#include <fcntl.h>
#include "anv_private.h"
+#include "brw_nir.h"
+#include "anv_nir.h"
+#include "glsl/nir/nir_spirv.h"
+
+/* Needed for SWIZZLE macros */
+#include "program/prog_instruction.h"
// Shader functions
VkResult anv_CreateShaderModule(
VkDevice _device,
const VkShaderModuleCreateInfo* pCreateInfo,
+ const VkAllocationCallbacks* pAllocator,
VkShaderModule* pShaderModule)
{
ANV_FROM_HANDLE(anv_device, device, _device);
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
assert(pCreateInfo->flags == 0);
- module = anv_device_alloc(device, sizeof(*module) + pCreateInfo->codeSize, 8,
- VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
+ module = anv_alloc2(&device->alloc, pAllocator,
+ sizeof(*module) + pCreateInfo->codeSize, 8,
+ VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (module == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ module->nir = NULL;
module->size = pCreateInfo->codeSize;
memcpy(module->data, pCreateInfo->pCode, module->size);
return VK_SUCCESS;
}
-VkResult anv_DestroyShaderModule(
+void anv_DestroyShaderModule(
VkDevice _device,
- VkShaderModule _module)
+ VkShaderModule _module,
+ const VkAllocationCallbacks* pAllocator)
{
ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_shader_module, module, _module);
- anv_device_free(device, module);
-
- return VK_SUCCESS;
+ anv_free2(&device->alloc, pAllocator, module);
}
-VkResult anv_CreateShader(
- VkDevice _device,
- const VkShaderCreateInfo* pCreateInfo,
- VkShader* pShader)
-{
- ANV_FROM_HANDLE(anv_device, device, _device);
- ANV_FROM_HANDLE(anv_shader_module, module, pCreateInfo->module);
- struct anv_shader *shader;
+#define SPIR_V_MAGIC_NUMBER 0x07230203
- assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_CREATE_INFO);
- assert(pCreateInfo->flags == 0);
-
- size_t name_len = strlen(pCreateInfo->pName);
-
- if (strcmp(pCreateInfo->pName, "main") != 0) {
+/* Eventually, this will become part of anv_CreateShader. Unfortunately,
+ * we can't do that yet because we don't have the ability to copy nir.
+ */
+static nir_shader *
+anv_shader_compile_to_nir(struct anv_device *device,
+ struct anv_shader_module *module,
+ const char *entrypoint_name,
+ gl_shader_stage stage)
+{
+ if (strcmp(entrypoint_name, "main") != 0) {
anv_finishme("Multiple shaders per module not really supported");
}
- shader = anv_device_alloc(device, sizeof(*shader) + name_len + 1, 8,
- VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
- if (shader == NULL)
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ const struct brw_compiler *compiler =
+ device->instance->physicalDevice.compiler;
+ const nir_shader_compiler_options *nir_options =
+ compiler->glsl_compiler_options[stage].NirOptions;
+
+ nir_shader *nir;
+ if (module->nir) {
+ /* Some things such as our meta clear/blit code will give us a NIR
+ * shader directly. In that case, we just ignore the SPIR-V entirely
+ * and just use the NIR shader */
+ nir = module->nir;
+ nir->options = nir_options;
+ } else {
+ uint32_t *spirv = (uint32_t *) module->data;
+ assert(spirv[0] == SPIR_V_MAGIC_NUMBER);
+ assert(module->size % 4 == 0);
- shader->module = module;
- memcpy(shader->entrypoint, pCreateInfo->pName, name_len + 1);
+ nir = spirv_to_nir(spirv, module->size / 4, stage, nir_options);
+ }
+ nir_validate_shader(nir);
+
+ /* Vulkan uses the separate-shader linking model */
+ nir->info.separate_shader = true;
+
+ /* Make sure the provided shader has exactly one entrypoint and that the
+ * name matches the name that came in from the VkShader.
+ */
+ nir_function_impl *entrypoint = NULL;
+ nir_foreach_overload(nir, overload) {
+ if (strcmp(entrypoint_name, overload->function->name) == 0 &&
+ overload->impl) {
+ assert(entrypoint == NULL);
+ entrypoint = overload->impl;
+ }
+ }
+ assert(entrypoint != NULL);
- *pShader = anv_shader_to_handle(shader);
+ nir = brw_preprocess_nir(nir, compiler->scalar_stage[stage]);
- return VK_SUCCESS;
-}
+ nir_shader_gather_info(nir, entrypoint);
-VkResult anv_DestroyShader(
- VkDevice _device,
- VkShader _shader)
-{
- ANV_FROM_HANDLE(anv_device, device, _device);
- ANV_FROM_HANDLE(anv_shader, shader, _shader);
-
- anv_device_free(device, shader);
-
- return VK_SUCCESS;
+ return nir;
}
-
VkResult anv_CreatePipelineCache(
VkDevice device,
const VkPipelineCacheCreateInfo* pCreateInfo,
+ const VkAllocationCallbacks* pAllocator,
VkPipelineCache* pPipelineCache)
{
- pPipelineCache->handle = 1;
+ *pPipelineCache = (VkPipelineCache)1;
stub_return(VK_SUCCESS);
}
-VkResult anv_DestroyPipelineCache(
+void anv_DestroyPipelineCache(
VkDevice _device,
- VkPipelineCache _cache)
-{
- /* VkPipelineCache is a dummy object. */
- return VK_SUCCESS;
-}
-
-size_t anv_GetPipelineCacheSize(
- VkDevice device,
- VkPipelineCache pipelineCache)
+ VkPipelineCache _cache,
+ const VkAllocationCallbacks* pAllocator)
{
- stub_return(0);
}
VkResult anv_GetPipelineCacheData(
VkDevice device,
VkPipelineCache pipelineCache,
+ size_t* pDataSize,
void* pData)
{
- stub_return(VK_UNSUPPORTED);
+ *pDataSize = 0;
+ stub_return(VK_SUCCESS);
}
VkResult anv_MergePipelineCaches(
uint32_t srcCacheCount,
const VkPipelineCache* pSrcCaches)
{
- stub_return(VK_UNSUPPORTED);
+ stub_return(VK_SUCCESS);
}
-// Pipeline functions
-
-static void
-emit_vertex_input(struct anv_pipeline *pipeline,
- const VkPipelineVertexInputStateCreateInfo *info)
+void anv_DestroyPipeline(
+ VkDevice _device,
+ VkPipeline _pipeline,
+ const VkAllocationCallbacks* pAllocator)
{
- const uint32_t num_dwords = 1 + info->attributeCount * 2;
- uint32_t *p;
- bool instancing_enable[32];
+ ANV_FROM_HANDLE(anv_device, device, _device);
+ ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
- pipeline->vb_used = 0;
- for (uint32_t i = 0; i < info->bindingCount; i++) {
- const VkVertexInputBindingDescription *desc =
- &info->pVertexBindingDescriptions[i];
+ anv_reloc_list_finish(&pipeline->batch_relocs,
+ pAllocator ? pAllocator : &device->alloc);
+ anv_state_stream_finish(&pipeline->program_stream);
+ if (pipeline->blend_state.map)
+ anv_state_pool_free(&device->dynamic_state_pool, pipeline->blend_state);
+ anv_free2(&device->alloc, pAllocator, pipeline);
+}
- pipeline->vb_used |= 1 << desc->binding;
- pipeline->binding_stride[desc->binding] = desc->strideInBytes;
+static const uint32_t vk_to_gen_primitive_type[] = {
+ [VK_PRIMITIVE_TOPOLOGY_POINT_LIST] = _3DPRIM_POINTLIST,
+ [VK_PRIMITIVE_TOPOLOGY_LINE_LIST] = _3DPRIM_LINELIST,
+ [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP] = _3DPRIM_LINESTRIP,
+ [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST] = _3DPRIM_TRILIST,
+ [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP] = _3DPRIM_TRISTRIP,
+ [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN] = _3DPRIM_TRIFAN,
+ [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY] = _3DPRIM_LINELIST_ADJ,
+ [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY] = _3DPRIM_LINESTRIP_ADJ,
+ [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY] = _3DPRIM_TRILIST_ADJ,
+ [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY] = _3DPRIM_TRISTRIP_ADJ,
+/* [VK_PRIMITIVE_TOPOLOGY_PATCH_LIST] = _3DPRIM_PATCHLIST_1 */
