The-Homebrew-Cloud/skyline
2
0
Fork 0
mirror of https://github.com/skyline-emu/skyline.git synced 2024-11-27 04:04:14 +01:00
Code Issues Packages Projects Releases Wiki Activity

Adapt Maxwell 3D engine to new interconnect code

Browse Source 
Removes all usage of graphics_context.h from the codebase, exclusively using the new interconnect and its dirty tracking system. While porting the code a number of bugs were discovered such as not respecting the base instance or primitive type override, which have all been fixed. Currently only clears and constant buffer updates are implemented but due to the dirty state system allowing register handling on the interconnect end there shouldn't end up being many more changes.
This commit is contained in:
Billy Laws 2022-09-02 12:24:58 +01:00
parent ef11900a39
commit 3649d4c779
2 changed files with 146 additions and 636 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

756
app/src/main/cpp/skyline/soc/gm20b/engines/maxwell_3d.cpp
View File

@ -3,92 +3,135 @@
// Copyright © 2018-2020 fincs (https://github.com/devkitPro/deko3d) // Copyright © 2018-2020 fincs (https://github.com/devkitPro/deko3d)
#include <boost/preprocessor/repeat.hpp> #include <boost/preprocessor/repeat.hpp>
#include <gpu/interconnect/command_executor.h>
#include <soc/gm20b/channel.h>
#include <soc.h> #include <soc.h>
#include "maxwell/types.h"
#include "maxwell_3d.h" #include "maxwell_3d.h"
namespace skyline::soc::gm20b::engine::maxwell3d { namespace skyline::soc::gm20b::engine::maxwell3d {
#define REGTYPE(state) gpu::interconnect::maxwell3d::state::EngineRegisters
static gpu::interconnect::maxwell3d::PipelineState::EngineRegisters MakePipelineStateRegisters(const Maxwell3D::Registers &registers) {
return {
.colorRenderTargetsRegisters = util::MergeInto<REGTYPE(ColorRenderTargetState), type::ColorTargetCount>(*registers.colorTargets),
.depthRenderTargetRegisters = {*registers.ztSize, *registers.ztOffset, *registers.ztFormat, *registers.ztBlockSize, *registers.ztArrayPitch, *registers.ztSelect, *registers.ztLayer},
.vertexInputRegisters = {*registers.vertexStreams, *registers.vertexStreamInstance, *registers.vertexAttributes},
.inputAssemblyRegisters = {*registers.primitiveRestartEnable}
};
}
static gpu::interconnect::maxwell3d::ActiveState::EngineRegisters MakeActiveStateRegisters(const Maxwell3D::Registers &registers) {
return {
.pipelineRegisters = MakePipelineStateRegisters(registers),
.vertexBuffersRegisters = util::MergeInto<REGTYPE(VertexBufferState), type::VertexStreamCount>(*registers.vertexStreams, *registers.vertexStreamLimits),
.indexBufferRegisters = {*registers.indexBuffer},
.transformFeedbackBuffersRegisters = util::MergeInto<REGTYPE(TransformFeedbackBufferState), type::StreamOutBufferCount>(*registers.streamOutBuffers, *registers.streamOutEnable),
.viewportsRegisters = util::MergeInto<REGTYPE(ViewportState), type::ViewportCount>(*registers.viewports, *registers.viewportClips, *registers.windowOrigin, *registers.viewportScaleOffsetEnable),
.scissorsRegisters = util::MergeInto<REGTYPE(ScissorState), type::ViewportCount>(*registers.scissors),
.lineWidthRegisters = {*registers.lineWidth, *registers.lineWidthAliased, *registers.aliasedLineWidthEnable},
.depthBiasRegisters = {*registers.depthBias, *registers.depthBiasClamp, *registers.slopeScaleDepthBias},
.blendConstantsRegisters = {*registers.blendConsts},
.depthBoundsRegisters = {*registers.depthBoundsMin, *registers.depthBoundsMin},
.stencilValuesRegisters = {*registers.stencilValues, *registers.backStencilValues, *registers.twoSidedStencilTestEnable},
};
}
static gpu::interconnect::maxwell3d::Maxwell3D::EngineRegisterBundle MakeEngineRegisters(const Maxwell3D::Registers &registers) {
return {
.activeStateRegisters = MakeActiveStateRegisters(registers),
.clearRegisters = {registers.scissors[0], registers.viewportClips[0], *registers.clearRect, *registers.colorClearValue, *registers.zClearValue, *registers.stencilClearValue, *registers.surfaceClip, *registers.clearSurfaceControl},
.constantBufferSelectorRegisters = {*registers.constantBufferSelector}
};
}
#undef REGTYPE
type::DrawTopology Maxwell3D::GetCurrentTopology() {
return registers.primitiveTopologyControl->override == type::PrimitiveTopologyControl::Override::UseTopologyInBeginMethods ?
