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 "nir/nir_xfb_info.h"
36 #include "spirv/nir_spirv.h"
39 /* Needed for SWIZZLE macros */
40 #include "program/prog_instruction.h"
44 VkResult
anv_CreateShaderModule(
46 const VkShaderModuleCreateInfo
* pCreateInfo
,
47 const VkAllocationCallbacks
* pAllocator
,
48 VkShaderModule
* pShaderModule
)
50 ANV_FROM_HANDLE(anv_device
, device
, _device
);
51 struct anv_shader_module
*module
;
53 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
54 assert(pCreateInfo
->flags
== 0);
56 module
= vk_alloc2(&device
->alloc
, pAllocator
,
57 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
58 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
60 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
62 module
->size
= pCreateInfo
->codeSize
;
63 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
65 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
67 *pShaderModule
= anv_shader_module_to_handle(module
);
72 void anv_DestroyShaderModule(
74 VkShaderModule _module
,
75 const VkAllocationCallbacks
* pAllocator
)
77 ANV_FROM_HANDLE(anv_device
, device
, _device
);
78 ANV_FROM_HANDLE(anv_shader_module
, module
, _module
);
83 vk_free2(&device
->alloc
, pAllocator
, module
);
86 #define SPIR_V_MAGIC_NUMBER 0x07230203
88 static const uint64_t stage_to_debug
[] = {
89 [MESA_SHADER_VERTEX
] = DEBUG_VS
,
90 [MESA_SHADER_TESS_CTRL
] = DEBUG_TCS
,
91 [MESA_SHADER_TESS_EVAL
] = DEBUG_TES
,
92 [MESA_SHADER_GEOMETRY
] = DEBUG_GS
,
93 [MESA_SHADER_FRAGMENT
] = DEBUG_WM
,
94 [MESA_SHADER_COMPUTE
] = DEBUG_CS
,
97 /* Eventually, this will become part of anv_CreateShader. Unfortunately,
98 * we can't do that yet because we don't have the ability to copy nir.
101 anv_shader_compile_to_nir(struct anv_device
*device
,
103 const struct anv_shader_module
*module
,
104 const char *entrypoint_name
,
105 gl_shader_stage stage
,
106 const VkSpecializationInfo
*spec_info
)
108 const struct anv_physical_device
*pdevice
=
109 &device
->instance
->physicalDevice
;
110 const struct brw_compiler
*compiler
= pdevice
->compiler
;
111 const nir_shader_compiler_options
*nir_options
=
112 compiler
->glsl_compiler_options
[stage
].NirOptions
;
114 uint32_t *spirv
= (uint32_t *) module
->data
;
115 assert(spirv
[0] == SPIR_V_MAGIC_NUMBER
);
116 assert(module
->size
% 4 == 0);
118 uint32_t num_spec_entries
= 0;
119 struct nir_spirv_specialization
*spec_entries
= NULL
;
120 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
121 num_spec_entries
= spec_info
->mapEntryCount
;
122 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
123 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
124 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
125 const void *data
= spec_info
->pData
+ entry
.offset
;
126 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
128 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
129 if (spec_info
->dataSize
== 8)
130 spec_entries
[i
].data64
= *(const uint64_t *)data
;
132 spec_entries
[i
].data32
= *(const uint32_t *)data
;
136 struct spirv_to_nir_options spirv_options
= {
137 .lower_workgroup_access_to_offsets
= true,
139 .device_group
= true,
140 .draw_parameters
= true,
141 .float64
= pdevice
->info
.gen
>= 8,
142 .geometry_streams
= true,
143 .image_write_without_format
= true,
144 .int16
= pdevice
->info
.gen
>= 8,
145 .int64
= pdevice
->info
.gen
>= 8,
148 .physical_storage_buffer_address
= pdevice
->info
.gen
>= 8 &&
149 pdevice
->use_softpin
,
150 .post_depth_coverage
= pdevice
->info
.gen
>= 9,
151 .shader_viewport_index_layer
= true,
152 .stencil_export
= pdevice
->info
.gen
>= 9,
153 .storage_8bit
= pdevice
->info
.gen
>= 8,
154 .storage_16bit
= pdevice
->info
.gen
>= 8,
155 .subgroup_arithmetic
= true,
156 .subgroup_basic
= true,
157 .subgroup_ballot
= true,
158 .subgroup_quad
= true,
159 .subgroup_shuffle
= true,
160 .subgroup_vote
= true,
161 .tessellation
= true,
162 .transform_feedback
= pdevice
->info
.gen
>= 8,
163 .variable_pointers
= true,
165 .ubo_ptr_type
= glsl_vector_type(GLSL_TYPE_UINT
, 2),
166 .ssbo_ptr_type
= glsl_vector_type(GLSL_TYPE_UINT
, 2),
167 .phys_ssbo_ptr_type
= glsl_vector_type(GLSL_TYPE_UINT64
, 1),
168 .push_const_ptr_type
= glsl_uint_type(),
169 .shared_ptr_type
= glsl_uint_type(),
172 nir_function
*entry_point
=
173 spirv_to_nir(spirv
, module
->size
/ 4,
174 spec_entries
, num_spec_entries
,
175 stage
, entrypoint_name
, &spirv_options
, nir_options
);
176 nir_shader
*nir
= entry_point
->shader
;
177 assert(nir
->info
.stage
== stage
);
178 nir_validate_shader(nir
, "after spirv_to_nir");
179 ralloc_steal(mem_ctx
, nir
);
183 if (unlikely(INTEL_DEBUG
& stage_to_debug
[stage
])) {
184 fprintf(stderr
, "NIR (from SPIR-V) for %s shader:\n",
185 gl_shader_stage_name(stage
));
186 nir_print_shader(nir
, stderr
);
189 /* We have to lower away local constant initializers right before we
190 * inline functions. That way they get properly initialized at the top
191 * of the function and not at the top of its caller.
193 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
194 NIR_PASS_V(nir
, nir_lower_returns
);
195 NIR_PASS_V(nir
, nir_inline_functions
);
196 NIR_PASS_V(nir
, nir_opt_deref
);
198 /* Pick off the single entrypoint that we want */
199 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
200 if (func
!= entry_point
)
201 exec_node_remove(&func
->node
);
203 assert(exec_list_length(&nir
->functions
) == 1);
205 /* Now that we've deleted all but the main function, we can go ahead and
206 * lower the rest of the constant initializers. We do this here so that
207 * nir_remove_dead_variables and split_per_member_structs below see the
208 * corresponding stores.
210 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
212 /* Split member structs. We do this before lower_io_to_temporaries so that
213 * it doesn't lower system values to temporaries by accident.