+};
- /* Step rate is programmed per vertex element (attribute), not
- * binding. Set up a map of which bindings step per instance, for
- * reference by vertex element setup. */
- switch (desc->stepRate) {
- default:
- case VK_VERTEX_INPUT_STEP_RATE_VERTEX:
- instancing_enable[desc->binding] = false;
- break;
- case VK_VERTEX_INPUT_STEP_RATE_INSTANCE:
- instancing_enable[desc->binding] = true;
- break;
- }
+static void
+populate_sampler_prog_key(const struct brw_device_info *devinfo,
+ struct brw_sampler_prog_key_data *key)
+{
+ /* XXX: Handle texture swizzle on HSW- */
+ for (int i = 0; i < MAX_SAMPLERS; i++) {
+ /* Assume color sampler, no swizzling. (Works for BDW+) */
+ key->swizzles[i] = SWIZZLE_XYZW;
}
-
- p = anv_batch_emitn(&pipeline->batch, num_dwords,
- GEN8_3DSTATE_VERTEX_ELEMENTS);
-
- for (uint32_t i = 0; i < info->attributeCount; i++) {
- const VkVertexInputAttributeDescription *desc =
- &info->pVertexAttributeDescriptions[i];
- const struct anv_format *format = anv_format_for_vk_format(desc->format);
-
- struct GEN8_VERTEX_ELEMENT_STATE element = {
- .VertexBufferIndex = desc->binding,
- .Valid = true,
- .SourceElementFormat = format->surface_format,
- .EdgeFlagEnable = false,
- .SourceElementOffset = desc->offsetInBytes,
- .Component0Control = VFCOMP_STORE_SRC,
- .Component1Control = format->num_channels >= 2 ? VFCOMP_STORE_SRC : VFCOMP_STORE_0,
- .Component2Control = format->num_channels >= 3 ? VFCOMP_STORE_SRC : VFCOMP_STORE_0,
- .Component3Control = format->num_channels >= 4 ? VFCOMP_STORE_SRC : VFCOMP_STORE_1_FP
- };
- GEN8_VERTEX_ELEMENT_STATE_pack(NULL, &p[1 + i * 2], &element);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_VF_INSTANCING,
- .InstancingEnable = instancing_enable[desc->binding],
- .VertexElementIndex = i,
- /* Vulkan so far doesn't have an instance divisor, so
- * this is always 1 (ignored if not instancing). */
- .InstanceDataStepRate = 1);
- }
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_VF_SGVS,
- .VertexIDEnable = pipeline->vs_prog_data.uses_vertexid,
- .VertexIDComponentNumber = 2,
- .VertexIDElementOffset = info->bindingCount,
- .InstanceIDEnable = pipeline->vs_prog_data.uses_instanceid,
- .InstanceIDComponentNumber = 3,
- .InstanceIDElementOffset = info->bindingCount);
}
static void
-emit_ia_state(struct anv_pipeline *pipeline,
- const VkPipelineInputAssemblyStateCreateInfo *info,
- const struct anv_pipeline_create_info *extra)
+populate_vs_prog_key(const struct brw_device_info *devinfo,
+ struct brw_vs_prog_key *key)
{
- static const uint32_t vk_to_gen_primitive_type[] = {
- [VK_PRIMITIVE_TOPOLOGY_POINT_LIST] = _3DPRIM_POINTLIST,
- [VK_PRIMITIVE_TOPOLOGY_LINE_LIST] = _3DPRIM_LINELIST,
- [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP] = _3DPRIM_LINESTRIP,
- [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST] = _3DPRIM_TRILIST,
- [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP] = _3DPRIM_TRISTRIP,
- [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN] = _3DPRIM_TRIFAN,
- [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_ADJ] = _3DPRIM_LINELIST_ADJ,
- [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_ADJ] = _3DPRIM_LINESTRIP_ADJ,
- [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_ADJ] = _3DPRIM_TRILIST_ADJ,
- [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_ADJ] = _3DPRIM_TRISTRIP_ADJ,
- [VK_PRIMITIVE_TOPOLOGY_PATCH] = _3DPRIM_PATCHLIST_1
- };
- uint32_t topology = vk_to_gen_primitive_type[info->topology];
+ memset(key, 0, sizeof(*key));
- if (extra && extra->use_rectlist)
- topology = _3DPRIM_RECTLIST;
+ populate_sampler_prog_key(devinfo, &key->tex);
- struct GEN8_3DSTATE_VF vf = {
- GEN8_3DSTATE_VF_header,
- .IndexedDrawCutIndexEnable = info->primitiveRestartEnable,
- };
- GEN8_3DSTATE_VF_pack(NULL, pipeline->state_vf, &vf);
+ /* XXX: Handle vertex input work-arounds */
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_VF_TOPOLOGY,
- .PrimitiveTopologyType = topology);
+ /* XXX: Handle sampler_prog_key */
}
static void
-emit_rs_state(struct anv_pipeline *pipeline,
- const VkPipelineRasterStateCreateInfo *info,
- const struct anv_pipeline_create_info *extra)
+populate_gs_prog_key(const struct brw_device_info *devinfo,
+ struct brw_gs_prog_key *key)
{
- static const uint32_t vk_to_gen_cullmode[] = {
- [VK_CULL_MODE_NONE] = CULLMODE_NONE,
- [VK_CULL_MODE_FRONT] = CULLMODE_FRONT,
- [VK_CULL_MODE_BACK] = CULLMODE_BACK,
- [VK_CULL_MODE_FRONT_AND_BACK] = CULLMODE_BOTH
- };
+ memset(key, 0, sizeof(*key));
- static const uint32_t vk_to_gen_fillmode[] = {
- [VK_FILL_MODE_POINTS] = RASTER_POINT,
- [VK_FILL_MODE_WIREFRAME] = RASTER_WIREFRAME,
- [VK_FILL_MODE_SOLID] = RASTER_SOLID
- };
+ populate_sampler_prog_key(devinfo, &key->tex);
+}
- static const uint32_t vk_to_gen_front_face[] = {
- [VK_FRONT_FACE_CCW] = CounterClockwise,
- [VK_FRONT_FACE_CW] = Clockwise
- };
+static void
+populate_wm_prog_key(const struct brw_device_info *devinfo,
+ const VkGraphicsPipelineCreateInfo *info,
+ struct brw_wm_prog_key *key)
+{
+ ANV_FROM_HANDLE(anv_render_pass, render_pass, info->renderPass);
- struct GEN8_3DSTATE_SF sf = {
- GEN8_3DSTATE_SF_header,
- .ViewportTransformEnable = !(extra && extra->disable_viewport),
- .TriangleStripListProvokingVertexSelect = 0,
- .LineStripListProvokingVertexSelect = 0,
- .TriangleFanProvokingVertexSelect = 0,
- .PointWidthSource = pipeline->writes_point_size ? Vertex : State,
- .PointWidth = 1.0,
- };
+ memset(key, 0, sizeof(*key));
- /* FINISHME: VkBool32 rasterizerDiscardEnable; */
+ populate_sampler_prog_key(devinfo, &key->tex);
- GEN8_3DSTATE_SF_pack(NULL, pipeline->state_sf, &sf);
+ /* TODO: Fill out key->input_slots_valid */
- struct GEN8_3DSTATE_RASTER raster = {
- GEN8_3DSTATE_RASTER_header,
- .FrontWinding = vk_to_gen_front_face[info->frontFace],
- .CullMode = vk_to_gen_cullmode[info->cullMode],
- .FrontFaceFillMode = vk_to_gen_fillmode[info->fillMode],
- .BackFaceFillMode = vk_to_gen_fillmode[info->fillMode],
- .ScissorRectangleEnable = !(extra && extra->disable_scissor),
- .ViewportZClipTestEnable = info->depthClipEnable
- };
+ /* Vulkan doesn't specify a default */
+ key->high_quality_derivatives = false;
+
+ /* XXX Vulkan doesn't appear to specify */
+ key->clamp_fragment_color = false;
+
+ /* Vulkan always specifies upper-left coordinates */
+ key->drawable_height = 0;
+ key->render_to_fbo = false;
+
+ key->nr_color_regions = render_pass->subpasses[info->subpass].color_count;
- GEN8_3DSTATE_RASTER_pack(NULL, pipeline->state_raster, &raster);
+ key->replicate_alpha = key->nr_color_regions > 1 &&
+ info->pMultisampleState &&
+ info->pMultisampleState->alphaToCoverageEnable;
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_SBE,
- .ForceVertexURBEntryReadLength = false,
- .ForceVertexURBEntryReadOffset = false,
- .PointSpriteTextureCoordinateOrigin = UPPERLEFT,
- .NumberofSFOutputAttributes =
- pipeline->wm_prog_data.num_varying_inputs);
+ if (info->pMultisampleState && info->pMultisampleState->rasterizationSamples > 1) {
+ /* We should probably pull this out of the shader, but it's fairly
+ * harmless to compute it and then let dead-code take care of it.
