[PATCH 0/1] MR695: vkd3d-shader/spirv: Implement support for rasteriser-ordered views.
From: Evan Tang <etang(a)codeweavers.com> Using SPV_EXT_fragment_shader_interlock. --- include/vkd3d_shader.h | 2 + libs/vkd3d-shader/spirv.c | 46 ++++- .../hlsl/rasteriser-ordered-views.shader_test | 2 +- tests/shader_runner_vulkan.c | 162 +++++++++++++++--- tests/vulkan_procs.h | 1 + 5 files changed, 190 insertions(+), 23 deletions(-) diff --git a/include/vkd3d_shader.h b/include/vkd3d_shader.h index 9e663919c..7e76e5b02 100644 --- a/include/vkd3d_shader.h +++ b/include/vkd3d_shader.h @@ -886,6 +886,8 @@ enum vkd3d_shader_spirv_extension VKD3D_SHADER_SPIRV_EXTENSION_EXT_STENCIL_EXPORT, /** \since 1.11 */ VKD3D_SHADER_SPIRV_EXTENSION_EXT_VIEWPORT_INDEX_LAYER, + /** \since 1.12 */ + VKD3D_SHADER_SPIRV_EXTENSION_EXT_FRAGMENT_SHADER_INTERLOCK, VKD3D_FORCE_32_BIT_ENUM(VKD3D_SHADER_SPIRV_EXTENSION), }; diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c index c0cbdccf6..2e7116be0 100644 --- a/libs/vkd3d-shader/spirv.c +++ b/libs/vkd3d-shader/spirv.c @@ -1940,6 +1940,9 @@ static bool vkd3d_spirv_compile_module(struct vkd3d_spirv_builder *builder, || vkd3d_spirv_capability_is_enabled(builder, SpvCapabilityStorageImageArrayDynamicIndexing) || vkd3d_spirv_capability_is_enabled(builder, SpvCapabilityShaderNonUniformEXT)) vkd3d_spirv_build_op_extension(&stream, "SPV_EXT_descriptor_indexing"); + if (vkd3d_spirv_capability_is_enabled(builder, SpvCapabilityFragmentShaderPixelInterlockEXT) + || vkd3d_spirv_capability_is_enabled(builder, SpvCapabilityFragmentShaderSampleInterlockEXT)) + vkd3d_spirv_build_op_extension(&stream, "SPV_EXT_fragment_shader_interlock"); if (vkd3d_spirv_capability_is_enabled(builder, SpvCapabilityStencilExportEXT)) vkd3d_spirv_build_op_extension(&stream, "SPV_EXT_shader_stencil_export"); if (vkd3d_spirv_capability_is_enabled(builder, SpvCapabilityShaderViewportIndexLayerEXT)) @@ -2346,6 +2349,7 @@ struct spirv_compiler unsigned int output_control_point_count; bool use_vocp; + bool use_invocation_interlock; bool emit_point_size; enum vkd3d_shader_opcode phase; @@ -6272,9 +6276,24 @@ static void spirv_compiler_emit_resource_declaration(struct spirv_compiler *comp if (!(d->flags & VKD3D_SHADER_DESCRIPTOR_INFO_FLAG_UAV_READ)) vkd3d_spirv_build_op_decorate(builder, var_id, SpvDecorationNonReadable, NULL, 0); - if (d->uav_flags & VKD3DSUF_GLOBALLY_COHERENT) + /* ROVs are implicitly globally coherent. */ + if (d->uav_flags & (VKD3DSUF_GLOBALLY_COHERENT | VKD3DSUF_RASTERISER_ORDERED_VIEW)) vkd3d_spirv_build_op_decorate(builder, var_id, SpvDecorationCoherent, NULL, 0); + if (d->uav_flags & VKD3DSUF_RASTERISER_ORDERED_VIEW) + { + if (compiler->shader_type != VKD3D_SHADER_TYPE_PIXEL) + spirv_compiler_error(compiler, VKD3D_SHADER_ERROR_SPV_UNSUPPORTED_FEATURE, + "Rasteriser-ordered views are only supported in fragment shaders."); + else if (!spirv_compiler_is_target_extension_supported(compiler, + VKD3D_SHADER_SPIRV_EXTENSION_EXT_FRAGMENT_SHADER_INTERLOCK)) + spirv_compiler_error(compiler, VKD3D_SHADER_ERROR_SPV_UNSUPPORTED_FEATURE, + "Cannot enable fragment shader interlock. " + "The target environment does not support fragment shader interlock."); + else + compiler->use_invocation_interlock = true; + } + if (d->flags & VKD3D_SHADER_DESCRIPTOR_INFO_FLAG_UAV_COUNTER) { assert(structure_stride); /* counters