[MESA_SHADER_COMPUTE] = DEBUG_CS,
};
+struct anv_spirv_debug_data {
+ struct anv_device *device;
+ const struct anv_shader_module *module;
+};
+
+static void anv_spirv_nir_debug(void *private_data,
+ enum nir_spirv_debug_level level,
+ size_t spirv_offset,
+ const char *message)
+{
+ struct anv_spirv_debug_data *debug_data = private_data;
+ static const VkDebugReportFlagsEXT vk_flags[] = {
+ [NIR_SPIRV_DEBUG_LEVEL_INFO] = VK_DEBUG_REPORT_INFORMATION_BIT_EXT,
+ [NIR_SPIRV_DEBUG_LEVEL_WARNING] = VK_DEBUG_REPORT_WARNING_BIT_EXT,
+ [NIR_SPIRV_DEBUG_LEVEL_ERROR] = VK_DEBUG_REPORT_ERROR_BIT_EXT,
+ };
+ char buffer[256];
+
+ snprintf(buffer, sizeof(buffer), "SPIR-V offset %lu: %s", (unsigned long) spirv_offset, message);
+
+ vk_debug_report(&debug_data->device->instance->debug_report_callbacks,
+ vk_flags[level],
+ VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT,
+ (uint64_t) (uintptr_t) debug_data->module,
+ 0, 0, "anv", buffer);
+}
+
/* 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.
*/
}
}
+ struct anv_spirv_debug_data spirv_debug_data = {
+ .device = device,
+ .module = module,
+ };
struct spirv_to_nir_options spirv_options = {
.lower_workgroup_access_to_offsets = true,
.caps = {
+ .demote_to_helper_invocation = true,
+ .derivative_group = true,
+ .descriptor_array_dynamic_indexing = true,
+ .descriptor_array_non_uniform_indexing = true,
+ .descriptor_indexing = true,
.device_group = true,
.draw_parameters = true,
+ .float16 = pdevice->info.gen >= 8,
.float64 = pdevice->info.gen >= 8,
+ .fragment_shader_sample_interlock = pdevice->info.gen >= 9,
+ .fragment_shader_pixel_interlock = pdevice->info.gen >= 9,
.geometry_streams = true,
.image_write_without_format = true,
+ .int8 = pdevice->info.gen >= 8,
.int16 = pdevice->info.gen >= 8,
.int64 = pdevice->info.gen >= 8,
+ .int64_atomics = pdevice->info.gen >= 9 && pdevice->use_softpin,
.min_lod = true,
.multiview = true,
- .physical_storage_buffer_address = pdevice->info.gen >= 8 &&
- pdevice->use_softpin,
+ .physical_storage_buffer_address = pdevice->has_a64_buffer_access,
.post_depth_coverage = pdevice->info.gen >= 9,
+ .runtime_descriptor_array = true,
.shader_viewport_index_layer = true,
.stencil_export = pdevice->info.gen >= 9,
.storage_8bit = pdevice->info.gen >= 8,
.transform_feedback = pdevice->info.gen >= 8,
.variable_pointers = true,
},
- .ubo_ptr_type = glsl_vector_type(GLSL_TYPE_UINT, 2),
- .ssbo_ptr_type = glsl_vector_type(GLSL_TYPE_UINT, 2),
- .phys_ssbo_ptr_type = glsl_vector_type(GLSL_TYPE_UINT64, 1),
- .push_const_ptr_type = glsl_uint_type(),
- .shared_ptr_type = glsl_uint_type(),
+ .ubo_addr_format = nir_address_format_32bit_index_offset,
+ .ssbo_addr_format =
+ anv_nir_ssbo_addr_format(pdevice, device->robust_buffer_access),
+ .phys_ssbo_addr_format = nir_address_format_64bit_global,
+ .push_const_addr_format = nir_address_format_logical,
+
+ /* TODO: Consider changing this to an address format that has the NULL
+ * pointer equals to 0. That might be a better format to play nice
+ * with certain code / code generators.