+ */
+ key->persample_shading = info->pMultisampleState->sampleShadingEnable;
+ if (key->persample_shading)
+ key->persample_2x = info->pMultisampleState->rasterizationSamples == 2;
+ key->compute_pos_offset = info->pMultisampleState->sampleShadingEnable;
+ key->compute_sample_id = info->pMultisampleState->sampleShadingEnable;
+ }
}
static void
-emit_cb_state(struct anv_pipeline *pipeline,
- const VkPipelineColorBlendStateCreateInfo *info)
+populate_cs_prog_key(const struct brw_device_info *devinfo,
+ struct brw_cs_prog_key *key)
{
- struct anv_device *device = pipeline->device;
-
- static const uint32_t vk_to_gen_logic_op[] = {
- [VK_LOGIC_OP_COPY] = LOGICOP_COPY,
- [VK_LOGIC_OP_CLEAR] = LOGICOP_CLEAR,
- [VK_LOGIC_OP_AND] = LOGICOP_AND,
- [VK_LOGIC_OP_AND_REVERSE] = LOGICOP_AND_REVERSE,
- [VK_LOGIC_OP_AND_INVERTED] = LOGICOP_AND_INVERTED,
- [VK_LOGIC_OP_NOOP] = LOGICOP_NOOP,
- [VK_LOGIC_OP_XOR] = LOGICOP_XOR,
- [VK_LOGIC_OP_OR] = LOGICOP_OR,
- [VK_LOGIC_OP_NOR] = LOGICOP_NOR,
- [VK_LOGIC_OP_EQUIV] = LOGICOP_EQUIV,
- [VK_LOGIC_OP_INVERT] = LOGICOP_INVERT,
- [VK_LOGIC_OP_OR_REVERSE] = LOGICOP_OR_REVERSE,
- [VK_LOGIC_OP_COPY_INVERTED] = LOGICOP_COPY_INVERTED,
- [VK_LOGIC_OP_OR_INVERTED] = LOGICOP_OR_INVERTED,
- [VK_LOGIC_OP_NAND] = LOGICOP_NAND,
- [VK_LOGIC_OP_SET] = LOGICOP_SET,
- };
+ memset(key, 0, sizeof(*key));
- static const uint32_t vk_to_gen_blend[] = {
- [VK_BLEND_ZERO] = BLENDFACTOR_ZERO,
- [VK_BLEND_ONE] = BLENDFACTOR_ONE,
- [VK_BLEND_SRC_COLOR] = BLENDFACTOR_SRC_COLOR,
- [VK_BLEND_ONE_MINUS_SRC_COLOR] = BLENDFACTOR_INV_SRC_COLOR,
- [VK_BLEND_DEST_COLOR] = BLENDFACTOR_DST_COLOR,
- [VK_BLEND_ONE_MINUS_DEST_COLOR] = BLENDFACTOR_INV_DST_COLOR,
- [VK_BLEND_SRC_ALPHA] = BLENDFACTOR_SRC_ALPHA,
- [VK_BLEND_ONE_MINUS_SRC_ALPHA] = BLENDFACTOR_INV_SRC_ALPHA,
- [VK_BLEND_DEST_ALPHA] = BLENDFACTOR_DST_ALPHA,
- [VK_BLEND_ONE_MINUS_DEST_ALPHA] = BLENDFACTOR_INV_DST_ALPHA,
- [VK_BLEND_CONSTANT_COLOR] = BLENDFACTOR_CONST_COLOR,
- [VK_BLEND_ONE_MINUS_CONSTANT_COLOR] = BLENDFACTOR_INV_CONST_COLOR,
- [VK_BLEND_CONSTANT_ALPHA] = BLENDFACTOR_CONST_ALPHA,
- [VK_BLEND_ONE_MINUS_CONSTANT_ALPHA] = BLENDFACTOR_INV_CONST_ALPHA,
- [VK_BLEND_SRC_ALPHA_SATURATE] = BLENDFACTOR_SRC_ALPHA_SATURATE,
- [VK_BLEND_SRC1_COLOR] = BLENDFACTOR_SRC1_COLOR,
- [VK_BLEND_ONE_MINUS_SRC1_COLOR] = BLENDFACTOR_INV_SRC1_COLOR,
- [VK_BLEND_SRC1_ALPHA] = BLENDFACTOR_SRC1_ALPHA,
- [VK_BLEND_ONE_MINUS_SRC1_ALPHA] = BLENDFACTOR_INV_SRC1_ALPHA,
- };
+ populate_sampler_prog_key(devinfo, &key->tex);
+}
- static const uint32_t vk_to_gen_blend_op[] = {
- [VK_BLEND_OP_ADD] = BLENDFUNCTION_ADD,
- [VK_BLEND_OP_SUBTRACT] = BLENDFUNCTION_SUBTRACT,
- [VK_BLEND_OP_REVERSE_SUBTRACT] = BLENDFUNCTION_REVERSE_SUBTRACT,
- [VK_BLEND_OP_MIN] = BLENDFUNCTION_MIN,
- [VK_BLEND_OP_MAX] = BLENDFUNCTION_MAX,
- };
+static nir_shader *
+anv_pipeline_compile(struct anv_pipeline *pipeline,
+ struct anv_shader_module *module,
+ const char *entrypoint,
+ gl_shader_stage stage,
+ struct brw_stage_prog_data *prog_data)
+{
+ const struct brw_compiler *compiler =
+ pipeline->device->instance->physicalDevice.compiler;
- uint32_t num_dwords = 1 + info->attachmentCount * 2;
- pipeline->blend_state =
- anv_state_pool_alloc(&device->dynamic_state_pool, num_dwords * 4, 64);
+ nir_shader *nir = anv_shader_compile_to_nir(pipeline->device,
+ module, entrypoint, stage);
+ if (nir == NULL)
+ return NULL;
- struct GEN8_BLEND_STATE blend_state = {
- .AlphaToCoverageEnable = info->alphaToCoverageEnable,
- };
+ anv_nir_lower_push_constants(nir, compiler->scalar_stage[stage]);
+
+ /* Figure out the number of parameters */
+ prog_data->nr_params = 0;
+
+ if (nir->num_uniforms > 0) {
+ /* If the shader uses any push constants at all, we'll just give
+ * them the maximum possible number
+ */
+ prog_data->nr_params += MAX_PUSH_CONSTANTS_SIZE / sizeof(float);
+ }
+
+ if (pipeline->layout && pipeline->layout->stage[stage].has_dynamic_offsets)
+ prog_data->nr_params += MAX_DYNAMIC_BUFFERS * 2;
+
+ if (pipeline->layout && pipeline->layout->stage[stage].image_count > 0)
+ prog_data->nr_params += pipeline->layout->stage[stage].image_count *
+ BRW_IMAGE_PARAM_SIZE;
+
+ if (prog_data->nr_params > 0) {
+ /* XXX: I think we're leaking this */
+ prog_data->param = (const union gl_constant_value **)
+ malloc(prog_data->nr_params * sizeof(union gl_constant_value *));
- uint32_t *state = pipeline->blend_state.map;
- GEN8_BLEND_STATE_pack(NULL, state, &blend_state);
-
- for (uint32_t i = 0; i < info->attachmentCount; i++) {
- const VkPipelineColorBlendAttachmentState *a = &info->pAttachments[i];
-
- struct GEN8_BLEND_STATE_ENTRY entry = {
- .LogicOpEnable = info->logicOpEnable,
- .LogicOpFunction = vk_to_gen_logic_op[info->logicOp],
- .ColorBufferBlendEnable = a->blendEnable,
- .PreBlendSourceOnlyClampEnable = false,
- .PreBlendColorClampEnable = false,
- .PostBlendColorClampEnable = false,
- .SourceBlendFactor = vk_to_gen_blend[a->srcBlendColor],
- .DestinationBlendFactor = vk_to_gen_blend[a->destBlendColor],
- .ColorBlendFunction = vk_to_gen_blend_op[a->blendOpColor],
- .SourceAlphaBlendFactor = vk_to_gen_blend[a->srcBlendAlpha],
- .DestinationAlphaBlendFactor = vk_to_gen_blend[a->destBlendAlpha],
- .AlphaBlendFunction = vk_to_gen_blend_op[a->blendOpAlpha],
- .WriteDisableAlpha = !(a->channelWriteMask & VK_CHANNEL_A_BIT),
- .WriteDisableRed = !(a->channelWriteMask & VK_CHANNEL_R_BIT),
- .WriteDisableGreen = !(a->channelWriteMask & VK_CHANNEL_G_BIT),
- .WriteDisableBlue = !(a->channelWriteMask & VK_CHANNEL_B_BIT),
- };
-
- GEN8_BLEND_STATE_ENTRY_pack(NULL, state + i * 2 + 1, &entry);
- }
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_BLEND_STATE_POINTERS,
- .BlendStatePointer = pipeline->blend_state.offset,
- .BlendStatePointerValid = true);
+ /* We now set the param values to be offsets into a
+ * anv_push_constant_data structure. Since the compiler doesn't
+ * actually dereference any of the gl_constant_value pointers in the
+ * params array, it doesn't really matter what we put here.
+ */
+ struct anv_push_constants *null_data = NULL;
+ if (nir->num_uniforms > 0) {
+ /* Fill out the push constants section of the param array */
+ for (unsigned i = 0; i < MAX_PUSH_CONSTANTS_SIZE / sizeof(float); i++)
+ prog_data->param[i] = (const union gl_constant_value *)
+ &null_data->client_data[i * sizeof(float)];
+ }
+ }
+
+ /* Set up dynamic offsets */
+ anv_nir_apply_dynamic_offsets(pipeline, nir, prog_data);
+
+ /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
+ if (pipeline->layout)
+ anv_nir_apply_pipeline_layout(nir, prog_data, pipeline->layout);
+
+ /* All binding table offsets provided by apply_pipeline_layout() are
+ * relative to the start of the bindint table (plus MAX_RTS for VS).
+ */
+ unsigned bias = stage == MESA_SHADER_FRAGMENT ? MAX_RTS : 0;
+ prog_data->binding_table.size_bytes = 0;
+ prog_data->binding_table.texture_start = bias;
+ prog_data->binding_table.ubo_start = bias;
+ prog_data->binding_table.ssbo_start = bias;
+ prog_data->binding_table.image_start = bias;
+
+ /* Finish the optimization and compilation process */
+ nir = brw_lower_nir(nir, &pipeline->device->info, NULL,
+ compiler->scalar_stage[stage]);
+
+ /* nir_lower_io will only handle the push constants; we need to set this
+ * to the full number of possible uniforms.