215 NIR_PASS_V(nir
, nir_split_var_copies
);
216 NIR_PASS_V(nir
, nir_split_per_member_structs
);
218 NIR_PASS_V(nir
, nir_remove_dead_variables
,
219 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
221 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_global
,
222 nir_address_format_64bit_global
);
224 NIR_PASS_V(nir
, nir_propagate_invariant
);
225 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
226 entry_point
->impl
, true, false);
228 /* Vulkan uses the separate-shader linking model */
229 nir
->info
.separate_shader
= true;
231 nir
= brw_preprocess_nir(compiler
, nir
, NULL
);
236 void anv_DestroyPipeline(
238 VkPipeline _pipeline
,
239 const VkAllocationCallbacks
* pAllocator
)
241 ANV_FROM_HANDLE(anv_device
, device
, _device
);
242 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, _pipeline
);
247 anv_reloc_list_finish(&pipeline
->batch_relocs
,
248 pAllocator
? pAllocator
: &device
->alloc
);
249 if (pipeline
->blend_state
.map
)
250 anv_state_pool_free(&device
->dynamic_state_pool
, pipeline
->blend_state
);
252 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
253 if (pipeline
->shaders
[s
])
254 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
257 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
260 static const uint32_t vk_to_gen_primitive_type
[] = {
261 [VK_PRIMITIVE_TOPOLOGY_POINT_LIST
] = _3DPRIM_POINTLIST
,
262 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST
] = _3DPRIM_LINELIST
,
263 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
264 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
] = _3DPRIM_TRILIST
,
265 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
266 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
267 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
268 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
269 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
270 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
274 populate_sampler_prog_key(const struct gen_device_info
*devinfo
,
275 struct brw_sampler_prog_key_data
*key
)
277 /* Almost all multisampled textures are compressed. The only time when we
278 * don't compress a multisampled texture is for 16x MSAA with a surface
279 * width greater than 8k which is a bit of an edge case. Since the sampler
280 * just ignores the MCS parameter to ld2ms when MCS is disabled, it's safe
281 * to tell the compiler to always assume compression.
283 key
->compressed_multisample_layout_mask
= ~0;
285 /* SkyLake added support for 16x MSAA. With this came a new message for
286 * reading from a 16x MSAA surface with compression. The new message was
287 * needed because now the MCS data is 64 bits instead of 32 or lower as is
288 * the case for 8x, 4x, and 2x. The key->msaa_16 bit-field controls which
289 * message we use. Fortunately, the 16x message works for 8x, 4x, and 2x
290 * so we can just use it unconditionally. This may not be quite as
291 * efficient but it saves us from recompiling.
293 if (devinfo
->gen
>= 9)
296 /* XXX: Handle texture swizzle on HSW- */
297 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
298 /* Assume color sampler, no swizzling. (Works for BDW+) */
299 key
->swizzles
[i
] = SWIZZLE_XYZW
;
304 populate_vs_prog_key(const struct gen_device_info
*devinfo
,
305 struct brw_vs_prog_key
*key
)
307 memset(key
, 0, sizeof(*key
));
309 populate_sampler_prog_key(devinfo
, &key
->tex
);
311 /* XXX: Handle vertex input work-arounds */
313 /* XXX: Handle sampler_prog_key */
317 populate_tcs_prog_key(const struct gen_device_info
*devinfo
,
318 unsigned input_vertices
,
319 struct brw_tcs_prog_key
*key
)
321 memset(key
, 0, sizeof(*key
));
323 populate_sampler_prog_key(devinfo
, &key
->tex
);
325 key
->input_vertices
= input_vertices
;
329 populate_tes_prog_key(const struct gen_device_info
*devinfo
,
330 struct brw_tes_prog_key
*key
)
332 memset(key
, 0, sizeof(*key
));
334 populate_sampler_prog_key(devinfo
, &key
->tex
);
338 populate_gs_prog_key(const struct gen_device_info
*devinfo
,
339 struct brw_gs_prog_key
*key
)
341 memset(key
, 0, sizeof(*key
));
343 populate_sampler_prog_key(devinfo
, &key
->tex
);
347 populate_wm_prog_key(const struct gen_device_info
*devinfo
,
348 const struct anv_subpass
*subpass
,
349 const VkPipelineMultisampleStateCreateInfo
*ms_info
,
350 struct brw_wm_prog_key
*key
)
352 memset(key
, 0, sizeof(*key
));
354 populate_sampler_prog_key(devinfo
, &key
->tex
);
356 /* We set this to 0 here and set to the actual value before we call
359 key
->input_slots_valid
= 0;
361 /* Vulkan doesn't specify a default */
362 key
->high_quality_derivatives
= false;
364 /* XXX Vulkan doesn't appear to specify */
365 key
->clamp_fragment_color
= false;
367 assert(subpass
->color_count
<= MAX_RTS
);
368 for (uint32_t i
= 0; i
< subpass
->color_count
; i
++) {
369 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
370 key
->color_outputs_valid
|= (1 << i
);
373 key
->nr_color_regions
= util_bitcount(key
->color_outputs_valid
);
375 key
->replicate_alpha
= key
->nr_color_regions
> 1 &&
376 ms_info
&& ms_info
->alphaToCoverageEnable
;
379 /* We should probably pull this out of the shader, but it's fairly
380 * harmless to compute it and then let dead-code take care of it.
382 if (ms_info
->rasterizationSamples
> 1) {
383 key
->persample_interp
=
384 (ms_info
->minSampleShading
* ms_info
->rasterizationSamples
) > 1;
385 key
->multisample_fbo
= true;
388 key
->frag_coord_adds_sample_pos
= ms_info
->sampleShadingEnable
;
393 populate_cs_prog_key(const struct gen_device_info
*devinfo
,
394 struct brw_cs_prog_key
*key
)
396 memset(key
, 0, sizeof(*key
));
398 populate_sampler_prog_key(devinfo
, &key
->tex
);
401 struct anv_pipeline_stage
{
402 gl_shader_stage stage
;
404 const struct anv_shader_module
*module
;
405 const char *entrypoint
;
406 const VkSpecializationInfo
*spec_info
;
408 unsigned char shader_sha1
[20];
410 union brw_any_prog_key key
;
413 gl_shader_stage stage
;
414 unsigned char sha1
[20];
419 struct anv_pipeline_binding surface_to_descriptor
[256];
420 struct anv_pipeline_binding sampler_to_descriptor
[256];
421 struct anv_pipeline_bind_map bind_map
;
423 union brw_any_prog_data prog_data
;
427 anv_pipeline_hash_shader(const struct anv_shader_module
*module
,
428 const char *entrypoint
,
429 gl_shader_stage stage
,
430 const VkSpecializationInfo
*spec_info
,
