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 "common/gen_l3_config.h"
32 #include "anv_private.h"
33 #include "compiler/brw_nir.h"
35 #include "spirv/nir_spirv.h"
37 /* Needed for SWIZZLE macros */
38 #include "program/prog_instruction.h"
42 VkResult
anv_CreateShaderModule(
44 const VkShaderModuleCreateInfo
* pCreateInfo
,
45 const VkAllocationCallbacks
* pAllocator
,
46 VkShaderModule
* pShaderModule
)
48 ANV_FROM_HANDLE(anv_device
, device
, _device
);
49 struct anv_shader_module
*module
;
51 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
52 assert(pCreateInfo
->flags
== 0);
54 module
= vk_alloc2(&device
->alloc
, pAllocator
,
55 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
56 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
58 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
60 module
->size
= pCreateInfo
->codeSize
;
61 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
63 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
65 *pShaderModule
= anv_shader_module_to_handle(module
);
70 void anv_DestroyShaderModule(
72 VkShaderModule _module
,
73 const VkAllocationCallbacks
* pAllocator
)
75 ANV_FROM_HANDLE(anv_device
, device
, _device
);
76 ANV_FROM_HANDLE(anv_shader_module
, module
, _module
);
81 vk_free2(&device
->alloc
, pAllocator
, module
);
84 #define SPIR_V_MAGIC_NUMBER 0x07230203
86 static const uint64_t stage_to_debug
[] = {
87 [MESA_SHADER_VERTEX
] = DEBUG_VS
,
88 [MESA_SHADER_TESS_CTRL
] = DEBUG_TCS
,
89 [MESA_SHADER_TESS_EVAL
] = DEBUG_TES
,
90 [MESA_SHADER_GEOMETRY
] = DEBUG_GS
,
91 [MESA_SHADER_FRAGMENT
] = DEBUG_WM
,
92 [MESA_SHADER_COMPUTE
] = DEBUG_CS
,
95 /* Eventually, this will become part of anv_CreateShader. Unfortunately,
96 * we can't do that yet because we don't have the ability to copy nir.
99 anv_shader_compile_to_nir(struct anv_pipeline
*pipeline
,
101 struct anv_shader_module
*module
,
102 const char *entrypoint_name
,
103 gl_shader_stage stage
,
104 const VkSpecializationInfo
*spec_info
)
106 const struct anv_device
*device
= pipeline
->device
;
108 const struct brw_compiler
*compiler
=
109 device
->instance
->physicalDevice
.compiler
;
110 const nir_shader_compiler_options
*nir_options
=
111 compiler
->glsl_compiler_options
[stage
].NirOptions
;
113 uint32_t *spirv
= (uint32_t *) module
->data
;
114 assert(spirv
[0] == SPIR_V_MAGIC_NUMBER
);
115 assert(module
->size
% 4 == 0);
117 uint32_t num_spec_entries
= 0;
118 struct nir_spirv_specialization
*spec_entries
= NULL
;
119 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
120 num_spec_entries
= spec_info
->mapEntryCount
;
121 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
122 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
123 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
124 const void *data
= spec_info
->pData
+ entry
.offset
;
125 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
127 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
128 if (spec_info
->dataSize
== 8)
129 spec_entries
[i
].data64
= *(const uint64_t *)data
;
131 spec_entries
[i
].data32
= *(const uint32_t *)data
;
135 struct spirv_to_nir_options spirv_options
= {
136 .lower_workgroup_access_to_offsets
= true,
138 .float64
= device
->instance
->physicalDevice
.info
.gen
>= 8,
139 .int64
= device
->instance
->physicalDevice
.info
.gen
>= 8,
140 .tessellation
= true,
141 .device_group
= true,
142 .draw_parameters
= true,
143 .image_write_without_format
= true,
145 .variable_pointers
= true,
146 .storage_16bit
= device
->instance
->physicalDevice
.info
.gen
>= 8,
150 nir_function
*entry_point
=
151 spirv_to_nir(spirv
, module
->size
/ 4,
152 spec_entries
, num_spec_entries
,
153 stage
, entrypoint_name
, &spirv_options
, nir_options
);
154 nir_shader
*nir
= entry_point
->shader
;
155 assert(nir
->info
.stage
== stage
);
156 nir_validate_shader(nir
);
157 ralloc_steal(mem_ctx
, nir
);
161 if (unlikely(INTEL_DEBUG
& stage_to_debug
[stage
])) {
162 fprintf(stderr
, "NIR (from SPIR-V) for %s shader:\n",
163 gl_shader_stage_name(stage
));
164 nir_print_shader(nir
, stderr
);
167 /* We have to lower away local constant initializers right before we
168 * inline functions. That way they get properly initialized at the top
169 * of the function and not at the top of its caller.
171 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_local
);
172 NIR_PASS_V(nir
, nir_lower_returns
);
173 NIR_PASS_V(nir
, nir_inline_functions
);
175 /* Pick off the single entrypoint that we want */
176 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
177 if (func
!= entry_point
)
178 exec_node_remove(&func
->node
);
180 assert(exec_list_length(&nir
->functions
) == 1);
181 entry_point
->name
= ralloc_strdup(entry_point
, "main");
183 /* Make sure we lower constant initializers on output variables so that
184 * nir_remove_dead_variables below sees the corresponding stores
186 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_shader_out
);
188 NIR_PASS_V(nir
, nir_remove_dead_variables
,
189 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
191 if (stage
== MESA_SHADER_FRAGMENT
)
192 NIR_PASS_V(nir
, nir_lower_wpos_center
, pipeline
->sample_shading_enable
);
194 /* Now that we've deleted all but the main function, we can go ahead and
195 * lower the rest of the constant initializers.
