End stream-out queries in ClearState(). Closes #1412

* Also add tests of stream-out and DrawAuto()

renderdoc/driver/d3d11/d3d11_context_wrap.cpp

@@ -6427,6 +6427,20 @@ bool WrappedID3D11DeviceContext::Serialise_ClearState(SerialiserType &ser)
 {
   if(IsReplayingAndReading())
   {
+    // end stream-out queries for outgoing targets
+    for(UINT b = 0; b < D3D11_SO_STREAM_COUNT; b++)
+    {
+      ID3D11Buffer *buf = m_CurrentPipelineState->SO.Buffers[b];
+
+      if(buf)
+      {
+        ResourceId id = GetIDForResource(buf);
+
+        m_pRealContext->End(m_StreamOutCounters[id].query);
+        m_StreamOutCounters[id].running = false;
+      }
+    }
+
     m_CurrentPipelineState->Clear();
     m_pRealContext->ClearState();
     VerifyState();
@@ -6454,6 +6468,20 @@ void WrappedID3D11DeviceContext::ClearState()
     m_ContextRecord->AddChunk(scope.Get());
   }
 
+  // end stream-out queries for outgoing targets
+  for(UINT b = 0; b < D3D11_SO_STREAM_COUNT; b++)
+  {
+    ID3D11Buffer *buf = m_CurrentPipelineState->SO.Buffers[b];
+
+    if(buf)
+    {
+      ResourceId id = GetIDForResource(buf);
+
+      m_pRealContext->End(m_StreamOutCounters[id].query);
+      m_StreamOutCounters[id].running = false;
+    }
+  }
+
   m_CurrentPipelineState->Clear();
   VerifyState();
 }

util/test/demos/d3d11/d3d11_stream_out.cpp

@@ -24,7 +24,7 @@
 
 #include "d3d11_test.h"
 
-TEST(D3D11_StreamOut, D3D11GraphicsTest)
+TEST(D3D11_Stream_Out, D3D11GraphicsTest)
 {
   static constexpr const char *Description = "Test using D3D11's streamout feature";
 
@@ -74,9 +74,25 @@ TEST(D3D11_StreamOut, D3D11GraphicsTest)
     ID3D11BufferPtr vb = MakeBuffer().Vertex().Data(DefaultTri);
 
     ID3D11BufferPtr so[2] = {
-        MakeBuffer().StreamOut().Size(2048), MakeBuffer().StreamOut().Size(2048),
+        MakeBuffer().StreamOut().Vertex().Size(2048), MakeBuffer().StreamOut().Vertex().Size(2048),
     };
 
+    D3D11_INPUT_ELEMENT_DESC layoutdesc[] = {
+        {
+            "POSITION", 0, DXGI_FORMAT_R32G32B32A32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0,
+        },
+        {
+            "COLOR", 0, DXGI_FORMAT_R32G32B32A32_FLOAT, 1, 0, D3D11_INPUT_PER_VERTEX_DATA, 0,
+        },
+        {
+            "TEXCOORD", 0, DXGI_FORMAT_R32G32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0,
+        },
+    };
+
+    ID3D11InputLayoutPtr streamoutLayout;
+    CHECK_HR(dev->CreateInputLayout(layoutdesc, ARRAY_COUNT(layoutdesc), vsblob->GetBufferPointer(),
+                                    vsblob->GetBufferSize(), &streamoutLayout));
+
     while(Running())
     {
       ctx->ClearState();
@@ -124,6 +140,51 @@ TEST(D3D11_StreamOut, D3D11GraphicsTest)
 
       ctx->Draw(3, 0);
 
+      ctx->UpdateSubresource(so[0], 0, NULL, empty, 2048, 2048);
+      ctx->UpdateSubresource(so[1], 0, NULL, empty, 2048, 2048);
+
+      // test DrawAuto()
+
+      RSSetViewport({0.0f, 0.0f, (float)screenWidth, (float)screenHeight, 0.0f, 1.0f});
+
+      // draw with streamout and explicitly unbind
+      ctx->SOSetTargets(2, bufs, offs);
+      ctx->Draw(3, 0);
+      ctx->SOSetTargets(0, NULL, NULL);
+
+      RSSetViewport({0.0f, 0.0f, (float)screenWidth / 4.0f, (float)screenHeight / 4.0f, 0.0f, 1.0f});
+
+      ctx->IASetVertexBuffers(0, 2, bufs, &strides[0], offs);
+      ctx->IASetInputLayout(streamoutLayout);
+      ctx->DrawAuto();
+
+      RSSetViewport({0.0f, 0.0f, (float)screenWidth, (float)screenHeight, 0.0f, 1.0f});
+
+      ID3D11Buffer *emptyBuf[2] = {};
+      ctx->IASetVertexBuffers(0, 2, emptyBuf, &strides[0], offs);
+      IASetVertexBuffer(vb, sizeof(DefaultA2V), 0);
+      ctx->IASetInputLayout(defaultLayout);
+
+      // draw with streamout and clear state
+      ctx->SOSetTargets(2, bufs, offs);
+      ctx->DrawInstanced(3, 2, 0, 0);
+      ctx->ClearState();
+
+      ctx->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
+
+      ctx->VSSetShader(vs, NULL, 0);
+      ctx->GSSetShader(gs, NULL, 0);
+      ctx->PSSetShader(ps, NULL, 0);
+
+      ctx->OMSetRenderTargets(1, &bbRTV.GetInterfacePtr(), NULL);
+
+      RSSetViewport({screenWidth / 4.0f, 0.0f, (float)screenWidth / 4.0f,
+                     (float)screenHeight / 4.0f, 0.0f, 1.0f});
+
+      ctx->IASetVertexBuffers(0, 2, bufs, &strides[0], offs);
+      ctx->IASetInputLayout(streamoutLayout);
+      ctx->DrawAuto();
+
       Present();
     }
 