431 unsigned char *sha1_out
)
433 struct mesa_sha1 ctx
;
434 _mesa_sha1_init(&ctx
);
436 _mesa_sha1_update(&ctx
, module
->sha1
, sizeof(module
->sha1
));
437 _mesa_sha1_update(&ctx
, entrypoint
, strlen(entrypoint
));
438 _mesa_sha1_update(&ctx
, &stage
, sizeof(stage
));
440 _mesa_sha1_update(&ctx
, spec_info
->pMapEntries
,
441 spec_info
->mapEntryCount
*
442 sizeof(*spec_info
->pMapEntries
));
443 _mesa_sha1_update(&ctx
, spec_info
->pData
,
444 spec_info
->dataSize
);
447 _mesa_sha1_final(&ctx
, sha1_out
);
451 anv_pipeline_hash_graphics(struct anv_pipeline
*pipeline
,
452 struct anv_pipeline_layout
*layout
,
453 struct anv_pipeline_stage
*stages
,
454 unsigned char *sha1_out
)
456 struct mesa_sha1 ctx
;
457 _mesa_sha1_init(&ctx
);
459 _mesa_sha1_update(&ctx
, &pipeline
->subpass
->view_mask
,
460 sizeof(pipeline
->subpass
->view_mask
));
463 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
465 const bool rba
= pipeline
->device
->robust_buffer_access
;
466 _mesa_sha1_update(&ctx
, &rba
, sizeof(rba
));
468 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
469 if (stages
[s
].entrypoint
) {
470 _mesa_sha1_update(&ctx
, stages
[s
].shader_sha1
,
471 sizeof(stages
[s
].shader_sha1
));
472 _mesa_sha1_update(&ctx
, &stages
[s
].key
, brw_prog_key_size(s
));
476 _mesa_sha1_final(&ctx
, sha1_out
);
480 anv_pipeline_hash_compute(struct anv_pipeline
*pipeline
,
481 struct anv_pipeline_layout
*layout
,
482 struct anv_pipeline_stage
*stage
,
483 unsigned char *sha1_out
)
485 struct mesa_sha1 ctx
;
486 _mesa_sha1_init(&ctx
);
489 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
491 const bool rba
= pipeline
->device
->robust_buffer_access
;
492 _mesa_sha1_update(&ctx
, &rba
, sizeof(rba
));
494 _mesa_sha1_update(&ctx
, stage
->shader_sha1
,
495 sizeof(stage
->shader_sha1
));
496 _mesa_sha1_update(&ctx
, &stage
->key
.cs
, sizeof(stage
->key
.cs
));
498 _mesa_sha1_final(&ctx
, sha1_out
);
502 anv_pipeline_stage_get_nir(struct anv_pipeline
*pipeline
,
503 struct anv_pipeline_cache
*cache
,
505 struct anv_pipeline_stage
*stage
)
507 const struct brw_compiler
*compiler
=
508 pipeline
->device
->instance
->physicalDevice
.compiler
;
509 const nir_shader_compiler_options
*nir_options
=
510 compiler
->glsl_compiler_options
[stage
->stage
].NirOptions
;
513 nir
= anv_device_search_for_nir(pipeline
->device
, cache
,
518 assert(nir
->info
.stage
== stage
->stage
);
522 nir
= anv_shader_compile_to_nir(pipeline
->device
,
529 anv_device_upload_nir(pipeline
->device
, cache
, nir
, stage
->shader_sha1
);
537 anv_pipeline_lower_nir(struct anv_pipeline
*pipeline
,
539 struct anv_pipeline_stage
*stage
,
540 struct anv_pipeline_layout
*layout
)
542 const struct brw_compiler
*compiler
=
543 pipeline
->device
->instance
->physicalDevice
.compiler
;
545 struct brw_stage_prog_data
*prog_data
= &stage
->prog_data
.base
;
546 nir_shader
*nir
= stage
->nir
;
548 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
549 NIR_PASS_V(nir
, nir_lower_wpos_center
, pipeline
->sample_shading_enable
);
550 NIR_PASS_V(nir
, anv_nir_lower_input_attachments
);
553 NIR_PASS_V(nir
, anv_nir_lower_ycbcr_textures
, layout
);
555 NIR_PASS_V(nir
, anv_nir_lower_push_constants
);
557 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
558 NIR_PASS_V(nir
, anv_nir_lower_multiview
, pipeline
->subpass
->view_mask
);
560 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
)
561 prog_data
->total_shared
= nir
->num_shared
;
563 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
565 if (nir
->num_uniforms
> 0) {
566 assert(prog_data
->nr_params
== 0);
568 /* If the shader uses any push constants at all, we'll just give
569 * them the maximum possible number
571 assert(nir
->num_uniforms
<= MAX_PUSH_CONSTANTS_SIZE
);
572 nir
->num_uniforms
= MAX_PUSH_CONSTANTS_SIZE
;
573 prog_data
->nr_params
+= MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float);
574 prog_data
->param
= ralloc_array(mem_ctx
, uint32_t, prog_data
->nr_params
);
576 /* We now set the param values to be offsets into a
577 * anv_push_constant_data structure. Since the compiler doesn't
578 * actually dereference any of the gl_constant_value pointers in the
579 * params array, it doesn't really matter what we put here.
581 struct anv_push_constants
*null_data
= NULL
;
582 /* Fill out the push constants section of the param array */
583 for (unsigned i
= 0; i
< MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float); i
++) {
584 prog_data
->param
[i
] = ANV_PARAM_PUSH(
585 (uintptr_t)&null_data
->client_data
[i
* sizeof(float)]);
589 if (nir
->info
.num_ssbos
> 0 || nir
->info
.num_images
> 0)
590 pipeline
->needs_data_cache
= true;
592 NIR_PASS_V(nir
, brw_nir_lower_image_load_store
, compiler
->devinfo
);
594 /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
596 anv_nir_apply_pipeline_layout(&pipeline
->device
->instance
->physicalDevice
,
597 pipeline
->device
->robust_buffer_access
,
598 layout
, nir
, prog_data
,
601 NIR_PASS_V(nir
, nir_lower_explicit_io
,
602 nir_var_mem_ubo
| nir_var_mem_ssbo
,
603 nir_address_format_vk_index_offset
);
605 NIR_PASS_V(nir
, nir_opt_constant_folding
);
608 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
609 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
611 assert(nir
->num_uniforms
== prog_data
->nr_params
* 4);
617 anv_pipeline_link_vs(const struct brw_compiler
*compiler
,
618 struct anv_pipeline_stage
*vs_stage
,
619 struct anv_pipeline_stage
*next_stage
)
622 brw_nir_link_shaders(compiler
, &vs_stage
->nir
, &next_stage
->nir
);
625 static const unsigned *
626 anv_pipeline_compile_vs(const struct brw_compiler
*compiler
,
628 struct anv_device
*device
,
629 struct anv_pipeline_stage
*vs_stage
)
631 brw_compute_vue_map(compiler
->devinfo
,
632 &vs_stage
->prog_data
.vs
.base
.vue_map
,
633 vs_stage
->nir
->info
.outputs_written
,
634 vs_stage
->nir
->info
.separate_shader
);
636 return brw_compile_vs(compiler
, device
, mem_ctx
, &vs_stage
->key
.vs
,
637 &vs_stage
->prog_data
.vs
, vs_stage
->nir
, -1, NULL
);
641 merge_tess_info(struct shader_info
*tes_info
,
642 const struct shader_info
*tcs_info
)
644 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
646 * "PointMode. Controls generation of points rather than triangles
647 * or lines. This functionality defaults to disabled, and is
648 * enabled if either shader stage includes the execution mode.