197 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
198 NIR_PASS_V(nir
, nir_propagate_invariant
);
199 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
200 entry_point
->impl
, true, false);
202 /* Vulkan uses the separate-shader linking model */
203 nir
->info
.separate_shader
= true;
205 nir
= brw_preprocess_nir(compiler
, nir
);
207 if (stage
== MESA_SHADER_FRAGMENT
)
208 NIR_PASS_V(nir
, anv_nir_lower_input_attachments
);
213 void anv_DestroyPipeline(
215 VkPipeline _pipeline
,
216 const VkAllocationCallbacks
* pAllocator
)
218 ANV_FROM_HANDLE(anv_device
, device
, _device
);
219 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, _pipeline
);
224 anv_reloc_list_finish(&pipeline
->batch_relocs
,
225 pAllocator
? pAllocator
: &device
->alloc
);
226 if (pipeline
->blend_state
.map
)
227 anv_state_pool_free(&device
->dynamic_state_pool
, pipeline
->blend_state
);
229 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
230 if (pipeline
->shaders
[s
])
231 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
234 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
237 static const uint32_t vk_to_gen_primitive_type
[] = {
238 [VK_PRIMITIVE_TOPOLOGY_POINT_LIST
] = _3DPRIM_POINTLIST
,
239 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST
] = _3DPRIM_LINELIST
,
240 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
241 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
] = _3DPRIM_TRILIST
,
242 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
243 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
244 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
245 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
246 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
247 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
251 populate_sampler_prog_key(const struct gen_device_info
*devinfo
,
252 struct brw_sampler_prog_key_data
*key
)
254 /* Almost all multisampled textures are compressed. The only time when we
255 * don't compress a multisampled texture is for 16x MSAA with a surface
256 * width greater than 8k which is a bit of an edge case. Since the sampler
257 * just ignores the MCS parameter to ld2ms when MCS is disabled, it's safe
258 * to tell the compiler to always assume compression.
260 key
->compressed_multisample_layout_mask
= ~0;
262 /* SkyLake added support for 16x MSAA. With this came a new message for
263 * reading from a 16x MSAA surface with compression. The new message was
264 * needed because now the MCS data is 64 bits instead of 32 or lower as is
265 * the case for 8x, 4x, and 2x. The key->msaa_16 bit-field controls which
266 * message we use. Fortunately, the 16x message works for 8x, 4x, and 2x
267 * so we can just use it unconditionally. This may not be quite as
268 * efficient but it saves us from recompiling.
270 if (devinfo
->gen
>= 9)
273 /* XXX: Handle texture swizzle on HSW- */
274 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
275 /* Assume color sampler, no swizzling. (Works for BDW+) */
276 key
->swizzles
[i
] = SWIZZLE_XYZW
;
281 populate_vs_prog_key(const struct gen_device_info
*devinfo
,
282 struct brw_vs_prog_key
*key
)
284 memset(key
, 0, sizeof(*key
));
286 populate_sampler_prog_key(devinfo
, &key
->tex
);
288 /* XXX: Handle vertex input work-arounds */
290 /* XXX: Handle sampler_prog_key */
294 populate_gs_prog_key(const struct gen_device_info
*devinfo
,
295 struct brw_gs_prog_key
*key
)
297 memset(key
, 0, sizeof(*key
));
299 populate_sampler_prog_key(devinfo
, &key
->tex
);
303 populate_wm_prog_key(const struct anv_pipeline
*pipeline
,
304 const VkGraphicsPipelineCreateInfo
*info
,
305 struct brw_wm_prog_key
*key
)
307 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
309 memset(key
, 0, sizeof(*key
));
311 populate_sampler_prog_key(devinfo
, &key
->tex
);
313 /* TODO: we could set this to 0 based on the information in nir_shader, but
314 * this function is called before spirv_to_nir. */
315 const struct brw_vue_map
*vue_map
=
316 &anv_pipeline_get_last_vue_prog_data(pipeline
)->vue_map
;
317 key
->input_slots_valid
= vue_map
->slots_valid
;
319 /* Vulkan doesn't specify a default */
320 key
->high_quality_derivatives
= false;
322 /* XXX Vulkan doesn't appear to specify */
323 key
->clamp_fragment_color
= false;
325 key
->nr_color_regions
= pipeline
->subpass
->color_count
;
327 key
->replicate_alpha
= key
->nr_color_regions
> 1 &&
328 info
->pMultisampleState
&&
329 info
->pMultisampleState
->alphaToCoverageEnable
;
331 if (info
->pMultisampleState
) {
332 /* We should probably pull this out of the shader, but it's fairly
333 * harmless to compute it and then let dead-code take care of it.
335 if (info
->pMultisampleState
->rasterizationSamples
> 1) {
336 key
->persample_interp
=
337 (info
->pMultisampleState
->minSampleShading
*
338 info
->pMultisampleState
->rasterizationSamples
) > 1;
339 key
->multisample_fbo
= true;
342 key
->frag_coord_adds_sample_pos
=
343 info
->pMultisampleState
->sampleShadingEnable
;
348 populate_cs_prog_key(const struct gen_device_info
*devinfo
,
349 struct brw_cs_prog_key
*key
)
351 memset(key
, 0, sizeof(*key
));
353 populate_sampler_prog_key(devinfo
, &key
->tex
);
357 anv_pipeline_hash_shader(struct anv_pipeline
*pipeline
,
358 struct anv_pipeline_layout
*layout
,
359 struct anv_shader_module
*module
,
360 const char *entrypoint
,
361 gl_shader_stage stage
,
362 const VkSpecializationInfo
*spec_info
,
363 const void *key
, size_t key_size
,
364 unsigned char *sha1_out
)
366 struct mesa_sha1 ctx
;
368 _mesa_sha1_init(&ctx
);
369 if (stage
!= MESA_SHADER_COMPUTE
) {
370 _mesa_sha1_update(&ctx
, &pipeline
->subpass
->view_mask
,
371 sizeof(pipeline
->subpass
->view_mask
));
374 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
375 _mesa_sha1_update(&ctx
, module
->sha1
, sizeof(module
->sha1
));
376 _mesa_sha1_update(&ctx
, entrypoint
, strlen(entrypoint
));
377 _mesa_sha1_update(&ctx
, &stage
, sizeof(stage
));
379 _mesa_sha1_update(&ctx
, spec_info
->pMapEntries
,
380 spec_info
->mapEntryCount
* sizeof(*spec_info
->pMapEntries
));
381 _mesa_sha1_update(&ctx
, spec_info
->pData
, spec_info
->dataSize
);
383 _mesa_sha1_update(&ctx
, key
, key_size
);
384 _mesa_sha1_final(&ctx
, sha1_out
);