+ */
+ nir->num_uniforms = prog_data->nr_params * 4;
+
+ return nir;
}
-static const uint32_t vk_to_gen_compare_op[] = {
- [VK_COMPARE_OP_NEVER] = COMPAREFUNCTION_NEVER,
- [VK_COMPARE_OP_LESS] = COMPAREFUNCTION_LESS,
- [VK_COMPARE_OP_EQUAL] = COMPAREFUNCTION_EQUAL,
- [VK_COMPARE_OP_LESS_EQUAL] = COMPAREFUNCTION_LEQUAL,
- [VK_COMPARE_OP_GREATER] = COMPAREFUNCTION_GREATER,
- [VK_COMPARE_OP_NOT_EQUAL] = COMPAREFUNCTION_NOTEQUAL,
- [VK_COMPARE_OP_GREATER_EQUAL] = COMPAREFUNCTION_GEQUAL,
- [VK_COMPARE_OP_ALWAYS] = COMPAREFUNCTION_ALWAYS,
-};
+static uint32_t
+anv_pipeline_upload_kernel(struct anv_pipeline *pipeline,
+ const void *data, size_t size)
+{
+ struct anv_state state =
+ anv_state_stream_alloc(&pipeline->program_stream, size, 64);
-static const uint32_t vk_to_gen_stencil_op[] = {
- [VK_STENCIL_OP_KEEP] = 0,
- [VK_STENCIL_OP_ZERO] = 0,
- [VK_STENCIL_OP_REPLACE] = 0,
- [VK_STENCIL_OP_INC_CLAMP] = 0,
- [VK_STENCIL_OP_DEC_CLAMP] = 0,
- [VK_STENCIL_OP_INVERT] = 0,
- [VK_STENCIL_OP_INC_WRAP] = 0,
- [VK_STENCIL_OP_DEC_WRAP] = 0
-};
+ assert(size < pipeline->program_stream.block_pool->block_size);
+
+ memcpy(state.map, data, size);
+
+ if (!pipeline->device->info.has_llc)
+ anv_state_clflush(state);
+
+ return state.offset;
+}
static void
-emit_ds_state(struct anv_pipeline *pipeline,
- const VkPipelineDepthStencilStateCreateInfo *info)
+anv_pipeline_add_compiled_stage(struct anv_pipeline *pipeline,
+ gl_shader_stage stage,
+ struct brw_stage_prog_data *prog_data)
{
- if (info == NULL) {
- /* We're going to OR this together with the dynamic state. We need
- * to make sure it's initialized to something useful.
- */
- memset(pipeline->state_wm_depth_stencil, 0,
- sizeof(pipeline->state_wm_depth_stencil));
- return;
- }
-
- /* VkBool32 depthBoundsEnable; // optional (depth_bounds_test) */
-
- struct GEN8_3DSTATE_WM_DEPTH_STENCIL wm_depth_stencil = {
- .DepthTestEnable = info->depthTestEnable,
- .DepthBufferWriteEnable = info->depthWriteEnable,
- .DepthTestFunction = vk_to_gen_compare_op[info->depthCompareOp],
- .DoubleSidedStencilEnable = true,
-
- .StencilTestEnable = info->stencilTestEnable,
- .StencilFailOp = vk_to_gen_stencil_op[info->front.stencilFailOp],
- .StencilPassDepthPassOp = vk_to_gen_stencil_op[info->front.stencilPassOp],
- .StencilPassDepthFailOp = vk_to_gen_stencil_op[info->front.stencilDepthFailOp],
- .StencilTestFunction = vk_to_gen_compare_op[info->front.stencilCompareOp],
- .BackfaceStencilFailOp = vk_to_gen_stencil_op[info->back.stencilFailOp],
- .BackfaceStencilPassDepthPassOp = vk_to_gen_stencil_op[info->back.stencilPassOp],
- .BackfaceStencilPassDepthFailOp =vk_to_gen_stencil_op[info->back.stencilDepthFailOp],
- .BackfaceStencilTestFunction = vk_to_gen_compare_op[info->back.stencilCompareOp],
+ struct brw_device_info *devinfo = &pipeline->device->info;
+ uint32_t max_threads[] = {
+ [MESA_SHADER_VERTEX] = devinfo->max_vs_threads,
+ [MESA_SHADER_TESS_CTRL] = 0,
+ [MESA_SHADER_TESS_EVAL] = 0,
+ [MESA_SHADER_GEOMETRY] = devinfo->max_gs_threads,
+ [MESA_SHADER_FRAGMENT] = devinfo->max_wm_threads,
+ [MESA_SHADER_COMPUTE] = devinfo->max_cs_threads,
};
- GEN8_3DSTATE_WM_DEPTH_STENCIL_pack(NULL, pipeline->state_wm_depth_stencil, &wm_depth_stencil);
+ pipeline->prog_data[stage] = prog_data;
+ pipeline->active_stages |= mesa_to_vk_shader_stage(stage);
+ pipeline->scratch_start[stage] = pipeline->total_scratch;
+ pipeline->total_scratch =
+ align_u32(pipeline->total_scratch, 1024) +
+ prog_data->total_scratch * max_threads[stage];
}
-VkResult
-anv_pipeline_create(
- VkDevice _device,
- const VkGraphicsPipelineCreateInfo* pCreateInfo,
- const struct anv_pipeline_create_info * extra,
- VkPipeline* pPipeline)
+static VkResult
+anv_pipeline_compile_vs(struct anv_pipeline *pipeline,
+ const VkGraphicsPipelineCreateInfo *info,
+ struct anv_shader_module *module,
+ const char *entrypoint)
{
- ANV_FROM_HANDLE(anv_device, device, _device);
- struct anv_pipeline *pipeline;
- VkResult result;
- uint32_t offset, length;
-
- assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO);
-
- pipeline = anv_device_alloc(device, sizeof(*pipeline), 8,
- VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
- if (pipeline == NULL)
+ const struct brw_compiler *compiler =
+ pipeline->device->instance->physicalDevice.compiler;
+ struct brw_vs_prog_data *prog_data = &pipeline->vs_prog_data;
+ struct brw_vs_prog_key key;
+
+ populate_vs_prog_key(&pipeline->device->info, &key);
+
+ /* TODO: Look up shader in cache */
+
+ memset(prog_data, 0, sizeof(*prog_data));
+
+ nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
+ MESA_SHADER_VERTEX,
+ &prog_data->base.base);
+ if (nir == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- pipeline->device = device;
- pipeline->layout = anv_pipeline_layout_from_handle(pCreateInfo->layout);
- memset(pipeline->shaders, 0, sizeof(pipeline->shaders));
+ void *mem_ctx = ralloc_context(NULL);
+
+ if (module->nir == NULL)
+ ralloc_steal(mem_ctx, nir);
+
+ prog_data->inputs_read = nir->info.inputs_read;
+ pipeline->writes_point_size = nir->info.outputs_written & VARYING_SLOT_PSIZ;
- result = anv_reloc_list_init(&pipeline->batch.relocs, device);
- if (result != VK_SUCCESS) {
- anv_device_free(device, pipeline);
- return result;
+ brw_compute_vue_map(&pipeline->device->info,
+ &prog_data->base.vue_map,
+ nir->info.outputs_written,
+ nir->info.separate_shader);
+
+ unsigned code_size;
+ const unsigned *shader_code =
+ brw_compile_vs(compiler, NULL, mem_ctx, &key, prog_data, nir,
+ NULL, false, -1, &code_size, NULL);
+ if (shader_code == NULL) {
+ ralloc_free(mem_ctx);
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
- pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
- pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
- anv_state_stream_init(&pipeline->program_stream,
- &device->instruction_block_pool);
+ const uint32_t offset =
+ anv_pipeline_upload_kernel(pipeline, shader_code, code_size);
+ if (prog_data->base.dispatch_mode == DISPATCH_MODE_SIMD8) {
+ pipeline->vs_simd8 = offset;
+ pipeline->vs_vec4 = NO_KERNEL;
+ } else {
+ pipeline->vs_simd8 = NO_KERNEL;
+ pipeline->vs_vec4 = offset;
+ }
- for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
- pipeline->shaders[pCreateInfo->pStages[i].stage] =
- anv_shader_from_handle(pCreateInfo->pStages[i].shader);
+ ralloc_free(mem_ctx);
+
+ anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_VERTEX,
+ &prog_data->base.base);
+
+ return VK_SUCCESS;
+}
+
+static VkResult
+anv_pipeline_compile_gs(struct anv_pipeline *pipeline,
+ const VkGraphicsPipelineCreateInfo *info,
+ struct anv_shader_module *module,
+ const char *entrypoint)
+{
+ const struct brw_compiler *compiler =
+ pipeline->device->instance->physicalDevice.compiler;
+ struct brw_gs_prog_data *prog_data = &pipeline->gs_prog_data;
+ struct brw_gs_prog_key key;
+
+ populate_gs_prog_key(&pipeline->device->info, &key);
+
+ /* TODO: Look up shader in cache */
+
+ memset(prog_data, 0, sizeof(*prog_data));
+
+ nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
+ MESA_SHADER_GEOMETRY,
+ &prog_data->base.base);
+ if (nir == NULL)
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+ void *mem_ctx = ralloc_context(NULL);
+
+ if (module->nir == NULL)
+ ralloc_steal(mem_ctx, nir);
+
+ brw_compute_vue_map(&pipeline->device->info,
+ &prog_data->base.vue_map,
+ nir->info.outputs_written,
+ nir->info.separate_shader);
+
+ unsigned code_size;
+ const unsigned *shader_code =
+ brw_compile_gs(compiler, NULL, mem_ctx, &key, prog_data, nir,
+ NULL, -1, &code_size, NULL);
+ if (shader_code == NULL) {
+ ralloc_free(mem_ctx);
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
- if (pCreateInfo->pTessellationState)
- anv_finishme("VK_STRUCTURE_TYPE_PIPELINE_TESSELLATION_STATE_CREATE_INFO");
- if (pCreateInfo->pViewportState)
- anv_finishme("VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO");
- if (pCreateInfo->pMultisampleState)
- anv_finishme("VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO");
+ /* TODO: SIMD8 GS */
+ pipeline->gs_kernel =
+ anv_pipeline_upload_kernel(pipeline, shader_code, code_size);
+ pipeline->gs_vertex_count = nir->info.gs.vertices_in;
- pipeline->use_repclear = extra && extra->use_repclear;
+ ralloc_free(mem_ctx);
- anv_compiler_run(device->compiler, pipeline);
-
- /* FIXME: The compiler dead-codes FS inputs when we don't have a VS, so we
- * hard code this to num_attributes - 2. This is because the attributes
- * include VUE header and position, which aren't counted as varying
- * inputs. */
- if (pipeline->vs_simd8 == NO_KERNEL) {
- pipeline->wm_prog_data.num_varying_inputs =
- pCreateInfo->pVertexInputState->attributeCount - 2;
- }
-
- assert(pCreateInfo->pVertexInputState);
- emit_vertex_input(pipeline, pCreateInfo->pVertexInputState);
- assert(pCreateInfo->pInputAssemblyState);
- emit_ia_state(pipeline, pCreateInfo->pInputAssemblyState, extra);
- assert(pCreateInfo->pRasterState);
- emit_rs_state(pipeline, pCreateInfo->pRasterState, extra);
- emit_ds_state(pipeline, pCreateInfo->pDepthStencilState);
- emit_cb_state(pipeline, pCreateInfo->pColorBlendState);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_VF_STATISTICS,