650 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
651 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
652 * and OutputVertices, it says:
654 * "One mode must be set in at least one of the tessellation
657 * So, the fields can be set in either the TCS or TES, but they must
658 * agree if set in both. Our backend looks at TES, so bitwise-or in
659 * the values from the TCS.
661 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
662 tes_info
->tess
.tcs_vertices_out
== 0 ||
663 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
664 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
666 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
667 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
668 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
669 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
671 assert(tcs_info
->tess
.primitive_mode
== 0 ||
672 tes_info
->tess
.primitive_mode
== 0 ||
673 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
674 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
675 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
676 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
680 anv_pipeline_link_tcs(const struct brw_compiler
*compiler
,
681 struct anv_pipeline_stage
*tcs_stage
,
682 struct anv_pipeline_stage
*tes_stage
)
684 assert(tes_stage
&& tes_stage
->stage
== MESA_SHADER_TESS_EVAL
);
686 brw_nir_link_shaders(compiler
, &tcs_stage
->nir
, &tes_stage
->nir
);
688 nir_lower_patch_vertices(tes_stage
->nir
,
689 tcs_stage
->nir
->info
.tess
.tcs_vertices_out
,
692 /* Copy TCS info into the TES info */
693 merge_tess_info(&tes_stage
->nir
->info
, &tcs_stage
->nir
->info
);
695 /* Whacking the key after cache lookup is a bit sketchy, but all of
696 * this comes from the SPIR-V, which is part of the hash used for the
697 * pipeline cache. So it should be safe.
699 tcs_stage
->key
.tcs
.tes_primitive_mode
=
700 tes_stage
->nir
->info
.tess
.primitive_mode
;
701 tcs_stage
->key
.tcs
.quads_workaround
=
702 compiler
->devinfo
->gen
< 9 &&
703 tes_stage
->nir
->info
.tess
.primitive_mode
== 7 /* GL_QUADS */ &&
704 tes_stage
->nir
->info
.tess
.spacing
== TESS_SPACING_EQUAL
;
707 static const unsigned *
708 anv_pipeline_compile_tcs(const struct brw_compiler
*compiler
,
710 struct anv_device
*device
,
711 struct anv_pipeline_stage
*tcs_stage
,
712 struct anv_pipeline_stage
*prev_stage
)
714 tcs_stage
->key
.tcs
.outputs_written
=
715 tcs_stage
->nir
->info
.outputs_written
;
716 tcs_stage
->key
.tcs
.patch_outputs_written
=
717 tcs_stage
->nir
->info
.patch_outputs_written
;
719 return brw_compile_tcs(compiler
, device
, mem_ctx
, &tcs_stage
->key
.tcs
,
720 &tcs_stage
->prog_data
.tcs
, tcs_stage
->nir
,
725 anv_pipeline_link_tes(const struct brw_compiler
*compiler
,
726 struct anv_pipeline_stage
*tes_stage
,
727 struct anv_pipeline_stage
*next_stage
)
730 brw_nir_link_shaders(compiler
, &tes_stage
->nir
, &next_stage
->nir
);
733 static const unsigned *
734 anv_pipeline_compile_tes(const struct brw_compiler
*compiler
,
736 struct anv_device
*device
,
737 struct anv_pipeline_stage
*tes_stage
,
738 struct anv_pipeline_stage
*tcs_stage
)
740 tes_stage
->key
.tes
.inputs_read
=
741 tcs_stage
->nir
->info
.outputs_written
;
742 tes_stage
->key
.tes
.patch_inputs_read
=
743 tcs_stage
->nir
->info
.patch_outputs_written
;
745 return brw_compile_tes(compiler
, device
, mem_ctx
, &tes_stage
->key
.tes
,
746 &tcs_stage
->prog_data
.tcs
.base
.vue_map
,
747 &tes_stage
->prog_data
.tes
, tes_stage
->nir
,
752 anv_pipeline_link_gs(const struct brw_compiler
*compiler
,
753 struct anv_pipeline_stage
*gs_stage
,
754 struct anv_pipeline_stage
*next_stage
)
757 brw_nir_link_shaders(compiler
, &gs_stage
->nir
, &next_stage
->nir
);
760 static const unsigned *
761 anv_pipeline_compile_gs(const struct brw_compiler
*compiler
,
763 struct anv_device
*device
,
764 struct anv_pipeline_stage
*gs_stage
,
765 struct anv_pipeline_stage
*prev_stage
)
767 brw_compute_vue_map(compiler
->devinfo
,
768 &gs_stage
->prog_data
.gs
.base
.vue_map
,
769 gs_stage
->nir
->info
.outputs_written
,
770 gs_stage
->nir
->info
.separate_shader
);
772 return brw_compile_gs(compiler
, device
, mem_ctx
, &gs_stage
->key
.gs
,
773 &gs_stage
->prog_data
.gs
, gs_stage
->nir
,
778 anv_pipeline_link_fs(const struct brw_compiler
*compiler
,
779 struct anv_pipeline_stage
*stage
)
781 unsigned num_rts
= 0;
782 const int max_rt
= FRAG_RESULT_DATA7
- FRAG_RESULT_DATA0
+ 1;
783 struct anv_pipeline_binding rt_bindings
[max_rt
];
784 nir_function_impl
*impl
= nir_shader_get_entrypoint(stage
->nir
);
785 int rt_to_bindings
[max_rt
];
786 memset(rt_to_bindings
, -1, sizeof(rt_to_bindings
));
787 bool rt_used
[max_rt
];
788 memset(rt_used
, 0, sizeof(rt_used
));
790 /* Flag used render targets */
791 nir_foreach_variable_safe(var
, &stage
->nir
->outputs
) {
792 if (var
->data
.location
< FRAG_RESULT_DATA0
)
795 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
796 /* Unused or out-of-bounds */
797 if (rt
>= MAX_RTS
|| !(stage
->key
.wm
.color_outputs_valid
& (1 << rt
)))
800 const unsigned array_len
=
801 glsl_type_is_array(var
->type
) ? glsl_get_length(var
->type
) : 1;
802 assert(rt
+ array_len
<= max_rt
);
804 for (unsigned i
= 0; i
< array_len
; i
++)
805 rt_used
[rt
+ i
] = true;
808 /* Set new, compacted, location */
809 for (unsigned i
= 0; i
< max_rt
; i
++) {
813 rt_to_bindings
[i
] = num_rts
;
814 rt_bindings
[rt_to_bindings
[i
]] = (struct anv_pipeline_binding
) {
815 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
822 bool deleted_output
= false;
823 nir_foreach_variable_safe(var
, &stage
->nir
->outputs
) {
824 if (var
->data
.location
< FRAG_RESULT_DATA0
)
827 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
829 !(stage
->key
.wm
.color_outputs_valid
& (1 << rt
))) {
830 /* Unused or out-of-bounds, throw it away */
831 deleted_output
= true;
832 var
->data
.mode
= nir_var_function_temp
;
833 exec_node_remove(&var
->node
);
834 exec_list_push_tail(&impl
->locals
, &var
->node
);
838 /* Give it the new location */
839 assert(rt_to_bindings
[rt
] != -1);
840 var
->data
.location
= rt_to_bindings
[rt
] + FRAG_RESULT_DATA0
;
844 nir_fixup_deref_modes(stage
->nir
);
847 /* If we have no render targets, we need a null render target */
848 rt_bindings
[0] = (struct anv_pipeline_binding
) {
849 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
856 /* Now that we've determined the actual number of render targets, adjust
857 * the key accordingly.