+ */
+ .shared_addr_format = nir_address_format_32bit_offset,
+ .debug = {
+ .func = anv_spirv_nir_debug,
+ .private_data = &spirv_debug_data,
+ },
};
- nir_function *entry_point =
+
+ nir_shader *nir =
spirv_to_nir(spirv, module->size / 4,
spec_entries, num_spec_entries,
stage, entrypoint_name, &spirv_options, nir_options);
- nir_shader *nir = entry_point->shader;
assert(nir->info.stage == stage);
nir_validate_shader(nir, "after spirv_to_nir");
ralloc_steal(mem_ctx, nir);
/* Pick off the single entrypoint that we want */
foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
- if (func != entry_point)
+ if (!func->is_entrypoint)
exec_node_remove(&func->node);
}
assert(exec_list_length(&nir->functions) == 1);
NIR_PASS_V(nir, nir_remove_dead_variables,
nir_var_shader_in | nir_var_shader_out | nir_var_system_value);
- NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_global,
- nir_address_format_64bit_global);
-
NIR_PASS_V(nir, nir_propagate_invariant);
NIR_PASS_V(nir, nir_lower_io_to_temporaries,
- entry_point->impl, true, false);
+ nir_shader_get_entrypoint(nir), true, false);
NIR_PASS_V(nir, nir_lower_frexp);
/* Vulkan uses the separate-shader linking model */
nir->info.separate_shader = true;
- nir = brw_preprocess_nir(compiler, nir, NULL);
+ brw_preprocess_nir(compiler, nir, NULL);
return nir;
}
}
}
+static void
+populate_base_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
+ struct brw_base_prog_key *key)
+{
+ if (flags & VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT_EXT)
+ key->subgroup_size_type = BRW_SUBGROUP_SIZE_VARYING;
+ else
+ key->subgroup_size_type = BRW_SUBGROUP_SIZE_API_CONSTANT;
+
+ populate_sampler_prog_key(devinfo, &key->tex);
+}
+
static void
populate_vs_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
struct brw_vs_prog_key *key)
{
memset(key, 0, sizeof(*key));
- populate_sampler_prog_key(devinfo, &key->tex);
+ populate_base_prog_key(devinfo, flags, &key->base);
/* XXX: Handle vertex input work-arounds */
static void
populate_tcs_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
unsigned input_vertices,
struct brw_tcs_prog_key *key)
{
memset(key, 0, sizeof(*key));
- populate_sampler_prog_key(devinfo, &key->tex);
+ populate_base_prog_key(devinfo, flags, &key->base);
key->input_vertices = input_vertices;
}
static void
populate_tes_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
struct brw_tes_prog_key *key)
{
memset(key, 0, sizeof(*key));
- populate_sampler_prog_key(devinfo, &key->tex);
+ populate_base_prog_key(devinfo, flags, &key->base);
}
static void
populate_gs_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
struct brw_gs_prog_key *key)
{
memset(key, 0, sizeof(*key));
- populate_sampler_prog_key(devinfo, &key->tex);
+ populate_base_prog_key(devinfo, flags, &key->base);
}
static void
populate_wm_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
const struct anv_subpass *subpass,
const VkPipelineMultisampleStateCreateInfo *ms_info,
struct brw_wm_prog_key *key)
{
memset(key, 0, sizeof(*key));
- populate_sampler_prog_key(devinfo, &key->tex);
+ populate_base_prog_key(devinfo, flags, &key->base);
/* We set this to 0 here and set to the actual value before we call
* brw_compile_fs.
key->nr_color_regions = util_bitcount(key->color_outputs_valid);
- key->replicate_alpha = key->nr_color_regions > 1 &&
- ms_info && ms_info->alphaToCoverageEnable;
+ /* To reduce possible shader recompilations we would need to know if
+ * there is a SampleMask output variable to compute if we should emit
+ * code to workaround the issue that hardware disables alpha to coverage
+ * when there is SampleMask output.
+ */
+ key->alpha_to_coverage = ms_info && ms_info->alphaToCoverageEnable;
+
+ /* Vulkan doesn't support fixed-function alpha test */
+ key->alpha_test_replicate_alpha = false;
if (ms_info) {
/* 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.
*/
if (ms_info->rasterizationSamples > 1) {
- key->persample_interp =
+ key->persample_interp = ms_info->sampleShadingEnable &&
(ms_info->minSampleShading * ms_info->rasterizationSamples) > 1;
key->multisample_fbo = true;
}
- key->frag_coord_adds_sample_pos = ms_info->sampleShadingEnable;
+ key->frag_coord_adds_sample_pos = key->persample_interp;
}
}
static void
populate_cs_prog_key(const struct gen_device_info *devinfo,
+ VkPipelineShaderStageCreateFlags flags,
+ const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT *rss_info,
struct brw_cs_prog_key *key)
{
memset(key, 0, sizeof(*key));
- populate_sampler_prog_key(devinfo, &key->tex);
+ populate_base_prog_key(devinfo, flags, &key->base);
+
+ if (rss_info) {
+ assert(key->base.subgroup_size_type != BRW_SUBGROUP_SIZE_VARYING);
+
+ /* These enum values are expressly chosen to be equal to the subgroup
+ * size that they require.