registers.begin->op : type::ConvertPrimitiveTopologyToDrawTopology(*registers.primitiveTopology);
}
Maxwell3D::Maxwell3D(const DeviceState &state, ChannelContext &channelCtx, MacroState &macroState, gpu::interconnect::CommandExecutor &executor) Maxwell3D::Maxwell3D(const DeviceState &state, ChannelContext &channelCtx, MacroState &macroState, gpu::interconnect::CommandExecutor &executor)
: MacroEngineBase(macroState), : MacroEngineBase{macroState},
syncpoints(state.soc->host1x.syncpoints), syncpoints{state.soc->host1x.syncpoints},
context(*state.gpu, channelCtx, executor), i2m{channelCtx},
i2m(channelCtx), dirtyManager{registers},
channelCtx(channelCtx) { interconnect{*state.gpu, channelCtx, executor, dirtyManager, MakeEngineRegisters(registers)},
channelCtx{channelCtx} {
executor.AddFlushCallback([this]() { FlushEngineState(); }); executor.AddFlushCallback([this]() { FlushEngineState(); });
InitializeRegisters(); InitializeRegisters();
} }
void Maxwell3D::FlushDeferredDraw() { __attribute__((always_inline)) void Maxwell3D::FlushDeferredDraw() {
if (deferredDraw.pending) { if (deferredDraw.pending) {
deferredDraw.pending = false; deferredDraw.pending = false;
interconnect.Draw(deferredDraw.drawTopology, deferredDraw.indexed, deferredDraw.drawCount, deferredDraw.drawFirst, deferredDraw.instanceCount, deferredDraw.drawBaseVertex, deferredDraw.drawBaseInstance);
if (deferredDraw.indexed)
context.DrawIndexed(deferredDraw.drawCount, deferredDraw.drawFirst, deferredDraw.instanceCount, deferredDraw.drawBaseVertex);
else
context.Draw(deferredDraw.drawCount, deferredDraw.drawFirst, deferredDraw.instanceCount);
deferredDraw.instanceCount = 1; deferredDraw.instanceCount = 1;
} }
} }
void Maxwell3D::HandleMethod(u32 method, u32 argument) { __attribute__((always_inline)) void Maxwell3D::HandleMethod(u32 method, u32 argument) {
if (method != ENGINE_STRUCT_OFFSET(mme, shadowRamControl)) { if (method == ENGINE_STRUCT_OFFSET(mme, shadowRamControl)) [[unlikely]] {
if (shadowRegisters.mme->shadowRamControl == type::MmeShadowRamControl::MethodTrack || shadowRegisters.mme->shadowRamControl == type::MmeShadowRamControl::MethodTrackWithFilter) shadowRegisters.raw[method] = registers.raw[method] = argument;
shadowRegisters.raw[method] = argument; return;
else if (shadowRegisters.mme->shadowRamControl == type::MmeShadowRamControl::MethodReplay)
argument = shadowRegisters.raw[method];
} }
if (shadowRegisters.mme->shadowRamControl == type::MmeShadowRamControl::MethodTrack || shadowRegisters.mme->shadowRamControl == type::MmeShadowRamControl::MethodTrackWithFilter) [[unlikely]]
shadowRegisters.raw[method] = argument;
else if (shadowRegisters.mme->shadowRamControl == type::MmeShadowRamControl::MethodReplay) [[unlikely]]
argument = shadowRegisters.raw[method];
bool redundant{registers.raw[method] == argument}; bool redundant{registers.raw[method] == argument};
registers.raw[method] = argument; registers.raw[method] = argument;
if (batchConstantBufferUpdate.Active()) { // TODO COMBINE THESE
if (batchLoadConstantBuffer.Active()) {
switch (method) { switch (method) {
// Add to the batch constant buffer update buffer // Add to the batch constant buffer update buffer
// Return early here so that any code below can rely on the fact that any cbuf updates will always be the first of a batch // Return early here so that any code below can rely on the fact that any cbuf updates will always be the first of a batch
#define CBUF_UPDATE_CALLBACKS(z, index, data_) \ #define LOAD_CONSTANT_BUFFER_CALLBACKS(z, index, data_) \
ENGINE_STRUCT_ARRAY_CASE(constantBufferUpdate, data, index, { \ ENGINE_STRUCT_ARRAY_CASE(loadConstantBuffer, data, index, { \
batchConstantBufferUpdate.buffer.push_back(data); \ batchLoadConstantBuffer.buffer.push_back(argument); \
registers.constantBufferUpdate->offset += 4; \ registers.loadConstantBuffer->offset += 4; \
return; \ return; \
}) })
BOOST_PP_REPEAT(16, CBUF_UPDATE_CALLBACKS, 0) BOOST_PP_REPEAT(16, LOAD_CONSTANT_BUFFER_CALLBACKS, 0)
#undef CBUF_UPDATE_CALLBACKS #undef LOAD_CONSTANT_BUFFER_CALLBACKS
default: default:
// When a method other than constant buffer update is called submit our submit the previously built-up update as a batch // When a method other than constant buffer update is called submit our submit the previously built-up update as a batch
context.ConstantBufferUpdate(batchConstantBufferUpdate.buffer, batchConstantBufferUpdate.Invalidate()); interconnect.LoadConstantBuffer(batchLoadConstantBuffer.buffer, batchLoadConstantBuffer.Invalidate());
batchConstantBufferUpdate.Reset(); batchLoadConstantBuffer.Reset();
break; // Continue on here to handle the actual method break; // Continue on here to handle the actual method
} }
} } else if (deferredDraw.pending) { // See DeferredDrawState comment for full details
// See DeferredDrawState comment for full details
if (deferredDraw.pending) {
switch (method) { switch (method) {
ENGINE_CASE(vertexBeginGl, { ENGINE_CASE(begin, {
if (vertexBeginGl.instanceNext) { if (begin.instanceId == Registers::Begin::InstanceId::Subsequent) {
