(d
, stage
->code
, code_offset
, stream
);
1054 gen_disasm_destroy(d
);
1058 /* Copy it to a ralloc'd thing */
1059 disasm
= ralloc_size(pipeline
->mem_ctx
, stream_size
+ 1);
1060 memcpy(disasm
, stream_data
, stream_size
);
1061 disasm
[stream_size
] = 0;
1066 pipeline
->executables
[pipeline
->num_executables
++] =
1067 (struct anv_pipeline_executable
) {
1068 .stage
= stage
->stage
,
1076 anv_pipeline_add_executables(struct anv_pipeline
*pipeline
,
1077 struct anv_pipeline_stage
*stage
,
1078 struct anv_shader_bin
*bin
)
1080 if (stage
->stage
== MESA_SHADER_FRAGMENT
) {
1081 /* We pull the prog data and stats out of the anv_shader_bin because
1082 * the anv_pipeline_stage may not be fully populated if we successfully
1083 * looked up the shader in a cache.
1085 const struct brw_wm_prog_data
*wm_prog_data
=
1086 (const struct brw_wm_prog_data
*)bin
->prog_data
;
1087 struct brw_compile_stats
*stats
= bin
->stats
;
1089 if (wm_prog_data
->dispatch_8
) {
1090 anv_pipeline_add_executable(pipeline
, stage
, stats
++, 0);
1093 if (wm_prog_data
->dispatch_16
) {
1094 anv_pipeline_add_executable(pipeline
, stage
, stats
++,
1095 wm_prog_data
->prog_offset_16
);
1098 if (wm_prog_data
->dispatch_32
) {
1099 anv_pipeline_add_executable(pipeline
, stage
, stats
++,
1100 wm_prog_data
->prog_offset_32
);
1103 anv_pipeline_add_executable(pipeline
, stage
, bin
->stats
, 0);
1108 anv_pipeline_compile_graphics(struct anv_pipeline
*pipeline
,
1109 struct anv_pipeline_cache
*cache
,
1110 const VkGraphicsPipelineCreateInfo
*info
)
1112 VkPipelineCreationFeedbackEXT pipeline_feedback
= {
1113 .flags
= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
,
1115 int64_t pipeline_start
= os_time_get_nano();
1117 const struct brw_compiler
*compiler
= pipeline
->device
->physical
->compiler
;
1118 struct anv_pipeline_stage stages
[MESA_SHADER_STAGES
] = {};
1120 pipeline
->active_stages
= 0;
1123 for (uint32_t i
= 0; i
< info
->stageCount
; i
++) {
1124 const VkPipelineShaderStageCreateInfo
*sinfo
= &info
->pStages
[i
];
1125 gl_shader_stage stage
= vk_to_mesa_shader_stage(sinfo
->stage
);
1127 pipeline
->active_stages
|= sinfo
->stage
;
1129 int64_t stage_start
= os_time_get_nano();
1131 stages
[stage
].stage
= stage
;
1132 stages
[stage
].module
= anv_shader_module_from_handle(sinfo
->module
);
1133 stages
[stage
].entrypoint
= sinfo
->pName
;
1134 stages
[stage
].spec_info
= sinfo
->pSpecializationInfo
;
1135 anv_pipeline_hash_shader(stages
[stage
].module
,
1136 stages
[stage
].entrypoint
,
1138 stages
[stage
].spec_info
,
1139 stages
[stage
].shader_sha1
);
1141 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1143 case MESA_SHADER_VERTEX
:
1144 populate_vs_prog_key(devinfo
, sinfo
->flags
, &stages
[stage
].key
.vs
);
1146 case MESA_SHADER_TESS_CTRL
:
1147 populate_tcs_prog_key(devinfo
, sinfo
->flags
,
1148 info
->pTessellationState
->patchControlPoints
,
1149 &stages
[stage
].key
.tcs
);
1151 case MESA_SHADER_TESS_EVAL
:
1152 populate_tes_prog_key(devinfo
, sinfo
->flags
, &stages
[stage
].key
.tes
);
1154 case MESA_SHADER_GEOMETRY
:
1155 populate_gs_prog_key(devinfo
, sinfo
->flags
, &stages
[stage
].key
.gs
);
1157 case MESA_SHADER_FRAGMENT
: {
1158 const bool raster_enabled
=
1159 !info
->pRasterizationState
->rasterizerDiscardEnable
;
1160 populate_wm_prog_key(devinfo
, sinfo
->flags
,
1162 raster_enabled
? info
->pMultisampleState
: NULL
,
1163 &stages
[stage
].key
.wm
);
1167 unreachable("Invalid graphics shader stage");
1170 stages
[stage
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1171 stages
[stage
].feedback
.flags
|= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
;
1174 if (pipeline
->active_stages
& VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
)
1175 pipeline
->active_stages
|= VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
;
1177 assert(pipeline
->active_stages
& VK_SHADER_STAGE_VERTEX_BIT
);
1179 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
1181 unsigned char sha1
[20];
1182 anv_pipeline_hash_graphics(pipeline
, layout
, stages
, sha1
);
1184 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1185 if (!stages
[s
].entrypoint
)
1188 stages
[s
].cache_key
.stage
= s
;
1189 memcpy(stages
[s
].cache_key
.sha1
, sha1
, sizeof(sha1
));
1192 const bool skip_cache_lookup
=
1193 (pipeline
->flags
& VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
);
1195 if (!skip_cache_lookup
) {
1197 unsigned cache_hits
= 0;
1198 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1199 if (!stages
[s
].entrypoint
)
1202 int64_t stage_start
= os_time_get_nano();
1205 struct anv_shader_bin
*bin
=
1206 anv_device_search_for_kernel(pipeline
->device
, cache
,
1207 &stages
[s
].cache_key
,
1208 sizeof(stages
[s
].cache_key
), &cache_hit
);
1211 pipeline
->shaders
[s
] = bin
;
1216 stages
[s
].feedback
.flags
|=
1217 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1219 stages
[s
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1222 if (found
== __builtin_popcount(pipeline
->active_stages
)) {
1223 if (cache_hits
== found
) {
1224 pipeline_feedback
.flags
|=
1225 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1227 /* We found all our shaders in the cache. We're done. */
1228 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1229 if (!stages
[s
].entrypoint
)
1232 anv_pipeline_add_executables(pipeline
, &stages
[s
],
1233 pipeline
->shaders
[s
]);
1236 } else if (found
> 0) {
1237 /* We found some but not all of our shaders. This shouldn't happen
1238 * most of the time but it can if we have a partially populated
1241 assert(found
< __builtin_popcount(pipeline
->active_stages
));
1243 vk_debug_report(&pipeline
->device
->physical
->instance
->debug_report_callbacks
,
1244 VK_DEBUG_REPORT_WARNING_BIT_EXT
|
1245 VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT
,
1246 VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_CACHE_EXT
,
1247 (uint64_t)(uintptr_t)cache
,
1249 "Found a partial pipeline in the cache. This is "
1250 "most likely caused by an incomplete pipeline cache "
1251 "import or export");
1253 /* We're going to have to recompile anyway, so just throw away our
1254 * references to the shaders in the cache. We'll get them out of the
1255 * cache again as part of the compilation process.
1257 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1258 stages
[s
].feedback
.flags
= 0;
1259 if (pipeline
->shaders
[s
]) {
1260 anv_shader_bin_unref(pipeline
->device
, pipeline
->shaders
[s
]);
1261 pipeline
->shaders
[s
] = NULL
;
1267 void *pipeline_ctx
= ralloc_context(NULL
);
1269 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1270 if (!stages
[s
].entrypoint
)
1273 int64_t stage_start
= os_time_get_nano();
1275 assert(stages
[s
].stage
== s
);
1276 assert(pipeline
->shaders
[s
] == NULL
);
1278 stages
[s
].bind_map
= (struct anv_pipeline_bind_map
) {
1279 .surface_to_descriptor
= stages
[s
].surface_to_descriptor
,
1280 .sampler_to_descriptor
= stages
[s
].sampler_to_descriptor
1283 stages
[s
].nir
= anv_pipeline_stage_get_nir(pipeline
, cache
,
1286 if (stages
[s
].nir
== NULL
) {
1287 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1291 stages
[s
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1294 /* Walk backwards to link */
1295 struct anv_pipeline_stage
*next_stage
= NULL
;
1296 for (int s
= MESA_SHADER_STAGES
- 1; s
>= 0; s
--) {
1297 if (!stages
[s
].entrypoint
)
1301 case MESA_SHADER_VERTEX
:
1302 anv_pipeline_link_vs(compiler
, &stages
[s
], next_stage
);
1304 case MESA_SHADER_TESS_CTRL
:
1305 anv_pipeline_link_tcs(compiler
, &stages
[s
], next_stage
);
1307 case MESA_SHADER_TESS_EVAL
:
1308 anv_pipeline_link_tes(compiler
, &stages
[s
], next_stage
);
1310 case MESA_SHADER_GEOMETRY
:
1311 anv_pipeline_link_gs(compiler
, &stages
[s
], next_stage
);
1313 case MESA_SHADER_FRAGMENT
:
1314 anv_pipeline_link_fs(compiler
, &stages
[s
]);
1317 unreachable("Invalid graphics shader stage");
1320 next_stage
= &stages
[s
];
1323 struct anv_pipeline_stage
*prev_stage
= NULL
;
1324 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1325 if (!stages
[s
].entrypoint
)
1328 int64_t stage_start
= os_time_get_nano();
1330 void *stage_ctx
= ralloc_context(NULL
);
1332 nir_xfb_info
*xfb_info
= NULL
;
1333 if (s
== MESA_SHADER_VERTEX
||
1334 s
== MESA_SHADER_TESS_EVAL
||
1335 s
== MESA_SHADER_GEOMETRY
)
1336 xfb_info
= nir_gather_xfb_info(stages
[s
].nir
, stage_ctx
);
1338 anv_pipeline_lower_nir(pipeline
, stage_ctx
, &stages
[s
], layout
);
1341 case MESA_SHADER_VERTEX
:
1342 anv_pipeline_compile_vs(compiler
, stage_ctx
, pipeline
->device
,
1345 case MESA_SHADER_TESS_CTRL
:
1346 anv_pipeline_compile_tcs(compiler
, stage_ctx