388 anv_pipeline_compile(struct anv_pipeline
*pipeline
,
390 struct anv_pipeline_layout
*layout
,
391 struct anv_shader_module
*module
,
392 const char *entrypoint
,
393 gl_shader_stage stage
,
394 const VkSpecializationInfo
*spec_info
,
395 struct brw_stage_prog_data
*prog_data
,
396 struct anv_pipeline_bind_map
*map
)
398 const struct brw_compiler
*compiler
=
399 pipeline
->device
->instance
->physicalDevice
.compiler
;
401 nir_shader
*nir
= anv_shader_compile_to_nir(pipeline
, mem_ctx
,
402 module
, entrypoint
, stage
,
407 NIR_PASS_V(nir
, anv_nir_lower_ycbcr_textures
, layout
);
409 NIR_PASS_V(nir
, anv_nir_lower_push_constants
);
411 if (stage
!= MESA_SHADER_COMPUTE
)
412 NIR_PASS_V(nir
, anv_nir_lower_multiview
, pipeline
->subpass
->view_mask
);
414 if (stage
== MESA_SHADER_COMPUTE
)
415 prog_data
->total_shared
= nir
->num_shared
;
417 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
419 if (nir
->num_uniforms
> 0) {
420 assert(prog_data
->nr_params
== 0);
422 /* If the shader uses any push constants at all, we'll just give
423 * them the maximum possible number
425 assert(nir
->num_uniforms
<= MAX_PUSH_CONSTANTS_SIZE
);
426 nir
->num_uniforms
= MAX_PUSH_CONSTANTS_SIZE
;
427 prog_data
->nr_params
+= MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float);
428 prog_data
->param
= ralloc_array(mem_ctx
, uint32_t, prog_data
->nr_params
);
430 /* We now set the param values to be offsets into a
431 * anv_push_constant_data structure. Since the compiler doesn't
432 * actually dereference any of the gl_constant_value pointers in the
433 * params array, it doesn't really matter what we put here.
435 struct anv_push_constants
*null_data
= NULL
;
436 /* Fill out the push constants section of the param array */
437 for (unsigned i
= 0; i
< MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float); i
++) {
438 prog_data
->param
[i
] = ANV_PARAM_PUSH(
439 (uintptr_t)&null_data
->client_data
[i
* sizeof(float)]);
443 if (nir
->info
.num_ssbos
> 0 || nir
->info
.num_images
> 0)
444 pipeline
->needs_data_cache
= true;
446 /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
448 anv_nir_apply_pipeline_layout(pipeline
, layout
, nir
, prog_data
, map
);
450 if (stage
!= MESA_SHADER_COMPUTE
)
451 brw_nir_analyze_ubo_ranges(compiler
, nir
, prog_data
->ubo_ranges
);
453 assert(nir
->num_uniforms
== prog_data
->nr_params
* 4);
459 anv_fill_binding_table(struct brw_stage_prog_data
*prog_data
, unsigned bias
)
461 prog_data
->binding_table
.size_bytes
= 0;
462 prog_data
->binding_table
.texture_start
= bias
;
463 prog_data
->binding_table
.gather_texture_start
= bias
;
464 prog_data
->binding_table
.ubo_start
= bias
;
465 prog_data
->binding_table
.ssbo_start
= bias
;
466 prog_data
->binding_table
.image_start
= bias
;
469 static struct anv_shader_bin
*
470 anv_pipeline_upload_kernel(struct anv_pipeline
*pipeline
,
471 struct anv_pipeline_cache
*cache
,
472 const void *key_data
, uint32_t key_size
,
473 const void *kernel_data
, uint32_t kernel_size
,
474 const struct brw_stage_prog_data
*prog_data
,
475 uint32_t prog_data_size
,
476 const struct anv_pipeline_bind_map
*bind_map
)
479 return anv_pipeline_cache_upload_kernel(cache
, key_data
, key_size
,
480 kernel_data
, kernel_size
,
481 prog_data
, prog_data_size
,
484 return anv_shader_bin_create(pipeline
->device
, key_data
, key_size
,
485 kernel_data
, kernel_size
,
486 prog_data
, prog_data_size
,
487 prog_data
->param
, bind_map
);
493 anv_pipeline_add_compiled_stage(struct anv_pipeline
*pipeline
,
494 gl_shader_stage stage
,
495 struct anv_shader_bin
*shader
)
497 pipeline
->shaders
[stage
] = shader
;
498 pipeline
->active_stages
|= mesa_to_vk_shader_stage(stage
);
502 anv_pipeline_compile_vs(struct anv_pipeline
*pipeline
,
503 struct anv_pipeline_cache
*cache
,
504 const VkGraphicsPipelineCreateInfo
*info
,
505 struct anv_shader_module
*module
,
506 const char *entrypoint
,
507 const VkSpecializationInfo
*spec_info
)
509 const struct brw_compiler
*compiler
=
510 pipeline
->device
->instance
->physicalDevice
.compiler
;
511 struct brw_vs_prog_key key
;
512 struct anv_shader_bin
*bin
= NULL
;
513 unsigned char sha1
[20];
515 populate_vs_prog_key(&pipeline
->device
->info
, &key
);
517 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
520 anv_pipeline_hash_shader(pipeline
, layout
, module
, entrypoint
,
521 MESA_SHADER_VERTEX
, spec_info
,
522 &key
, sizeof(key
), sha1
);
523 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
527 struct brw_vs_prog_data prog_data
= {};
528 struct anv_pipeline_binding surface_to_descriptor
[256];
529 struct anv_pipeline_binding sampler_to_descriptor
[256];
531 struct anv_pipeline_bind_map map
= {
532 .surface_to_descriptor
= surface_to_descriptor
,
533 .sampler_to_descriptor
= sampler_to_descriptor
536 void *mem_ctx
= ralloc_context(NULL
);
538 nir_shader
*nir
= anv_pipeline_compile(pipeline
, mem_ctx
, layout
,
540 MESA_SHADER_VERTEX
, spec_info
,
541 &prog_data
.base
.base
, &map
);
543 ralloc_free(mem_ctx
);
544 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
547 anv_fill_binding_table(&prog_data
.base
.base
, 0);
549 brw_compute_vue_map(&pipeline
->device
->info
,
550 &prog_data
.base
.vue_map
,
551 nir
->info
.outputs_written
,
552 nir
->info
.separate_shader
);
554 const unsigned *shader_code
=
555 brw_compile_vs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
557 if (shader_code
== NULL
) {
558 ralloc_free(mem_ctx
);
559 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
562 unsigned code_size
= prog_data
.base
.base
.program_size
;
563 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
564 shader_code
, code_size
,
565 &prog_data
.base
.base
, sizeof(prog_data
),
568 ralloc_free(mem_ctx
);
569 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
572 ralloc_free(mem_ctx
);
575 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_VERTEX
, bin
);
581 merge_tess_info(struct shader_info
*tes_info
,
582 const struct shader_info
*tcs_info
)
584 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
586 * "PointMode. Controls generation of points rather than triangles
587 * or lines. This functionality defaults to disabled, and is
588 * enabled if either shader stage includes the execution mode.