- .StatisticsEnable = true);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_HS, .Enable = false);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_TE, .TEEnable = false);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_DS, .FunctionEnable = false);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_STREAMOUT, .SOFunctionEnable = false);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_PUSH_CONSTANT_ALLOC_VS,
- .ConstantBufferOffset = 0,
- .ConstantBufferSize = 4);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_PUSH_CONSTANT_ALLOC_GS,
- .ConstantBufferOffset = 4,
- .ConstantBufferSize = 4);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_PUSH_CONSTANT_ALLOC_PS,
- .ConstantBufferOffset = 8,
- .ConstantBufferSize = 4);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_WM_CHROMAKEY,
- .ChromaKeyKillEnable = false);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_SBE_SWIZ);
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_AA_LINE_PARAMETERS);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_CLIP,
- .ClipEnable = true,
- .ViewportXYClipTestEnable = !(extra && extra->disable_viewport),
- .MinimumPointWidth = 0.125,
- .MaximumPointWidth = 255.875);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_WM,
- .StatisticsEnable = true,
- .LineEndCapAntialiasingRegionWidth = _05pixels,
- .LineAntialiasingRegionWidth = _10pixels,
- .EarlyDepthStencilControl = NORMAL,
- .ForceThreadDispatchEnable = NORMAL,
- .PointRasterizationRule = RASTRULE_UPPER_RIGHT,
- .BarycentricInterpolationMode =
- pipeline->wm_prog_data.barycentric_interp_modes);
-
- uint32_t samples = 1;
- uint32_t log2_samples = __builtin_ffs(samples) - 1;
- bool enable_sampling = samples > 1 ? true : false;
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_MULTISAMPLE,
- .PixelPositionOffsetEnable = enable_sampling,
- .PixelLocation = CENTER,
- .NumberofMultisamples = log2_samples);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_SAMPLE_MASK,
- .SampleMask = 0xffff);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_URB_VS,
- .VSURBStartingAddress = pipeline->urb.vs_start,
- .VSURBEntryAllocationSize = pipeline->urb.vs_size - 1,
- .VSNumberofURBEntries = pipeline->urb.nr_vs_entries);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_URB_GS,
- .GSURBStartingAddress = pipeline->urb.gs_start,
- .GSURBEntryAllocationSize = pipeline->urb.gs_size - 1,
- .GSNumberofURBEntries = pipeline->urb.nr_gs_entries);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_URB_HS,
- .HSURBStartingAddress = pipeline->urb.vs_start,
- .HSURBEntryAllocationSize = 0,
- .HSNumberofURBEntries = 0);
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_URB_DS,
- .DSURBStartingAddress = pipeline->urb.vs_start,
- .DSURBEntryAllocationSize = 0,
- .DSNumberofURBEntries = 0);
-
- const struct brw_gs_prog_data *gs_prog_data = &pipeline->gs_prog_data;
- offset = 1;
- length = (gs_prog_data->base.vue_map.num_slots + 1) / 2 - offset;
-
- if (pipeline->gs_vec4 == NO_KERNEL)
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_GS, .Enable = false);
- else
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_GS,
- .SingleProgramFlow = false,
- .KernelStartPointer = pipeline->gs_vec4,
- .VectorMaskEnable = Vmask,
- .SamplerCount = 0,
- .BindingTableEntryCount = 0,
- .ExpectedVertexCount = pipeline->gs_vertex_count,
-
- .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_GEOMETRY],
- .PerThreadScratchSpace = ffs(gs_prog_data->base.base.total_scratch / 2048),
-
- .OutputVertexSize = gs_prog_data->output_vertex_size_hwords * 2 - 1,
- .OutputTopology = gs_prog_data->output_topology,
- .VertexURBEntryReadLength = gs_prog_data->base.urb_read_length,
- .DispatchGRFStartRegisterForURBData =
- gs_prog_data->base.base.dispatch_grf_start_reg,
-
- .MaximumNumberofThreads = device->info.max_gs_threads,
- .ControlDataHeaderSize = gs_prog_data->control_data_header_size_hwords,
- //pipeline->gs_prog_data.dispatch_mode |
- .StatisticsEnable = true,
- .IncludePrimitiveID = gs_prog_data->include_primitive_id,
- .ReorderMode = TRAILING,
- .Enable = true,
-
- .ControlDataFormat = gs_prog_data->control_data_format,
-
- /* FIXME: mesa sets this based on ctx->Transform.ClipPlanesEnabled:
- * UserClipDistanceClipTestEnableBitmask_3DSTATE_GS(v)
- * UserClipDistanceCullTestEnableBitmask(v)
- */
-
- .VertexURBEntryOutputReadOffset = offset,
- .VertexURBEntryOutputLength = length);
-
- const struct brw_vue_prog_data *vue_prog_data = &pipeline->vs_prog_data.base;
- /* Skip the VUE header and position slots */
- offset = 1;
- length = (vue_prog_data->vue_map.num_slots + 1) / 2 - offset;
-
- if (pipeline->vs_simd8 == NO_KERNEL || (extra && extra->disable_vs))
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_VS,
- .FunctionEnable = false,
- .VertexURBEntryOutputReadOffset = 1,
- /* Even if VS is disabled, SBE still gets the amount of
- * vertex data to read from this field. We use attribute
- * count - 1, as we don't count the VUE header here. */
- .VertexURBEntryOutputLength =
- DIV_ROUND_UP(pCreateInfo->pVertexInputState->attributeCount - 1, 2));
+ anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_GEOMETRY,
+ &prog_data->base.base);
+
+ return VK_SUCCESS;
+}
+
+static VkResult
+anv_pipeline_compile_fs(struct anv_pipeline *pipeline,
+ const VkGraphicsPipelineCreateInfo *info,
+ struct anv_shader_module *module,
+ const char *entrypoint)
+{
+ const struct brw_compiler *compiler =
+ pipeline->device->instance->physicalDevice.compiler;
+ struct brw_wm_prog_data *prog_data = &pipeline->wm_prog_data;
+ struct brw_wm_prog_key key;
+
+ populate_wm_prog_key(&pipeline->device->info, info, &key);
+
+ if (pipeline->use_repclear)
+ key.nr_color_regions = 1;
+
+ /* TODO: Look up shader in cache */
+
+ memset(prog_data, 0, sizeof(*prog_data));
+
+ prog_data->binding_table.render_target_start = 0;
+
+ nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
+ MESA_SHADER_FRAGMENT,
+ &prog_data->base);
+ if (nir == NULL)
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+ void *mem_ctx = ralloc_context(NULL);
+
+ if (module->nir == NULL)
+ ralloc_steal(mem_ctx, nir);
+
+ unsigned code_size;
+ const unsigned *shader_code =
+ brw_compile_fs(compiler, NULL, mem_ctx, &key, prog_data, nir,
+ NULL, -1, -1, pipeline->use_repclear, &code_size, NULL);
+ if (shader_code == NULL) {
+ ralloc_free(mem_ctx);
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ }
+
+ uint32_t offset = anv_pipeline_upload_kernel(pipeline,
+ shader_code, code_size);
+ if (prog_data->no_8)
+ pipeline->ps_simd8 = NO_KERNEL;
else
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_VS,
- .KernelStartPointer = pipeline->vs_simd8,
- .SingleVertexDispatch = Multiple,
- .VectorMaskEnable = Dmask,
- .SamplerCount = 0,
- .BindingTableEntryCount =
- vue_prog_data->base.binding_table.size_bytes / 4,
- .ThreadDispatchPriority = Normal,
- .FloatingPointMode = IEEE754,
- .IllegalOpcodeExceptionEnable = false,
- .AccessesUAV = false,
- .SoftwareExceptionEnable = false,
-
- .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_VERTEX],
- .PerThreadScratchSpace = ffs(vue_prog_data->base.total_scratch / 2048),
-
- .DispatchGRFStartRegisterForURBData =
- vue_prog_data->base.dispatch_grf_start_reg,
- .VertexURBEntryReadLength = vue_prog_data->urb_read_length,
- .VertexURBEntryReadOffset = 0,
-
- .MaximumNumberofThreads = device->info.max_vs_threads - 1,
- .StatisticsEnable = false,
- .SIMD8DispatchEnable = true,
- .VertexCacheDisable = false,
- .FunctionEnable = true,
-
- .VertexURBEntryOutputReadOffset = offset,
- .VertexURBEntryOutputLength = length,
- .UserClipDistanceClipTestEnableBitmask = 0,
- .UserClipDistanceCullTestEnableBitmask = 0);
-
- const struct brw_wm_prog_data *wm_prog_data = &pipeline->wm_prog_data;
- uint32_t ksp0, ksp2, grf_start0, grf_start2;
-
- ksp2 = 0;
- grf_start2 = 0;
+ pipeline->ps_simd8 = offset;
+
+ if (prog_data->no_8 || prog_data->prog_offset_16) {
+ pipeline->ps_simd16 = offset + prog_data->prog_offset_16;
+ } else {
+ pipeline->ps_simd16 = NO_KERNEL;
+ }
+
+ pipeline->ps_ksp2 = 0;
+ pipeline->ps_grf_start2 = 0;
if (pipeline->ps_simd8 != NO_KERNEL) {
- ksp0 = pipeline->ps_simd8;
- grf_start0 = wm_prog_data->base.dispatch_grf_start_reg;
+ pipeline->ps_ksp0 = pipeline->ps_simd8;
+ pipeline->ps_grf_start0 = prog_data->base.dispatch_grf_start_reg;
if (pipeline->ps_simd16 != NO_KERNEL) {
- ksp2 = pipeline->ps_simd16;
- grf_start2 = wm_prog_data->dispatch_grf_start_reg_16;
+ pipeline->ps_ksp2 = pipeline->ps_simd16;
+ pipeline->ps_grf_start2 = prog_data->dispatch_grf_start_reg_16;
}
} else if (pipeline->ps_simd16 != NO_KERNEL) {
- ksp0 = pipeline->ps_simd16;
- grf_start0 = wm_prog_data->dispatch_grf_start_reg_16;
- } else {
- unreachable("no ps shader");
- }
-
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_PS,
- .KernelStartPointer0 = ksp0,
-
- .SingleProgramFlow = false,
- .VectorMaskEnable = true,
- .SamplerCount = 1,
-
- .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_FRAGMENT],
- .PerThreadScratchSpace = ffs(wm_prog_data->base.total_scratch / 2048),
-
- .MaximumNumberofThreadsPerPSD = 64 - 2,
- .PositionXYOffsetSelect = wm_prog_data->uses_pos_offset ?
