The-Homebrew-Cloud/Cemu
2
0
Fork 0
mirror of https://github.com/cemu-project/Cemu.git synced 2024-12-01 21:44:17 +01:00
Code Issues Packages Projects Releases Wiki Activity

fix: streamout

Browse Source
This commit is contained in:
Samuliak 2024-08-19 13:07:30 +02:00
parent 9aa72e6295
commit e2f66b8aa3
3 changed files with 27 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerEmitMSL.cpp
View File

@ -3284,8 +3284,6 @@ static void _emitCFRingWriteCode(LatteDecompilerShaderContext* shaderContext, La
continue; continue;
uint32 u32Offset = streamWrite->exportArrayBase + i; uint32 u32Offset = streamWrite->exportArrayBase + i;
// HACK: disable streamout temporarily, since it causes GPU hangs
continue;
src->addFmt("sb[sbBase{} + {}]", streamWrite->bufferIndex, u32Offset); src->addFmt("sb[sbBase{} + {}]", streamWrite->bufferIndex, u32Offset);
src->add(" = "); src->add(" = ");
@ -3399,8 +3397,6 @@ static void _emitStreamWriteCode(LatteDecompilerShaderContext* shaderContext, La
continue; continue;
uint32 u32Offset = cfInstruction->exportArrayBase + i; uint32 u32Offset = cfInstruction->exportArrayBase + i;
// HACK: disable streamout temporarily, since it causes GPU hangs
continue;
src->addFmt("sb[sbBase{} + {}]", streamoutBufferIndex, u32Offset); src->addFmt("sb[sbBase{} + {}]", streamoutBufferIndex, u32Offset);
src->add(" = "); src->add(" = ");

