Add experimental PerVertexKHR capture and replay (test entry point only)

ExternalRevisions/SPIRV-Cross_repo_revision

@@ -1 +1 @@
-aa217aeb6c9f0ace7a0ab233b28807edf45eb165
+81739edf516c2bef043ac61f510b4bd78e126a22

MoltenVK/MoltenVK/Commands/MVKCmdDebug.h

@@ -46,6 +46,7 @@ class MVKCmdDebugMarkerBegin : public MVKCmdDebugMarker {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder*) override { return true; }
 
 protected:
 	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
@@ -62,6 +63,7 @@ class MVKCmdDebugMarkerEnd : public MVKCommand {
 	VkResult setContent(MVKCommandBuffer* cmdBuff);
 
 	void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder*) override { return true; }
 
 protected:
 	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
@@ -76,6 +78,7 @@ class MVKCmdDebugMarkerInsert : public MVKCmdDebugMarker {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder*) override { return true; }
 
 protected:
 	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;

MoltenVK/MoltenVK/Commands/MVKCmdDebug.mm

@@ -46,7 +46,7 @@
 	if (mtlCmdEnc) {
 		[mtlCmdEnc pushDebugGroup: _markerName];
 	} else {
-		[cmdEncoder->_mtlCmdBuffer pushDebugGroup: _markerName];
+		cmdEncoder->pushCommandBufferDebugGroup(_markerName);
 	}
 }
 
@@ -63,7 +63,7 @@
 	if (mtlCmdEnc) {
 		[mtlCmdEnc popDebugGroup];
 	} else {
-		[cmdEncoder->_mtlCmdBuffer popDebugGroup];
+		cmdEncoder->popCommandBufferDebugGroup();
 	}
 }
 

MoltenVK/MoltenVK/Commands/MVKCmdDraw.h

@@ -21,9 +21,31 @@
 #include "MVKCommand.h"
 #include "MVKMTLResourceBindings.h"
 #include "MVKSmallVector.h"
+#include "MVKPerVertexReplay.h"
 
 #import <Metal/Metal.h>
 
+class MVKGraphicsPipeline;
+
+// Preserve the conservative power-of-two scratch ceiling used by recording validation.
+static inline bool mvkCanEncodePerVertexDraw(uint32_t vertexCount, uint32_t instanceCount, uint32_t stride, VkPrimitiveTopology topology, uint64_t maxMTLBufferSize) {
+	if (!mvkPerVertexReplayVertexCount(topology)) { return false; }
+	if (!vertexCount || !instanceCount) { return true; }
+	if (!maxMTLBufferSize) { return false; }
+	uint64_t maxBufferSize = 1;
+	while (maxBufferSize <= maxMTLBufferSize / 2) { maxBufferSize *= 2; }
+	uint64_t recordCount = uint64_t(vertexCount) * instanceCount;
+	if (!stride || recordCount > UINT32_MAX || recordCount > maxBufferSize / stride) { return false; }
+	uint64_t primitiveCount = uint64_t(mvkPerVertexPrimitiveCount(vertexCount, topology)) * instanceCount;
+	uint64_t occurrenceCount = primitiveCount * mvkPerVertexReplayVertexCount(topology);
+	return primitiveCount <= UINT32_MAX / 3 && primitiveCount <= maxBufferSize / (3 * sizeof(uint32_t)) && occurrenceCount <= UINT32_MAX / 2 && occurrenceCount <= maxBufferSize / (2 * sizeof(uint32_t));
+}
+
+// PerVertex TES plan: status, replay arguments and constants, per-patch counts and offsets, then, 256-byte
+// aligned for the fragment binding, the patch index of each generated triangle (at most 13 per patch).
+static inline uint64_t mvkPerVertexTessPrimitiveIdOffset(uint64_t patches) { return mvkAlignByteCount(512 + patches * 2 * sizeof(uint32_t), 256); }
+static inline uint64_t mvkPerVertexTessPlanSize(uint64_t patches) { return mvkPerVertexTessPrimitiveIdOffset(patches) + patches * 13 * sizeof(uint32_t); }
+
 
 #pragma mark -
 #pragma mark MVKCmdBindVertexBuffers
@@ -45,6 +67,7 @@ class MVKCmdBindVertexBuffers : public MVKCommand {
 						const VkDeviceSize* pStrides);
 
     void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder* cmdEncoder) override { encode(cmdEncoder); return true; }
 
 protected:
 	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
@@ -77,6 +100,7 @@ class MVKCmdBindIndexBuffer : public MVKCommand {
 						VkIndexType indexType);
 
 	void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder* cmdEncoder) override { encode(cmdEncoder); return true; }
 
 protected:
 	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
@@ -99,6 +123,8 @@ class MVKCmdDraw : public MVKCommand {
 						uint32_t drawIndex = 0);
 
     void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder* cmdEncoder) override;
+	void encodePerVertexInput(MVKCommandEncoder* cmdEncoder, MVKGraphicsPipeline* pipeline);
 	void encodeIndexedIndirect(MVKCommandEncoder* cmdEncoder);
 
 protected:
@@ -127,6 +153,8 @@ class MVKCmdDrawIndexed : public MVKCommand {
 						uint32_t drawIndex = 0);
 
 	void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder* cmdEncoder) override;
+	void encodePerVertexInput(MVKCommandEncoder* cmdEncoder, MVKGraphicsPipeline* pipeline);
 
 protected:
 	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
@@ -218,3 +246,49 @@ class MVKCmdDrawIndexedIndirect : public MVKCommand {
 	id<MTLBuffer> _mtlCountBuffer;
 	VkDeviceSize _mtlCountBufferOffset;
 };
+
+
+#pragma mark -
+#pragma mark MVKCmdDrawMeshTasks
+
+/** Vulkan command to draw mesh workgroups (vkCmdDrawMeshTasksEXT). */
+class MVKCmdDrawMeshTasks : public MVKCommand {
+
+public:
+	VkResult setContent(MVKCommandBuffer* cmdBuff, uint32_t groupCountX, uint32_t groupCountY, uint32_t groupCountZ);
+
+	void encode(MVKCommandEncoder* cmdEncoder) override;
+	bool encodePerVertexPrepass(MVKCommandEncoder* cmdEncoder) override;
+
+protected:
+	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
+	bool encodeMesh(MVKCommandEncoder* cmdEncoder, bool prepass);
+
+	uint32_t _groupCountX;
+	uint32_t _groupCountY;
+	uint32_t _groupCountZ;
+};
+
+
+#pragma mark -
+#pragma mark MVKCmdDrawMeshTasksIndirect
+
+/**
+ * Vulkan command to draw one indirect grid of mesh workgroups (vkCmdDrawMeshTasksIndirectEXT, drawCount 1), with
+ * MVK_CONFIG_MESH_INDIRECT_GPU=1. The grid is read by Metal from the indirect buffer, with no CPU readback.
+ */
+class MVKCmdDrawMeshTasksIndirect : public MVKCommand {
+
+public:
+	VkResult setContent(MVKCommandBuffer* cmdBuff, VkBuffer buffer, VkDeviceSize offset);
+
+	void encode(MVKCommandEncoder* cmdEncoder) override;
+	// A native mesh draw has nothing to capture: nothing to do in the PerVertex prepass, which continues.
+	bool encodePerVertexPrepass(MVKCommandEncoder*) override { return true; }
+
+protected:
+	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
+
+	id<MTLBuffer> _mtlIndirectBuffer;
+	VkDeviceSize _mtlIndirectBufferOffset;
+};