[d3d9] WIP: only copy smaller used float constants

doitsujin · Sep 9, 2021 · 3c8d284 · 3c8d284
1 parent 4759c6f
commit 3c8d284
Show file tree

Hide file tree

Showing 2 changed files with 39 additions and 5 deletions.
diff --git a/src/d3d9/d3d9_device.cpp b/src/d3d9/d3d9_device.cpp
@@ -3209,6 +3209,8 @@ namespace dxvk {
     UINT   Vector4fCount) {
     D3D9DeviceLock lock = LockDevice();
 
+    m_psFloatConstsCount = std::max(m_psFloatConstsCount, StartRegister + Vector4fCount);
+
     return SetShaderConstants <
       DxsoProgramTypes::PixelShader,
       D3D9ConstantType::Float>(
@@ -4836,17 +4838,21 @@ namespace dxvk {
 
 
   void D3D9DeviceEx::CreateConstantBuffers() {
+    auto& robustness2 = m_dxvkDevice->adapter()->devicePropertiesExt().extRobustness2;
+    const uint32_t ssboSizeAlignment = robustness2.robustStorageBufferAccessSizeAlignment;
+    const uint32_t uboSizeAlignment = robustness2.robustUniformBufferAccessSizeAlignment;
+
     if (!m_isSWVP) {
       m_consts[DxsoProgramTypes::VertexShader].buffer =
       CreateConstantBuffer(m_dxsoOptions.vertexConstantBufferAsSSBO,
-                          m_vsLayout.totalSize(),
+                          align(m_vsLayout.totalSize(), m_dxsoOptions.vertexConstantBufferAsSSBO ? ssboSizeAlignment : uboSizeAlignment),
                           DxsoProgramType::VertexShader,
                           DxsoConstantBuffers::VSConstantBuffer);
     }
     // SWVP constant buffers are created late based on the amount of constants set by the application
     m_consts[DxsoProgramTypes::PixelShader].buffer =
     CreateConstantBuffer(false,
-                        m_psLayout.totalSize(),
+                        align(m_psLayout.totalSize(), uboSizeAlignment),
                         DxsoProgramType::PixelShader,
                         DxsoConstantBuffers::PSConstantBuffer);
 
@@ -4981,6 +4987,31 @@ namespace dxvk {
 
     constSet.dirty = false;
 
+    const uint32_t floatCount = ShaderStage == DxsoProgramType::VertexShader ? m_definedVSFloatConstsCount : m_psFloatConstsCount;
+
+    const uint32_t intRange = caps::MaxOtherConstants * sizeof(Vector4i);
+    const uint32_t intDataSize = std::min(constSet.meta.maxConstIndexI + 1u, caps::MaxOtherConstants) * sizeof(Vector4i);
+    uint32_t floatDataSize = std::min(constSet.meta.maxConstIndexF + 1u, floatCount) * sizeof(Vector4);
+
+    auto& robustness2 = m_dxvkDevice->adapter()->devicePropertiesExt().extRobustness2;
+    uint32_t alignment = robustness2.robustUniformBufferAccessSizeAlignment;
+    alignment = std::max(alignment, 64u); // Make sure we do not recreate the buffer because the new one has to be a tiny bit larger
+
+    const uint32_t bufferSize = std::max(align(floatDataSize + intRange, alignment), alignment);
+    floatDataSize = bufferSize - intRange; // Read additional floats for padding so we don't end up with garbage data
+
+    if (m_boundConstantsBufferSize != bufferSize) {
+      constexpr uint32_t slotId = computeResourceSlotId(ShaderStage, DxsoBindingType::ConstantBuffer, 0);
+      EmitCs([
+        cBuffer = constSet.buffer,
+        cSlotId = slotId,
+        cSize   = bufferSize
+      ] (DxvkContext* ctx) {
+        ctx->bindResourceBuffer(cSlotId,
+          DxvkBufferSlice(cBuffer, 0, cSize));
+      });
+    }
+
     DxvkBufferSliceHandle slice = constSet.buffer->allocSlice();
 
     EmitCs([
@@ -4992,10 +5023,10 @@ namespace dxvk {
 
     auto* dst = reinterpret_cast<HardwareLayoutType*>(slice.mapPtr);
 
-    if (constSet.meta.maxConstIndexF)
-      std::memcpy(dst->fConsts, Src.fConsts, constSet.meta.maxConstIndexF * sizeof(Vector4));
     if (constSet.meta.maxConstIndexI)
-      std::memcpy(dst->iConsts, Src.iConsts, constSet.meta.maxConstIndexI * sizeof(Vector4i));
+      std::memcpy(dst->iConsts, Src.iConsts, intDataSize);
+    if (constSet.meta.maxConstIndexF)
+      std::memcpy(dst->fConsts, Src.fConsts, floatDataSize);
 
     if (constSet.meta.needsConstantCopies) {
       Vector4* data = reinterpret_cast<Vector4*>(dst->fConsts);

diff --git a/src/d3d9/d3d9_device.h b/src/d3d9/d3d9_device.h
@@ -1227,6 +1227,9 @@ namespace dxvk {
     uint32_t                        m_definedVSIntConstsCount   = 0;
     uint32_t                        m_definedVSBoolConstsCount  = 0;
 
+    uint32_t                        m_psFloatConstsCount = 0;
+    uint32_t                        m_boundConstantsBufferSize = 0;
+
     D3D9ConstantLayout              m_vsLayout;
     D3D9ConstantLayout              m_psLayout;
     D3D9ConstantSets                m_consts[DxsoProgramTypes::Count];