590 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
591 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
592 * and OutputVertices, it says:
594 * "One mode must be set in at least one of the tessellation
597 * So, the fields can be set in either the TCS or TES, but they must
598 * agree if set in both. Our backend looks at TES, so bitwise-or in
599 * the values from the TCS.
601 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
602 tes_info
->tess
.tcs_vertices_out
== 0 ||
603 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
604 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
606 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
607 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
608 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
609 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
611 assert(tcs_info
->tess
.primitive_mode
== 0 ||
612 tes_info
->tess
.primitive_mode
== 0 ||
613 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
614 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
615 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
616 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
620 anv_pipeline_compile_tcs_tes(struct anv_pipeline
*pipeline
,
621 struct anv_pipeline_cache
*cache
,
622 const VkGraphicsPipelineCreateInfo
*info
,
623 struct anv_shader_module
*tcs_module
,
624 const char *tcs_entrypoint
,
625 const VkSpecializationInfo
*tcs_spec_info
,
626 struct anv_shader_module
*tes_module
,
627 const char *tes_entrypoint
,
628 const VkSpecializationInfo
*tes_spec_info
)
630 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
631 const struct brw_compiler
*compiler
=
632 pipeline
->device
->instance
->physicalDevice
.compiler
;
633 struct brw_tcs_prog_key tcs_key
= {};
634 struct brw_tes_prog_key tes_key
= {};
635 struct anv_shader_bin
*tcs_bin
= NULL
;
636 struct anv_shader_bin
*tes_bin
= NULL
;
637 unsigned char tcs_sha1
[40];
638 unsigned char tes_sha1
[40];
640 populate_sampler_prog_key(&pipeline
->device
->info
, &tcs_key
.tex
);
641 populate_sampler_prog_key(&pipeline
->device
->info
, &tes_key
.tex
);
642 tcs_key
.input_vertices
= info
->pTessellationState
->patchControlPoints
;
644 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
647 anv_pipeline_hash_shader(pipeline
, layout
, tcs_module
, tcs_entrypoint
,
648 MESA_SHADER_TESS_CTRL
, tcs_spec_info
,
649 &tcs_key
, sizeof(tcs_key
), tcs_sha1
);
650 anv_pipeline_hash_shader(pipeline
, layout
, tes_module
, tes_entrypoint
,
651 MESA_SHADER_TESS_EVAL
, tes_spec_info
,
652 &tes_key
, sizeof(tes_key
), tes_sha1
);
653 memcpy(&tcs_sha1
[20], tes_sha1
, 20);
654 memcpy(&tes_sha1
[20], tcs_sha1
, 20);
655 tcs_bin
= anv_pipeline_cache_search(cache
, tcs_sha1
, sizeof(tcs_sha1
));
656 tes_bin
= anv_pipeline_cache_search(cache
, tes_sha1
, sizeof(tes_sha1
));
659 if (tcs_bin
== NULL
|| tes_bin
== NULL
) {
660 struct brw_tcs_prog_data tcs_prog_data
= {};
661 struct brw_tes_prog_data tes_prog_data
= {};
662 struct anv_pipeline_binding tcs_surface_to_descriptor
[256];
663 struct anv_pipeline_binding tcs_sampler_to_descriptor
[256];
664 struct anv_pipeline_binding tes_surface_to_descriptor
[256];
665 struct anv_pipeline_binding tes_sampler_to_descriptor
[256];
667 struct anv_pipeline_bind_map tcs_map
= {
668 .surface_to_descriptor
= tcs_surface_to_descriptor
,
669 .sampler_to_descriptor
= tcs_sampler_to_descriptor
671 struct anv_pipeline_bind_map tes_map
= {
672 .surface_to_descriptor
= tes_surface_to_descriptor
,
673 .sampler_to_descriptor
= tes_sampler_to_descriptor
676 void *mem_ctx
= ralloc_context(NULL
);
678 nir_shader
*tcs_nir
=
679 anv_pipeline_compile(pipeline
, mem_ctx
, layout
,
680 tcs_module
, tcs_entrypoint
,
681 MESA_SHADER_TESS_CTRL
, tcs_spec_info
,
682 &tcs_prog_data
.base
.base
, &tcs_map
);
683 nir_shader
*tes_nir
=
684 anv_pipeline_compile(pipeline
, mem_ctx
, layout
,
685 tes_module
, tes_entrypoint
,
686 MESA_SHADER_TESS_EVAL
, tes_spec_info
,
687 &tes_prog_data
.base
.base
, &tes_map
);
688 if (tcs_nir
== NULL
|| tes_nir
== NULL
) {
689 ralloc_free(mem_ctx
);
690 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
693 nir_lower_tes_patch_vertices(tes_nir
,
694 tcs_nir
->info
.tess
.tcs_vertices_out
);
696 /* Copy TCS info into the TES info */
697 merge_tess_info(&tes_nir
->info
, &tcs_nir
->info
);
699 anv_fill_binding_table(&tcs_prog_data
.base
.base
, 0);
700 anv_fill_binding_table(&tes_prog_data
.base
.base
, 0);
702 /* Whacking the key after cache lookup is a bit sketchy, but all of
703 * this comes from the SPIR-V, which is part of the hash used for the
704 * pipeline cache. So it should be safe.
