|
| 1 | +VK: drop the stages of shader features the device lacks from stage masks. |
| 2 | + |
| 3 | +GetPipelineStageFlags and GetShaderStageFlags (ConversionVK.h) turn a mask naming every shader |
| 4 | +stage ("StageBits::ALL_SHADERS" in a barrier or a fence wait, a pipeline layout whose stages are |
| 5 | +"ALL") into every Vulkan stage bit, whatever the device has. On a device without tessellation, |
| 6 | +geometry, task, mesh or ray tracing shaders the validation layer then refuses the barrier |
| 7 | +(VUID-VkMemoryBarrier2-srcStageMask-07946 / -dstStageMask-07946 for the ray tracing stage, 03929, |
| 8 | +03934 and 03935 for geometry, mesh and task), the push constant range and the push |
| 9 | +(VUID-VkPushConstantRange-stageFlags-parameter, VUID-VkPushConstantsInfo-stageFlags-parameter: ray |
| 10 | +tracing, task and mesh shader stage bits without their extension). Measured on v191: MoltenVK |
| 11 | +(no geometry, mesh or ray tracing) and NVIDIA with "disableVKRayTracing". CmdSetDescriptorSet |
| 12 | +already adds the mesh stages only with "features.meshShader". |
| 13 | + |
| 14 | +DeviceVK records the shader stages it lacks once, from the features it enabled; barriers and |
| 15 | +fence waits drop their pipeline stages, and a pipeline layout drops them from its stages, which |
| 16 | +"ALL" and "ALL_SHADERS" resolve to. |
| 17 | + |
| 18 | +diff --git a/Source/VK/CommandBufferVK.hpp b/Source/VK/CommandBufferVK.hpp |
| 19 | +index 8dcb4c25..0bacce0b 100644 |
| 20 | +--- a/Source/VK/CommandBufferVK.hpp |
| 21 | ++++ b/Source/VK/CommandBufferVK.hpp |
| 22 | +@@ -2638,6 +2638,8 @@ NRI_INLINE void CommandBufferVK::DispatchIndirect(const Buffer& buffer, uint64_t |
| 23 | + } |
| 24 | + |
| 25 | + NRI_INLINE void CommandBufferVK::Barrier(const BarrierDesc& barrierDesc) { |
| 26 | ++ const VkPipelineStageFlags2 unsupportedStages = m_Device.GetUnsupportedPipelineStages(); |
| 27 | ++ |
| 28 | + // Global |
| 29 | + Scratch<VkMemoryBarrier2> memoryBarriers = NRI_ALLOCATE_SCRATCH(m_Device, VkMemoryBarrier2, barrierDesc.globalNum); |
| 30 | + for (uint32_t i = 0; i < barrierDesc.globalNum; i++) { |
| 31 | +@@ -2645,9 +2647,9 @@ NRI_INLINE void CommandBufferVK::Barrier(const BarrierDesc& barrierDesc) { |
| 32 | + |
| 33 | + VkMemoryBarrier2& out = memoryBarriers[i]; |
| 34 | + out = {VK_STRUCTURE_TYPE_MEMORY_BARRIER_2}; |
| 35 | +- out.srcStageMask = GetPipelineStageFlags(in.before.stages); |
| 36 | ++ out.srcStageMask = GetPipelineStageFlags(in.before.stages) & ~unsupportedStages; |
| 37 | + out.srcAccessMask = GetAccessFlags(in.before.access); |
| 38 | +- out.dstStageMask = GetPipelineStageFlags(in.after.stages); |
| 39 | ++ out.dstStageMask = GetPipelineStageFlags(in.after.stages) & ~unsupportedStages; |
| 40 | + out.dstAccessMask = GetAccessFlags(in.after.access); |