are valid only for structured buffers */ @@ -7704,11 +7723,31 @@ static uint32_t spirv_compiler_emit_conditional_branch(struct spirv_compiler *co return merge_block_id; } +static void spirv_compiler_end_invocation_interlock(struct spirv_compiler *compiler) +{ + struct vkd3d_spirv_builder *builder = &compiler->spirv_builder; + + if (vkd3d_spirv_capability_is_enabled(builder, SpvCapabilitySampleRateShading)) + { + spirv_compiler_emit_execution_mode(compiler, SpvExecutionModeSampleInterlockOrderedEXT, NULL, 0); + vkd3d_spirv_enable_capability(builder, SpvCapabilityFragmentShaderSampleInterlockEXT); + } + else + { + spirv_compiler_emit_execution_mode(compiler, SpvExecutionModePixelInterlockOrderedEXT, NULL, 0); + vkd3d_spirv_enable_capability(builder, SpvCapabilityFragmentShaderPixelInterlockEXT); + } + vkd3d_spirv_build_op(&builder->function_stream, SpvOpEndInvocationInterlockEXT); +} + static void spirv_compiler_emit_return(struct spirv_compiler *compiler, const struct vkd3d_shader_instruction *instruction) { struct vkd3d_spirv_builder *builder = &compiler->spirv_builder; + if (compiler->use_invocation_interlock) + spirv_compiler_end_invocation_interlock(compiler); + if (compiler->shader_type != VKD3D_SHADER_TYPE_GEOMETRY && (is_in_default_phase(compiler) || is_in_control_point_phase(compiler))) spirv_compiler_emit_shader_epilogue_invocation(compiler); @@ -9477,6 +9516,11 @@ static void spirv_compiler_emit_main_prolog(struct spirv_compiler *compiler) if (compiler->emit_point_size) spirv_compiler_emit_point_size(compiler); + + /* Maybe in the future we can try to shrink the size of the interlocked + * section. */ + if (compiler->use_invocation_interlock) + vkd3d_spirv_build_op(&compiler->spirv_builder.function_stream, SpvOpBeginInvocationInterlockEXT); } static int spirv_compiler_handle_instruction(struct spirv_compiler *compiler, diff --git a/tests/hlsl/rasteriser-ordered-views.shader_test b/tests/hlsl/rasteriser-ordered-views.shader_test index d273b05b1..a7aa7827b 100644 --- a/tests/hlsl/rasteriser-ordered-views.shader_test +++ b/tests/hlsl/rasteriser-ordered-views.shader_test @@ -64,6 +64,6 @@ shader model >= 5.0 rov [test] -todo draw triangle list 93 +draw triangle list 93 probe all rui(0x6aaaaaaa) probe uav 1 all rui(0x2aaaaaaa) diff --git a/tests/shader_runner_vulkan.c b/tests/shader_runner_vulkan.c index 913cb9da0..d2cfd7eea 100644 --- a/tests/shader_runner_vulkan.c +++ b/tests/shader_runner_vulkan.c @@ -73,11 +73,24 @@ struct vulkan_shader_runner } samplers[MAX_SAMPLERS]; DECLARE_VK_PFN(vkCreateInstance); + DECLARE_VK_PFN(vkEnumerateInstanceExtensionProperties); #define VK_INSTANCE_PFN DECLARE_VK_PFN #define VK_DEVICE_PFN DECLARE_VK_PFN #include "vulkan_procs.h" }; +struct extension_list +{ + const char **names; + size_t count; +}; + +struct physical_device_info +{ + VkPhysicalDeviceFeatures2 features2; + VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT interlock_features; +}; + static struct vulkan_shader_runner *vulkan_shader_runner(struct shader_runner *r) { return CONTAINING_RECORD(r, struct vulkan_shader_runner, r); @@ -504,6 +517,16 @@ static bool compile_shader(struct vulkan_shader_runner *runner, const char *sour spirv_info.next = &interface_info; spirv_info.environment = VKD3D_SHADER_SPIRV_ENVIRONMENT_VULKAN_1_0; + if (runner->caps.rov) + { + static const enum vkd3d_shader_spirv_extension extensions[] = + { + VKD3D_SHADER_SPIRV_EXTENSION_EXT_FRAGMENT_SHADER_INTERLOCK, + }; + + spirv_info.extensions = extensions; + spirv_info.extension_count = ARRAY_SIZE(extensions); + } push_constants.register_space = 0; push_constants.register_index = 0; @@ -1253,12 +1276,6 @@ static bool get_graphics_queue_index(const struct vulkan_shader_runner *runner, return false; } -static const char *const required_device_extensions[] = -{ - VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME, - VK_KHR_MAINTENANCE1_EXTENSION_NAME, -}; - static bool has_extension(const VkExtensionProperties *extensions, uint32_t count, const char *extension_name) { uint32_t i; @@ -1271,21 +1288,75 @@ static bool has_extension(const VkExtensionProperties *extensions, uint32_t coun return false; } -static bool check_device_extensions(struct vulkan_shader_runner *runner) +static void check_instance_extensions(struct vulkan_shader_runner *runner, struct extension_list *enabled_extensions) +{ + VkExtensionProperties *extensions; + uint32_t count, i; + + static const char *instance_extensions[] = + { + VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME, + }; + + enabled_extensions->names = calloc(ARRAY_SIZE(instance_extensions), sizeof(*enabled_extensions->names)); + enabled_extensions->count = 0; + + VK_CALL(vkEnumerateInstanceExtensionProperties(NULL, &count, NULL)); + extensions = calloc(count, sizeof(*extensions)); + VK_CALL(vkEnumerateInstanceExtensionProperties(NULL, &count, extensions)); + + for (i = 0; i < ARRAY_SIZE(instance_extensions); ++i) + { + const char *name = instance_extensions[i]; + + if (has_extension(extensions, count, name)) + enabled_extensions->names[enabled_extensions->count++] = name; + } + + free(extensions); +} + +static bool check_device_extensions(struct vulkan_shader_runner *runner, struct extension_list *enabled_extensions) { VkPhysicalDevice phys_device = runner->phys_device; VkExtensionProperties *extensions; uint32_t i, count; + static const struct + { + const char *name; + bool required; + } + device_extensions[] = + { + {VK_EXT_FRAGMENT_SHADER_INTERLOCK_EXTENSION_NAME}, + {VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME, true}, + {VK_KHR_MAINTENANCE1_EXTENSION_NAME, true}, + }; + + enabled_extensions->names = calloc(ARRAY_SIZE(device_extensions), sizeof(*enabled_extensions->names)); + enabled_extensions->count = 0; + VK_CALL(vkEnumerateDeviceExtensionProperties(phys_device, NULL, &count, NULL)); extensions = calloc(count, sizeof(*extensions)); VK_CALL(vkEnumerateDeviceExtensionProperties(phys_device, NULL, &count, extensions)); - for (i = 0; i < ARRAY_SIZE(required_device_extensions); ++i) + for (i = 0; i < ARRAY_SIZE(device_extensions); ++i) { - if (!has_extension(extensions, count, required_device_extensions[i])) + const char *name = device_extensions[i].name; + + if (has_extension(extensions, count, name)) { - skip("The selected Vulkan device does not support %s.\n", required_device_extensions[i]); + enabled_extensions->names[enabled_extensions->count++] = name; + if (!strcmp(name, VK_EXT_FRAGMENT_SHADER_INTERLOCK_EXTENSION_NAME)) + runner->caps.rov = true; + continue; + } + + if (device_extensions[i].required) + { + skip("The selected Vulkan device does not support %s.