if (deferredDraw.drawTopology != vertexBeginGl.topology && !vertexBeginGl.instanceContinue) if (deferredDraw.drawTopology != begin.op &&
registers.primitiveTopologyControl->override == type::PrimitiveTopologyControl::Override::UseTopologyInBeginMethods)
Logger::Warn("Vertex topology changed partway through instanced draw!"); Logger::Warn("Vertex topology changed partway through instanced draw!");
deferredDraw.instanceCount++; deferredDraw.instanceCount++;
} else if (vertexBeginGl.instanceContinue) { } else {
FlushDeferredDraw(); FlushDeferredDraw();
break; // This instanced draw is finished, continue on to handle the actual method break; // This instanced draw is finished, continue on to handle the next draw
} }
return; return;
}) })
// Can be ignored since we handle drawing in draw{Vertex,Index}Count // Can be ignored since we handle drawing in draw{Vertex,Index}Count
ENGINE_CASE(vertexEndGl, { return; }) ENGINE_CASE(end, { return; })
// Draws here can be ignored since they're just repeats of the original instanced draw // Draws here can be ignored since they're just repeats of the original instanced draw
ENGINE_CASE(drawVertexCount, { ENGINE_CASE(drawVertexArray, {
if (!redundant) if (!redundant)
Logger::Warn("Vertex count changed partway through instanced draw!"); Logger::Warn("Vertex count changed partway through instanced draw!");
return; return;
}) })
ENGINE_CASE(drawIndexCount, { ENGINE_CASE(drawIndexBuffer, {
if (!redundant) if (!redundant)
Logger::Warn("Index count changed partway through instanced draw!"); Logger::Warn("Index count changed partway through instanced draw!");
return; return;
@ -101,568 +144,39 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
} }
} }
if (!redundant) { if (!redundant) {
dirtyManager.MarkDirty(method);
switch (method) { switch (method) {
ENGINE_STRUCT_CASE(mme, shadowRamControl, { #define VERTEX_STREAM_CALLBACKS(z, idx, data) \
shadowRegisters.mme->shadowRamControl = shadowRamControl; ENGINE_ARRAY_STRUCT_CASE(vertexStreams, idx, format, { \
}) interconnect.directState.vertexInput.SetStride(idx, format.stride); \
#define RENDER_TARGET_ARRAY(z, index, data) \
ENGINE_ARRAY_STRUCT_STRUCT_CASE(renderTargets, index, address, high, { \
context.SetColorRenderTargetAddressHigh(index, high); \
}) \ }) \
ENGINE_ARRAY_STRUCT_STRUCT_CASE(renderTargets, index, address, low, { \ ENGINE_ARRAY_STRUCT_CASE(vertexStreams, idx, frequency, { \
context.SetColorRenderTargetAddressLow(index, low); \ interconnect.directState.vertexInput.SetDivisor(idx, frequency); \
}) \ }) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, width, { \ ENGINE_ARRAY_CASE(vertexStreamInstance, idx, { \
context.SetColorRenderTargetWidth(index, width); \ interconnect.directState.vertexInput.SetInputRate(idx, vertexStreamInstance); \
}) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, height, { \
context.SetColorRenderTargetHeight(index, height); \
}) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, format, { \
context.SetColorRenderTargetFormat(index, format); \
}) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, tileMode, { \
context.SetColorRenderTargetTileMode(index, tileMode); \
}) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, arrayMode, { \
context.SetColorRenderTargetArrayMode(index, arrayMode); \
}) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, layerStrideLsr2, { \
context.SetColorRenderTargetLayerStride(index, layerStrideLsr2); \
}) \
ENGINE_ARRAY_STRUCT_CASE(renderTargets, index, baseLayer, { \
context.SetColorRenderTargetBaseLayer(index, baseLayer); \
}) })
BOOST_PP_REPEAT(8, RENDER_TARGET_ARRAY, 0) BOOST_PP_REPEAT(16, VERTEX_STREAM_CALLBACKS, 0)
static_assert(type::RenderTargetCount == 8 && type::RenderTargetCount < BOOST_PP_LIMIT_REPEAT); static_assert(type::VertexStreamCount == 16 && type::VertexStreamCount < BOOST_PP_LIMIT_REPEAT);
#undef RENDER_TARGET_ARRAY #undef VERTEX_STREAM_CALLBACKS
ENGINE_CASE(pointSpriteSize, {
context.SetPointSpriteSize(pointSpriteSize);
})
ENGINE_CASE(depthTargetEnable, {
context.SetDepthRenderTargetEnabled(depthTargetEnable);
})
ENGINE_STRUCT_CASE(depthTargetAddress, high, {
context.SetDepthRenderTargetAddressHigh(high);
})
ENGINE_STRUCT_CASE(depthTargetAddress, low, {
context.SetDepthRenderTargetAddressLow(low);
})
ENGINE_CASE(depthTargetFormat, {
context.SetDepthRenderTargetFormat(depthTargetFormat);
})
ENGINE_CASE(depthTargetTileMode, {
context.SetDepthRenderTargetTileMode(depthTargetTileMode);
})
ENGINE_CASE(depthTargetLayerStride, {
context.SetDepthRenderTargetLayerStride(depthTargetLayerStride);
})
ENGINE_CASE(depthTargetWidth, {
context.SetDepthRenderTargetWidth(depthTargetWidth);
})
ENGINE_CASE(depthTargetHeight, {
context.SetDepthRenderTargetHeight(depthTargetHeight);
})
ENGINE_CASE(depthTargetArrayMode, {
context.SetDepthRenderTargetArrayMode(depthTargetArrayMode);
})
ENGINE_CASE(linkedTscHandle, {
context.SetTscIndexLinked(linkedTscHandle);
});
#define VIEWPORT_TRANSFORM_CALLBACKS(_z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, scaleX, { \
