Use thread group count x/y/z for mesh draw commands (API253012)

Graphics/GraphicsEngine/include/DeviceContextBase.hpp

@@ -62,7 +62,7 @@ bool VerifyDispatchComputeAttribs        (const DispatchComputeAttribs&
 bool VerifyDispatchComputeIndirectAttribs(const DispatchComputeIndirectAttribs& Attribs);
 // clang-format on
 
-bool VerifyDrawMeshAttribs(Uint32 MaxDrawMeshTasksCount, const DrawMeshAttribs& Attribs);
+bool VerifyDrawMeshAttribs(const MeshShaderProperties& MeshShaderProps, const DrawMeshAttribs& Attribs);
 bool VerifyDrawMeshIndirectAttribs(const DrawMeshIndirectAttribs& Attribs, Uint32 IndirectCmdStride);
 
 bool VerifyResolveTextureSubresourceAttribs(const ResolveTextureSubresourceAttribs& ResolveAttribs,
@@ -2170,7 +2170,7 @@ inline void DeviceContextBase<ImplementationTraits>::DvpVerifyDrawMeshArguments(
                   "DrawMesh command arguments are invalid: pipeline state '",
                   m_pPipelineState->GetDesc().Name, "' is not a mesh pipeline.");
 
-    DEV_CHECK_ERR(VerifyDrawMeshAttribs(m_pDevice->GetAdapterInfo().MeshShader.MaxTaskCount, Attribs), "DrawMeshAttribs are invalid");
+    DEV_CHECK_ERR(VerifyDrawMeshAttribs(m_pDevice->GetAdapterInfo().MeshShader, Attribs), "DrawMeshAttribs are invalid");
 }
 
 template <typename ImplementationTraits>

Graphics/GraphicsEngine/interface/APIInfo.h

@@ -30,7 +30,7 @@
 /// \file
 /// Diligent API information
 
-#define DILIGENT_API_VERSION 253011
+#define DILIGENT_API_VERSION 253012
 
 #include "../../../Primitives/interface/BasicTypes.h"
 

Graphics/GraphicsEngine/interface/DeviceContext.h

@@ -540,20 +540,43 @@ typedef struct DrawIndexedIndirectAttribs DrawIndexedIndirectAttribs;
 /// This structure is used by IDeviceContext::DrawMesh().
 struct DrawMeshAttribs
 {
-    /// The number of dispatched groups
-    Uint32 ThreadGroupCount DEFAULT_INITIALIZER(1);
+    /// The number of groups dispatched in X direction.
+    Uint32 ThreadGroupCountX DEFAULT_INITIALIZER(1);
+
+    /// The number of groups dispatched in Y direction.
+    Uint32 ThreadGroupCountY DEFAULT_INITIALIZER(1);
+
+    /// The number of groups dispatched in Y direction.
+    Uint32 ThreadGroupCountZ DEFAULT_INITIALIZER(1);
 
     /// Additional flags, see Diligent::DRAW_FLAGS.
-    DRAW_FLAGS Flags        DEFAULT_INITIALIZER(DRAW_FLAG_NONE);
+    DRAW_FLAGS Flags         DEFAULT_INITIALIZER(DRAW_FLAG_NONE);
 
 #if DILIGENT_CPP_INTERFACE
     /// Initializes the structure members with default values.
     constexpr DrawMeshAttribs() noexcept {}
 
-    /// Initializes the structure with user-specified values.
-    constexpr DrawMeshAttribs(Uint32     _ThreadGroupCount,
-                              DRAW_FLAGS _Flags) noexcept :
-        ThreadGroupCount {_ThreadGroupCount},
+    explicit constexpr DrawMeshAttribs(Uint32     _ThreadGroupCountX,
+                                       DRAW_FLAGS _Flags = DRAW_FLAG_NONE) noexcept :
+        ThreadGroupCountX{_ThreadGroupCountX},
+        Flags            {_Flags}
+    {}
+
+    constexpr DrawMeshAttribs(Uint32     _ThreadGroupCountX,
+                              Uint32     _ThreadGroupCountY,
+                              DRAW_FLAGS _Flags = DRAW_FLAG_NONE) noexcept :
+        ThreadGroupCountX{_ThreadGroupCountX},
+        ThreadGroupCountY{_ThreadGroupCountY},
+        Flags            {_Flags}
+    {}
+
+    constexpr DrawMeshAttribs(Uint32     _ThreadGroupCountX,
+                              Uint32     _ThreadGroupCountY,
+                              Uint32     _ThreadGroupCountZ,
+                              DRAW_FLAGS _Flags = DRAW_FLAG_NONE) noexcept :
+        ThreadGroupCountX{_ThreadGroupCountX},
+        ThreadGroupCountY{_ThreadGroupCountY},
+        ThreadGroupCountZ{_ThreadGroupCountZ},
         Flags            {_Flags}
     {}
 #endif