\n", name); + free(enabled_extensions->names); free(extensions); return false; } @@ -1295,6 +1366,23 @@ static bool check_device_extensions(struct vulkan_shader_runner *runner) return true; } +static void get_physical_device_info(struct vulkan_shader_runner *runner, struct physical_device_info *info) +{ + memset(info, 0, sizeof(*info)); + + info->features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2; + if (runner->caps.rov) + { + info->features2.pNext = &info->interlock_features; + info->interlock_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT; + } + + if (runner->vkGetPhysicalDeviceFeatures2KHR) + VK_CALL(vkGetPhysicalDeviceFeatures2KHR(runner->phys_device, &info->features2)); + else + VK_CALL(vkGetPhysicalDeviceFeatures(runner->phys_device, &info->features2.features)); +} + static bool init_vulkan_runner(struct vulkan_shader_runner *runner) { VkDescriptorPoolCreateInfo descriptor_pool_desc = {.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO}; @@ -1303,9 +1391,13 @@ static bool init_vulkan_runner(struct vulkan_shader_runner *runner) VkDeviceQueueCreateInfo queue_desc = {.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO}; VkInstanceCreateInfo instance_desc = {.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO}; VkDeviceCreateInfo device_desc = {.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO}; - VkPhysicalDeviceFeatures ret_features, features; + VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT interlock_features; VkDescriptorPoolSize descriptor_pool_sizes[5]; + const VkPhysicalDeviceFeatures *ret_features; + struct extension_list enabled_extensions; static const float queue_priority = 1.0f; + struct physical_device_info device_info; + VkPhysicalDeviceFeatures features; VkFormatProperties format_props; uint32_t count, graphics_index; VkDevice device; @@ -1320,8 +1412,15 @@ static bool init_vulkan_runner(struct vulkan_shader_runner *runner) vkGetInstanceProcAddr = vkd3d_dlsym(libvulkan, "vkGetInstanceProcAddr"); runner->vkCreateInstance = (void *)vkGetInstanceProcAddr(NULL, "vkCreateInstance"); - - if ((vr = VK_CALL(vkCreateInstance(&instance_desc, NULL, &runner->instance))) < 0) + runner->vkEnumerateInstanceExtensionProperties = (void *)vkGetInstanceProcAddr(NULL, + "vkEnumerateInstanceExtensionProperties"); + + check_instance_extensions(runner, &enabled_extensions); + instance_desc.ppEnabledExtensionNames = enabled_extensions.names; + instance_desc.enabledExtensionCount = enabled_extensions.count; + vr = VK_CALL(vkCreateInstance(&instance_desc, NULL, &runner->instance)); + free(enabled_extensions.names); + if (vr < 0) { skip("Failed to create a Vulkan instance, vr %d.\n", vr); return false; @@ -1351,15 +1450,15 @@ static bool init_vulkan_runner(struct vulkan_shader_runner *runner) device_desc.pQueueCreateInfos = &queue_desc; device_desc.queueCreateInfoCount = 1; - device_desc.enabledExtensionCount = ARRAY_SIZE(required_device_extensions); - device_desc.ppEnabledExtensionNames = required_device_extensions; queue_desc.queueFamilyIndex = graphics_index; queue_desc.queueCount = 1; queue_desc.pQueuePriorities = &queue_priority; - if (!check_device_extensions(runner)) + if (!check_device_extensions(runner, &enabled_extensions)) goto out_destroy_instance; + device_desc.ppEnabledExtensionNames = enabled_extensions.names; + device_desc.enabledExtensionCount = enabled_extensions.count; VK_CALL(vkGetPhysicalDeviceFormatProperties(runner->phys_device, VK_FORMAT_R32G32B32A32_SFLOAT, &format_props)); if (!