context.SetViewportX(index, scaleX, registers.viewportTransforms[index].translateX); \
}) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, translateX, { \
context.SetViewportX(index, registers.viewportTransforms[index].scaleX, translateX); \
}) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, scaleY, { \
context.SetViewportY(index, scaleY, registers.viewportTransforms[index].translateY); \
}) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, translateY, { \
context.SetViewportY(index, registers.viewportTransforms[index].scaleY, translateY); \
}) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, scaleZ, { \
context.SetViewportZ(index, scaleZ, registers.viewportTransforms[index].translateZ); \
}) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, translateZ, { \
context.SetViewportZ(index, registers.viewportTransforms[index].scaleZ, translateZ); \
}) \
ENGINE_ARRAY_STRUCT_CASE(viewportTransforms, index, swizzles, { \
context.SetViewportSwizzle(index, swizzles.x, swizzles.y, swizzles.z, swizzles.w); \
})
BOOST_PP_REPEAT(16, VIEWPORT_TRANSFORM_CALLBACKS, 0)
static_assert(type::ViewportCount == 16 && type::ViewportCount < BOOST_PP_LIMIT_REPEAT);
#undef VIEWPORT_TRANSFORM_CALLBACKS
#define COLOR_CLEAR_CALLBACKS(z, index, data) \
ENGINE_ARRAY_CASE(clearColorValue, index, { \
context.UpdateClearColorValue(index, clearColorValue); \
})
BOOST_PP_REPEAT(4, COLOR_CLEAR_CALLBACKS, 0)
static_assert(4 < BOOST_PP_LIMIT_REPEAT);
#undef COLOR_CLEAR_CALLBACKS
ENGINE_CASE(clearDepthValue, {
context.UpdateClearDepthValue(clearDepthValue);
})
ENGINE_CASE(clearStencilValue, {
context.UpdateClearStencilValue(clearStencilValue);
})
ENGINE_STRUCT_CASE(polygonMode, front, {
context.SetPolygonModeFront(front);
})
ENGINE_CASE(tessellationPatchSize, {
context.SetTessellationPatchSize(tessellationPatchSize);
})
ENGINE_CASE(tessellationMode, {
context.SetTessellationMode(tessellationMode.primitive, tessellationMode.spacing, tessellationMode.winding);
})
#define TRANSFORM_FEEDBACK_CALLBACKS(z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(transformFeedbackBuffers, index, enable, { \
context.SetTransformFeedbackBufferEnabled(index, enable); \
}) \
ENGINE_ARRAY_STRUCT_STRUCT_CASE(transformFeedbackBuffers, index, iova, high, { \
context.SetTransformFeedbackBufferIovaHigh(index, high); \
}) \
ENGINE_ARRAY_STRUCT_STRUCT_CASE(transformFeedbackBuffers, index, iova, low, { \
context.SetTransformFeedbackBufferIovaLow(index, low); \
}) \
ENGINE_ARRAY_STRUCT_CASE(transformFeedbackBuffers, index, size, { \
context.SetTransformFeedbackBufferSize(index, size); \
}) \
ENGINE_ARRAY_STRUCT_CASE(transformFeedbackBuffers, index, offset, { \
context.SetTransformFeedbackBufferOffset(index, offset); \
}) \
ENGINE_ARRAY_STRUCT_CASE(transformFeedbackBufferStates, index, varyingCount, { \
context.SetTransformFeedbackBufferVaryingCount(index, varyingCount); \
}) \
ENGINE_ARRAY_STRUCT_CASE(transformFeedbackBufferStates, index, stride, { \
context.SetTransformFeedbackBufferStride(index, stride); \
})
BOOST_PP_REPEAT(4, TRANSFORM_FEEDBACK_CALLBACKS, 0)
static_assert(type::TransformFeedbackBufferCount == 4 && type::TransformFeedbackBufferCount < BOOST_PP_LIMIT_REPEAT);
#undef TRANSFORM_FEEDBACK_CALLBACKS
ENGINE_CASE(transformFeedbackEnable, {
context.SetTransformFeedbackEnabled(transformFeedbackEnable);
})
ENGINE_STRUCT_CASE(depthBiasEnable, point, {
context.SetDepthBiasPointEnabled(point);
})
ENGINE_STRUCT_CASE(depthBiasEnable, line, {
context.SetDepthBiasLineEnabled(line);
})
ENGINE_STRUCT_CASE(depthBiasEnable, fill, {
context.SetDepthBiasFillEnabled(fill);
})
#define SCISSOR_CALLBACKS(z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(scissors, index, enable, { \
context.SetScissor(index, enable ? registers.scissors[index] : std::optional<type::Scissor>{}); \
}) \
ENGINE_ARRAY_STRUCT_CASE(scissors, index, horizontal, { \
context.SetScissorHorizontal(index, horizontal); \
}) \
ENGINE_ARRAY_STRUCT_CASE(scissors, index, vertical, { \
context.SetScissorVertical(index, vertical); \
})
BOOST_PP_REPEAT(16, SCISSOR_CALLBACKS, 0)
static_assert(type::ViewportCount == 16 && type::ViewportCount < BOOST_PP_LIMIT_REPEAT);
#undef SCISSOR_CALLBACKS
ENGINE_CASE(commonColorWriteMask, {
if (commonColorWriteMask) {
auto colorWriteMask{registers.colorWriteMask[0]};
for (u32 index{}; index != type::RenderTargetCount; index++)
context.SetColorWriteMask(index, colorWriteMask);
} else {
for (u32 index{}; index != type::RenderTargetCount; index++)
context.SetColorWriteMask(index, registers.colorWriteMask[index]);
}
})
ENGINE_CASE(renderTargetControl, {
context.UpdateRenderTargetControl(renderTargetControl);
})
ENGINE_CASE(depthTestEnable, {
context.SetDepthTestEnabled(depthTestEnable);
})
ENGINE_CASE(depthTestFunc, {
context.SetDepthTestFunction(depthTestFunc);
})
ENGINE_CASE(depthWriteEnable, {
context.SetDepthWriteEnabled(depthWriteEnable);
})
ENGINE_CASE(depthBoundsEnable, {
context.SetDepthBoundsTestEnabled(depthBoundsEnable);
})
ENGINE_CASE(depthBoundsNear, {
context.SetMinDepthBounds(depthBoundsNear);
})
ENGINE_CASE(depthBoundsFar, {