- POSOFFSET_SAMPLE: POSOFFSET_NONE,
- .PushConstantEnable = wm_prog_data->base.nr_params > 0,
- ._8PixelDispatchEnable = pipeline->ps_simd8 != NO_KERNEL,
- ._16PixelDispatchEnable = pipeline->ps_simd16 != NO_KERNEL,
- ._32PixelDispatchEnable = false,
-
- .DispatchGRFStartRegisterForConstantSetupData0 = grf_start0,
- .DispatchGRFStartRegisterForConstantSetupData1 = 0,
- .DispatchGRFStartRegisterForConstantSetupData2 = grf_start2,
-
- .KernelStartPointer1 = 0,
- .KernelStartPointer2 = ksp2);
-
- bool per_sample_ps = false;
- anv_batch_emit(&pipeline->batch, GEN8_3DSTATE_PS_EXTRA,
- .PixelShaderValid = true,
- .PixelShaderKillsPixel = wm_prog_data->uses_kill,
- .PixelShaderComputedDepthMode = wm_prog_data->computed_depth_mode,
- .AttributeEnable = wm_prog_data->num_varying_inputs > 0,
- .oMaskPresenttoRenderTarget = wm_prog_data->uses_omask,
- .PixelShaderIsPerSample = per_sample_ps);
-
- *pPipeline = anv_pipeline_to_handle(pipeline);
+ pipeline->ps_ksp0 = pipeline->ps_simd16;
+ pipeline->ps_grf_start0 = prog_data->dispatch_grf_start_reg_16;
+ }
+
+ ralloc_free(mem_ctx);
+
+ anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_FRAGMENT,
+ &prog_data->base);
return VK_SUCCESS;
}
-VkResult anv_DestroyPipeline(
- VkDevice _device,
- VkPipeline _pipeline)
+VkResult
+anv_pipeline_compile_cs(struct anv_pipeline *pipeline,
+ const VkComputePipelineCreateInfo *info,
+ struct anv_shader_module *module,
+ const char *entrypoint)
{
- ANV_FROM_HANDLE(anv_device, device, _device);
- ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
+ const struct brw_compiler *compiler =
+ pipeline->device->instance->physicalDevice.compiler;
+ struct brw_cs_prog_data *prog_data = &pipeline->cs_prog_data;
+ struct brw_cs_prog_key key;
- anv_compiler_free(pipeline);
- anv_reloc_list_finish(&pipeline->batch.relocs, pipeline->device);
- anv_state_stream_finish(&pipeline->program_stream);
- anv_state_pool_free(&device->dynamic_state_pool, pipeline->blend_state);
- anv_device_free(pipeline->device, pipeline);
+ populate_cs_prog_key(&pipeline->device->info, &key);
+
+ /* TODO: Look up shader in cache */
+
+ memset(prog_data, 0, sizeof(*prog_data));
+
+ nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
+ MESA_SHADER_COMPUTE,
+ &prog_data->base);
+ if (nir == NULL)
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+ void *mem_ctx = ralloc_context(NULL);
+
+ if (module->nir == NULL)
+ ralloc_steal(mem_ctx, nir);
+
+ unsigned code_size;
+ const unsigned *shader_code =
+ brw_compile_cs(compiler, NULL, mem_ctx, &key, prog_data, nir,
+ -1, &code_size, NULL);
+ if (shader_code == NULL) {
+ ralloc_free(mem_ctx);
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ }
+
+ pipeline->cs_simd = anv_pipeline_upload_kernel(pipeline,
+ shader_code, code_size);
+ ralloc_free(mem_ctx);
+
+ anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_COMPUTE,
+ &prog_data->base);
return VK_SUCCESS;
}
-VkResult anv_CreateGraphicsPipelines(
- VkDevice _device,
- VkPipelineCache pipelineCache,
- uint32_t count,
- const VkGraphicsPipelineCreateInfo* pCreateInfos,
- VkPipeline* pPipelines)
+static const int gen8_push_size = 32 * 1024;
+
+static void
+gen7_compute_urb_partition(struct anv_pipeline *pipeline)
{
- VkResult result = VK_SUCCESS;
+ const struct brw_device_info *devinfo = &pipeline->device->info;
+ bool vs_present = pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT;
+ unsigned vs_size = vs_present ? pipeline->vs_prog_data.base.urb_entry_size : 1;
+ unsigned vs_entry_size_bytes = vs_size * 64;
+ bool gs_present = pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT;
+ unsigned gs_size = gs_present ? pipeline->gs_prog_data.base.urb_entry_size : 1;
+ unsigned gs_entry_size_bytes = gs_size * 64;
+
+ /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
+ *
+ * VS Number of URB Entries must be divisible by 8 if the VS URB Entry
+ * Allocation Size is less than 9 512-bit URB entries.
+ *
+ * Similar text exists for GS.
+ */
+ unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
+ unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
+
+ /* URB allocations must be done in 8k chunks. */
+ unsigned chunk_size_bytes = 8192;
+
+ /* Determine the size of the URB in chunks. */
+ unsigned urb_chunks = devinfo->urb.size * 1024 / chunk_size_bytes;
+
+ /* Reserve space for push constants */
+ unsigned push_constant_bytes = gen8_push_size;
+ unsigned push_constant_chunks =
+ push_constant_bytes / chunk_size_bytes;
+
+ /* Initially, assign each stage the minimum amount of URB space it needs,
+ * and make a note of how much additional space it "wants" (the amount of
+ * additional space it could actually make use of).
+ */
+
+ /* VS has a lower limit on the number of URB entries */
+ unsigned vs_chunks =
+ ALIGN(devinfo->urb.min_vs_entries * vs_entry_size_bytes,
+ chunk_size_bytes) / chunk_size_bytes;
+ unsigned vs_wants =
+ ALIGN(devinfo->urb.max_vs_entries * vs_entry_size_bytes,
+ chunk_size_bytes) / chunk_size_bytes - vs_chunks;
+
+ unsigned gs_chunks = 0;
+ unsigned gs_wants = 0;
+ if (gs_present) {
+ /* There are two constraints on the minimum amount of URB space we can
+ * allocate:
+ *
+ * (1) We need room for at least 2 URB entries, since we always operate
+ * the GS in DUAL_OBJECT mode.
+ *
+ * (2) We can't allocate less than nr_gs_entries_granularity.
+ */
+ gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
+ chunk_size_bytes) / chunk_size_bytes;
+ gs_wants =
+ ALIGN(devinfo->urb.max_gs_entries * gs_entry_size_bytes,
+ chunk_size_bytes) / chunk_size_bytes - gs_chunks;
+ }
- unsigned i = 0;
- for (; i < count; i++) {
- result = anv_pipeline_create(_device, &pCreateInfos[i],
- NULL, &pPipelines[i]);
- if (result != VK_SUCCESS) {
- for (unsigned j = 0; j < i; j++) {
- anv_DestroyPipeline(_device, pPipelines[j]);
- }
+ /* There should always be enough URB space to satisfy the minimum
+ * requirements of each stage.
+ */
+ unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
+ assert(total_needs <= urb_chunks);
+
+ /* Mete out remaining space (if any) in proportion to "wants". */
+ unsigned total_wants = vs_wants + gs_wants;
+ unsigned remaining_space = urb_chunks - total_needs;
+ if (remaining_space > total_wants)
+ remaining_space = total_wants;
+ if (remaining_space > 0) {
+ unsigned vs_additional = (unsigned)
+ round(vs_wants * (((double) remaining_space) / total_wants));
+ vs_chunks += vs_additional;
+ remaining_space -= vs_additional;
+ gs_chunks += remaining_space;
+ }
- return result;
+ /* Sanity check that we haven't over-allocated. */
+ assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
+
+ /* Finally, compute the number of entries that can fit in the space
+ * allocated to each stage.