(format_props.optimalTilingFeatures & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT)) @@ -1369,7 +1468,8 @@ static bool init_vulkan_runner(struct vulkan_shader_runner *runner) } runner->caps.runner = "Vulkan"; - VK_CALL(vkGetPhysicalDeviceFeatures(runner->phys_device, &ret_features)); + get_physical_device_info(runner, &device_info); + ret_features = &device_info.features2.features; device_desc.pEnabledFeatures = &features; memset(&features, 0, sizeof(features)); @@ -1377,7 +1477,7 @@ static bool init_vulkan_runner(struct vulkan_shader_runner *runner) /* FIXME: Probably make these optional. */ #define ENABLE_FEATURE(x) \ - if (!ret_features.x) \ + if (!ret_features->x) \ { \ skip("The selected Vulkan device does not support " #x ".\n"); \ goto out_destroy_instance; \ @@ -1385,21 +1485,41 @@ static bool init_vulkan_runner(struct vulkan_shader_runner *runner) features.x = VK_TRUE ENABLE_FEATURE(fragmentStoresAndAtomics); + /* For SV_PrimitiveID/SpvBuiltInPrimitiveId in fragment shaders. */ + ENABLE_FEATURE(geometryShader); ENABLE_FEATURE(shaderImageGatherExtended); ENABLE_FEATURE(shaderStorageImageWriteWithoutFormat); - if (ret_features.shaderFloat64) + if (ret_features->shaderFloat64) { features.shaderFloat64 = VK_TRUE; runner->caps.float64 = true; } - if (ret_features.shaderInt64) + + if (ret_features->shaderInt64) { features.shaderInt64 = VK_TRUE; runner->caps.int64 = true; } - if ((vr = VK_CALL(vkCreateDevice(runner->phys_device, &device_desc, NULL, &device)))) + if (device_info.interlock_features.fragmentShaderSampleInterlock + && device_info.interlock_features.fragmentShaderPixelInterlock) + { + memset(&interlock_features, 0, sizeof(interlock_features)); + interlock_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT; + interlock_features.pNext = (void *)device_desc.pNext; + interlock_features.fragmentShaderSampleInterlock = VK_TRUE; + interlock_features.fragmentShaderPixelInterlock = VK_TRUE; + device_desc.pNext = &interlock_features; + } + else + { + runner->caps.rov = false; + } + + vr = VK_CALL(vkCreateDevice(runner->phys_device, &device_desc, NULL, &device)); + free(enabled_extensions.names); + if (vr) { skip("Failed to create device, vr %d.\n", vr); goto out_destroy_instance; diff --git a/tests/vulkan_procs.h b/tests/vulkan_procs.h index 1829133b2..1f7065e59 100644 --- a/tests/vulkan_procs.h +++ b/tests/vulkan_procs.h @@ -39,6 +39,7 @@ VK_INSTANCE_PFN(vkEnumerateDeviceExtensionProperties) VK_INSTANCE_PFN(vkEnumeratePhysicalDevices) VK_INSTANCE_PFN(vkGetDeviceProcAddr) VK_INSTANCE_PFN(vkGetPhysicalDeviceFeatures) +VK_INSTANCE_PFN(vkGetPhysicalDeviceFeatures2KHR) VK_INSTANCE_PFN(vkGetPhysicalDeviceFormatProperties) VK_INSTANCE_PFN(vkGetPhysicalDeviceMemoryProperties) VK_INSTANCE_PFN(vkGetPhysicalDeviceQueueFamilyProperties) -- GitLab https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/695
This merge request was approved by Giovanni Mascellani. -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/695
This merge request was approved by Henri Verbeet. -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/695
participants (3)
-
Evan Tang -
Giovanni Mascellani (＠giomasce)
-
Henri Verbeet (＠hverbeet)