| 41 | + } |
| 42 | + |
| 43 | +@@ -2660,9 +2662,9 @@ NRI_INLINE void CommandBufferVK::Barrier(const BarrierDesc& barrierDesc) { |
| 44 | + |
| 45 | + VkBufferMemoryBarrier2& out = bufferBarriers[i]; |
| 46 | + out = {VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2}; |
| 47 | +- out.srcStageMask = GetPipelineStageFlags(in.before.stages); |
| 48 | ++ out.srcStageMask = GetPipelineStageFlags(in.before.stages) & ~unsupportedStages; |
| 49 | + out.srcAccessMask = GetAccessFlags(in.before.access); |
| 50 | +- out.dstStageMask = GetPipelineStageFlags(in.after.stages); |
| 51 | ++ out.dstStageMask = GetPipelineStageFlags(in.after.stages) & ~unsupportedStages; |
| 52 | + out.dstAccessMask = GetAccessFlags(in.after.access); |
| 53 | + out.srcQueueFamilyIndex = GetQueueFamilyIndex(in.srcQueue, isConcurrent); |
| 54 | + out.dstQueueFamilyIndex = GetQueueFamilyIndex(in.dstQueue, isConcurrent); |
| 55 | +@@ -2683,9 +2685,9 @@ NRI_INLINE void CommandBufferVK::Barrier(const BarrierDesc& barrierDesc) { |
| 56 | + |
| 57 | + VkImageMemoryBarrier2& out = textureBarriers[i]; |
| 58 | + out = {VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER_2}; |
| 59 | +- out.srcStageMask = GetPipelineStageFlags(in.before.stages); |
| 60 | ++ out.srcStageMask = GetPipelineStageFlags(in.before.stages) & ~unsupportedStages; |
| 61 | + out.srcAccessMask = GetAccessFlags(in.before.access); |
| 62 | +- out.dstStageMask = GetPipelineStageFlags(in.after.stages); |
| 63 | ++ out.dstStageMask = GetPipelineStageFlags(in.after.stages) & ~unsupportedStages; |
| 64 | + out.dstAccessMask = GetAccessFlags(in.after.access); |
| 65 | + out.oldLayout = GetImageLayout(in.before.layout); |
| 66 | + out.newLayout = GetImageLayout(in.after.layout); |
| 67 | +diff --git a/Source/VK/DeviceVK.h b/Source/VK/DeviceVK.h |
| 68 | +index 741c61fa..5848dd8c 100644 |
| 69 | +--- a/Source/VK/DeviceVK.h |
| 70 | ++++ b/Source/VK/DeviceVK.h |
| 71 | +@@ -199,6 +199,14 @@ struct DeviceVK final : public DeviceBase { |
| 72 | + return m_Vma; |
| 73 | + } |
| 74 | + |
| 75 | ++ inline StageBits GetSupportedShaderStages() const { |
| 76 | ++ return m_SupportedShaderStages; |
| 77 | ++ } |
| 78 | ++ |
| 79 | ++ inline VkPipelineStageFlags2 GetUnsupportedPipelineStages() const { |
| 80 | ++ return m_UnsupportedPipelineStages; |
| 81 | ++ } |
| 82 | ++ |
| 83 | + template <typename Implementation, typename Interface, typename... Args> |
| 84 | + [[nodiscard]] inline Result CreateImplementation(Interface*& entity, Args&&... args) { |
| 85 | + return nri::CreateImplementation(GetAllocationCallbacks(), entity, Allocate<Implementation>(GetAllocationCallbacks(), *this), std::forward<Args>(args)...); |
| 86 | +@@ -333,6 +341,8 @@ private: |
| 87 | + uint32_t m_MinorVersion = 0; |