24
src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp
View File

@ -634,7 +634,9 @@ void MetalRenderer::bufferCache_copy(uint32 srcOffset, uint32 dstOffset, uint32
void MetalRenderer::bufferCache_copyStreamoutToMainBuffer(uint32 srcOffset, uint32 dstOffset, uint32 size) void MetalRenderer::bufferCache_copyStreamoutToMainBuffer(uint32 srcOffset, uint32 dstOffset, uint32 size)
{ {
debug_printf("MetalRenderer::bufferCache_copyStreamoutToMainBuffer not implemented\n"); auto blitCommandEncoder = GetBlitCommandEncoder();
blitCommandEncoder->copyFromBuffer(m_xfbRingBuffer, srcOffset, m_memoryManager->GetBufferCache(), dstOffset, size);
} }
void MetalRenderer::buffer_bindVertexBuffer(uint32 bufferIndex, uint32 offset, uint32 size) void MetalRenderer::buffer_bindVertexBuffer(uint32 bufferIndex, uint32 offset, uint32 size)
@ -669,17 +671,18 @@ RendererShader* MetalRenderer::shader_create(RendererShader::ShaderType type, ui
void MetalRenderer::streamout_setupXfbBuffer(uint32 bufferIndex, sint32 ringBufferOffset, uint32 rangeAddr, uint32 rangeSize) void MetalRenderer::streamout_setupXfbBuffer(uint32 bufferIndex, sint32 ringBufferOffset, uint32 rangeAddr, uint32 rangeSize)
{ {
debug_printf("MetalRenderer::streamout_setupXfbBuffer not implemented\n"); m_state.m_streamoutState.buffers[bufferIndex].enabled = true;
m_state.m_streamoutState.buffers[bufferIndex].ringBufferOffset = ringBufferOffset;
} }
void MetalRenderer::streamout_begin() void MetalRenderer::streamout_begin()
{ {
debug_printf("MetalRenderer::streamout_begin not implemented\n"); // Do nothing
} }
void MetalRenderer::streamout_rendererFinishDrawcall() void MetalRenderer::streamout_rendererFinishDrawcall()
{ {
debug_printf("MetalRenderer::streamout_rendererFinishDrawcall not implemented\n"); // Do nothing
} }
void MetalRenderer::draw_beginSequence() void MetalRenderer::draw_beginSequence()
@ -966,6 +969,10 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
encoderState.m_renderPipelineState = renderPipelineState; encoderState.m_renderPipelineState = renderPipelineState;
} }
// Prepare streamout
m_state.m_streamoutState.verticesPerInstance = count;
LatteStreamout_PrepareDrawcall(count, instanceCount);
// Uniform buffers, textures and samplers // Uniform buffers, textures and samplers
BindStageResources(renderCommandEncoder, vertexShader); BindStageResources(renderCommandEncoder, vertexShader);
BindStageResources(renderCommandEncoder, pixelShader); BindStageResources(renderCommandEncoder, pixelShader);
@ -981,6 +988,8 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
renderCommandEncoder->drawPrimitives(mtlPrimitiveType, baseVertex, count, instanceCount, baseInstance); renderCommandEncoder->drawPrimitives(mtlPrimitiveType, baseVertex, count, instanceCount, baseInstance);
} }
LatteStreamout_FinishDrawcall(false);
LatteGPUState.drawCallCounter++; LatteGPUState.drawCallCounter++;
} }
@ -1498,20 +1507,17 @@ void MetalRenderer::BindStageResources(MTL::RenderCommandEncoder* renderCommandE
{ {
LatteMRT::GetCurrentFragCoordScale(GET_UNIFORM_DATA_PTR(shader->uniform.loc_fragCoordScale)); LatteMRT::GetCurrentFragCoordScale(GET_UNIFORM_DATA_PTR(shader->uniform.loc_fragCoordScale));
} }
// TODO: uncomment
/*
if (shader->uniform.loc_verticesPerInstance >= 0) if (shader->uniform.loc_verticesPerInstance >= 0)
{ {
*(int*)(supportBufferData + ((size_t)shader->uniform.loc_verticesPerInstance / 4)) = m_streamoutState.verticesPerInstance; *(int*)(supportBufferData + ((size_t)shader->uniform.loc_verticesPerInstance / 4)) = m_state.m_streamoutState.verticesPerInstance;
for (sint32 b = 0; b < LATTE_NUM_STREAMOUT_BUFFER; b++) for (sint32 b = 0; b < LATTE_NUM_STREAMOUT_BUFFER; b++)
{ {
if (shader->uniform.loc_streamoutBufferBase[b] >= 0) if (shader->uniform.loc_streamoutBufferBase[b] >= 0)
{ {
*(uint32*)GET_UNIFORM_DATA_PTR(shader->uniform.loc_streamoutBufferBase[b]) = m_streamoutState.buffer[b].ringBufferOffset; *(uint32*)GET_UNIFORM_DATA_PTR(shader->uniform.loc_streamoutBufferBase[b]) = m_state.m_streamoutState.buffers[b].ringBufferOffset;
} }
} }
} }
*/
auto& bufferAllocator = m_memoryManager->GetTemporaryBufferAllocator(); auto& bufferAllocator = m_memoryManager->GetTemporaryBufferAllocator();
size_t size = shader->uniform.uniformRangeSize; size_t size = shader->uniform.uniformRangeSize;

12
src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h
View File

@ -80,6 +80,16 @@ struct MetalEncoderState
size_t m_uniformBufferOffsets[METAL_SHADER_TYPE_TOTAL][MAX_MTL_BUFFERS]; size_t m_uniformBufferOffsets[METAL_SHADER_TYPE_TOTAL][MAX_MTL_BUFFERS];
}; };
struct MetalStreamoutState
{
struct
{
bool enabled;
uint32 ringBufferOffset;
} buffers[LATTE_NUM_STREAMOUT_BUFFER];
sint32 verticesPerInstance;
};
struct MetalState struct MetalState
{ {
MetalEncoderState m_encoderState{}; MetalEncoderState m_encoderState{};
@ -99,6 +109,8 @@ struct MetalState
MTL::Viewport m_viewport; MTL::Viewport m_viewport;
MTL::ScissorRect m_scissor; MTL::ScissorRect m_scissor;
MetalStreamoutState m_streamoutState;
}; };
struct MetalCommandBuffer struct MetalCommandBuffer