706 tcs_key
.tes_primitive_mode
= tes_nir
->info
.tess
.primitive_mode
;
707 tcs_key
.outputs_written
= tcs_nir
->info
.outputs_written
;
708 tcs_key
.patch_outputs_written
= tcs_nir
->info
.patch_outputs_written
;
709 tcs_key
.quads_workaround
=
711 tes_nir
->info
.tess
.primitive_mode
== 7 /* GL_QUADS */ &&
712 tes_nir
->info
.tess
.spacing
== TESS_SPACING_EQUAL
;
714 tes_key
.inputs_read
= tcs_key
.outputs_written
;
715 tes_key
.patch_inputs_read
= tcs_key
.patch_outputs_written
;
717 const int shader_time_index
= -1;
718 const unsigned *shader_code
;
721 brw_compile_tcs(compiler
, NULL
, mem_ctx
, &tcs_key
, &tcs_prog_data
,
722 tcs_nir
, shader_time_index
, NULL
);
723 if (shader_code
== NULL
) {
724 ralloc_free(mem_ctx
);
725 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
728 unsigned code_size
= tcs_prog_data
.base
.base
.program_size
;
729 tcs_bin
= anv_pipeline_upload_kernel(pipeline
, cache
,
730 tcs_sha1
, sizeof(tcs_sha1
),
731 shader_code
, code_size
,
732 &tcs_prog_data
.base
.base
,
733 sizeof(tcs_prog_data
),
736 ralloc_free(mem_ctx
);
737 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
741 brw_compile_tes(compiler
, NULL
, mem_ctx
, &tes_key
,
742 &tcs_prog_data
.base
.vue_map
, &tes_prog_data
, tes_nir
,
743 NULL
, shader_time_index
, NULL
);
744 if (shader_code
== NULL
) {
745 ralloc_free(mem_ctx
);
746 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
749 code_size
= tes_prog_data
.base
.base
.program_size
;
750 tes_bin
= anv_pipeline_upload_kernel(pipeline
, cache
,
751 tes_sha1
, sizeof(tes_sha1
),
752 shader_code
, code_size
,
753 &tes_prog_data
.base
.base
,
754 sizeof(tes_prog_data
),
757 ralloc_free(mem_ctx
);
758 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
761 ralloc_free(mem_ctx
);
764 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_TESS_CTRL
, tcs_bin
);
765 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_TESS_EVAL
, tes_bin
);
771 anv_pipeline_compile_gs(struct anv_pipeline
*pipeline
,
772 struct anv_pipeline_cache
*cache
,
773 const VkGraphicsPipelineCreateInfo
*info
,
774 struct anv_shader_module
*module
,
775 const char *entrypoint
,
776 const VkSpecializationInfo
*spec_info
)
778 const struct brw_compiler
*compiler
=
779 pipeline
->device
->instance
->physicalDevice
.compiler
;
780 struct brw_gs_prog_key key
;
781 struct anv_shader_bin
*bin
= NULL
;
782 unsigned char sha1
[20];
784 populate_gs_prog_key(&pipeline
->device
->info
, &key
);
786 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
789 anv_pipeline_hash_shader(pipeline
, layout
, module
, entrypoint
,
790 MESA_SHADER_GEOMETRY
, spec_info
,
791 &key
, sizeof(key
), sha1
);
792 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
796 struct brw_gs_prog_data prog_data
= {};
797 struct anv_pipeline_binding surface_to_descriptor
[256];
798 struct anv_pipeline_binding sampler_to_descriptor
[256];
800 struct anv_pipeline_bind_map map
= {
801 .surface_to_descriptor
= surface_to_descriptor
,
802 .sampler_to_descriptor
= sampler_to_descriptor
805 void *mem_ctx
= ralloc_context(NULL
);
807 nir_shader
*nir
= anv_pipeline_compile(pipeline
, mem_ctx
, layout
,
809 MESA_SHADER_GEOMETRY
, spec_info
,
810 &prog_data
.base
.base
, &map
);
812 ralloc_free(mem_ctx
);
813 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
816 anv_fill_binding_table(&prog_data
.base
.base
, 0);
818 brw_compute_vue_map(&pipeline
->device
->info
,
819 &prog_data
.base
.vue_map
,
820 nir
->info
.outputs_written
,
821 nir
->info
.separate_shader
);
823 const unsigned *shader_code
=
824 brw_compile_gs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
826 if (shader_code
== NULL
) {
827 ralloc_free(mem_ctx
);
828 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
832 const unsigned code_size
= prog_data
.base
.base
.program_size
;
833 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
834 shader_code
, code_size
,
835 &prog_data
.base
.base
, sizeof(prog_data
),
838 ralloc_free(mem_ctx
);
839 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
842 ralloc_free(mem_ctx
);
845 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_GEOMETRY
, bin
);
851 anv_pipeline_compile_fs(struct anv_pipeline
*pipeline
,
852 struct anv_pipeline_cache
*cache
,
853 const VkGraphicsPipelineCreateInfo
*info
,
854 struct anv_shader_module
*module
,
855 const char *entrypoint
,
856 const VkSpecializationInfo
*spec_info
)
858 const struct brw_compiler
*compiler
=
859 pipeline
->device
->instance
->physicalDevice
.compiler
;
860 struct brw_wm_prog_key key
;
861 struct anv_shader_bin
*bin
= NULL
;
862 unsigned char sha1
[20];
864 populate_wm_prog_key(pipeline
, info
, &key
);
866 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
869 anv_pipeline_hash_shader(pipeline
, layout
, module
, entrypoint
,
870 MESA_SHADER_FRAGMENT
, spec_info
,
871 &key
, sizeof(key
), sha1
);
872 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
876 struct brw_wm_prog_data prog_data
= {};
877 struct anv_pipeline_binding surface_to_descriptor
[256];
878 struct anv_pipeline_binding sampler_to_descriptor
[256];
880 struct anv_pipeline_bind_map map
= {
881 .surface_to_descriptor
= surface_to_descriptor
+ 8,
882 .sampler_to_descriptor
= sampler_to_descriptor
885 void *mem_ctx
= ralloc_context(NULL
);
887 nir_shader
*nir
= anv_pipeline_compile(pipeline
, mem_ctx
, layout
,
889 MESA_SHADER_FRAGMENT
, spec_info
,
890 &prog_data
.base
, &map
);
892 ralloc_free(mem_ctx
);