859 stage
->key
.wm
.nr_color_regions
= num_rts
;
860 stage
->key
.wm
.color_outputs_valid
= (1 << num_rts
) - 1;
862 assert(num_rts
<= max_rt
);
863 assert(stage
->bind_map
.surface_count
== 0);
864 typed_memcpy(stage
->bind_map
.surface_to_descriptor
,
865 rt_bindings
, num_rts
);
866 stage
->bind_map
.surface_count
+= num_rts
;
869 static const unsigned *
870 anv_pipeline_compile_fs(const struct brw_compiler
*compiler
,
872 struct anv_device
*device
,
873 struct anv_pipeline_stage
*fs_stage
,
874 struct anv_pipeline_stage
*prev_stage
)
876 /* TODO: we could set this to 0 based on the information in nir_shader, but
877 * we need this before we call spirv_to_nir.
880 fs_stage
->key
.wm
.input_slots_valid
=
881 prev_stage
->prog_data
.vue
.vue_map
.slots_valid
;
883 const unsigned *code
=
884 brw_compile_fs(compiler
, device
, mem_ctx
, &fs_stage
->key
.wm
,
885 &fs_stage
->prog_data
.wm
, fs_stage
->nir
,
886 NULL
, -1, -1, -1, true, false, NULL
, NULL
);
888 if (fs_stage
->key
.wm
.nr_color_regions
== 0 &&
889 !fs_stage
->prog_data
.wm
.has_side_effects
&&
890 !fs_stage
->prog_data
.wm
.uses_kill
&&
891 fs_stage
->prog_data
.wm
.computed_depth_mode
== BRW_PSCDEPTH_OFF
&&
892 !fs_stage
->prog_data
.wm
.computed_stencil
) {
893 /* This fragment shader has no outputs and no side effects. Go ahead
894 * and return the code pointer so we don't accidentally think the
895 * compile failed but zero out prog_data which will set program_size to
896 * zero and disable the stage.
898 memset(&fs_stage
->prog_data
, 0, sizeof(fs_stage
->prog_data
));
905 anv_pipeline_compile_graphics(struct anv_pipeline
*pipeline
,
906 struct anv_pipeline_cache
*cache
,
907 const VkGraphicsPipelineCreateInfo
*info
)
909 const struct brw_compiler
*compiler
=
910 pipeline
->device
->instance
->physicalDevice
.compiler
;
911 struct anv_pipeline_stage stages
[MESA_SHADER_STAGES
] = {};
913 pipeline
->active_stages
= 0;
916 for (uint32_t i
= 0; i
< info
->stageCount
; i
++) {
917 const VkPipelineShaderStageCreateInfo
*sinfo
= &info
->pStages
[i
];
918 gl_shader_stage stage
= vk_to_mesa_shader_stage(sinfo
->stage
);
920 pipeline
->active_stages
|= sinfo
->stage
;
922 stages
[stage
].stage
= stage
;
923 stages
[stage
].module
= anv_shader_module_from_handle(sinfo
->module
);
924 stages
[stage
].entrypoint
= sinfo
->pName
;
925 stages
[stage
].spec_info
= sinfo
->pSpecializationInfo
;
926 anv_pipeline_hash_shader(stages
[stage
].module
,
927 stages
[stage
].entrypoint
,
929 stages
[stage
].spec_info
,
930 stages
[stage
].shader_sha1
);
932 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
934 case MESA_SHADER_VERTEX
:
935 populate_vs_prog_key(devinfo
, &stages
[stage
].key
.vs
);
937 case MESA_SHADER_TESS_CTRL
:
938 populate_tcs_prog_key(devinfo
,
939 info
->pTessellationState
->patchControlPoints
,
940 &stages
[stage
].key
.tcs
);
942 case MESA_SHADER_TESS_EVAL
:
943 populate_tes_prog_key(devinfo
, &stages
[stage
].key
.tes
);
945 case MESA_SHADER_GEOMETRY
:
946 populate_gs_prog_key(devinfo
, &stages
[stage
].key
.gs
);
948 case MESA_SHADER_FRAGMENT
:
949 populate_wm_prog_key(devinfo
, pipeline
->subpass
,
950 info
->pMultisampleState
,
951 &stages
[stage
].key
.wm
);
954 unreachable("Invalid graphics shader stage");
958 if (pipeline
->active_stages
& VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
)
959 pipeline
->active_stages
|= VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
;
961 assert(pipeline
->active_stages
& VK_SHADER_STAGE_VERTEX_BIT
);
963 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
965 unsigned char sha1
[20];
966 anv_pipeline_hash_graphics(pipeline
, layout
, stages
, sha1
);
969 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
970 if (!stages
[s
].entrypoint
)
973 stages
[s
].cache_key
.stage
= s
;
974 memcpy(stages
[s
].cache_key
.sha1
, sha1
, sizeof(sha1
));
976 struct anv_shader_bin
*bin
=
977 anv_device_search_for_kernel(pipeline
->device
, cache
,
978 &stages
[s
].cache_key
,
979 sizeof(stages
[s
].cache_key
));
982 pipeline
->shaders
[s
] = bin
;
986 if (found
== __builtin_popcount(pipeline
->active_stages
)) {
987 /* We found all our shaders in the cache. We're done. */
989 } else if (found
> 0) {
990 /* We found some but not all of our shaders. This shouldn't happen
991 * most of the time but it can if we have a partially populated
994 assert(found
< __builtin_popcount(pipeline
->active_stages
));
996 vk_debug_report(&pipeline
->device
->instance
->debug_report_callbacks
,
997 VK_DEBUG_REPORT_WARNING_BIT_EXT
|
998 VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT
,
999 VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_CACHE_EXT
,
1000 (uint64_t)(uintptr_t)cache
,
1002 "Found a partial pipeline in the cache. This is "
1003 "most likely caused by an incomplete pipeline cache "
1004 "import or export");
1006 /* We're going to have to recompile anyway, so just throw away our
1007 * references to the shaders in the cache. We'll get them out of the
1008 * cache again as part of the compilation process.