+ */
+ unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
+ unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
+
+ /* Since we rounded up when computing *_wants, this may be slightly more
+ * than the maximum allowed amount, so correct for that.
+ */
+ nr_vs_entries = MIN2(nr_vs_entries, devinfo->urb.max_vs_entries);
+ nr_gs_entries = MIN2(nr_gs_entries, devinfo->urb.max_gs_entries);
+
+ /* Ensure that we program a multiple of the granularity. */
+ nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
+ nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
+
+ /* Finally, sanity check to make sure we have at least the minimum number
+ * of entries needed for each stage.
+ */
+ assert(nr_vs_entries >= devinfo->urb.min_vs_entries);
+ if (gs_present)
+ assert(nr_gs_entries >= 2);
+
+ /* Lay out the URB in the following order:
+ * - push constants
+ * - VS
+ * - GS
+ */
+ pipeline->urb.vs_start = push_constant_chunks;
+ pipeline->urb.vs_size = vs_size;
+ pipeline->urb.nr_vs_entries = nr_vs_entries;
+
+ pipeline->urb.gs_start = push_constant_chunks + vs_chunks;
+ pipeline->urb.gs_size = gs_size;
+ pipeline->urb.nr_gs_entries = nr_gs_entries;
+}
+
+static void
+anv_pipeline_init_dynamic_state(struct anv_pipeline *pipeline,
+ const VkGraphicsPipelineCreateInfo *pCreateInfo)
+{
+ anv_cmd_dirty_mask_t states = ANV_CMD_DIRTY_DYNAMIC_ALL;
+ ANV_FROM_HANDLE(anv_render_pass, pass, pCreateInfo->renderPass);
+ struct anv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
+
+ pipeline->dynamic_state = default_dynamic_state;
+
+ if (pCreateInfo->pDynamicState) {
+ /* Remove all of the states that are marked as dynamic */
+ uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
+ for (uint32_t s = 0; s < count; s++)
+ states &= ~(1 << pCreateInfo->pDynamicState->pDynamicStates[s]);
+ }
+
+ struct anv_dynamic_state *dynamic = &pipeline->dynamic_state;
+
+ dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount;
+ if (states & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
+ typed_memcpy(dynamic->viewport.viewports,
+ pCreateInfo->pViewportState->pViewports,
+ pCreateInfo->pViewportState->viewportCount);
+ }
+
+ dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount;
+ if (states & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
+ typed_memcpy(dynamic->scissor.scissors,
+ pCreateInfo->pViewportState->pScissors,
+ pCreateInfo->pViewportState->scissorCount);
+ }
+
+ if (states & (1 << VK_DYNAMIC_STATE_LINE_WIDTH)) {
+ assert(pCreateInfo->pRasterizationState);
+ dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth;
+ }
+
+ if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS)) {
+ assert(pCreateInfo->pRasterizationState);
+ dynamic->depth_bias.bias =
+ pCreateInfo->pRasterizationState->depthBiasConstantFactor;
+ dynamic->depth_bias.clamp =
+ pCreateInfo->pRasterizationState->depthBiasClamp;
+ dynamic->depth_bias.slope =
+ pCreateInfo->pRasterizationState->depthBiasSlopeFactor;
+ }
+
+ if (states & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS)) {
+ assert(pCreateInfo->pColorBlendState);
+ typed_memcpy(dynamic->blend_constants,
+ pCreateInfo->pColorBlendState->blendConstants, 4);
+ }
+
+ /* If there is no depthstencil attachment, then don't read
+ * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
+ * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
+ * no need to override the depthstencil defaults in
+ * anv_pipeline::dynamic_state when there is no depthstencil attachment.
+ *
+ * From the Vulkan spec (20 Oct 2015, git-aa308cb):
+ *
+ * pDepthStencilState [...] may only be NULL if renderPass and subpass
+ * specify a subpass that has no depth/stencil attachment.
+ */
+ if (subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED) {
+ if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS)) {
+ assert(pCreateInfo->pDepthStencilState);
+ dynamic->depth_bounds.min =
+ pCreateInfo->pDepthStencilState->minDepthBounds;
+ dynamic->depth_bounds.max =
+ pCreateInfo->pDepthStencilState->maxDepthBounds;
+ }
+
+ if (states & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK)) {
+ assert(pCreateInfo->pDepthStencilState);
+ dynamic->stencil_compare_mask.front =
+ pCreateInfo->pDepthStencilState->front.compareMask;
+ dynamic->stencil_compare_mask.back =
+ pCreateInfo->pDepthStencilState->back.compareMask;
+ }
+
+ if (states & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK)) {
+ assert(pCreateInfo->pDepthStencilState);
+ dynamic->stencil_write_mask.front =
+ pCreateInfo->pDepthStencilState->front.writeMask;
+ dynamic->stencil_write_mask.back =
+ pCreateInfo->pDepthStencilState->back.writeMask;
+ }
+
+ if (states & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE)) {
+ assert(pCreateInfo->pDepthStencilState);
+ dynamic->stencil_reference.front =
+ pCreateInfo->pDepthStencilState->front.reference;
+ dynamic->stencil_reference.back =
+ pCreateInfo->pDepthStencilState->back.reference;
}
}
- return VK_SUCCESS;
+ pipeline->dynamic_state_mask = states;
}
-static VkResult anv_compute_pipeline_create(
- VkDevice _device,
- const VkComputePipelineCreateInfo* pCreateInfo,
- VkPipeline* pPipeline)
+static void
+anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo *info)
{
- ANV_FROM_HANDLE(anv_device, device, _device);
- struct anv_pipeline *pipeline;
- VkResult result;
+ struct anv_render_pass *renderpass = NULL;
+ struct anv_subpass *subpass = NULL;
- assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO);
+ /* Assert that all required members of VkGraphicsPipelineCreateInfo are
+ * present, as explained by the Vulkan (20 Oct 2015, git-aa308cb), Section
+ * 4.2 Graphics Pipeline.
+ */
+ assert(info->sType == VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO);
- pipeline = anv_device_alloc(device, sizeof(*pipeline), 8,
- VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
- if (pipeline == NULL)
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ renderpass = anv_render_pass_from_handle(info->renderPass);
+ assert(renderpass);
+
+ if (renderpass != &anv_meta_dummy_renderpass) {
+ assert(info->subpass < renderpass->subpass_count);
+ subpass = &renderpass->subpasses[info->subpass];
+ }
+
+ assert(info->stageCount >= 1);
+ assert(info->pVertexInputState);
+ assert(info->pInputAssemblyState);
+ assert(info->pViewportState);
+ assert(info->pRasterizationState);
+
+ if (subpass && subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED)
+ assert(info->pDepthStencilState);
+
+ if (subpass && subpass->color_count > 0)
+ assert(info->pColorBlendState);
+
+ for (uint32_t i = 0; i < info->stageCount; ++i) {
+ switch (info->pStages[i].stage) {
+ case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
+ case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
+ assert(info->pTessellationState);
+ break;
+ default:
+ break;
+ }
+ }
+}
+
+VkResult
+anv_pipeline_init(struct anv_pipeline *pipeline, struct anv_device *device,
+ const VkGraphicsPipelineCreateInfo *pCreateInfo,
+ const struct anv_graphics_pipeline_create_info *extra,
+ const VkAllocationCallbacks *alloc)
+{
+ anv_validate {
+ anv_pipeline_validate_create_info(pCreateInfo);
+ }
+
+ if (alloc == NULL)
+ alloc = &device->alloc;
pipeline->device = device;
pipeline->layout = anv_pipeline_layout_from_handle(pCreateInfo->layout);
- result = anv_reloc_list_init(&pipeline->batch.relocs, device);
- if (result != VK_SUCCESS) {
- anv_device_free(device, pipeline);
- return result;
- }
+ anv_reloc_list_init(&pipeline->batch_relocs, alloc);
+ /* TODO: Handle allocation fail */
+
+ pipeline->batch.alloc = alloc;
pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
+ pipeline->batch.relocs = &pipeline->batch_relocs;
anv_state_stream_init(&pipeline->program_stream,
&device->instruction_block_pool);
- memset(pipeline->shaders, 0, sizeof(pipeline->shaders));
+ anv_pipeline_init_dynamic_state(pipeline, pCreateInfo);
- pipeline->shaders[VK_SHADER_STAGE_COMPUTE] =
- anv_shader_from_handle(pCreateInfo->cs.shader);
+ if (pCreateInfo->pTessellationState)
+ anv_finishme("VK_STRUCTURE_TYPE_PIPELINE_TESSELLATION_STATE_CREATE_INFO");
+ if (pCreateInfo->pMultisampleState &&
+ pCreateInfo->pMultisampleState->rasterizationSamples > 1)
+ anv_finishme("VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO");
- pipeline->use_repclear = false;
+ pipeline->use_repclear = extra && extra->use_repclear;
+ pipeline->writes_point_size = false;
- anv_compiler_run(device->compiler, pipeline);
+ /* When we free the pipeline, we detect stages based on the NULL status
+ * of various prog_data pointers. Make them NULL by default.