Graphics/GraphicsEngine/interface/GraphicsTypes.h

@@ -2254,8 +2254,17 @@ typedef struct RayTracingProperties RayTracingProperties;
 /// Mesh Shader Properties
 struct MeshShaderProperties
 {
-    /// The maximum number of mesh shader tasks per draw command.
-    Uint32 MaxTaskCount DEFAULT_INITIALIZER(0);
+    /// The maximum number of mesh shader thread groups in X direction.
+    Uint32 MaxThreadGroupCountX DEFAULT_INITIALIZER(0);
+
+    /// The maximum number of mesh shader thread groups in Y direction.
+    Uint32 MaxThreadGroupCountY DEFAULT_INITIALIZER(0);
+
+    /// The maximum number of mesh shader thread groups in Z direction.
+    Uint32 MaxThreadGroupCountZ DEFAULT_INITIALIZER(0);
+
+    /// The total maximum number of mesh shader groups per draw command.
+    Uint32 MaxThreadGroupTotalCount DEFAULT_INITIALIZER(0);
 
 #if DILIGENT_CPP_INTERFACE
     /// Comparison operator tests if two structures are equivalent
@@ -2266,7 +2275,10 @@ struct MeshShaderProperties
     /// - False otherwise.
     constexpr bool operator==(const MeshShaderProperties& RHS) const
     {
-        return MaxTaskCount == RHS.MaxTaskCount;
+        return MaxThreadGroupCountX     == RHS.MaxThreadGroupCountX &&
+               MaxThreadGroupCountY     == RHS.MaxThreadGroupCountY &&
+               MaxThreadGroupCountZ     == RHS.MaxThreadGroupCountZ &&
+               MaxThreadGroupTotalCount == RHS.MaxThreadGroupTotalCount;
     }
 #endif
 };

Graphics/GraphicsEngine/src/DeviceContextBase.cpp

@@ -1,5 +1,5 @@
 /*
- *  Copyright 2019-2022 Diligent Graphics LLC
+ *  Copyright 2019-2023 Diligent Graphics LLC
  *  Copyright 2015-2019 Egor Yusov
  *
  *  Licensed under the Apache License, Version 2.0 (the "License");
@@ -77,15 +77,31 @@ bool VerifyDrawIndexedAttribs(const DrawIndexedAttribs& Attribs)
     return true;
 }
 
-bool VerifyDrawMeshAttribs(Uint32 MaxDrawMeshTasksCount, const DrawMeshAttribs& Attribs)
+bool VerifyDrawMeshAttribs(const MeshShaderProperties& MeshShaderProps, const DrawMeshAttribs& Attribs)
 {
 #define CHECK_DRAW_MESH_ATTRIBS(Expr, ...) CHECK_PARAMETER(Expr, "Draw mesh attribs are invalid: ", __VA_ARGS__)
 