893 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
896 unsigned num_rts
= 0;
897 const int max_rt
= FRAG_RESULT_DATA7
- FRAG_RESULT_DATA0
+ 1;
898 struct anv_pipeline_binding rt_bindings
[max_rt
];
899 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
900 int rt_to_bindings
[max_rt
];
901 memset(rt_to_bindings
, -1, sizeof(rt_to_bindings
));
902 bool rt_used
[max_rt
];
903 memset(rt_used
, 0, sizeof(rt_used
));
905 /* Flag used render targets */
906 nir_foreach_variable_safe(var
, &nir
->outputs
) {
907 if (var
->data
.location
< FRAG_RESULT_DATA0
)
910 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
912 if (rt
>= key
.nr_color_regions
)
915 const unsigned array_len
=
916 glsl_type_is_array(var
->type
) ? glsl_get_length(var
->type
) : 1;
917 assert(rt
+ array_len
<= max_rt
);
919 for (unsigned i
= 0; i
< array_len
; i
++)
920 rt_used
[rt
+ i
] = true;
923 /* Set new, compacted, location */
924 for (unsigned i
= 0; i
< max_rt
; i
++) {
928 rt_to_bindings
[i
] = num_rts
;
929 rt_bindings
[rt_to_bindings
[i
]] = (struct anv_pipeline_binding
) {
930 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
937 nir_foreach_variable_safe(var
, &nir
->outputs
) {
938 if (var
->data
.location
< FRAG_RESULT_DATA0
)
941 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
942 if (rt
>= key
.nr_color_regions
) {
943 /* Out-of-bounds, throw it away */
944 var
->data
.mode
= nir_var_local
;
945 exec_node_remove(&var
->node
);
946 exec_list_push_tail(&impl
->locals
, &var
->node
);
950 /* Give it the new location */
951 assert(rt_to_bindings
[rt
] != -1);
952 var
->data
.location
= rt_to_bindings
[rt
] + FRAG_RESULT_DATA0
;
956 /* If we have no render targets, we need a null render target */
957 rt_bindings
[0] = (struct anv_pipeline_binding
) {
958 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
965 assert(num_rts
<= max_rt
);
966 map
.surface_to_descriptor
-= num_rts
;
967 map
.surface_count
+= num_rts
;
968 assert(map
.surface_count
<= 256);
969 memcpy(map
.surface_to_descriptor
, rt_bindings
,
970 num_rts
* sizeof(*rt_bindings
));
972 anv_fill_binding_table(&prog_data
.base
, num_rts
);
974 const unsigned *shader_code
=
975 brw_compile_fs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
976 NULL
, -1, -1, true, false, NULL
, NULL
);
977 if (shader_code
== NULL
) {
978 ralloc_free(mem_ctx
);
979 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
982 unsigned code_size
= prog_data
.base
.program_size
;
983 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
984 shader_code
, code_size
,
985 &prog_data
.base
, sizeof(prog_data
),
988 ralloc_free(mem_ctx
);
989 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
992 ralloc_free(mem_ctx
);
995 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_FRAGMENT
, bin
);
1001 anv_pipeline_compile_cs(struct anv_pipeline
*pipeline
,
1002 struct anv_pipeline_cache
*cache
,
1003 const VkComputePipelineCreateInfo
*info
,
1004 struct anv_shader_module
*module
,
1005 const char *entrypoint
,
1006 const VkSpecializationInfo
*spec_info
)
1008 const struct brw_compiler
*compiler
=
1009 pipeline
->device
->instance
->physicalDevice
.compiler
;
1010 struct brw_cs_prog_key key
;
1011 struct anv_shader_bin
*bin
= NULL
;
1012 unsigned char sha1
[20];
1014 populate_cs_prog_key(&pipeline
->device
->info
, &key
);
1016 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
1019 anv_pipeline_hash_shader(pipeline
, layout
, module
, entrypoint
,
1020 MESA_SHADER_COMPUTE
, spec_info
,
1021 &key
, sizeof(key
), sha1
);
1022 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
1026 struct brw_cs_prog_data prog_data
= {};
1027 struct anv_pipeline_binding surface_to_descriptor
[256];
1028 struct anv_pipeline_binding sampler_to_descriptor
[256];
1030 struct anv_pipeline_bind_map map
= {
1031 .surface_to_descriptor
= surface_to_descriptor
,
1032 .sampler_to_descriptor
= sampler_to_descriptor
1035 void *mem_ctx
= ralloc_context(NULL
);
1037 nir_shader
*nir
= anv_pipeline_compile(pipeline
, mem_ctx
, layout
,
1039 MESA_SHADER_COMPUTE
, spec_info
,
1040 &prog_data
.base
, &map
);
1042 ralloc_free(mem_ctx
);
1043 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1046 NIR_PASS_V(nir
, anv_nir_add_base_work_group_id
, &prog_data
);
1048 anv_fill_binding_table(&prog_data
.base
, 1);
1050 const unsigned *shader_code
=
1051 brw_compile_cs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
1053 if (shader_code
== NULL
) {
1054 ralloc_free(mem_ctx
);
1055 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1058 const unsigned code_size
= prog_data
.base
.program_size
;
1059 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
1060 shader_code
, code_size
,
1061 &prog_data
.base
, sizeof(prog_data
),
1064 ralloc_free(mem_ctx
);
1065 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1068 ralloc_free(mem_ctx
);
1071 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_COMPUTE
, bin
);
1077 * Copy pipeline state not marked as dynamic.
1078 * Dynamic state is pipeline state which hasn't been provided at pipeline
1079 * creation time, but is dynamically provided afterwards using various
1080 * vkCmdSet* functions.
1082 * The set of state considered "non_dynamic" is determined by the pieces of
1083 * state that have their corresponding VkDynamicState enums omitted from
1084 * VkPipelineDynamicStateCreateInfo::pDynamicStates.
1086 * @param[out] pipeline Destination non_dynamic state.
1087 * @param[in] pCreateInfo Source of non_dynamic state to be copied.