1010 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1011 if (pipeline
->shaders
[s
]) {
1012 anv_shader_bin_unref(pipeline
->device
, pipeline
->shaders
[s
]);
1013 pipeline
->shaders
[s
] = NULL
;
1018 void *pipeline_ctx
= ralloc_context(NULL
);
1020 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1021 if (!stages
[s
].entrypoint
)
1024 assert(stages
[s
].stage
== s
);
1025 assert(pipeline
->shaders
[s
] == NULL
);
1027 stages
[s
].bind_map
= (struct anv_pipeline_bind_map
) {
1028 .surface_to_descriptor
= stages
[s
].surface_to_descriptor
,
1029 .sampler_to_descriptor
= stages
[s
].sampler_to_descriptor
1032 stages
[s
].nir
= anv_pipeline_stage_get_nir(pipeline
, cache
,
1035 if (stages
[s
].nir
== NULL
) {
1036 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1041 /* Walk backwards to link */
1042 struct anv_pipeline_stage
*next_stage
= NULL
;
1043 for (int s
= MESA_SHADER_STAGES
- 1; s
>= 0; s
--) {
1044 if (!stages
[s
].entrypoint
)
1048 case MESA_SHADER_VERTEX
:
1049 anv_pipeline_link_vs(compiler
, &stages
[s
], next_stage
);
1051 case MESA_SHADER_TESS_CTRL
:
1052 anv_pipeline_link_tcs(compiler
, &stages
[s
], next_stage
);
1054 case MESA_SHADER_TESS_EVAL
:
1055 anv_pipeline_link_tes(compiler
, &stages
[s
], next_stage
);
1057 case MESA_SHADER_GEOMETRY
:
1058 anv_pipeline_link_gs(compiler
, &stages
[s
], next_stage
);
1060 case MESA_SHADER_FRAGMENT
:
1061 anv_pipeline_link_fs(compiler
, &stages
[s
]);
1064 unreachable("Invalid graphics shader stage");
1067 next_stage
= &stages
[s
];
1070 struct anv_pipeline_stage
*prev_stage
= NULL
;
1071 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1072 if (!stages
[s
].entrypoint
)
1075 void *stage_ctx
= ralloc_context(NULL
);
1077 nir_xfb_info
*xfb_info
= NULL
;
1078 if (s
== MESA_SHADER_VERTEX
||
1079 s
== MESA_SHADER_TESS_EVAL
||
1080 s
== MESA_SHADER_GEOMETRY
)
1081 xfb_info
= nir_gather_xfb_info(stages
[s
].nir
, stage_ctx
);
1083 anv_pipeline_lower_nir(pipeline
, stage_ctx
, &stages
[s
], layout
);
1085 const unsigned *code
;
1087 case MESA_SHADER_VERTEX
:
1088 code
= anv_pipeline_compile_vs(compiler
, stage_ctx
, pipeline
->device
,
1091 case MESA_SHADER_TESS_CTRL
:
1092 code
= anv_pipeline_compile_tcs(compiler
, stage_ctx
, pipeline
->device
,
1093 &stages
[s
], prev_stage
);
1095 case MESA_SHADER_TESS_EVAL
:
1096 code
= anv_pipeline_compile_tes(compiler
, stage_ctx
, pipeline
->device
,
1097 &stages
[s
], prev_stage
);
1099 case MESA_SHADER_GEOMETRY
:
1100 code
= anv_pipeline_compile_gs(compiler
, stage_ctx
, pipeline
->device
,
1101 &stages
[s
], prev_stage
);
1103 case MESA_SHADER_FRAGMENT
:
1104 code
= anv_pipeline_compile_fs(compiler
, stage_ctx
, pipeline
->device
,
1105 &stages
[s
], prev_stage
);
1108 unreachable("Invalid graphics shader stage");
1111 ralloc_free(stage_ctx
);
1112 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1116 struct anv_shader_bin
*bin
=
1117 anv_device_upload_kernel(pipeline
->device
, cache
,
1118 &stages
[s
].cache_key
,
1119 sizeof(stages
[s
].cache_key
),
1120 code
, stages
[s
].prog_data
.base
.program_size
,
1121 stages
[s
].nir
->constant_data
,
1122 stages
[s
].nir
->constant_data_size
,
1123 &stages
[s
].prog_data
.base
,
1124 brw_prog_data_size(s
),
1125 xfb_info
, &stages
[s
].bind_map
);
1127 ralloc_free(stage_ctx
);
1128 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1132 pipeline
->shaders
[s
] = bin
;
1133 ralloc_free(stage_ctx
);
1135 prev_stage
= &stages
[s
];
1138 ralloc_free(pipeline_ctx
);
1142 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
] &&
1143 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->prog_data
->program_size
== 0) {
1144 /* This can happen if we decided to implicitly disable the fragment
1145 * shader. See anv_pipeline_compile_fs().
1147 anv_shader_bin_unref(pipeline
->device
,
1148 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
1149 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] = NULL
;
1150 pipeline
->active_stages
&= ~VK_SHADER_STAGE_FRAGMENT_BIT
;
1156 ralloc_free(pipeline_ctx
);
1158 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1159 if (pipeline
->shaders
[s
])
1160 anv_shader_bin_unref(pipeline
->device
, pipeline
->shaders
[s
]);
1167 anv_pipeline_compile_cs(struct anv_pipeline
*pipeline
,
1168 struct anv_pipeline_cache
*cache
,
1169 const VkComputePipelineCreateInfo
*info
,
1170 const struct anv_shader_module
*module
,
1171 const char *entrypoint
,
1172 const VkSpecializationInfo
*spec_info
)
1174 const struct brw_compiler
*compiler
=
1175 pipeline
->device
->instance
->physicalDevice
.compiler
;
1177 struct anv_pipeline_stage stage
= {
1178 .stage
= MESA_SHADER_COMPUTE
,
1180 .entrypoint
= entrypoint
,
1181 .spec_info
= spec_info
,
1183 .stage
= MESA_SHADER_COMPUTE
,
1186 anv_pipeline_hash_shader(stage
.module
,
1188 MESA_SHADER_COMPUTE
,
1192 struct anv_shader_bin
*bin
= NULL
;
1194 populate_cs_prog_key(&pipeline
->device
->info
, &stage
.key
.cs
);
1196 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
1198 anv_pipeline_hash_compute(pipeline
, layout
, &stage
, stage
.cache_key
.sha1
);
1199 bin
= anv_device_search_for_kernel(pipeline
->device
, cache
, &stage
.cache_key
,
1200 sizeof(stage
.cache_key
));
1203 stage
.bind_map
= (struct anv_pipeline_bind_map
) {
1204 .surface_to_descriptor
= stage
.surface_to_descriptor
,
1205 .sampler_to_descriptor
= stage
.sampler_to_descriptor
1208 /* Set up a binding for the gl_NumWorkGroups */
1209 stage
.bind_map
.surface_count
= 1;
1210 stage
.bind_map
.surface_to_descriptor
[0] = (struct anv_pipeline_binding
) {
1211 .set
= ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS
,
1214 void *mem_ctx
= ralloc_context(NULL
);
1216 stage
.nir
= anv_pipeline_stage_get_nir(pipeline
, cache
, mem_ctx
, &stage
);
1217 if (stage
.nir
== NULL
) {
1218 ralloc_free(mem_ctx
);
1219 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1222 anv_pipeline_lower_nir(pipeline
, mem_ctx
, &stage
, layout
);
1224 NIR_PASS_V(stage
.nir
, anv_nir_add_base_work_group_id
,
1225 &stage
.prog_data
.cs
);
1227 const unsigned *shader_code
=
1228 brw_compile_cs(compiler
, pipeline
->device
, mem_ctx
, &stage
.key
.cs
,
1229 &stage
.prog_data
.cs
, stage
.nir
, -1, NULL
);
1230 if (shader_code
== NULL
) {
1231 ralloc_free(mem_ctx
);
1232 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1235 const unsigned code_size
= stage
.prog_data
.base
.program_size
;
1236 bin
= anv_device_upload_kernel(pipeline
->device
, cache
,
1237 &stage
.cache_key
, sizeof(stage
.cache_key
),
1238 shader_code
, code_size
,
1239 stage
.nir
->constant_data
,
1240 stage
.nir
->constant_data_size
,
1241 &stage
.prog_data
.base
,
1242 sizeof(stage
.prog_data
.cs
),
1243 NULL
, &stage
.bind_map
);
1245 ralloc_free(mem_ctx
);
1246 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1249 ralloc_free(mem_ctx
);
1252 pipeline
->active_stages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1253 pipeline
->shaders
[MESA_SHADER_COMPUTE
] = bin
;
1259 * Copy pipeline state not marked as dynamic.