, pipeline
->device
,
1347 &stages
[s
], prev_stage
);
1349 case MESA_SHADER_TESS_EVAL
:
1350 anv_pipeline_compile_tes(compiler
, stage_ctx
, pipeline
->device
,
1351 &stages
[s
], prev_stage
);
1353 case MESA_SHADER_GEOMETRY
:
1354 anv_pipeline_compile_gs(compiler
, stage_ctx
, pipeline
->device
,
1355 &stages
[s
], prev_stage
);
1357 case MESA_SHADER_FRAGMENT
:
1358 anv_pipeline_compile_fs(compiler
, stage_ctx
, pipeline
->device
,
1359 &stages
[s
], prev_stage
);
1362 unreachable("Invalid graphics shader stage");
1364 if (stages
[s
].code
== NULL
) {
1365 ralloc_free(stage_ctx
);
1366 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1370 anv_nir_validate_push_layout(&stages
[s
].prog_data
.base
,
1371 &stages
[s
].bind_map
);
1373 struct anv_shader_bin
*bin
=
1374 anv_device_upload_kernel(pipeline
->device
, cache
,
1375 &stages
[s
].cache_key
,
1376 sizeof(stages
[s
].cache_key
),
1378 stages
[s
].prog_data
.base
.program_size
,
1379 stages
[s
].nir
->constant_data
,
1380 stages
[s
].nir
->constant_data_size
,
1381 &stages
[s
].prog_data
.base
,
1382 brw_prog_data_size(s
),
1383 stages
[s
].stats
, stages
[s
].num_stats
,
1384 xfb_info
, &stages
[s
].bind_map
);
1386 ralloc_free(stage_ctx
);
1387 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1391 anv_pipeline_add_executables(pipeline
, &stages
[s
], bin
);
1393 pipeline
->shaders
[s
] = bin
;
1394 ralloc_free(stage_ctx
);
1396 stages
[s
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1398 prev_stage
= &stages
[s
];
1401 ralloc_free(pipeline_ctx
);
1405 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
] &&
1406 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->prog_data
->program_size
== 0) {
1407 /* This can happen if we decided to implicitly disable the fragment
1408 * shader. See anv_pipeline_compile_fs().
1410 anv_shader_bin_unref(pipeline
->device
,
1411 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
1412 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] = NULL
;
1413 pipeline
->active_stages
&= ~VK_SHADER_STAGE_FRAGMENT_BIT
;
1416 pipeline_feedback
.duration
= os_time_get_nano() - pipeline_start
;
1418 const VkPipelineCreationFeedbackCreateInfoEXT
*create_feedback
=
1419 vk_find_struct_const(info
->pNext
, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT
);
1420 if (create_feedback
) {
1421 *create_feedback
->pPipelineCreationFeedback
= pipeline_feedback
;
1423 assert(info
->stageCount
== create_feedback
->pipelineStageCreationFeedbackCount
);
1424 for (uint32_t i
= 0; i
< info
->stageCount
; i
++) {
1425 gl_shader_stage s
= vk_to_mesa_shader_stage(info
->pStages
[i
].stage
);
1426 create_feedback
->pPipelineStageCreationFeedbacks
[i
] = stages
[s
].feedback
;
1433 ralloc_free(pipeline_ctx
);
1435 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1436 if (pipeline
->shaders
[s
])
1437 anv_shader_bin_unref(pipeline
->device
, pipeline
->shaders
[s
]);
1444 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
1446 assert(glsl_type_is_vector_or_scalar(type
));
1448 uint32_t comp_size
= glsl_type_is_boolean(type
)
1449 ? 4 : glsl_get_bit_size(type
) / 8;
1450 unsigned length
= glsl_get_vector_elements(type
);
1451 *size
= comp_size
* length
,
1452 *align
= comp_size
* (length
== 3 ? 4 : length
);
1456 anv_pipeline_compile_cs(struct anv_pipeline
*pipeline
,
1457 struct anv_pipeline_cache
*cache
,
1458 const VkComputePipelineCreateInfo
*info
,
1459 const struct anv_shader_module
*module
,
1460 const char *entrypoint
,
1461 const VkSpecializationInfo
*spec_info
)
1463 VkPipelineCreationFeedbackEXT pipeline_feedback
= {
1464 .flags
= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
,
1466 int64_t pipeline_start
= os_time_get_nano();
1468 const struct brw_compiler
*compiler
= pipeline
->device
->physical
->compiler
;
1470 struct anv_pipeline_stage stage
= {
1471 .stage
= MESA_SHADER_COMPUTE
,
1473 .entrypoint
= entrypoint
,
1474 .spec_info
= spec_info
,
1476 .stage
= MESA_SHADER_COMPUTE
,
1479 .flags
= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
,
1482 anv_pipeline_hash_shader(stage
.module
,
1484 MESA_SHADER_COMPUTE
,
1488 struct anv_shader_bin
*bin
= NULL
;
1490 const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT
*rss_info
=
1491 vk_find_struct_const(info
->stage
.pNext
,
1492 PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT
);
1494 populate_cs_prog_key(&pipeline
->device
->info
, info
->stage
.flags
,
1495 rss_info
, &stage
.key
.cs
);
1497 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
1499 const bool skip_cache_lookup
=
1500 (pipeline
->flags
& VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
);
1502 anv_pipeline_hash_compute(pipeline
, layout
, &stage
, stage
.cache_key
.sha1
);
1504 bool cache_hit
= false;
1505 if (!skip_cache_lookup
) {
1506 bin
= anv_device_search_for_kernel(pipeline
->device
, cache
,
1508 sizeof(stage
.cache_key
),
1512 void *mem_ctx
= ralloc_context(NULL
);
1514 int64_t stage_start
= os_time_get_nano();
1516 stage
.bind_map
= (struct anv_pipeline_bind_map
) {
1517 .surface_to_descriptor
= stage
.surface_to_descriptor
,
1518 .sampler_to_descriptor
= stage
.sampler_to_descriptor
1521 /* Set up a binding for the gl_NumWorkGroups */
1522 stage
.bind_map
.surface_count
= 1;
1523 stage
.bind_map
.surface_to_descriptor
[0] = (struct anv_pipeline_binding
) {
1524 .set
= ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS
,
1527 stage
.nir
= anv_pipeline_stage_get_nir(pipeline
, cache
, mem_ctx
, &stage
);
1528 if (stage
.nir
== NULL
) {
1529 ralloc_free(mem_ctx
);
1530 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1533 NIR_PASS_V(stage
.nir
, anv_nir_add_base_work_group_id
);
1535 anv_pipeline_lower_nir(pipeline
, mem_ctx
, &stage
, layout
);
1537 NIR_PASS_V(stage
.nir
, nir_lower_vars_to_explicit_types
,
1538 nir_var_mem_shared
, shared_type_info
);
1539 NIR_PASS_V(stage
.nir
, nir_lower_explicit_io
,
1540 nir_var_mem_shared
, nir_address_format_32bit_offset
);
1542 stage
.num_stats
= 1;
1543 stage
.code
= brw_compile_cs(compiler
, pipeline
->device
, mem_ctx
,
1544 &stage
.key
.cs
, &stage
.prog_data
.cs
,
1545 stage
.nir
, -1, stage
.stats
, NULL
);
1546 if (stage
.code
== NULL
) {
1547 ralloc_free(mem_ctx
);
1548 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1551 anv_nir_validate_push_layout(&stage
.prog_data
.base
, &stage
.bind_map
);
1553 if (!stage
.prog_data
.cs
.uses_num_work_groups
) {
1554 assert(stage
.bind_map
.surface_to_descriptor
[0].set
==
1555 ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS
);
1556 stage
.bind_map
.surface_to_descriptor
[0].set
= ANV_DESCRIPTOR_SET_NULL
;
1559 const unsigned code_size
= stage
.prog_data
.base
.program_size
;
1560 bin
= anv_device_upload_kernel(pipeline
->device
, cache
,
1561 &stage
.cache_key
, sizeof(stage
.cache_key
),
1562 stage
.code
, code_size
,
1563 stage
.nir
->constant_data
,
1564 stage
.nir
->constant_data_size
,
1565 &stage
.prog_data
.base
,
1566 sizeof(stage
.prog_data
.cs
),
1567 stage
.stats
, stage
.num_stats
,
1568 NULL
, &stage
.bind_map
);
1570 ralloc_free(mem_ctx
);
1571 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1574 stage
.feedback
.duration
= os_time_get_nano() - stage_start
;
1577 anv_pipeline_add_executables(pipeline
, &stage
, bin
);
1579 ralloc_free(mem_ctx
);
1582 stage
.feedback
.flags
|=
1583 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1584 pipeline_feedback
.flags
|=
1585 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1587 pipeline_feedback
.duration
= os_time_get_nano() - pipeline_start
;
1589 const VkPipelineCreationFeedbackCreateInfoEXT
*create_feedback
=
1590 vk_find_struct_const(info
->pNext
, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT
);
1591 if (create_feedback
) {
1592 *create_feedback
->pPipelineCreationFeedback
= pipeline_feedback
;
1594 assert(create_feedback
->pipelineStageCreationFeedbackCount
== 1);
1595 create_feedback
->pPipelineStageCreationFeedbacks
[0] = stage
.feedback
;
1598 pipeline
->active_stages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1599 pipeline
->shaders
[MESA_SHADER_COMPUTE
] = bin
;
1605 * Copy pipeline state not marked as dynamic.
1606 * Dynamic state is pipeline state which hasn't been provided at pipeline
1607 * creation time, but is dynamically provided afterwards using various
1608 * vkCmdSet* functions.
1610 * The set of state considered "non_dynamic" is determined by the pieces of
1611 * state that have their corresponding VkDynamicState enums omitted from
1612 * VkPipelineDynamicStateCreateInfo::pDynamicStates.
1614 * @param[out] pipeline Destination non_dynamic state.
1615 * @param[in] pCreateInfo Source of non_dynamic state to be copied.
