GPU: Fix incorrect block size when D3D12 uniform buffer is rotated (#13469)

(cherry picked from commit 07af4b237bccf56cc793ac4dc9280c2f97b4c2f3)

src/gpu/d3d12/SDL_gpu_d3d12.c

@@ -1194,7 +1194,6 @@ struct D3D12UniformBuffer
     D3D12Buffer *buffer;
     Uint32 writeOffset;
     Uint32 drawOffset;
-    Uint32 currentBlockSize;
 };
 
 // Forward function declarations
@@ -4498,7 +4497,6 @@ static D3D12UniformBuffer *D3D12_INTERNAL_AcquireUniformBufferFromPool(
 
     SDL_UnlockMutex(renderer->acquireUniformBufferLock);
 
-    uniformBuffer->currentBlockSize = 0;
     uniformBuffer->drawOffset = 0;
     uniformBuffer->writeOffset = 0;
 
@@ -4537,6 +4535,7 @@ static void D3D12_INTERNAL_PushUniformData(
     Uint32 length)
 {
     D3D12UniformBuffer *uniformBuffer;
+    Uint32 blockSize;
 
     if (shaderStage == SDL_GPU_SHADERSTAGE_VERTEX) {
         if (commandBuffer->vertexUniformBuffers[slotIndex] == NULL) {
@@ -4561,13 +4560,13 @@ static void D3D12_INTERNAL_PushUniformData(
         return;
     }
 
-    uniformBuffer->currentBlockSize =
+    blockSize =
         D3D12_INTERNAL_Align(
             length,
             256);
 
     // If there is no more room, acquire a new uniform buffer
-    if (uniformBuffer->writeOffset + uniformBuffer->currentBlockSize >= UNIFORM_BUFFER_SIZE) {
+    if (uniformBuffer->writeOffset + blockSize >= UNIFORM_BUFFER_SIZE) {
         ID3D12Resource_Unmap(
             uniformBuffer->buffer->handle,
             0,
@@ -4597,7 +4596,7 @@ static void D3D12_INTERNAL_PushUniformData(
         data,
         length);
 
-    uniformBuffer->writeOffset += uniformBuffer->currentBlockSize;
+    uniformBuffer->writeOffset += blockSize;
 
     if (shaderStage == SDL_GPU_SHADERSTAGE_VERTEX) {
         commandBuffer->needVertexUniformBufferBind[slotIndex] = true;