-    if (Attribs.ThreadGroupCount == 0)
-        LOG_INFO_MESSAGE("DrawMeshAttribs.ThreadGroupCount is 0. This is OK as the draw command will be ignored, but may be unintentional.");
-
-    CHECK_DRAW_MESH_ATTRIBS(Attribs.ThreadGroupCount <= MaxDrawMeshTasksCount,
-                            "ThreadGroupCount (", Attribs.ThreadGroupCount, ") must not exceed ", MaxDrawMeshTasksCount);
+    if (Attribs.ThreadGroupCountX == 0)
+        LOG_INFO_MESSAGE("DrawMeshAttribs.ThreadGroupCountX is 0. This is OK as the draw command will be ignored, but may be unintentional.");
+    if (Attribs.ThreadGroupCountY == 0)
+        LOG_INFO_MESSAGE("DrawMeshAttribs.ThreadGroupCountY is 0. This is OK as the draw command will be ignored, but may be unintentional.");
+    if (Attribs.ThreadGroupCountZ == 0)
+        LOG_INFO_MESSAGE("DrawMeshAttribs.ThreadGroupCountZ is 0. This is OK as the draw command will be ignored, but may be unintentional.");
+
+    CHECK_DRAW_MESH_ATTRIBS(Attribs.ThreadGroupCountX <= MeshShaderProps.MaxThreadGroupCountX,
+                            "ThreadGroupCountX (", Attribs.ThreadGroupCountX, ") must not exceed MeshShaderProps.MaxThreadGroupCountX (",
+                            MeshShaderProps.MaxThreadGroupCountX, ").");
+    CHECK_DRAW_MESH_ATTRIBS(Attribs.ThreadGroupCountY <= MeshShaderProps.MaxThreadGroupCountY,
+                            "ThreadGroupCountY (", Attribs.ThreadGroupCountY, ") must not exceed MeshShaderProps.MaxThreadGroupCountY (",
+                            MeshShaderProps.MaxThreadGroupCountY, ").");
+    CHECK_DRAW_MESH_ATTRIBS(Attribs.ThreadGroupCountZ <= MeshShaderProps.MaxThreadGroupCountZ,
+                            "ThreadGroupCountZ (", Attribs.ThreadGroupCountZ, ") must not exceed MeshShaderProps.MaxThreadGroupCountZ (",
+                            MeshShaderProps.MaxThreadGroupCountZ, ").");
+
+    const auto TotalGroups = Uint64{Attribs.ThreadGroupCountX} * Uint64{Attribs.ThreadGroupCountY} * Uint64{Attribs.ThreadGroupCountZ};
+    CHECK_DRAW_MESH_ATTRIBS(TotalGroups <= MeshShaderProps.MaxThreadGroupTotalCount,
+                            "Total thread group count (", TotalGroups, ") must not exceed MeshShaderProps.MaxThreadGroupTotalCount (",
+                            MeshShaderProps.MaxThreadGroupTotalCount, ").");
 
 #undef CHECK_DRAW_MESH_ATTRIBS
 

Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp

@@ -794,9 +794,9 @@ void DeviceContextD3D12Impl::DrawMesh(const DrawMeshAttribs& Attribs)
     auto& GraphCtx = GetCmdContext().AsGraphicsContext6();
     PrepareForDraw(GraphCtx, Attribs.Flags);
 
-    if (Attribs.ThreadGroupCount > 0)
+    if (Attribs.ThreadGroupCountX > 0 && Attribs.ThreadGroupCountY > 0 && Attribs.ThreadGroupCountZ > 0)
     {
-        GraphCtx.DrawMesh(Attribs.ThreadGroupCount, 1, 1);
+        GraphCtx.DrawMesh(Attribs.ThreadGroupCountX, Attribs.ThreadGroupCountY, Attribs.ThreadGroupCountZ);
         ++m_State.NumCommands;
     }
 }

Graphics/GraphicsEngineD3D12/src/EngineFactoryD3D12.cpp

@@ -1,5 +1,5 @@
 /*
- *  Copyright 2019-2022 Diligent Graphics LLC
+ *  Copyright 2019-2023 Diligent Graphics LLC
  *  Copyright 2015-2019 Egor Yusov
  *
  *  Licensed under the Apache License, Version 2.0 (the "License");
@@ -760,8 +760,11 @@ GraphicsAdapterInfo EngineFactoryD3D12Impl::GetGraphicsAdapterInfo(void*
             Features.MeshShaders = DEVICE_FEATURE_STATE_ENABLED;
 
             auto& MeshProps{AdapterInfo.MeshShader};
-            MeshProps.MaxTaskCount = 64000; // from specs: https://microsoft.github.io/DirectX-Specs/d3d/MeshShader.html#dispatchmesh-api
-            ASSERT_SIZEOF(MeshProps, 4, "Did you add a new member to MeshShaderProperties? Please initialize it here.");
+            MeshProps.MaxThreadGroupCountX     = 65536; // from specs: https://microsoft.github.io/DirectX-Specs/d3d/MeshShader.html#dispatchmesh-api
+            MeshProps.MaxThreadGroupCountY     = 65536;
+            MeshProps.MaxThreadGroupCountZ     = 65536;
+            MeshProps.MaxThreadGroupTotalCount = 1u << 22u;
+            ASSERT_SIZEOF(MeshProps, 16, "Did you add a new member to MeshShaderProperties? Please initialize it here.");
         }
 