1618 copy_non_dynamic_state(struct anv_pipeline
*pipeline
,
1619 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1621 anv_cmd_dirty_mask_t states
= ANV_CMD_DIRTY_DYNAMIC_ALL
;
1622 struct anv_subpass
*subpass
= pipeline
->subpass
;
1624 pipeline
->dynamic_state
= default_dynamic_state
;
1626 if (pCreateInfo
->pDynamicState
) {
1627 /* Remove all of the states that are marked as dynamic */
1628 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1629 for (uint32_t s
= 0; s
< count
; s
++) {
1630 states
&= ~anv_cmd_dirty_bit_for_vk_dynamic_state(
1631 pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1635 struct anv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1637 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1639 * pViewportState is [...] NULL if the pipeline
1640 * has rasterization disabled.
1642 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1643 assert(pCreateInfo
->pViewportState
);
1645 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1646 if (states
& ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
) {
1647 typed_memcpy(dynamic
->viewport
.viewports
,
1648 pCreateInfo
->pViewportState
->pViewports
,
1649 pCreateInfo
->pViewportState
->viewportCount
);
1652 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1653 if (states
& ANV_CMD_DIRTY_DYNAMIC_SCISSOR
) {
1654 typed_memcpy(dynamic
->scissor
.scissors
,
1655 pCreateInfo
->pViewportState
->pScissors
,
1656 pCreateInfo
->pViewportState
->scissorCount
);
1660 if (states
& ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH
) {
1661 assert(pCreateInfo
->pRasterizationState
);
1662 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1665 if (states
& ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
) {
1666 assert(pCreateInfo
->pRasterizationState
);
1667 dynamic
->depth_bias
.bias
=
1668 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1669 dynamic
->depth_bias
.clamp
=
1670 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1671 dynamic
->depth_bias
.slope
=
1672 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1675 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1677 * pColorBlendState is [...] NULL if the pipeline has rasterization
1678 * disabled or if the subpass of the render pass the pipeline is
1679 * created against does not use any color attachments.
1681 bool uses_color_att
= false;
1682 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1683 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1684 uses_color_att
= true;
1689 if (uses_color_att
&&
1690 !pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1691 assert(pCreateInfo
->pColorBlendState
);
1693 if (states
& ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS
)
1694 typed_memcpy(dynamic
->blend_constants
,
1695 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1698 /* If there is no depthstencil attachment, then don't read
1699 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1700 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1701 * no need to override the depthstencil defaults in
1702 * anv_pipeline::dynamic_state when there is no depthstencil attachment.
1704 * Section 9.2 of the Vulkan 1.0.15 spec says:
1706 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1707 * disabled or if the subpass of the render pass the pipeline is created
1708 * against does not use a depth/stencil attachment.
1710 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1711 subpass
->depth_stencil_attachment
) {
1712 assert(pCreateInfo
->pDepthStencilState
);
1714 if (states
& ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS
) {
1715 dynamic
->depth_bounds
.min
=
1716 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1717 dynamic
->depth_bounds
.max
=
1718 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1721 if (states
& ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK
) {
1722 dynamic
->stencil_compare_mask
.front
=
1723 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1724 dynamic
->stencil_compare_mask
.back
=
1725 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1728 if (states
& ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK
) {
1729 dynamic
->stencil_write_mask
.front
=
1730 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1731 dynamic
->stencil_write_mask
.back
=
1732 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1735 if (states
& ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE
) {
1736 dynamic
->stencil_reference
.front
=
1737 pCreateInfo
->pDepthStencilState
->front
.reference
;
1738 dynamic
->stencil_reference
.back
=
1739 pCreateInfo
->pDepthStencilState
->back
.reference
;
1743 const VkPipelineRasterizationLineStateCreateInfoEXT
*line_state
=
1744 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
,
1745 PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT
);
1747 if (states
& ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE
) {
1748 dynamic
->line_stipple
.factor
= line_state
->lineStippleFactor
;
1749 dynamic
->line_stipple
.pattern
= line_state
->lineStipplePattern
;
1753 pipeline
->dynamic_state_mask
= states
;
1757 anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo
*info
)
1760 struct anv_render_pass
*renderpass
= NULL
;
1761 struct anv_subpass
*subpass
= NULL
;
1763 /* Assert that all required members of VkGraphicsPipelineCreateInfo are
1764 * present. See the Vulkan 1.0.28 spec, Section 9.2 Graphics Pipelines.
1766 assert(info
->sType
== VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
);
1768 renderpass
= anv_render_pass_from_handle(info
->renderPass
);
1771 assert(info
->subpass
< renderpass
->subpass_count
);
1772 subpass
= &renderpass
->subpasses
[info
->subpass
];
1774 assert(info
->stageCount
>= 1);
1775 assert(info
->pVertexInputState
);
1776 assert(info
->pInputAssemblyState
);
1777 assert(info
->pRasterizationState
);
1778 if (!info
->pRasterizationState
->rasterizerDiscardEnable
) {
1779 assert(info
->pViewportState
);
1780 assert(info
->pMultisampleState
);
1782 if (subpass
&& subpass
->depth_stencil_attachment
)
1783 assert(info
->pDepthStencilState
);
1785 if (subpass
&& subpass
->color_count
> 0) {
1786 bool all_color_unused
= true;
1787 for (int i
= 0; i
< subpass
->color_count
; i
++) {
1788 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
1789 all_color_unused
= false;
1791 /* pColorBlendState is ignored if the pipeline has rasterization
1792 * disabled or if the subpass of the render pass the pipeline is
1793 * created against does not use any color attachments.