context.SetMaxDepthBounds(depthBoundsFar);
})
ENGINE_CASE(stencilEnable, {
context.SetStencilTestEnabled(stencilEnable);
})
ENGINE_STRUCT_CASE(stencilFront, failOp, {
context.SetStencilFrontFailOp(failOp);
})
ENGINE_STRUCT_CASE(stencilFront, zFailOp, {
context.SetStencilFrontDepthFailOp(zFailOp);
})
ENGINE_STRUCT_CASE(stencilFront, passOp, {
context.SetStencilFrontPassOp(passOp);
})
ENGINE_STRUCT_CASE(stencilFront, compareOp, {
context.SetStencilFrontCompareOp(compareOp);
})
ENGINE_STRUCT_CASE(stencilFront, compareReference, {
context.SetStencilFrontReference(compareReference);
})
ENGINE_STRUCT_CASE(stencilFront, compareMask, {
context.SetStencilFrontCompareMask(compareMask);
})
ENGINE_STRUCT_CASE(stencilFront, writeMask, {
context.SetStencilFrontWriteMask(writeMask);
})
ENGINE_CASE(stencilTwoSideEnable, {
context.SetStencilTwoSideEnabled(stencilTwoSideEnable);
})
ENGINE_STRUCT_CASE(stencilBack, failOp, {
context.SetStencilBackFailOp(failOp);
})
ENGINE_STRUCT_CASE(stencilBack, zFailOp, {
context.SetStencilBackDepthFailOp(zFailOp);
})
ENGINE_STRUCT_CASE(stencilBack, passOp, {
context.SetStencilBackPassOp(passOp);
})
ENGINE_STRUCT_CASE(stencilBack, compareOp, {
context.SetStencilBackCompareOp(compareOp);
})
ENGINE_STRUCT_CASE(stencilBackExtra, compareReference, {
context.SetStencilBackReference(compareReference);
})
ENGINE_STRUCT_CASE(stencilBackExtra, compareMask, {
context.SetStencilBackCompareMask(compareMask);
})
ENGINE_STRUCT_CASE(stencilBackExtra, writeMask, {
context.SetStencilBackWriteMask(writeMask);
})
ENGINE_CASE(windowOriginMode, {
context.SetViewportOrigin(windowOriginMode.isOriginLowerLeft);
context.SetFrontFaceFlipEnabled(windowOriginMode.flipFrontFace);
})
ENGINE_CASE(independentBlendEnable, {
context.SetIndependentBlendingEnabled(independentBlendEnable);
})
ENGINE_CASE(alphaTestEnable, {
context.SetAlphaTestEnabled(alphaTestEnable);
})
#define SET_COLOR_BLEND_CONSTANT_CALLBACK(z, index, data) \
ENGINE_ARRAY_CASE(blendConstant, index, { \
context.SetColorBlendConstant(index, blendConstant); \
})
BOOST_PP_REPEAT(4, SET_COLOR_BLEND_CONSTANT_CALLBACK, 0)
static_assert(type::BlendColorChannelCount == 4 && type::BlendColorChannelCount < BOOST_PP_LIMIT_REPEAT);
#undef SET_COLOR_BLEND_CONSTANT_CALLBACK
ENGINE_STRUCT_CASE(blendStateCommon, colorOp, {
context.SetColorBlendOp(colorOp);
})
ENGINE_STRUCT_CASE(blendStateCommon, colorSrcFactor, {
context.SetSrcColorBlendFactor(colorSrcFactor);
})
ENGINE_STRUCT_CASE(blendStateCommon, colorDstFactor, {
context.SetDstColorBlendFactor(colorDstFactor);
})
ENGINE_STRUCT_CASE(blendStateCommon, alphaOp, {
context.SetAlphaBlendOp(alphaOp);
})
ENGINE_STRUCT_CASE(blendStateCommon, alphaSrcFactor, {
context.SetSrcAlphaBlendFactor(alphaSrcFactor);
})
ENGINE_STRUCT_CASE(blendStateCommon, alphaDstFactor, {
context.SetDstAlphaBlendFactor(alphaDstFactor);
})
#define SET_COLOR_BLEND_ENABLE_CALLBACK(z, index, data) \
ENGINE_ARRAY_CASE(rtBlendEnable, index, { \
context.SetColorBlendEnabled(index, rtBlendEnable); \
})
BOOST_PP_REPEAT(8, SET_COLOR_BLEND_ENABLE_CALLBACK, 0)
static_assert(type::RenderTargetCount == 8 && type::RenderTargetCount < BOOST_PP_LIMIT_REPEAT);
#undef SET_COLOR_BLEND_ENABLE_CALLBACK
ENGINE_CASE(lineWidthSmooth, {
if (*registers.lineSmoothEnable)
context.SetLineWidth(lineWidthSmooth);
})
ENGINE_CASE(lineWidthAliased, {
if (!*registers.lineSmoothEnable)
context.SetLineWidth(lineWidthAliased);
})
ENGINE_CASE(depthBiasFactor, {
context.SetDepthBiasSlopeFactor(depthBiasFactor);
})
ENGINE_CASE(lineSmoothEnable, {
context.SetLineWidth(lineSmoothEnable ? *registers.lineWidthSmooth : *registers.lineWidthAliased);
})
ENGINE_CASE(depthBiasUnits, {
context.SetDepthBiasConstantFactor(depthBiasUnits / 2.0f);
})
ENGINE_STRUCT_CASE(setProgramRegion, high, {
context.SetShaderBaseIovaHigh(high);
})
ENGINE_STRUCT_CASE(setProgramRegion, low, {
context.SetShaderBaseIovaLow(low);
})
ENGINE_CASE(provokingVertexIsLast, {
context.SetProvokingVertex(provokingVertexIsLast);
})
ENGINE_CASE(depthBiasClamp, {
context.SetDepthBiasClamp(depthBiasClamp);
})
ENGINE_CASE(cullFaceEnable, {
context.SetCullFaceEnabled(cullFaceEnable);
})
ENGINE_CASE(frontFace, {
context.SetFrontFace(frontFace);
})
ENGINE_CASE(cullFace, {
context.SetCullFace(cullFace);
})
#define SET_COLOR_WRITE_MASK_CALLBACK(z, index, data) \
ENGINE_ARRAY_CASE(colorWriteMask, index, { \
if (*registers.commonColorWriteMask) \
if (index == 0) \
for (u32 idx{}; idx != type::RenderTargetCount; idx++) \
context.SetColorWriteMask(idx, colorWriteMask); \
else \
context.SetColorWriteMask(index, colorWriteMask); \
})
BOOST_PP_REPEAT(8, SET_COLOR_WRITE_MASK_CALLBACK, 2)
static_assert(type::RenderTargetCount == 8 && type::RenderTargetCount < BOOST_PP_LIMIT_REPEAT);
#undef SET_COLOR_WRITE_MASK_CALLBACK
ENGINE_CASE(viewVolumeClipControl, {
context.SetDepthClampEnabled(!viewVolumeClipControl.depthClampDisable);
})
ENGINE_STRUCT_CASE(colorLogicOp, enable, {
context.SetBlendLogicOpEnable(enable);
})
ENGINE_STRUCT_CASE(colorLogicOp, type, {
context.SetBlendLogicOpType(type);
})
