Merge branch 'dev/adunn/fix_corruption_shader' into 'main'

[REMIX-2795][REMIX-3426] Fix vertex corruption See merge request lightspeedrtx/dxvk-remix-nv!984
NVIDIAGameWorks · Sep 17, 2024 · 7bd91b0 · 7bd91b0
2 parents fdab8d7 + deaa10f
commit 7bd91b0
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 13 deletions.
diff --git a/src/d3d9/d3d9_rtx.cpp b/src/d3d9/d3d9_rtx.cpp
@@ -25,11 +25,6 @@ namespace dxvk {
 
   D3D9Rtx::D3D9Rtx(D3D9DeviceEx* d3d9Device, bool enableDrawCallConversion)
     : m_rtStagingData(d3d9Device->GetDXVKDevice(), (VkMemoryPropertyFlagBits) (VK_MEMORY_PROPERTY_HOST_CACHED_BIT | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT))
-    , m_vertexCaptureData(d3d9Device->GetDXVKDevice(), 
-                          (VkMemoryPropertyFlagBits) (VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT), 
-                          (VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT),
-                          (VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT),
-                          VK_ACCESS_TRANSFER_READ_BIT)
     , m_parent(d3d9Device)
     , m_enableDrawCallConversion(enableDrawCallConversion)
     , m_pGeometryWorkers(enableDrawCallConversion ? std::make_unique<GeometryProcessor>(numGeometryProcessingThreads(), "geometry-processing") : nullptr) {
@@ -123,6 +118,15 @@ namespace dxvk {
     return result.slice;
   }
 
+  DxvkBufferSlice allocVertexCaptureBuffer(DxvkDevice* pDevice, const VkDeviceSize size) {
+    DxvkBufferCreateInfo info;
+    info.usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT;
+    info.access = VK_ACCESS_TRANSFER_READ_BIT;
+    info.stages = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT;
+    info.size = size;
+    return DxvkBufferSlice(pDevice->createBuffer(info, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, DxvkMemoryStats::Category::AppBuffer));
+  }
+
   void D3D9Rtx::prepareVertexCapture(const int vertexIndexOffset) {
     ScopedCpuProfileZone();
 
@@ -154,8 +158,7 @@ namespace dxvk {
     const uint32_t stride = sizeof(CapturedVertex);
     const size_t vertexCaptureDataSize = align(geoData.vertexCount * stride, CACHE_LINE_SIZE);
 
-    DxvkBufferSlice slice = m_vertexCaptureData.alloc(CACHE_LINE_SIZE, vertexCaptureDataSize);
-    slice.buffer()->acquire(DxvkAccess::Write);
+    DxvkBufferSlice slice = allocVertexCaptureBuffer(m_parent->GetDXVKDevice().ptr(), vertexCaptureDataSize);
 
     geoData.positionBuffer = RasterBuffer(slice, 0, stride, VK_FORMAT_R32G32B32A32_SFLOAT);
     assert(geoData.positionBuffer.offset() % 4 == 0);

diff --git a/src/d3d9/d3d9_rtx.h b/src/d3d9/d3d9_rtx.h
@@ -180,7 +180,6 @@ namespace dxvk {
     DrawCallState m_activeDrawCallState;
 
     RtxStagingDataAlloc m_rtStagingData;
-    RtxStagingDataAlloc m_vertexCaptureData;
     D3D9DeviceEx* m_parent;
 
     std::optional<D3DPRESENT_PARAMETERS> m_activePresentParams;

diff --git a/src/dxvk/rtx_render/rtx_geometry_utils.cpp b/src/dxvk/rtx_render/rtx_geometry_utils.cpp
@@ -807,7 +807,7 @@ namespace dxvk {
 
     assert(output.buffer->info().size == align(output.stride * input.vertexCount, CACHE_LINE_SIZE));
 
-    bool pendingGpuWrites = input.positionBuffer.isPendingGpuWrite();
+    bool mustUseGPU = input.positionBuffer.isPendingGpuWrite() || input.positionBuffer.mapPtr() == nullptr;
 
     // Interleave vertex data
     InterleaveGeometryArgs args;
@@ -821,7 +821,7 @@ namespace dxvk {
     }
     args.hasNormals = input.normalBuffer.defined();
     if (args.hasNormals) {
-      pendingGpuWrites |= input.normalBuffer.isPendingGpuWrite();
+      mustUseGPU |= input.normalBuffer.isPendingGpuWrite() || input.normalBuffer.mapPtr() == nullptr;
       assert(input.normalBuffer.offsetFromSlice() % 4 == 0);
       args.normalOffset = input.normalBuffer.offsetFromSlice() / 4;
       args.normalStride = input.normalBuffer.stride() / 4;
@@ -832,7 +832,7 @@ namespace dxvk {
     }
     args.hasTexcoord = input.texcoordBuffer.defined();
     if (args.hasTexcoord) {
-      pendingGpuWrites |= input.texcoordBuffer.isPendingGpuWrite();
+      mustUseGPU |= input.texcoordBuffer.isPendingGpuWrite() || input.texcoordBuffer.mapPtr() == nullptr;
       assert(input.texcoordBuffer.offsetFromSlice() % 4 == 0);
       args.texcoordOffset = input.texcoordBuffer.offsetFromSlice() / 4;
       args.texcoordStride = input.texcoordBuffer.stride() / 4;
@@ -843,7 +843,7 @@ namespace dxvk {
     }
     args.hasColor0 = input.color0Buffer.defined();
     if (args.hasColor0) {
-      pendingGpuWrites |= input.color0Buffer.isPendingGpuWrite();
+      mustUseGPU |= input.color0Buffer.isPendingGpuWrite() || input.color0Buffer.mapPtr() == nullptr;
       assert(input.color0Buffer.offsetFromSlice() % 4 == 0);
       args.color0Offset = input.color0Buffer.offsetFromSlice() / 4;
       args.color0Stride = input.color0Buffer.stride() / 4;
@@ -859,7 +859,7 @@ namespace dxvk {
     args.vertexCount = input.vertexCount;
 
     const uint32_t kNumVerticesToProcessOnCPU = 1024;
-    const bool useGPU = input.vertexCount > kNumVerticesToProcessOnCPU || pendingGpuWrites;
+    const bool useGPU = input.vertexCount > kNumVerticesToProcessOnCPU || mustUseGPU;
 
     if (useGPU) {
       ctx->bindResourceBuffer(INTERLEAVE_GEOMETRY_BINDING_OUTPUT, DxvkBufferSlice(output.buffer));