1795 assert(info
->pColorBlendState
|| all_color_unused
);
1799 for (uint32_t i
= 0; i
< info
->stageCount
; ++i
) {
1800 switch (info
->pStages
[i
].stage
) {
1801 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
:
1802 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
:
1803 assert(info
->pTessellationState
);
1813 * Calculate the desired L3 partitioning based on the current state of the
1814 * pipeline. For now this simply returns the conservative defaults calculated
1815 * by get_default_l3_weights(), but we could probably do better by gathering
1816 * more statistics from the pipeline state (e.g. guess of expected URB usage
1817 * and bound surfaces), or by using feed-back from performance counters.
1820 anv_pipeline_setup_l3_config(struct anv_pipeline
*pipeline
, bool needs_slm
)
1822 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1824 const struct gen_l3_weights w
=
1825 gen_get_default_l3_weights(devinfo
, pipeline
->needs_data_cache
, needs_slm
);
1827 pipeline
->urb
.l3_config
= gen_get_l3_config(devinfo
, w
);
1828 pipeline
->urb
.total_size
=
1829 gen_get_l3_config_urb_size(devinfo
, pipeline
->urb
.l3_config
);
1833 anv_pipeline_init(struct anv_pipeline
*pipeline
,
1834 struct anv_device
*device
,
1835 struct anv_pipeline_cache
*cache
,
1836 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1837 const VkAllocationCallbacks
*alloc
)
1841 anv_pipeline_validate_create_info(pCreateInfo
);
1844 alloc
= &device
->alloc
;
1846 pipeline
->device
= device
;
1848 ANV_FROM_HANDLE(anv_render_pass
, render_pass
, pCreateInfo
->renderPass
);
1849 assert(pCreateInfo
->subpass
< render_pass
->subpass_count
);
1850 pipeline
->subpass
= &render_pass
->subpasses
[pCreateInfo
->subpass
];
1852 result
= anv_reloc_list_init(&pipeline
->batch_relocs
, alloc
);
1853 if (result
!= VK_SUCCESS
)
1856 pipeline
->batch
.alloc
= alloc
;
1857 pipeline
->batch
.next
= pipeline
->batch
.start
= pipeline
->batch_data
;
1858 pipeline
->batch
.end
= pipeline
->batch
.start
+ sizeof(pipeline
->batch_data
);
1859 pipeline
->batch
.relocs
= &pipeline
->batch_relocs
;
1860 pipeline
->batch
.status
= VK_SUCCESS
;
1862 pipeline
->mem_ctx
= ralloc_context(NULL
);
1863 pipeline
->flags
= pCreateInfo
->flags
;
1865 assert(pCreateInfo
->pRasterizationState
);
1867 copy_non_dynamic_state(pipeline
, pCreateInfo
);
1868 pipeline
->depth_clamp_enable
= pCreateInfo
->pRasterizationState
->depthClampEnable
;
1870 /* Previously we enabled depth clipping when !depthClampEnable.
1871 * DepthClipStateCreateInfo now makes depth clipping explicit so if the
1872 * clipping info is available, use its enable value to determine clipping,
1873 * otherwise fallback to the previous !depthClampEnable logic.
1875 const VkPipelineRasterizationDepthClipStateCreateInfoEXT
*clip_info
=
1876 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
,
1877 PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT
);
1878 pipeline
->depth_clip_enable
= clip_info
? clip_info
->depthClipEnable
: !pipeline
->depth_clamp_enable
;
1880 pipeline
->sample_shading_enable
=
1881 !pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1882 pCreateInfo
->pMultisampleState
&&
1883 pCreateInfo
->pMultisampleState
->sampleShadingEnable
;
1885 pipeline
->needs_data_cache
= false;
1887 /* When we free the pipeline, we detect stages based on the NULL status
1888 * of various prog_data pointers. Make them NULL by default.