+ */
+ memset(pipeline->prog_data, 0, sizeof(pipeline->prog_data));
+ memset(pipeline->scratch_start, 0, sizeof(pipeline->scratch_start));
- const struct brw_cs_prog_data *cs_prog_data = &pipeline->cs_prog_data;
+ pipeline->vs_simd8 = NO_KERNEL;
+ pipeline->vs_vec4 = NO_KERNEL;
+ pipeline->gs_kernel = NO_KERNEL;
- anv_batch_emit(&pipeline->batch, GEN8_MEDIA_VFE_STATE,
- .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_FRAGMENT],
- .PerThreadScratchSpace = ffs(cs_prog_data->base.total_scratch / 2048),
- .ScratchSpaceBasePointerHigh = 0,
- .StackSize = 0,
+ pipeline->active_stages = 0;
+ pipeline->total_scratch = 0;
- .MaximumNumberofThreads = device->info.max_cs_threads - 1,
- .NumberofURBEntries = 2,
- .ResetGatewayTimer = true,
- .BypassGatewayControl = true,
- .URBEntryAllocationSize = 2,
- .CURBEAllocationSize = 0);
+ for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
+ ANV_FROM_HANDLE(anv_shader_module, module,
+ pCreateInfo->pStages[i].module);
+ const char *entrypoint = pCreateInfo->pStages[i].pName;
- struct brw_cs_prog_data *prog_data = &pipeline->cs_prog_data;
- uint32_t group_size = prog_data->local_size[0] *
- prog_data->local_size[1] * prog_data->local_size[2];
- pipeline->cs_thread_width_max = DIV_ROUND_UP(group_size, prog_data->simd_size);
- uint32_t remainder = group_size & (prog_data->simd_size - 1);
+ switch (pCreateInfo->pStages[i].stage) {
+ case VK_SHADER_STAGE_VERTEX_BIT:
+ anv_pipeline_compile_vs(pipeline, pCreateInfo, module, entrypoint);
+ break;
+ case VK_SHADER_STAGE_GEOMETRY_BIT:
+ anv_pipeline_compile_gs(pipeline, pCreateInfo, module, entrypoint);
+ break;
+ case VK_SHADER_STAGE_FRAGMENT_BIT:
+ anv_pipeline_compile_fs(pipeline, pCreateInfo, module, entrypoint);
+ break;
+ default:
+ anv_finishme("Unsupported shader stage");
+ }
+ }
- if (remainder > 0)
- pipeline->cs_right_mask = ~0u >> (32 - remainder);
- else
- pipeline->cs_right_mask = ~0u >> (32 - prog_data->simd_size);
+ if (!(pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT)) {
+ /* Vertex is only optional if disable_vs is set */
+ assert(extra->disable_vs);
+ memset(&pipeline->vs_prog_data, 0, sizeof(pipeline->vs_prog_data));
+ }
+ gen7_compute_urb_partition(pipeline);
- *pPipeline = anv_pipeline_to_handle(pipeline);
+ const VkPipelineVertexInputStateCreateInfo *vi_info =
+ pCreateInfo->pVertexInputState;
+ pipeline->vb_used = 0;
+ for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
+ const VkVertexInputBindingDescription *desc =
+ &vi_info->pVertexBindingDescriptions[i];
+
+ pipeline->vb_used |= 1 << desc->binding;
+ pipeline->binding_stride[desc->binding] = desc->stride;
+
+ /* Step rate is programmed per vertex element (attribute), not
+ * binding. Set up a map of which bindings step per instance, for
+ * reference by vertex element setup. */
+ switch (desc->inputRate) {
+ default:
+ case VK_VERTEX_INPUT_RATE_VERTEX:
+ pipeline->instancing_enable[desc->binding] = false;
+ break;
+ case VK_VERTEX_INPUT_RATE_INSTANCE:
+ pipeline->instancing_enable[desc->binding] = true;
+ break;
+ }
+ }
+
+ const VkPipelineInputAssemblyStateCreateInfo *ia_info =
+ pCreateInfo->pInputAssemblyState;
+ pipeline->primitive_restart = ia_info->primitiveRestartEnable;
+ pipeline->topology = vk_to_gen_primitive_type[ia_info->topology];
+
+ if (extra && extra->use_rectlist)
+ pipeline->topology = _3DPRIM_RECTLIST;
return VK_SUCCESS;
}
-VkResult anv_CreateComputePipelines(
+VkResult
+anv_graphics_pipeline_create(
+ VkDevice _device,
+ const VkGraphicsPipelineCreateInfo *pCreateInfo,
+ const struct anv_graphics_pipeline_create_info *extra,
+ const VkAllocationCallbacks *pAllocator,
+ VkPipeline *pPipeline)
+{
+ ANV_FROM_HANDLE(anv_device, device, _device);
+
+ switch (device->info.gen) {
+ case 7:
+ if (device->info.is_haswell)
+ return gen75_graphics_pipeline_create(_device, pCreateInfo, extra, pAllocator, pPipeline);
+ else
+ return gen7_graphics_pipeline_create(_device, pCreateInfo, extra, pAllocator, pPipeline);
+ case 8:
+ return gen8_graphics_pipeline_create(_device, pCreateInfo, extra, pAllocator, pPipeline);
+ case 9:
+ return gen9_graphics_pipeline_create(_device, pCreateInfo, extra, pAllocator, pPipeline);
+ default:
+ unreachable("unsupported gen\n");
+ }
+}
+
+VkResult anv_CreateGraphicsPipelines(
VkDevice _device,
VkPipelineCache pipelineCache,
uint32_t count,
- const VkComputePipelineCreateInfo* pCreateInfos,
+ const VkGraphicsPipelineCreateInfo* pCreateInfos,
+ const VkAllocationCallbacks* pAllocator,
VkPipeline* pPipelines)
{
VkResult result = VK_SUCCESS;
unsigned i = 0;
for (; i < count; i++) {
- result = anv_compute_pipeline_create(_device, &pCreateInfos[i],
- &pPipelines[i]);
+ result = anv_graphics_pipeline_create(_device, &pCreateInfos[i],
+ NULL, pAllocator, &pPipelines[i]);
if (result != VK_SUCCESS) {
for (unsigned j = 0; j < i; j++) {
- anv_DestroyPipeline(_device, pPipelines[j]);
+ anv_DestroyPipeline(_device, pPipelines[j], pAllocator);
}
return result;
return VK_SUCCESS;
}
-// Pipeline layout functions
-
-VkResult anv_CreatePipelineLayout(
+static VkResult anv_compute_pipeline_create(
VkDevice _device,
- const VkPipelineLayoutCreateInfo* pCreateInfo,
- VkPipelineLayout* pPipelineLayout)
+ const VkComputePipelineCreateInfo* pCreateInfo,
+ const VkAllocationCallbacks* pAllocator,
+ VkPipeline* pPipeline)
{
ANV_FROM_HANDLE(anv_device, device, _device);
- struct anv_pipeline_layout *layout;
- assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO);
-
- layout = anv_device_alloc(device, sizeof(*layout), 8,
- VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
- if (layout == NULL)
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-
- layout->num_sets = pCreateInfo->descriptorSetCount;
-
- uint32_t surface_start[VK_SHADER_STAGE_NUM] = { 0, };
- uint32_t sampler_start[VK_SHADER_STAGE_NUM] = { 0, };
-
- for (uint32_t s = 0; s < VK_SHADER_STAGE_NUM; s++) {
- layout->stage[s].surface_count = 0;
- layout->stage[s].sampler_count = 0;
- }
-
- for (uint32_t i = 0; i < pCreateInfo->descriptorSetCount; i++) {
- ANV_FROM_HANDLE(anv_descriptor_set_layout, set_layout,
- pCreateInfo->pSetLayouts[i]);
-
- layout->set[i].layout = set_layout;
- for (uint32_t s = 0; s < VK_SHADER_STAGE_NUM; s++) {
- layout->set[i].surface_start[s] = surface_start[s];
- surface_start[s] += set_layout->stage[s].surface_count;
- layout->set[i].sampler_start[s] = sampler_start[s];
- sampler_start[s] += set_layout->stage[s].sampler_count;
-
- layout->stage[s].surface_count += set_layout->stage[s].surface_count;
- layout->stage[s].sampler_count += set_layout->stage[s].sampler_count;
- }
+ switch (device->info.gen) {
+ case 7:
+ if (device->info.is_haswell)
+ return gen75_compute_pipeline_create(_device, pCreateInfo, pAllocator, pPipeline);
+ else
+ return gen7_compute_pipeline_create(_device, pCreateInfo, pAllocator, pPipeline);
+ case 8:
+ return gen8_compute_pipeline_create(_device, pCreateInfo, pAllocator, pPipeline);
+ case 9:
+ return gen9_compute_pipeline_create(_device, pCreateInfo, pAllocator, pPipeline);
+ default:
+ unreachable("unsupported gen\n");
}
-
- *pPipelineLayout = anv_pipeline_layout_to_handle(layout);
-
- return VK_SUCCESS;
}
-VkResult anv_DestroyPipelineLayout(
+VkResult anv_CreateComputePipelines(
VkDevice _device,
- VkPipelineLayout _pipelineLayout)
+ VkPipelineCache pipelineCache,
+ uint32_t count,
+ const VkComputePipelineCreateInfo* pCreateInfos,
+ const VkAllocationCallbacks* pAllocator,
+ VkPipeline* pPipelines)
{
- ANV_FROM_HANDLE(anv_device, device, _device);
- ANV_FROM_HANDLE(anv_pipeline_layout, pipeline_layout, _pipelineLayout);
+ VkResult result = VK_SUCCESS;
- anv_device_free(device, pipeline_layout);
+ unsigned i = 0;
+ for (; i < count; i++) {
+ result = anv_compute_pipeline_create(_device, &pCreateInfos[i],
+ pAllocator, &pPipelines[i]);
+ if (result != VK_SUCCESS) {
+ for (unsigned j = 0; j < i; j++) {
+ anv_DestroyPipeline(_device, pPipelines[j], pAllocator);
+ }
+
+ return result;
+ }
+ }
return VK_SUCCESS;
}