[Vulkan] Non-GS point sprites + minor SPIR-V fixes

src/xenia/gpu/d3d12/d3d12_command_processor.cc

@@ -2268,7 +2268,7 @@ bool D3D12CommandProcessor::IssueDraw(xenos::PrimitiveType primitive_type,
   UpdateSystemConstantValues(
       memexport_used, primitive_polygonal,
       primitive_processing_result.line_loop_closing_index,
-      primitive_processing_result.host_index_endian, viewport_info,
+      primitive_processing_result.host_shader_index_endian, viewport_info,
       used_texture_mask, normalized_depth_control, normalized_color_mask);
 
   // Update constant buffers, descriptors and root parameters.
@@ -2513,7 +2513,7 @@ bool D3D12CommandProcessor::IssueDraw(xenos::PrimitiveType primitive_type,
     }
     ID3D12Resource* scratch_index_buffer = nullptr;
     switch (primitive_processing_result.index_buffer_type) {
-      case PrimitiveProcessor::ProcessedIndexBufferType::kGuest: {
+      case PrimitiveProcessor::ProcessedIndexBufferType::kGuestDMA: {
         if (memexport_used) {
           // If the shared memory is a UAV, it can't be used as an index buffer
           // (UAV is a read/write state, index buffer is a read-only state).
@@ -2545,7 +2545,8 @@ bool D3D12CommandProcessor::IssueDraw(xenos::PrimitiveType primitive_type,
             primitive_processor_->GetConvertedIndexBufferGpuAddress(
                 primitive_processing_result.host_index_buffer_handle);
         break;
-      case PrimitiveProcessor::ProcessedIndexBufferType::kHostBuiltin:
+      case PrimitiveProcessor::ProcessedIndexBufferType::kHostBuiltinForAuto:
+      case PrimitiveProcessor::ProcessedIndexBufferType::kHostBuiltinForDMA:
         index_buffer_view.BufferLocation =
             primitive_processor_->GetBuiltinIndexBufferGpuAddress(
                 primitive_processing_result.host_index_buffer_handle);

src/xenia/gpu/d3d12/d3d12_primitive_processor.cc

@@ -28,7 +28,7 @@ namespace d3d12 {
 D3D12PrimitiveProcessor::~D3D12PrimitiveProcessor() { Shutdown(true); }
 
 bool D3D12PrimitiveProcessor::Initialize() {
-  if (!InitializeCommon(true, false, false, true)) {
+  if (!InitializeCommon(true, false, false, true, true, true)) {
     Shutdown();
     return false;
   }
@@ -83,9 +83,9 @@ void D3D12PrimitiveProcessor::EndFrame() {
   frame_index_buffers_.clear();
 }
 
-bool D3D12PrimitiveProcessor::InitializeBuiltin16BitIndexBuffer(
-    uint32_t index_count, std::function<void(uint16_t*)> fill_callback) {
-  assert_not_zero(index_count);
+bool D3D12PrimitiveProcessor::InitializeBuiltinIndexBuffer(
+    size_t size_bytes, std::function<void(void*)> fill_callback) {
+  assert_not_zero(size_bytes);
   assert_null(builtin_index_buffer_);
   assert_null(builtin_index_buffer_upload_);
 
@@ -94,9 +94,8 @@ bool D3D12PrimitiveProcessor::InitializeBuiltin16BitIndexBuffer(
   ID3D12Device* device = provider.GetDevice();
 
   D3D12_RESOURCE_DESC resource_desc;
-  ui::d3d12::util::FillBufferResourceDesc(
-      resource_desc, UINT64(sizeof(uint16_t) * index_count),
-      D3D12_RESOURCE_FLAG_NONE);
+  ui::d3d12::util::FillBufferResourceDesc(resource_desc, UINT64(size_bytes),
+                                          D3D12_RESOURCE_FLAG_NONE);
   Microsoft::WRL::ComPtr<ID3D12Resource> draw_resource;
   if (FAILED(device->CreateCommittedResource(
           &ui::d3d12::util::kHeapPropertiesDefault,
@@ -105,8 +104,8 @@ bool D3D12PrimitiveProcessor::InitializeBuiltin16BitIndexBuffer(
           IID_PPV_ARGS(&draw_resource)))) {
     XELOGE(
         "D3D12 primitive processor: Failed to create the built-in index "
-        "buffer GPU resource with {} 16-bit indices",
-        index_count);
+        "buffer GPU resource with {} bytes",
+        size_bytes);
     return false;
   }
   Microsoft::WRL::ComPtr<ID3D12Resource> upload_resource;
@@ -117,8 +116,8 @@ bool D3D12PrimitiveProcessor::InitializeBuiltin16BitIndexBuffer(
           IID_PPV_ARGS(&upload_resource)))) {
     XELOGE(
         "D3D12 primitive processor: Failed to create the built-in index "
-        "buffer upload resource with {} 16-bit indices",
-        index_count);
+        "buffer upload resource with {} bytes",
+        size_bytes);
     return false;
   }
 
@@ -127,8 +126,8 @@ bool D3D12PrimitiveProcessor::InitializeBuiltin16BitIndexBuffer(
   if (FAILED(upload_resource->Map(0, &upload_read_range, &mapping))) {
     XELOGE(
         "D3D12 primitive processor: Failed to map the built-in index buffer "
-        "upload resource with {} 16-bit indices",
-        index_count);
+        "upload resource with {} bytes",
+        size_bytes);
     return false;
   }
   fill_callback(reinterpret_cast<uint16_t*>(mapping));

src/xenia/gpu/d3d12/d3d12_primitive_processor.h

@@ -56,9 +56,8 @@ class D3D12PrimitiveProcessor final : public PrimitiveProcessor {
   }
 
  protected:
-  bool InitializeBuiltin16BitIndexBuffer(
-      uint32_t index_count,
-      std::function<void(uint16_t*)> fill_callback) override;
+  bool InitializeBuiltinIndexBuffer(
+      size_t size_bytes, std::function<void(void*)> fill_callback) override;
 
   void* RequestHostConvertedIndexBufferForCurrentFrame(
       xenos::IndexFormat format, uint32_t index_count, bool coalign_for_simd,