1890 memset(pipeline
->shaders
, 0, sizeof(pipeline
->shaders
));
1891 pipeline
->num_executables
= 0;
1893 result
= anv_pipeline_compile_graphics(pipeline
, cache
, pCreateInfo
);
1894 if (result
!= VK_SUCCESS
) {
1895 ralloc_free(pipeline
->mem_ctx
);
1896 anv_reloc_list_finish(&pipeline
->batch_relocs
, alloc
);
1900 assert(pipeline
->shaders
[MESA_SHADER_VERTEX
]);
1902 anv_pipeline_setup_l3_config(pipeline
, false);
1904 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1905 pCreateInfo
->pVertexInputState
;
1907 const uint64_t inputs_read
= get_vs_prog_data(pipeline
)->inputs_read
;
1909 pipeline
->vb_used
= 0;
1910 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1911 const VkVertexInputAttributeDescription
*desc
=
1912 &vi_info
->pVertexAttributeDescriptions
[i
];
1914 if (inputs_read
& (1ull << (VERT_ATTRIB_GENERIC0
+ desc
->location
)))
1915 pipeline
->vb_used
|= 1 << desc
->binding
;
1918 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1919 const VkVertexInputBindingDescription
*desc
=
1920 &vi_info
->pVertexBindingDescriptions
[i
];
1922 pipeline
->vb
[desc
->binding
].stride
= desc
->stride
;
1924 /* Step rate is programmed per vertex element (attribute), not
1925 * binding. Set up a map of which bindings step per instance, for
1926 * reference by vertex element setup. */
1927 switch (desc
->inputRate
) {
1929 case VK_VERTEX_INPUT_RATE_VERTEX
:
1930 pipeline
->vb
[desc
->binding
].instanced
= false;
1932 case VK_VERTEX_INPUT_RATE_INSTANCE
:
1933 pipeline
->vb
[desc
->binding
].instanced
= true;
1937 pipeline
->vb
[desc
->binding
].instance_divisor
= 1;
1940 const VkPipelineVertexInputDivisorStateCreateInfoEXT
*vi_div_state
=
1941 vk_find_struct_const(vi_info
->pNext
,
1942 PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT
);
1944 for (uint32_t i
= 0; i
< vi_div_state
->vertexBindingDivisorCount
; i
++) {
1945 const VkVertexInputBindingDivisorDescriptionEXT
*desc
=
1946 &vi_div_state
->pVertexBindingDivisors
[i
];
1948 pipeline
->vb
[desc
->binding
].instance_divisor
= desc
->divisor
;
1952 /* Our implementation of VK_KHR_multiview uses instancing to draw the
1953 * different views. If the client asks for instancing, we need to multiply
1954 * the instance divisor by the number of views ensure that we repeat the
1955 * client's per-instance data once for each view.
1957 if (pipeline
->subpass
->view_mask
) {
1958 const uint32_t view_count
= anv_subpass_view_count(pipeline
->subpass
);
1959 for (uint32_t vb
= 0; vb
< MAX_VBS
; vb
++) {
1960 if (pipeline
->vb
[vb
].instanced
)
1961 pipeline
->vb
[vb
].instance_divisor
*= view_count
;
1965 const VkPipelineInputAssemblyStateCreateInfo
*ia_info
=
1966 pCreateInfo
->pInputAssemblyState
;
1967 const VkPipelineTessellationStateCreateInfo
*tess_info
=
1968 pCreateInfo
->pTessellationState
;
1969 pipeline
->primitive_restart
= ia_info
->primitiveRestartEnable
;
1971 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
))
1972 pipeline
->topology
= _3DPRIM_PATCHLIST(tess_info
->patchControlPoints
);
1974 pipeline
->topology
= vk_to_gen_primitive_type
[ia_info
->topology
];
1979 #define WRITE_STR(field, ...) ({ \
1980 memset(field, 0, sizeof(field)); \
1981 UNUSED int i = snprintf(field, sizeof(field), __VA_ARGS__); \
1982 assert(i > 0 && i < sizeof(field)); \
1985 VkResult
anv_GetPipelineExecutablePropertiesKHR(
1987 const VkPipelineInfoKHR
* pPipelineInfo
,
1988 uint32_t* pExecutableCount
,
1989 VkPipelineExecutablePropertiesKHR
* pProperties
)
1991 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, pPipelineInfo
->pipeline
);
1992 VK_OUTARRAY_MAKE(out
, pProperties
, pExecutableCount
);
1994 for (uint32_t i
= 0; i
< pipeline
->num_executables
; i
++) {
1995 vk_outarray_append(&out
, props
) {
1996 gl_shader_stage stage
= pipeline
->executables
[i
].stage
;
1997 props
->stages
= mesa_to_vk_shader_stage(stage
);
1999 unsigned simd_width
= pipeline
->executables
[i
].stats
.dispatch_width
;
2000 if (stage
== MESA_SHADER_FRAGMENT
) {
2001 WRITE_STR(props
->name
, "%s%d %s",
2002 simd_width
? "SIMD" : "vec",
2003 simd_width
? simd_width
: 4,
2004 _mesa_shader_stage_to_string(stage
));
2006 WRITE_STR(props
->name
, "%s", _mesa_shader_stage_to_string(stage
));
2008 WRITE_STR(props
->description
, "%s%d %s shader",
2009 simd_width
? "SIMD" : "vec",
2010 simd_width
? simd_width
: 4,
2011 _mesa_shader_stage_to_string(stage
));
2013 /* The compiler gives us a dispatch width of 0 for vec4 but Vulkan
2014 * wants a subgroup size of 1.