#define VERTEX_BUFFER_CALLBACKS(z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(vertexBuffers, index, config, { \
context.SetVertexBufferStride(index, config.stride); \
}) \
ENGINE_ARRAY_STRUCT_STRUCT_CASE(vertexBuffers, index, iova, high, { \
context.SetVertexBufferStartIovaHigh(index, high); \
}) \
ENGINE_ARRAY_STRUCT_STRUCT_CASE(vertexBuffers, index, iova, low, { \
context.SetVertexBufferStartIovaLow(index, low); \
}) \
ENGINE_ARRAY_STRUCT_CASE(vertexBuffers, index, divisor, { \
context.SetVertexBufferDivisor(index, divisor); \
}) \
ENGINE_ARRAY_CASE(isVertexInputRatePerInstance, index, { \
context.SetVertexBufferInputRate(index, isVertexInputRatePerInstance); \
}) \
ENGINE_ARRAY_STRUCT_CASE(vertexBufferLimits, index, high, { \
context.SetVertexBufferEndIovaHigh(index, high); \
}) \
ENGINE_ARRAY_STRUCT_CASE(vertexBufferLimits, index, low, { \
context.SetVertexBufferEndIovaLow(index, low); \
})
BOOST_PP_REPEAT(16, VERTEX_BUFFER_CALLBACKS, 0)
static_assert(type::VertexBufferCount == 16 && type::VertexBufferCount < BOOST_PP_LIMIT_REPEAT);
#undef VERTEX_BUFFER_CALLBACKS
#define VERTEX_ATTRIBUTES_CALLBACKS(z, index, data) \ #define VERTEX_ATTRIBUTE_CALLBACKS(z, idx, data) \
ENGINE_ARRAY_CASE(vertexAttributeState, index, { \ ENGINE_ARRAY_CASE(vertexAttributes, idx, { \
context.SetVertexAttributeState(index, vertexAttributeState); \ interconnect.directState.vertexInput.SetAttribute(idx, vertexAttributes); \
}) })
BOOST_PP_REPEAT(32, VERTEX_ATTRIBUTES_CALLBACKS, 0) BOOST_PP_REPEAT(16, VERTEX_ATTRIBUTE_CALLBACKS, 0)
static_assert(type::VertexAttributeCount == 32 && type::VertexAttributeCount < BOOST_PP_LIMIT_REPEAT); static_assert(type::VertexAttributeCount == 32 && type::VertexAttributeCount < BOOST_PP_LIMIT_REPEAT);
#undef VERTEX_BUFFER_CALLBACKS #undef VERTEX_ATTRIBUTE_CALLBACKS
#define SET_INDEPENDENT_COLOR_BLEND_CALLBACKS(z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(independentBlend, index, colorOp, { \
context.SetColorBlendOp(index, colorOp); \
}) \
ENGINE_ARRAY_STRUCT_CASE(independentBlend, index, colorSrcFactor, { \
context.SetSrcColorBlendFactor(index, colorSrcFactor); \
}) \
ENGINE_ARRAY_STRUCT_CASE(independentBlend, index, colorDstFactor, { \
context.SetDstColorBlendFactor(index, colorDstFactor); \
}) \
ENGINE_ARRAY_STRUCT_CASE(independentBlend, index, alphaOp, { \
context.SetAlphaBlendOp(index, alphaOp); \
}) \
ENGINE_ARRAY_STRUCT_CASE(independentBlend, index, alphaSrcFactor, { \
context.SetSrcAlphaBlendFactor(index, alphaSrcFactor); \
}) \
ENGINE_ARRAY_STRUCT_CASE(independentBlend, index, alphaDstFactor, { \
context.SetDstAlphaBlendFactor(index, alphaDstFactor); \
})
BOOST_PP_REPEAT(8, SET_INDEPENDENT_COLOR_BLEND_CALLBACKS, 0)
static_assert(type::RenderTargetCount == 8 && type::RenderTargetCount < BOOST_PP_LIMIT_REPEAT);
#undef SET_COLOR_BLEND_ENABLE_CALLBACK
#define SET_SHADER_ENABLE_CALLBACK(z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(setProgram, index, info, { \
context.SetShaderEnabled(info.stage, info.enable); \
})
BOOST_PP_REPEAT(6, SET_SHADER_ENABLE_CALLBACK, 0)
static_assert(type::ShaderStageCount == 6 && type::ShaderStageCount < BOOST_PP_LIMIT_REPEAT);
#undef SET_SHADER_ENABLE_CALLBACK
ENGINE_CASE(primitiveRestartEnable, { ENGINE_CASE(primitiveRestartEnable, {
context.SetPrimitiveRestartEnabled(primitiveRestartEnable); interconnect.directState.inputAssembly.SetPrimitiveRestart(primitiveRestartEnable != 0);
}) })
ENGINE_STRUCT_CASE(constantBufferSelector, size, {
context.SetConstantBufferSelectorSize(size);
})
ENGINE_STRUCT_STRUCT_CASE(constantBufferSelector, address, high, {
context.SetConstantBufferSelectorIovaHigh(high);
})
ENGINE_STRUCT_STRUCT_CASE(constantBufferSelector, address, low, {
context.SetConstantBufferSelectorIovaLow(low);
})
ENGINE_STRUCT_STRUCT_CASE(indexBuffer, start, high, {
context.SetIndexBufferStartIovaHigh(high);
})
ENGINE_STRUCT_STRUCT_CASE(indexBuffer, start, low, {
context.SetIndexBufferStartIovaLow(low);
})
ENGINE_STRUCT_STRUCT_CASE(indexBuffer, limit, high, {
context.SetIndexBufferEndIovaHigh(high);
})
ENGINE_STRUCT_STRUCT_CASE(indexBuffer, limit, low, {
context.SetIndexBufferEndIovaLow(low);
})
ENGINE_STRUCT_CASE(indexBuffer, format, {
context.SetIndexBufferFormat(format);
})
ENGINE_CASE(bindlessTextureConstantBufferIndex, {
context.SetBindlessTextureConstantBufferIndex(bindlessTextureConstantBufferIndex);
})
ENGINE_STRUCT_STRUCT_CASE(samplerPool, address, high, {
context.SetSamplerPoolIovaHigh(high);
})
ENGINE_STRUCT_STRUCT_CASE(samplerPool, address, low, {
context.SetSamplerPoolIovaLow(low);
})
ENGINE_STRUCT_CASE(samplerPool, maximumIndex, {
context.SetSamplerPoolMaximumIndex(maximumIndex);
})
ENGINE_STRUCT_STRUCT_CASE(texturePool, address, high, {
context.SetTexturePoolIovaHigh(high);
})
ENGINE_STRUCT_STRUCT_CASE(texturePool, address, low, {
context.SetTexturePoolIovaLow(low);
})
ENGINE_STRUCT_CASE(texturePool, maximumIndex, {
context.SetTexturePoolMaximumIndex(maximumIndex);
})
ENGINE_CASE(depthMode, {
context.SetDepthMode(depthMode);
})
#define TRANSFORM_FEEDBACK_VARYINGS_CALLBACK(z, index, data) \