| 88 | + VkTimeDomainKHR m_CalibratedTimestampCPUTimeDomain = VK_TIME_DOMAIN_DEVICE_KHR; |
| 89 | + uint64_t m_NonCoherentAtomSize = 1; |
| 90 | ++ VkPipelineStageFlags2 m_UnsupportedPipelineStages = 0; |
| 91 | ++ StageBits m_SupportedShaderStages = StageBits::ALL_SHADERS; |
| 92 | + DeviceLostDump m_DeviceLostDump = {}; |
| 93 | + bool m_OwnsNativeObjects = true; |
| 94 | + bool m_IsMemoryZeroInitializationEnabled = false; |
| 95 | +diff --git a/Source/VK/DeviceVK.hpp b/Source/VK/DeviceVK.hpp |
| 96 | +index f0d801bc..d3359ded 100644 |
| 97 | +--- a/Source/VK/DeviceVK.hpp |
| 98 | ++++ b/Source/VK/DeviceVK.hpp |
| 99 | +@@ -1663,6 +1663,33 @@ Result DeviceVK::Create(const DeviceCreationDesc& desc, const DeviceCreationVKDe |
| 100 | + m_Desc.features.tessellationShader = features.features.tessellationShader != 0; |
| 101 | + m_Desc.features.geometryShader = features.features.geometryShader != 0; |
| 102 | + m_Desc.features.meshShader = MeshShaderFeatures.meshShader != 0 && MeshShaderFeatures.taskShader != 0; |
| 103 | ++ |
| 104 | ++ // Shader stages the device lacks, dropped from stage masks naming them (like "ALL_SHADERS") |
| 105 | ++ if (!m_Desc.features.tessellationShader) { |
| 106 | ++ m_SupportedShaderStages &= ~StageBits::TESSELLATION_SHADERS; |
| 107 | ++ m_UnsupportedPipelineStages |= VK_PIPELINE_STAGE_2_TESSELLATION_CONTROL_SHADER_BIT | VK_PIPELINE_STAGE_2_TESSELLATION_EVALUATION_SHADER_BIT; |
| 108 | ++ } |
| 109 | ++ |
| 110 | ++ if (!m_Desc.features.geometryShader) { |
| 111 | ++ m_SupportedShaderStages &= ~StageBits::GEOMETRY_SHADER; |
| 112 | ++ m_UnsupportedPipelineStages |= VK_PIPELINE_STAGE_2_GEOMETRY_SHADER_BIT; |
| 113 | ++ } |
| 114 | ++ |
| 115 | ++ if (!MeshShaderFeatures.taskShader) { |
| 116 | ++ m_SupportedShaderStages &= ~StageBits::TASK_SHADER; |
| 117 | ++ m_UnsupportedPipelineStages |= VK_PIPELINE_STAGE_2_TASK_SHADER_BIT_EXT; |
| 118 | ++ } |
| 119 | ++ |
| 120 | ++ if (!MeshShaderFeatures.meshShader) { |
| 121 | ++ m_SupportedShaderStages &= ~StageBits::MESH_SHADER; |
| 122 | ++ m_UnsupportedPipelineStages |= VK_PIPELINE_STAGE_2_MESH_SHADER_BIT_EXT; |
| 123 | ++ } |
| 124 | ++ |
| 125 | ++ if (!RayTracingPipelineFeatures.rayTracingPipeline) { |
| 126 | ++ m_SupportedShaderStages &= ~StageBits::RAY_TRACING_SHADERS; |
| 127 | ++ m_UnsupportedPipelineStages |= VK_PIPELINE_STAGE_2_RAY_TRACING_SHADER_BIT_KHR; |
| 128 | ++ } |
| 129 | ++ |
| 130 | + m_Desc.features.lowLatency = m_IsSupported.presentId != 0 && IsExtensionSupported(VK_NV_LOW_LATENCY_2_EXTENSION_NAME, desiredDeviceExts); |
| 131 | + m_Desc.features.componentSwizzle = true; |
| 132 | + m_Desc.features.independentFrontAndBackStencilReferenceAndMasks = true; |
| 133 | +diff --git a/Source/VK/PipelineLayoutVK.hpp b/Source/VK/PipelineLayoutVK.hpp |