1090 copy_non_dynamic_state(struct anv_pipeline
*pipeline
,
1091 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1093 anv_cmd_dirty_mask_t states
= ANV_CMD_DIRTY_DYNAMIC_ALL
;
1094 struct anv_subpass
*subpass
= pipeline
->subpass
;
1096 pipeline
->dynamic_state
= default_dynamic_state
;
1098 if (pCreateInfo
->pDynamicState
) {
1099 /* Remove all of the states that are marked as dynamic */
1100 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1101 for (uint32_t s
= 0; s
< count
; s
++)
1102 states
&= ~(1 << pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1105 struct anv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1107 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1109 * pViewportState is [...] NULL if the pipeline
1110 * has rasterization disabled.
1112 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1113 assert(pCreateInfo
->pViewportState
);
1115 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1116 if (states
& (1 << VK_DYNAMIC_STATE_VIEWPORT
)) {
1117 typed_memcpy(dynamic
->viewport
.viewports
,
1118 pCreateInfo
->pViewportState
->pViewports
,
1119 pCreateInfo
->pViewportState
->viewportCount
);
1122 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1123 if (states
& (1 << VK_DYNAMIC_STATE_SCISSOR
)) {
1124 typed_memcpy(dynamic
->scissor
.scissors
,
1125 pCreateInfo
->pViewportState
->pScissors
,
1126 pCreateInfo
->pViewportState
->scissorCount
);
1130 if (states
& (1 << VK_DYNAMIC_STATE_LINE_WIDTH
)) {
1131 assert(pCreateInfo
->pRasterizationState
);
1132 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1135 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BIAS
)) {
1136 assert(pCreateInfo
->pRasterizationState
);
1137 dynamic
->depth_bias
.bias
=
1138 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1139 dynamic
->depth_bias
.clamp
=
1140 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1141 dynamic
->depth_bias
.slope
=
1142 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1145 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1147 * pColorBlendState is [...] NULL if the pipeline has rasterization
1148 * disabled or if the subpass of the render pass the pipeline is
1149 * created against does not use any color attachments.
1151 bool uses_color_att
= false;
1152 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1153 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1154 uses_color_att
= true;
1159 if (uses_color_att
&&
1160 !pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1161 assert(pCreateInfo
->pColorBlendState
);
1163 if (states
& (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS
))
1164 typed_memcpy(dynamic
->blend_constants
,
1165 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1168 /* If there is no depthstencil attachment, then don't read
1169 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1170 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1171 * no need to override the depthstencil defaults in
1172 * anv_pipeline::dynamic_state when there is no depthstencil attachment.
1174 * Section 9.2 of the Vulkan 1.0.15 spec says:
1176 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1177 * disabled or if the subpass of the render pass the pipeline is created
1178 * against does not use a depth/stencil attachment.
1180 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1181 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1182 assert(pCreateInfo
->pDepthStencilState
);
1184 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS
)) {
1185 dynamic
->depth_bounds
.min
=
1186 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1187 dynamic
->depth_bounds
.max
=
1188 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1191 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
)) {
1192 dynamic
->stencil_compare_mask
.front
=
1193 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1194 dynamic
->stencil_compare_mask
.back
=
1195 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1198 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
)) {
1199 dynamic
->stencil_write_mask
.front
=
1200 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1201 dynamic
->stencil_write_mask
.back
=
1202 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1205 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE
)) {
1206 dynamic
->stencil_reference
.front
=
1207 pCreateInfo
->pDepthStencilState
->front
.reference
;
1208 dynamic
->stencil_reference
.back
=
1209 pCreateInfo
->pDepthStencilState
->back
.reference
;
1213 pipeline
->dynamic_state_mask
= states
;
1217 anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo
*info
)
1220 struct anv_render_pass
*renderpass
= NULL
;
1221 struct anv_subpass
*subpass
= NULL
;
1223 /* Assert that all required members of VkGraphicsPipelineCreateInfo are
1224 * present. See the Vulkan 1.0.28 spec, Section 9.2 Graphics Pipelines.
1226 assert(info
->sType
== VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
);
1228 renderpass
= anv_render_pass_from_handle(info
->renderPass
);
1231 assert(info
->subpass
< renderpass
->subpass_count
);
1232 subpass
= &renderpass
->subpasses
[info
->subpass
];
1234 assert(info
->stageCount
>= 1);
1235 assert(info
->pVertexInputState
);
1236 assert(info
->pInputAssemblyState
);
1237 assert(info
->pRasterizationState
);
1238 if (!info
->pRasterizationState
->rasterizerDiscardEnable
) {
1239 assert(info
->pViewportState
);
1240 assert(info
->pMultisampleState
);
1242 if (subpass
&& subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
)
1243 assert(info
->pDepthStencilState
);
1245 if (subpass
&& subpass
->color_count
> 0)
1246 assert(info
->pColorBlendState
);
1249 for (uint32_t i
= 0; i
< info
->stageCount
; ++i
) {
1250 switch (info
->pStages
[i
].stage
) {
1251 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
:
1252 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
:
1253 assert(info
->pTessellationState
);
1263 * Calculate the desired L3 partitioning based on the current state of the
1264 * pipeline. For now this simply returns the conservative defaults calculated
1265 * by get_default_l3_weights(), but we could probably do better by gathering
1266 * more statistics from the pipeline state (e.g. guess of expected URB usage
1267 * and bound surfaces), or by using feed-back from performance counters.