2016 props
->subgroupSize
= MAX2(simd_width
, 1);
2020 return vk_outarray_status(&out
);
2023 VkResult
anv_GetPipelineExecutableStatisticsKHR(
2025 const VkPipelineExecutableInfoKHR
* pExecutableInfo
,
2026 uint32_t* pStatisticCount
,
2027 VkPipelineExecutableStatisticKHR
* pStatistics
)
2029 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, pExecutableInfo
->pipeline
);
2030 VK_OUTARRAY_MAKE(out
, pStatistics
, pStatisticCount
);
2032 assert(pExecutableInfo
->executableIndex
< pipeline
->num_executables
);
2033 const struct anv_pipeline_executable
*exe
=
2034 &pipeline
->executables
[pExecutableInfo
->executableIndex
];
2035 const struct brw_stage_prog_data
*prog_data
=
2036 pipeline
->shaders
[exe
->stage
]->prog_data
;
2038 vk_outarray_append(&out
, stat
) {
2039 WRITE_STR(stat
->name
, "Instruction Count");
2040 WRITE_STR(stat
->description
,
2041 "Number of GEN instructions in the final generated "
2042 "shader executable.");
2043 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2044 stat
->value
.u64
= exe
->stats
.instructions
;
2047 vk_outarray_append(&out
, stat
) {
2048 WRITE_STR(stat
->name
, "Loop Count");
2049 WRITE_STR(stat
->description
,
2050 "Number of loops (not unrolled) in the final generated "
2051 "shader executable.");
2052 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2053 stat
->value
.u64
= exe
->stats
.loops
;
2056 vk_outarray_append(&out
, stat
) {
2057 WRITE_STR(stat
->name
, "Cycle Count");
2058 WRITE_STR(stat
->description
,
2059 "Estimate of the number of EU cycles required to execute "
2060 "the final generated executable. This is an estimate only "
2061 "and may vary greatly from actual run-time performance.");
2062 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2063 stat
->value
.u64
= exe
->stats
.cycles
;
2066 vk_outarray_append(&out
, stat
) {
2067 WRITE_STR(stat
->name
, "Spill Count");
2068 WRITE_STR(stat
->description
,
2069 "Number of scratch spill operations. This gives a rough "
2070 "estimate of the cost incurred due to spilling temporary "
2071 "values to memory. If this is non-zero, you may want to "
2072 "adjust your shader to reduce register pressure.");
2073 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2074 stat
->value
.u64
= exe
->stats
.spills
;
2077 vk_outarray_append(&out
, stat
) {
2078 WRITE_STR(stat
->name
, "Fill Count");
2079 WRITE_STR(stat
->description
,
2080 "Number of scratch fill operations. This gives a rough "
2081 "estimate of the cost incurred due to spilling temporary "
2082 "values to memory. If this is non-zero, you may want to "
2083 "adjust your shader to reduce register pressure.");
2084 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2085 stat
->value
.u64
= exe
->stats
.fills
;
2088 vk_outarray_append(&out
, stat
) {
2089 WRITE_STR(stat
->name
, "Scratch Memory Size");
2090 WRITE_STR(stat
->description
,
2091 "Number of bytes of scratch memory required by the "
2092 "generated shader executable. If this is non-zero, you "
2093 "may want to adjust your shader to reduce register "
2095 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2096 stat
->value
.u64
= prog_data
->total_scratch
;
2099 if (exe
->stage
== MESA_SHADER_COMPUTE
) {
2100 vk_outarray_append(&out
, stat
) {
2101 WRITE_STR(stat
->name
, "Workgroup Memory Size");
2102 WRITE_STR(stat
->description
,
2103 "Number of bytes of workgroup shared memory used by this "
2104 "compute shader including any padding.");
2105 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2106 stat
->value
.u64
= prog_data
->total_scratch
;
2110 return vk_outarray_status(&out
);
2114 write_ir_text(VkPipelineExecutableInternalRepresentationKHR
* ir
,
2117 ir
->isText
= VK_TRUE
;
2119 size_t data_len
= strlen(data
) + 1;
2121 if (ir
->pData
== NULL
) {
2122 ir
->dataSize
= data_len
;
2126 strncpy(ir
->pData
, data
, ir
->dataSize
);
2127 if (ir
->dataSize
< data_len
)
2130 ir
->dataSize
= data_len
;
2134 VkResult
anv_GetPipelineExecutableInternalRepresentationsKHR(
2136 const VkPipelineExecutableInfoKHR
* pExecutableInfo
,
2137 uint32_t* pInternalRepresentationCount
,
2138 VkPipelineExecutableInternalRepresentationKHR
* pInternalRepresentations
)
2140 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, pExecutableInfo
->pipeline
);
2141 VK_OUTARRAY_MAKE(out
, pInternalRepresentations
,
2142 pInternalRepresentationCount
);
2143 bool incomplete_text
= false;
2145 assert(pExecutableInfo
->executableIndex
< pipeline
->num_executables
);
2146 const struct anv_pipeline_executable
*exe
=
2147 &pipeline
->executables
[pExecutableInfo
->executableIndex
];
2150 vk_outarray_append(&out
, ir
) {
2151 WRITE_STR(ir
->name
, "Final NIR");
2152 WRITE_STR(ir
->description
,
2153 "Final NIR before going into the back-end compiler");
2155 if (!write_ir_text(ir
, exe
->nir
))
2156 incomplete_text
= true;
2161 vk_outarray_append(&out
, ir
) {
2162 WRITE_STR(ir
->name
, "GEN Assembly");
2163 WRITE_STR(ir
->description
,
2164 "Final GEN assembly for the generated shader binary");
2166 if (!write_ir_text(ir
, exe
->disasm
))
2167 incomplete_text
= true;
2171 return incomplete_text
? VK_INCOMPLETE
: vk_outarray_status(&out
);