1260 * Dynamic state is pipeline state which hasn't been provided at pipeline
1261 * creation time, but is dynamically provided afterwards using various
1262 * vkCmdSet* functions.
1264 * The set of state considered "non_dynamic" is determined by the pieces of
1265 * state that have their corresponding VkDynamicState enums omitted from
1266 * VkPipelineDynamicStateCreateInfo::pDynamicStates.
1268 * @param[out] pipeline Destination non_dynamic state.
1269 * @param[in] pCreateInfo Source of non_dynamic state to be copied.
1272 copy_non_dynamic_state(struct anv_pipeline
*pipeline
,
1273 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1275 anv_cmd_dirty_mask_t states
= ANV_CMD_DIRTY_DYNAMIC_ALL
;
1276 struct anv_subpass
*subpass
= pipeline
->subpass
;
1278 pipeline
->dynamic_state
= default_dynamic_state
;
1280 if (pCreateInfo
->pDynamicState
) {
1281 /* Remove all of the states that are marked as dynamic */
1282 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1283 for (uint32_t s
= 0; s
< count
; s
++)
1284 states
&= ~(1 << pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1287 struct anv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1289 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1291 * pViewportState is [...] NULL if the pipeline
1292 * has rasterization disabled.
1294 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1295 assert(pCreateInfo
->pViewportState
);
1297 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1298 if (states
& (1 << VK_DYNAMIC_STATE_VIEWPORT
)) {
1299 typed_memcpy(dynamic
->viewport
.viewports
,
1300 pCreateInfo
->pViewportState
->pViewports
,
1301 pCreateInfo
->pViewportState
->viewportCount
);
1304 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1305 if (states
& (1 << VK_DYNAMIC_STATE_SCISSOR
)) {
1306 typed_memcpy(dynamic
->scissor
.scissors
,
1307 pCreateInfo
->pViewportState
->pScissors
,
1308 pCreateInfo
->pViewportState
->scissorCount
);
1312 if (states
& (1 << VK_DYNAMIC_STATE_LINE_WIDTH
)) {
1313 assert(pCreateInfo
->pRasterizationState
);
1314 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1317 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BIAS
)) {
1318 assert(pCreateInfo
->pRasterizationState
);
1319 dynamic
->depth_bias
.bias
=
1320 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1321 dynamic
->depth_bias
.clamp
=
1322 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1323 dynamic
->depth_bias
.slope
=
1324 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1327 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1329 * pColorBlendState is [...] NULL if the pipeline has rasterization
1330 * disabled or if the subpass of the render pass the pipeline is
1331 * created against does not use any color attachments.
1333 bool uses_color_att
= false;
1334 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1335 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1336 uses_color_att
= true;
1341 if (uses_color_att
&&
1342 !pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1343 assert(pCreateInfo
->pColorBlendState
);
1345 if (states
& (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS
))
1346 typed_memcpy(dynamic
->blend_constants
,
1347 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1350 /* If there is no depthstencil attachment, then don't read
1351 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1352 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1353 * no need to override the depthstencil defaults in
1354 * anv_pipeline::dynamic_state when there is no depthstencil attachment.
1356 * Section 9.2 of the Vulkan 1.0.15 spec says:
1358 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1359 * disabled or if the subpass of the render pass the pipeline is created
1360 * against does not use a depth/stencil attachment.
1362 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1363 subpass
->depth_stencil_attachment
) {
1364 assert(pCreateInfo
->pDepthStencilState
);
1366 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS
)) {
1367 dynamic
->depth_bounds
.min
=
1368 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1369 dynamic
->depth_bounds
.max
=
1370 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1373 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
)) {
1374 dynamic
->stencil_compare_mask
.front
=
1375 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1376 dynamic
->stencil_compare_mask
.back
=
1377 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1380 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
)) {
1381 dynamic
->stencil_write_mask
.front
=
1382 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1383 dynamic
->stencil_write_mask
.back
=
1384 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1387 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE
)) {
1388 dynamic
->stencil_reference
.front
=
1389 pCreateInfo
->pDepthStencilState
->front
.reference
;
1390 dynamic
->stencil_reference
.back
=
1391 pCreateInfo
->pDepthStencilState
->back
.reference
;
1395 pipeline
->dynamic_state_mask
= states
;
1399 anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo
*info
)
1402 struct anv_render_pass
*renderpass
= NULL
;
1403 struct anv_subpass
*subpass
= NULL
;
1405 /* Assert that all required members of VkGraphicsPipelineCreateInfo are
1406 * present. See the Vulkan 1.0.28 spec, Section 9.2 Graphics Pipelines.
1408 assert(info
->sType
== VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
);
1410 renderpass
= anv_render_pass_from_handle(info
->renderPass
);
1413 assert(info
->subpass
< renderpass
->subpass_count
);
1414 subpass
= &renderpass
->subpasses
[info
->subpass
];
1416 assert(info
->stageCount
>= 1);
1417 assert(info
->pVertexInputState
);
1418 assert(info
->pInputAssemblyState
);
1419 assert(info
->pRasterizationState
);
1420 if (!info
->pRasterizationState
->rasterizerDiscardEnable
) {
1421 assert(info
->pViewportState
);
1422 assert(info
->pMultisampleState
);
1424 if (subpass
&& subpass
->depth_stencil_attachment
)
1425 assert(info
->pDepthStencilState
);
1427 if (subpass
&& subpass
->color_count
> 0) {
1428 bool all_color_unused
= true;
1429 for (int i
= 0; i
< subpass
->color_count
; i
++) {
1430 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
1431 all_color_unused
= false;
1433 /* pColorBlendState is ignored if the pipeline has rasterization
1434 * disabled or if the subpass of the render pass the pipeline is
1435 * created against does not use any color attachments.