ENGINE_ARRAY_CASE(transformFeedbackVaryings, index, { \
context.SetTransformFeedbackBufferVarying(index / (type::TransformFeedbackVaryingCount / sizeof(u32)), \
index % (type::TransformFeedbackVaryingCount / sizeof(u32)), \
transformFeedbackVaryings); \
})
BOOST_PP_REPEAT(128, TRANSFORM_FEEDBACK_VARYINGS_CALLBACK, 0)
static_assert((type::TransformFeedbackVaryingCount / sizeof(u32)) * type::TransformFeedbackBufferCount < BOOST_PP_LIMIT_REPEAT);
#undef TRANSFORM_FEEDBACK_VARYINGS_CALLBACK
default: default:
break; break;
} }
@ -701,28 +215,26 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
syncpoints.at(syncpointAction.id).Increment(); syncpoints.at(syncpointAction.id).Increment();
}) })
ENGINE_CASE(clearBuffers, { ENGINE_CASE(clearSurface, {
context.ClearBuffers(clearBuffers); interconnect.Clear(clearSurface);
}) })
ENGINE_CASE(vertexBeginGl, { ENGINE_CASE(begin, {
context.SetPrimitiveTopology(vertexBeginGl.topology);
// If we reach here then we aren't in a deferred draw so theres no need to flush anything // If we reach here then we aren't in a deferred draw so theres no need to flush anything
if (vertexBeginGl.instanceNext) if (begin.instanceId == Registers::Begin::InstanceId::Subsequent)
deferredDraw.instanceCount++; deferredDraw.instanceCount++;
else if (vertexBeginGl.instanceContinue) else
deferredDraw.instanceCount = 1; deferredDraw.instanceCount = 1;
}) })
ENGINE_CASE(drawVertexCount, { ENGINE_STRUCT_CASE(drawVertexArray, count, {
// Defer the draw until the first non-draw operation to allow for detecting instanced draws (see DeferredDrawState comment) // Defer the draw until the first non-draw operation to allow for detecting instanced draws (see DeferredDrawState comment)
deferredDraw.Set(drawVertexCount, *registers.drawVertexFirst, 0, registers.vertexBeginGl->topology, false); deferredDraw.Set(count, *registers.vertexArrayStart, 0, *registers.globalBaseInstanceIndex, GetCurrentTopology(), false);
}) })
ENGINE_CASE(drawIndexCount, { ENGINE_STRUCT_CASE(drawIndexBuffer, count, {
// Defer the draw until the first non-draw operation to allow for detecting instanced draws (see DeferredDrawState comment) // Defer the draw until the first non-draw operation to allow for detecting instanced draws (see DeferredDrawState comment)
deferredDraw.Set(drawIndexCount, *registers.drawIndexFirst, *registers.drawBaseVertex, registers.vertexBeginGl->topology, true); deferredDraw.Set(count, registers.indexBuffer->first, *registers.globalBaseVertexIndex, *registers.globalBaseInstanceIndex, GetCurrentTopology(), true);
}) })
ENGINE_STRUCT_CASE(semaphore, info, { ENGINE_STRUCT_CASE(semaphore, info, {
@ -742,7 +254,7 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
break; break;
default: default:
Logger::Warn("Unsupported semaphore counter type: 0x{:X}", static_cast<u8>(info.counterType)); //Logger::Warn("Unsupported semaphore counter type: 0x{:X}", static_cast<u8>(info.counterType));
break; break;
} }
break; break;
@ -754,39 +266,29 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
} }
}) })
#define SHADER_CALLBACKS(z, index, data) \
ENGINE_ARRAY_STRUCT_CASE(setProgram, index, offset, { \
context.SetShaderOffset(static_cast<type::ShaderStage>(index), offset); \
})
BOOST_PP_REPEAT(6, SHADER_CALLBACKS, 0)
static_assert(type::ShaderStageCount == 6 && type::ShaderStageCount < BOOST_PP_LIMIT_REPEAT);
#undef SHADER_CALLBACKS
#define PIPELINE_CALLBACKS(z, idx, data) \
ENGINE_ARRAY_STRUCT_CASE(bind, idx, constantBuffer, { \
context.BindPipelineConstantBuffer(static_cast<type::PipelineStage>(idx), constantBuffer.valid, constantBuffer.index); \
})
BOOST_PP_REPEAT(5, PIPELINE_CALLBACKS, 0)
static_assert(type::PipelineStageCount == 5 && type::PipelineStageCount < BOOST_PP_LIMIT_REPEAT);
#undef PIPELINE_CALLBACKS
ENGINE_ARRAY_CASE(firmwareCall, 4, { ENGINE_ARRAY_CASE(firmwareCall, 4, {
registers.raw[0xD00] = 1; registers.raw[0xD00] = 1;
}) })
// Begin a batch constant buffer update, this case will never be reached if a batch update is currently active // Begin a batch constant buffer update, this case will never be reached if a batch update is currently active
#define CBUF_UPDATE_CALLBACKS(z, index, data_) \ #define LOAD_CONSTANT_BUFFER_CALLBACKS(z, index, data_) \
ENGINE_STRUCT_ARRAY_CASE(constantBufferUpdate, data, index, { \ ENGINE_STRUCT_ARRAY_CASE(loadConstantBuffer, data, index, { \
batchConstantBufferUpdate.startOffset = registers.constantBufferUpdate->offset; \ batchLoadConstantBuffer.startOffset = registers.loadConstantBuffer->offset; \
batchConstantBufferUpdate.buffer.push_back(data); \ batchLoadConstantBuffer.buffer.push_back(data); \
registers.constantBufferUpdate->offset += 4; \ registers.loadConstantBuffer->offset += 4; \
}) })
BOOST_PP_REPEAT(16, CBUF_UPDATE_CALLBACKS, 0) BOOST_PP_REPEAT(16, LOAD_CONSTANT_BUFFER_CALLBACKS, 0)
#undef CBUF_UPDATE_CALLBACKS #undef LOAD_CONSTANT_BUFFER_CALLBACKS
#define PIPELINE_CALLBACKS(z, idx, data) \
ENGINE_ARRAY_STRUCT_CASE(bindGroups, idx, constantBuffer, { \
interconnect.BindConstantBuffer(static_cast<type::PipelineStage>(idx), constantBuffer.shaderSlot, constantBuffer.valid); \