| 134 | +index cfb15cb9..2d6d1046 100644 |
| 135 | +--- a/Source/VK/PipelineLayoutVK.hpp |
| 136 | ++++ b/Source/VK/PipelineLayoutVK.hpp |
| 137 | +@@ -32,6 +32,7 @@ PipelineLayoutVK::~PipelineLayoutVK() { |
| 138 | + |
| 139 | + Result PipelineLayoutVK::Create(const PipelineLayoutDesc& pipelineLayoutDesc) { |
| 140 | + m_ShaderStages = GetEffectiveShaderStages(pipelineLayoutDesc.shaderStages); |
| 141 | ++ m_ShaderStages &= m_Device.GetSupportedShaderStages(); |
| 142 | + |
| 143 | + // Binding offsets |
| 144 | + bool ignoreGlobalSPIRVOffsets = (pipelineLayoutDesc.flags & PipelineLayoutBits::IGNORE_GLOBAL_SPIRV_OFFSETS) != 0; |
| 145 | +diff --git a/Source/VK/QueueVK.hpp b/Source/VK/QueueVK.hpp |
| 146 | +index 0e6f33ee..e53d448e 100644 |
| 147 | +--- a/Source/VK/QueueVK.hpp |
| 148 | ++++ b/Source/VK/QueueVK.hpp |
| 149 | +@@ -81,13 +81,14 @@ NRI_INLINE void QueueVK::GetCalibratedTimestamps(uint64_t& timestampGPU, uint64_ |
| 150 | + |
| 151 | + NRI_INLINE Result QueueVK::Submit(const QueueSubmitDesc& queueSubmitDesc) { |
| 152 | + ExclusiveScope lock(*m_Lock); |
| 153 | ++ const VkPipelineStageFlags2 unsupportedStages = m_Device.GetUnsupportedPipelineStages(); |
| 154 | + |
| 155 | + Scratch<VkSemaphoreSubmitInfo> waitSemaphores = NRI_ALLOCATE_SCRATCH(m_Device, VkSemaphoreSubmitInfo, queueSubmitDesc.waitFenceNum); |
| 156 | + for (uint32_t i = 0; i < queueSubmitDesc.waitFenceNum; i++) { |
| 157 | + waitSemaphores[i] = {VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO}; |
| 158 | + waitSemaphores[i].semaphore = *(FenceVK*)queueSubmitDesc.waitFences[i].fence; |
| 159 | + waitSemaphores[i].value = queueSubmitDesc.waitFences[i].value; |
| 160 | +- waitSemaphores[i].stageMask = GetPipelineStageFlags(queueSubmitDesc.waitFences[i].stages); |
| 161 | ++ waitSemaphores[i].stageMask = GetPipelineStageFlags(queueSubmitDesc.waitFences[i].stages) & ~unsupportedStages; |
| 162 | + } |
| 163 | + |
| 164 | + Scratch<VkCommandBufferSubmitInfo> commandBuffers = NRI_ALLOCATE_SCRATCH(m_Device, VkCommandBufferSubmitInfo, queueSubmitDesc.commandBufferNum); |
| 165 | +@@ -101,7 +102,7 @@ NRI_INLINE Result QueueVK::Submit(const QueueSubmitDesc& queueSubmitDesc) { |
| 166 | + signalSemaphores[i] = {VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO}; |
| 167 | + signalSemaphores[i].semaphore = *(FenceVK*)queueSubmitDesc.signalFences[i].fence; |
| 168 | + signalSemaphores[i].value = queueSubmitDesc.signalFences[i].value; |
| 169 | +- signalSemaphores[i].stageMask = GetPipelineStageFlags(queueSubmitDesc.signalFences[i].stages); |
| 170 | ++ signalSemaphores[i].stageMask = GetPipelineStageFlags(queueSubmitDesc.signalFences[i].stages) & ~unsupportedStages; |
| 171 | + } |
| 172 | + |
| 173 | + VkSubmitInfo2 submitInfo = {VK_STRUCTURE_TYPE_SUBMIT_INFO_2}; |
0 commit comments