         Features.ShaderResourceRuntimeArray = DEVICE_FEATURE_STATE_ENABLED;

Graphics/GraphicsEngineVulkan/include/VulkanUtilities/VulkanCommandBuffer.hpp

@@ -166,14 +166,14 @@ class VulkanCommandBuffer
 #endif
     }
 
-    __forceinline void DrawMesh(uint32_t TaskCount, uint32_t FirstTask)
+    __forceinline void DrawMesh(uint32_t TaskCountX, uint32_t TaskCountY, uint32_t TaskCountZ)
     {
 #if DILIGENT_USE_VOLK
         VERIFY_EXPR(m_VkCmdBuffer != VK_NULL_HANDLE);
         VERIFY(m_State.RenderPass != VK_NULL_HANDLE, "vkCmdDrawMeshTasksEXT() must be called inside render pass");
         VERIFY(m_State.GraphicsPipeline != VK_NULL_HANDLE, "No graphics pipeline bound");
 
-        vkCmdDrawMeshTasksEXT(m_VkCmdBuffer, TaskCount, 1, 1);
+        vkCmdDrawMeshTasksEXT(m_VkCmdBuffer, TaskCountX, TaskCountY, TaskCountZ);
 #else
         UNSUPPORTED("DrawMesh is not supported when vulkan library is linked statically");
 #endif

Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp

@@ -909,9 +909,9 @@ void DeviceContextVkImpl::DrawMesh(const DrawMeshAttribs& Attribs)
 
     PrepareForDraw(Attribs.Flags);
 
-    if (Attribs.ThreadGroupCount > 0)
+    if (Attribs.ThreadGroupCountX > 0 && Attribs.ThreadGroupCountY > 0 && Attribs.ThreadGroupCountZ > 0)
     {
-        m_CommandBuffer.DrawMesh(Attribs.ThreadGroupCount, 0);
+        m_CommandBuffer.DrawMesh(Attribs.ThreadGroupCountX, Attribs.ThreadGroupCountY, Attribs.ThreadGroupCountZ);
         ++m_State.NumCommands;
     }
 }

Graphics/GraphicsEngineVulkan/src/EngineFactoryVk.cpp

@@ -262,8 +262,11 @@ GraphicsAdapterInfo GetPhysicalDeviceGraphicsAdapterInfo(const VulkanUtilities::
     if (AdapterInfo.Features.MeshShaders)
     {
         auto& MeshProps{AdapterInfo.MeshShader};
-        MeshProps.MaxTaskCount = vkDeviceExtProps.MeshShader.maxTaskWorkGroupCount[0];
-        ASSERT_SIZEOF(MeshProps, 4, "Did you add a new member to MeshShaderProperties? Please initialize it here.");
+        MeshProps.MaxThreadGroupCountX     = vkDeviceExtProps.MeshShader.maxMeshWorkGroupCount[0];
+        MeshProps.MaxThreadGroupCountY     = vkDeviceExtProps.MeshShader.maxMeshWorkGroupCount[1];
+        MeshProps.MaxThreadGroupCountZ     = vkDeviceExtProps.MeshShader.maxMeshWorkGroupCount[2];
+        MeshProps.MaxThreadGroupTotalCount = vkDeviceExtProps.MeshShader.maxMeshWorkGroupTotalCount;
+        ASSERT_SIZEOF(MeshProps, 16, "Did you add a new member to MeshShaderProperties? Please initialize it here.");
     }
 
     // Compute shader properties

ReleaseHistory.md

@@ -1,5 +1,6 @@
 ## Current progress
 
+* Use thread group count X/Y/Z for mesh draw commands (API253012)
 * Added `ShaderMacroArray` struct (API253011)
   * The `Macros` member of `ShaderCreateInfo` struct is now of type `ShaderMacroArray`
 * Replaced `ResourceMappingDesc` with `ResourceMappingCreateInfo` (API253010)