1437 assert(info
->pColorBlendState
|| all_color_unused
);
1441 for (uint32_t i
= 0; i
< info
->stageCount
; ++i
) {
1442 switch (info
->pStages
[i
].stage
) {
1443 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
:
1444 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
:
1445 assert(info
->pTessellationState
);
1455 * Calculate the desired L3 partitioning based on the current state of the
1456 * pipeline. For now this simply returns the conservative defaults calculated
1457 * by get_default_l3_weights(), but we could probably do better by gathering
1458 * more statistics from the pipeline state (e.g. guess of expected URB usage
1459 * and bound surfaces), or by using feed-back from performance counters.
1462 anv_pipeline_setup_l3_config(struct anv_pipeline
*pipeline
, bool needs_slm
)
1464 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1466 const struct gen_l3_weights w
=
1467 gen_get_default_l3_weights(devinfo
, pipeline
->needs_data_cache
, needs_slm
);
1469 pipeline
->urb
.l3_config
= gen_get_l3_config(devinfo
, w
);
1470 pipeline
->urb
.total_size
=
1471 gen_get_l3_config_urb_size(devinfo
, pipeline
->urb
.l3_config
);
1475 anv_pipeline_init(struct anv_pipeline
*pipeline
,
1476 struct anv_device
*device
,
1477 struct anv_pipeline_cache
*cache
,
1478 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1479 const VkAllocationCallbacks
*alloc
)
1483 anv_pipeline_validate_create_info(pCreateInfo
);
1486 alloc
= &device
->alloc
;
1488 pipeline
->device
= device
;
1490 ANV_FROM_HANDLE(anv_render_pass
, render_pass
, pCreateInfo
->renderPass
);
1491 assert(pCreateInfo
->subpass
< render_pass
->subpass_count
);
1492 pipeline
->subpass
= &render_pass
->subpasses
[pCreateInfo
->subpass
];
1494 result
= anv_reloc_list_init(&pipeline
->batch_relocs
, alloc
);
1495 if (result
!= VK_SUCCESS
)
1498 pipeline
->batch
.alloc
= alloc
;
1499 pipeline
->batch
.next
= pipeline
->batch
.start
= pipeline
->batch_data
;
1500 pipeline
->batch
.end
= pipeline
->batch
.start
+ sizeof(pipeline
->batch_data
);
1501 pipeline
->batch
.relocs
= &pipeline
->batch_relocs
;
1502 pipeline
->batch
.status
= VK_SUCCESS
;
1504 copy_non_dynamic_state(pipeline
, pCreateInfo
);
1505 pipeline
->depth_clamp_enable
= pCreateInfo
->pRasterizationState
&&
1506 pCreateInfo
->pRasterizationState
->depthClampEnable
;
1508 /* Previously we enabled depth clipping when !depthClampEnable.
1509 * DepthClipStateCreateInfo now makes depth clipping explicit so if the
1510 * clipping info is available, use its enable value to determine clipping,
1511 * otherwise fallback to the previous !depthClampEnable logic.
1513 const VkPipelineRasterizationDepthClipStateCreateInfoEXT
*clip_info
=
1514 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
,
1515 PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT
);
1516 pipeline
->depth_clip_enable
= clip_info
? clip_info
->depthClipEnable
: !pipeline
->depth_clamp_enable
;
1518 pipeline
->sample_shading_enable
= pCreateInfo
->pMultisampleState
&&
1519 pCreateInfo
->pMultisampleState
->sampleShadingEnable
;
1521 pipeline
->needs_data_cache
= false;
1523 /* When we free the pipeline, we detect stages based on the NULL status
1524 * of various prog_data pointers. Make them NULL by default.
1526 memset(pipeline
->shaders
, 0, sizeof(pipeline
->shaders
));
1528 result
= anv_pipeline_compile_graphics(pipeline
, cache
, pCreateInfo
);
1529 if (result
!= VK_SUCCESS
) {
1530 anv_reloc_list_finish(&pipeline
->batch_relocs
, alloc
);
1534 assert(pipeline
->shaders
[MESA_SHADER_VERTEX
]);
1536 anv_pipeline_setup_l3_config(pipeline
, false);
1538 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1539 pCreateInfo
->pVertexInputState
;
1541 const uint64_t inputs_read
= get_vs_prog_data(pipeline
)->inputs_read
;
1543 pipeline
->vb_used
= 0;
1544 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1545 const VkVertexInputAttributeDescription
*desc
=
1546 &vi_info
->pVertexAttributeDescriptions
[i
];
1548 if (inputs_read
& (1ull << (VERT_ATTRIB_GENERIC0
+ desc
->location
)))
1549 pipeline
->vb_used
|= 1 << desc
->binding
;
1552 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1553 const VkVertexInputBindingDescription
*desc
=
1554 &vi_info
->pVertexBindingDescriptions
[i
];
1556 pipeline
->vb
[desc
->binding
].stride
= desc
->stride
;
1558 /* Step rate is programmed per vertex element (attribute), not
1559 * binding. Set up a map of which bindings step per instance, for
1560 * reference by vertex element setup. */
1561 switch (desc
->inputRate
) {
1563 case VK_VERTEX_INPUT_RATE_VERTEX
:
1564 pipeline
->vb
[desc
->binding
].instanced
= false;
1566 case VK_VERTEX_INPUT_RATE_INSTANCE
:
1567 pipeline
->vb
[desc
->binding
].instanced
= true;
1571 pipeline
->vb
[desc
->binding
].instance_divisor
= 1;
1574 const VkPipelineVertexInputDivisorStateCreateInfoEXT
*vi_div_state
=
1575 vk_find_struct_const(vi_info
->pNext
,
1576 PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT
);
1578 for (uint32_t i
= 0; i
< vi_div_state
->vertexBindingDivisorCount
; i
++) {
1579 const VkVertexInputBindingDivisorDescriptionEXT
*desc
=
1580 &vi_div_state
->pVertexBindingDivisors
[i
];
1582 pipeline
->vb
[desc
->binding
].instance_divisor
= desc
->divisor
;
1586 /* Our implementation of VK_KHR_multiview uses instancing to draw the
1587 * different views. If the client asks for instancing, we need to multiply
1588 * the instance divisor by the number of views ensure that we repeat the
1589 * client's per-instance data once for each view.
1591 if (pipeline
->subpass
->view_mask
) {
1592 const uint32_t view_count
= anv_subpass_view_count(pipeline
->subpass
);
1593 for (uint32_t vb
= 0; vb
< MAX_VBS
; vb
++) {
1594 if (pipeline
->vb
[vb
].instanced
)
1595 pipeline
->vb
[vb
].instance_divisor
*= view_count
;
1599 const VkPipelineInputAssemblyStateCreateInfo
*ia_info
=
1600 pCreateInfo
->pInputAssemblyState
;
1601 const VkPipelineTessellationStateCreateInfo
*tess_info
=
1602 pCreateInfo
->pTessellationState
;
1603 pipeline
->primitive_restart
= ia_info
->primitiveRestartEnable
;
1605 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
))
1606 pipeline
->topology
= _3DPRIM_PATCHLIST(tess_info
->patchControlPoints
);
1608 pipeline
->topology
= vk_to_gen_primitive_type
[ia_info
->topology
];