1270 anv_pipeline_setup_l3_config(struct anv_pipeline
*pipeline
, bool needs_slm
)
1272 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1274 const struct gen_l3_weights w
=
1275 gen_get_default_l3_weights(devinfo
, pipeline
->needs_data_cache
, needs_slm
);
1277 pipeline
->urb
.l3_config
= gen_get_l3_config(devinfo
, w
);
1278 pipeline
->urb
.total_size
=
1279 gen_get_l3_config_urb_size(devinfo
, pipeline
->urb
.l3_config
);
1283 anv_pipeline_init(struct anv_pipeline
*pipeline
,
1284 struct anv_device
*device
,
1285 struct anv_pipeline_cache
*cache
,
1286 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1287 const VkAllocationCallbacks
*alloc
)
1291 anv_pipeline_validate_create_info(pCreateInfo
);
1294 alloc
= &device
->alloc
;
1296 pipeline
->device
= device
;
1298 ANV_FROM_HANDLE(anv_render_pass
, render_pass
, pCreateInfo
->renderPass
);
1299 assert(pCreateInfo
->subpass
< render_pass
->subpass_count
);
1300 pipeline
->subpass
= &render_pass
->subpasses
[pCreateInfo
->subpass
];
1302 result
= anv_reloc_list_init(&pipeline
->batch_relocs
, alloc
);
1303 if (result
!= VK_SUCCESS
)
1306 pipeline
->batch
.alloc
= alloc
;
1307 pipeline
->batch
.next
= pipeline
->batch
.start
= pipeline
->batch_data
;
1308 pipeline
->batch
.end
= pipeline
->batch
.start
+ sizeof(pipeline
->batch_data
);
1309 pipeline
->batch
.relocs
= &pipeline
->batch_relocs
;
1310 pipeline
->batch
.status
= VK_SUCCESS
;
1312 copy_non_dynamic_state(pipeline
, pCreateInfo
);
1313 pipeline
->depth_clamp_enable
= pCreateInfo
->pRasterizationState
&&
1314 pCreateInfo
->pRasterizationState
->depthClampEnable
;
1316 pipeline
->sample_shading_enable
= pCreateInfo
->pMultisampleState
&&
1317 pCreateInfo
->pMultisampleState
->sampleShadingEnable
;
1319 pipeline
->needs_data_cache
= false;
1321 /* When we free the pipeline, we detect stages based on the NULL status
1322 * of various prog_data pointers. Make them NULL by default.
1324 memset(pipeline
->shaders
, 0, sizeof(pipeline
->shaders
));
1326 pipeline
->active_stages
= 0;
1328 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = {};
1329 struct anv_shader_module
*modules
[MESA_SHADER_STAGES
] = {};
1330 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
1331 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
1332 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
1333 modules
[stage
] = anv_shader_module_from_handle(pStages
[stage
]->module
);
1336 if (modules
[MESA_SHADER_VERTEX
]) {
1337 result
= anv_pipeline_compile_vs(pipeline
, cache
, pCreateInfo
,
1338 modules
[MESA_SHADER_VERTEX
],
1339 pStages
[MESA_SHADER_VERTEX
]->pName
,
1340 pStages
[MESA_SHADER_VERTEX
]->pSpecializationInfo
);
1341 if (result
!= VK_SUCCESS
)
1345 if (modules
[MESA_SHADER_TESS_EVAL
]) {
1346 anv_pipeline_compile_tcs_tes(pipeline
, cache
, pCreateInfo
,
1347 modules
[MESA_SHADER_TESS_CTRL
],
1348 pStages
[MESA_SHADER_TESS_CTRL
]->pName
,
1349 pStages
[MESA_SHADER_TESS_CTRL
]->pSpecializationInfo
,
1350 modules
[MESA_SHADER_TESS_EVAL
],
1351 pStages
[MESA_SHADER_TESS_EVAL
]->pName
,
1352 pStages
[MESA_SHADER_TESS_EVAL
]->pSpecializationInfo
);
1355 if (modules
[MESA_SHADER_GEOMETRY
]) {
1356 result
= anv_pipeline_compile_gs(pipeline
, cache
, pCreateInfo
,
1357 modules
[MESA_SHADER_GEOMETRY
],
1358 pStages
[MESA_SHADER_GEOMETRY
]->pName
,
1359 pStages
[MESA_SHADER_GEOMETRY
]->pSpecializationInfo
);
1360 if (result
!= VK_SUCCESS
)
1364 if (modules
[MESA_SHADER_FRAGMENT
]) {
1365 result
= anv_pipeline_compile_fs(pipeline
, cache
, pCreateInfo
,
1366 modules
[MESA_SHADER_FRAGMENT
],
1367 pStages
[MESA_SHADER_FRAGMENT
]->pName
,
1368 pStages
[MESA_SHADER_FRAGMENT
]->pSpecializationInfo
);
1369 if (result
!= VK_SUCCESS
)
1373 assert(pipeline
->active_stages
& VK_SHADER_STAGE_VERTEX_BIT
);
1375 anv_pipeline_setup_l3_config(pipeline
, false);
1377 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1378 pCreateInfo
->pVertexInputState
;
1380 const uint64_t inputs_read
= get_vs_prog_data(pipeline
)->inputs_read
;
1382 pipeline
->vb_used
= 0;
1383 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1384 const VkVertexInputAttributeDescription
*desc
=
1385 &vi_info
->pVertexAttributeDescriptions
[i
];
1387 if (inputs_read
& (1ull << (VERT_ATTRIB_GENERIC0
+ desc
->location
)))
1388 pipeline
->vb_used
|= 1 << desc
->binding
;
1391 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1392 const VkVertexInputBindingDescription
*desc
=
1393 &vi_info
->pVertexBindingDescriptions
[i
];
1395 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
1397 /* Step rate is programmed per vertex element (attribute), not
1398 * binding. Set up a map of which bindings step per instance, for
1399 * reference by vertex element setup. */
1400 switch (desc
->inputRate
) {
1402 case VK_VERTEX_INPUT_RATE_VERTEX
:
1403 pipeline
->instancing_enable
[desc
->binding
] = false;
1405 case VK_VERTEX_INPUT_RATE_INSTANCE
:
1406 pipeline
->instancing_enable
[desc
->binding
] = true;
1411 const VkPipelineInputAssemblyStateCreateInfo
*ia_info
=
1412 pCreateInfo
->pInputAssemblyState
;
1413 const VkPipelineTessellationStateCreateInfo
*tess_info
=
1414 pCreateInfo
->pTessellationState
;
1415 pipeline
->primitive_restart
= ia_info
->primitiveRestartEnable
;
1417 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
))
1418 pipeline
->topology
= _3DPRIM_PATCHLIST(tess_info
->patchControlPoints
);
1420 pipeline
->topology
= vk_to_gen_primitive_type
[ia_info
->topology
];
1425 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1426 if (pipeline
->shaders
[s
])
1427 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
1430 anv_reloc_list_finish(&pipeline
->batch_relocs
, alloc
);