+ */
+ assert(rss_info->requiredSubgroupSize == 8 ||
+ rss_info->requiredSubgroupSize == 16 ||
+ rss_info->requiredSubgroupSize == 32);
+ key->base.subgroup_size_type = rss_info->requiredSubgroupSize;
+ } else if (flags & VK_PIPELINE_SHADER_STAGE_CREATE_REQUIRE_FULL_SUBGROUPS_BIT_EXT) {
+ /* If the client expressly requests full subgroups and they don't
+ * specify a subgroup size, we need to pick one. If they're requested
+ * varying subgroup sizes, we set it to UNIFORM and let the back-end
+ * compiler pick. Otherwise, we specify the API value of 32.
+ * Performance will likely be terrible in this case but there's nothing
+ * we can do about that. The client should have chosen a size.
+ */
+ if (flags & VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT_EXT)
+ key->base.subgroup_size_type = BRW_SUBGROUP_SIZE_UNIFORM;
+ else
+ key->base.subgroup_size_type = BRW_SUBGROUP_SIZE_REQUIRE_32;
+ }
}
struct anv_pipeline_stage {
struct anv_pipeline_stage *stage,
struct anv_pipeline_layout *layout)
{
- const struct brw_compiler *compiler =
- pipeline->device->instance->physicalDevice.compiler;
+ const struct anv_physical_device *pdevice =
+ &pipeline->device->instance->physicalDevice;
+ const struct brw_compiler *compiler = pdevice->compiler;
struct brw_stage_prog_data *prog_data = &stage->prog_data.base;
nir_shader *nir = stage->nir;
if (nir->info.stage == MESA_SHADER_FRAGMENT) {
NIR_PASS_V(nir, nir_lower_wpos_center, pipeline->sample_shading_enable);
- NIR_PASS_V(nir, anv_nir_lower_input_attachments);
+ NIR_PASS_V(nir, nir_lower_input_attachments, false);
}
NIR_PASS_V(nir, anv_nir_lower_ycbcr_textures, layout);
NIR_PASS_V(nir, brw_nir_lower_image_load_store, compiler->devinfo);
+ NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_global,
+ nir_address_format_64bit_global);
+
/* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
if (layout) {
- anv_nir_apply_pipeline_layout(&pipeline->device->instance->physicalDevice,
+ anv_nir_apply_pipeline_layout(pdevice,
pipeline->device->robust_buffer_access,
layout, nir, prog_data,
&stage->bind_map);
- NIR_PASS_V(nir, nir_lower_explicit_io,
- nir_var_mem_ubo | nir_var_mem_ssbo,
+ NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_ubo,
nir_address_format_32bit_index_offset);
+ NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_ssbo,
+ anv_nir_ssbo_addr_format(pdevice,
+ pipeline->device->robust_buffer_access));
NIR_PASS_V(nir, nir_opt_constant_folding);
+
+ /* We don't support non-uniform UBOs and non-uniform SSBO access is
+ * handled naturally by falling back to A64 messages.