util/test/demos/demos.vcxproj

@@ -145,7 +145,7 @@
     <ClCompile Include="d3d11\d3d11_saturate.cpp" />
     <ClCompile Include="d3d11\d3d11_simple_dispatch.cpp" />
     <ClCompile Include="d3d11\d3d11_simple_triangle.cpp" />
-    <ClCompile Include="d3d11\d3d11_streamout.cpp" />
+    <ClCompile Include="d3d11\d3d11_stream_out.cpp" />
     <ClCompile Include="d3d11\d3d11_stripped_shaders.cpp" />
     <ClCompile Include="d3d11\d3d11_structured_buffer_misaligned_dirty.cpp" />
     <ClCompile Include="d3d11\d3d11_structured_buffer_nested.cpp" />

util/test/demos/demos.vcxproj.filters

@@ -70,9 +70,6 @@
     <ClCompile Include="gl\gl_vao_0.cpp">
       <Filter>OpenGL\demos</Filter>
     </ClCompile>
-    <ClCompile Include="d3d11\d3d11_streamout.cpp">
-      <Filter>D3D11\demos</Filter>
-    </ClCompile>
     <ClCompile Include="d3d11\d3d11_texture_3d.cpp">
       <Filter>D3D11\demos</Filter>
     </ClCompile>
@@ -309,6 +306,9 @@
     <ClCompile Include="vk\vk_image_layouts.cpp">
       <Filter>Vulkan\demos</Filter>
     </ClCompile>
+    <ClCompile Include="d3d11\d3d11_stream_out.cpp">
+      <Filter>D3D11\demos</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="D3D11">

util/test/tests/D3D11/D3D11_Stream_Out.py

@@ -0,0 +1,58 @@
+import rdtest
+import struct
+import renderdoc as rd
+
+
+class D3D11_Stream_Out(rdtest.TestCase):
+    demos_test_name = 'D3D11_Stream_Out'
+
+    def check_capture(self):
+        draw = self.find_draw("Draw")
+
+        self.check(draw is not None)
+
+        self.controller.SetFrameEvent(draw.eventId, False)
+
+        pipe: rd.PipeState = self.controller.GetPipelineState()
+
+        # Get the input data as our reference
+
+        # First draw should have data at offset 0. First buffer has positions, second has colors (with doubled stride)
+
+        vsin = self.get_vsin(draw)
+
+        pos = [(*v['POSITION'], 1.0) for v in vsin]
+        col = [v['COLOR'] for v in vsin]
+
+        d3d11pipe: rd.D3D11State = self.controller.GetD3D11PipelineState()
+
+        so: rd.D3D11StreamOut = d3d11pipe.streamOut
+        so_bytes = self.controller.GetBufferData(so.outputs[0].resourceId, so.outputs[0].byteOffset, 0)
+
+        for i,p in enumerate(pos):
+            so_p = struct.unpack_from("4f", so_bytes, 0 + 4*4*i)
+            if not rdtest.value_compare(p, so_p):
+                raise rdtest.TestFailureException("Streamed-out position {} doesn't match expected {}".format(so_p, p))
+
+        so_bytes = self.controller.GetBufferData(so.outputs[1].resourceId, so.outputs[1].byteOffset, 0)
+
+        for i,c in enumerate(col):
+            so_c = struct.unpack_from("4f", so_bytes, 0 + 8*4*i)
+            if not rdtest.value_compare(c, so_c):
+                raise rdtest.TestFailureException("Streamed-out color {} doesn't match expected {}".format(so_c, c))
+
+        draw_auto = self.find_draw("DrawAuto", draw.eventId)
+
+        # First draw should be 3 vertices
+        if not rdtest.value_compare(draw_auto.numIndices, 3):
+            raise rdtest.TestFailureException("First DrawAuto() draws {} vertices".format(draw_auto.numIndices))
+
+        draw_auto = self.find_draw("DrawAuto", draw_auto.eventId+1)
+
+        # Second draw should be 6 vertices (3 vertices, instanced twice
+        if not rdtest.value_compare(draw_auto.numIndices, 6):
+            raise rdtest.TestFailureException("Second DrawAuto() draws {} vertices".format(draw_auto.numIndices))
+        if not rdtest.value_compare(draw_auto.numInstances, 1):
+            raise rdtest.TestFailureException("Second DrawAuto() draws {} instances".format(draw_auto.numInstances))
+
+        rdtest.log.success("First draw stream-out data is correct")