})
BOOST_PP_REPEAT(5, PIPELINE_CALLBACKS, 0)
static_assert(type::PipelineStageCount == 5 && type::PipelineStageCount < BOOST_PP_LIMIT_REPEAT);
#undef PIPELINE_CALLBACKS
default: default:
break; break;
} }
@ -816,9 +318,9 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
void Maxwell3D::FlushEngineState() { void Maxwell3D::FlushEngineState() {
FlushDeferredDraw(); FlushDeferredDraw();
if (batchConstantBufferUpdate.Active()) { if (batchLoadConstantBuffer.Active()) {
context.ConstantBufferUpdate(batchConstantBufferUpdate.buffer, batchConstantBufferUpdate.Invalidate()); interconnect.LoadConstantBuffer(batchLoadConstantBuffer.buffer, batchLoadConstantBuffer.Invalidate());
batchConstantBufferUpdate.Reset(); batchLoadConstantBuffer.Reset();
} }
} }

22
app/src/main/cpp/skyline/soc/gm20b/engines/maxwell_3d.h
View File

@ -5,9 +5,12 @@
#pragma once #pragma once
#include "maxwell/types.h" #include <gpu/interconnect/maxwell_3d/maxwell_3d.h>
#include "engine.h" #include "engine.h"
#include <soc/host1x/syncpoint.h>
#include "gpu/interconnect/maxwell_3d/common.h"
#include "inline2memory.h" #include "inline2memory.h"
#include "maxwell/types.h"
namespace skyline::soc::gm20b { namespace skyline::soc::gm20b {
struct ChannelContext; struct ChannelContext;
@ -20,10 +23,11 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
class Maxwell3D : public MacroEngineBase { class Maxwell3D : public MacroEngineBase {
private: private:
host1x::SyncpointSet &syncpoints; host1x::SyncpointSet &syncpoints;
gpu::interconnect::GraphicsContext context;
Inline2MemoryBackend i2m; Inline2MemoryBackend i2m;
gpu::interconnect::maxwell3d::DirtyManager dirtyManager;
gpu::interconnect::maxwell3d::Maxwell3D interconnect;
struct BatchConstantBufferUpdateState { struct BatchLoadConstantBufferState {
std::vector<u32> buffer; std::vector<u32> buffer;
u32 startOffset{std::numeric_limits<u32>::max()}; u32 startOffset{std::numeric_limits<u32>::max()};
@ -38,7 +42,7 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
void Reset() { void Reset() {
buffer.clear(); buffer.clear();
} }
} batchConstantBufferUpdate; //!< Holds state for updating constant buffer data in a batch rather than word by word } batchLoadConstantBuffer; //!< Holds state for updating constant buffer data in a batch rather than word by word
/** /**
* @brief In the Maxwell 3D engine, instanced draws are implemented by repeating the exact same draw in sequence with special flag set in vertexBeginGl. This flag allows either incrementing the instance counter or resetting it, since we need to supply an instance count to the host API we defer all draws until state changes occur. If there are no state changes between draws we can skip them and count the occurences to get the number of instances to draw. * @brief In the Maxwell 3D engine, instanced draws are implemented by repeating the exact same draw in sequence with special flag set in vertexBeginGl. This flag allows either incrementing the instance counter or resetting it, since we need to supply an instance count to the host API we defer all draws until state changes occur. If there are no state changes between draws we can skip them and count the occurences to get the number of instances to draw.
@ -46,25 +50,29 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
struct DeferredDrawState { struct DeferredDrawState {
bool pending; bool pending;
bool indexed; //!< If the deferred draw is indexed bool indexed; //!< If the deferred draw is indexed
type::PrimitiveTopology drawTopology; //!< Topology of draw at draw time type::DrawTopology drawTopology; //!< Topology of draw at draw time
u32 instanceCount{1}; //!< Number of instances in the final draw u32 instanceCount{1}; //!< Number of instances in the final draw
u32 drawCount; //!< indexed ? drawIndexCount : drawVertexCount u32 drawCount; //!< indexed ? drawIndexCount : drawVertexCount
u32 drawFirst; //!< indexed ? drawIndexFirst : drawVertexFirst u32 drawFirst; //!< indexed ? drawIndexFirst : drawVertexFirst
i32 drawBaseVertex; //!< Only applicable to indexed draws u32 drawBaseVertex; //!< Only applicable to indexed draws
u32 drawBaseInstance;
/** /**
* @brief Sets up the state necessary to defer a new draw * @brief Sets up the state necessary to defer a new draw
*/ */
void Set(u32 pDrawCount, u32 pDrawFirst, i32 pDrawBaseVertex, type::PrimitiveTopology pDrawTopology, bool pIndexed) { void Set(u32 pDrawCount, u32 pDrawFirst, u32 pDrawBaseVertex, u32 pDrawBaseInstance, type::DrawTopology pDrawTopology, bool pIndexed) {
pending = true; pending = true;
indexed = pIndexed; indexed = pIndexed;
drawTopology = pDrawTopology; drawTopology = pDrawTopology;
drawCount = pDrawCount; drawCount = pDrawCount;
drawFirst = pDrawFirst; drawFirst = pDrawFirst;
drawBaseVertex = pDrawBaseVertex; drawBaseVertex = pDrawBaseVertex;
drawBaseInstance = pDrawBaseInstance;
} }
} deferredDraw{}; } deferredDraw{};
type::DrawTopology GetCurrentTopology();
void FlushDeferredDraw(); void FlushDeferredDraw();
/** /**