+ */
+ NIR_PASS_V(nir, nir_lower_non_uniform_access,
+ nir_lower_non_uniform_texture_access |
+ nir_lower_non_uniform_image_access);
}
if (nir->info.stage != MESA_SHADER_COMPUTE)
struct anv_pipeline_stage *next_stage)
{
if (next_stage)
- brw_nir_link_shaders(compiler, &vs_stage->nir, &next_stage->nir);
+ brw_nir_link_shaders(compiler, vs_stage->nir, next_stage->nir);
}
static const unsigned *
{
assert(tes_stage && tes_stage->stage == MESA_SHADER_TESS_EVAL);
- brw_nir_link_shaders(compiler, &tcs_stage->nir, &tes_stage->nir);
+ brw_nir_link_shaders(compiler, tcs_stage->nir, tes_stage->nir);
nir_lower_patch_vertices(tes_stage->nir,
tcs_stage->nir->info.tess.tcs_vertices_out,
struct anv_pipeline_stage *next_stage)
{
if (next_stage)
- brw_nir_link_shaders(compiler, &tes_stage->nir, &next_stage->nir);
+ brw_nir_link_shaders(compiler, tes_stage->nir, next_stage->nir);
}
static const unsigned *
struct anv_pipeline_stage *next_stage)
{
if (next_stage)
- brw_nir_link_shaders(compiler, &gs_stage->nir, &next_stage->nir);
+ brw_nir_link_shaders(compiler, gs_stage->nir, next_stage->nir);
}
static const unsigned *
continue;
const unsigned rt = var->data.location - FRAG_RESULT_DATA0;
- /* Unused or out-of-bounds */
- if (rt >= MAX_RTS || !(stage->key.wm.color_outputs_valid & (1 << rt)))
+ /* Out-of-bounds */
+ if (rt >= MAX_RTS)
continue;
const unsigned array_len =
glsl_type_is_array(var->type) ? glsl_get_length(var->type) : 1;
assert(rt + array_len <= max_rt);
+ /* Unused */
+ if (!(stage->key.wm.color_outputs_valid & BITFIELD_RANGE(rt, array_len))) {
+ /* If this is the RT at location 0 and we have alpha to coverage
+ * enabled we will have to create a null RT for it, so mark it as
+ * used.
+ */
+ if (rt > 0 || !stage->key.wm.alpha_to_coverage)
+ continue;
+ }
+
for (unsigned i = 0; i < array_len; i++)
rt_used[rt + i] = true;
}
continue;
rt_to_bindings[i] = num_rts;
- rt_bindings[rt_to_bindings[i]] = (struct anv_pipeline_binding) {
- .set = ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS,
- .binding = 0,
- .index = i,
- };
+
+ if (stage->key.wm.color_outputs_valid & (1 << i)) {
+ rt_bindings[rt_to_bindings[i]] = (struct anv_pipeline_binding) {
+ .set = ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS,
+ .binding = 0,
+ .index = i,
+ };
+ } else {
+ /* Setup a null render target */
+ rt_bindings[rt_to_bindings[i]] = (struct anv_pipeline_binding) {
+ .set = ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS,
+ .binding = 0,
+ .index = UINT32_MAX,
+ };
+ }
+
num_rts++;
}
continue;
const unsigned rt = var->data.location - FRAG_RESULT_DATA0;
- if (rt >= MAX_RTS ||
- !(stage->key.wm.color_outputs_valid & (1 << rt))) {
- /* Unused or out-of-bounds, throw it away */
+
+ if (rt >= MAX_RTS || !rt_used[rt]) {
+ /* Unused or out-of-bounds, throw it away, unless it is the first
+ * RT and we have alpha to coverage enabled.
+ */
deleted_output = true;
var->data.mode = nir_var_function_temp;
exec_node_remove(&var->node);
const struct gen_device_info *devinfo = &pipeline->device->info;
switch (stage) {
case MESA_SHADER_VERTEX:
- populate_vs_prog_key(devinfo, &stages[stage].key.vs);
+ populate_vs_prog_key(devinfo, sinfo->flags, &stages[stage].key.vs);
break;
case MESA_SHADER_TESS_CTRL:
- populate_tcs_prog_key(devinfo,
+ populate_tcs_prog_key(devinfo, sinfo->flags,
info->pTessellationState->patchControlPoints,
&stages[stage].key.tcs);
break;
case MESA_SHADER_TESS_EVAL:
- populate_tes_prog_key(devinfo, &stages[stage].key.tes);
+ populate_tes_prog_key(devinfo, sinfo->flags, &stages[stage].key.tes);
break;
case MESA_SHADER_GEOMETRY:
- populate_gs_prog_key(devinfo, &stages[stage].key.gs);
+ populate_gs_prog_key(devinfo, sinfo->flags, &stages[stage].key.gs);
break;
case MESA_SHADER_FRAGMENT:
- populate_wm_prog_key(devinfo, pipeline->subpass,
+ populate_wm_prog_key(devinfo, sinfo->flags,
+ pipeline->subpass,
info->pMultisampleState,
&stages[stage].key.wm);
break;
struct anv_shader_bin *bin = NULL;
- populate_cs_prog_key(&pipeline->device->info, &stage.key.cs);
+ const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT *rss_info =
+ vk_find_struct_const(info->stage.pNext,
+ PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT);
+
+ populate_cs_prog_key(&pipeline->device->info, info->stage.flags,
+ rss_info, &stage.key.cs);
ANV_FROM_HANDLE(anv_pipeline_layout, layout, info->layout);