Diamond75/Panda3DS
1
0
Fork 0
mirror of https://github.com/wheremyfoodat/Panda3DS.git synced 2025-06-07 19:41:38 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions

specialize shader

Browse source
This commit is contained in:
Samuliak 2024-07-05 20:29:05 +02:00
parent 7e8582d985
commit 0c19f5a3ea
6 changed files with 125 additions and 27 deletions
Download patch file Download diff file Expand all files Collapse all files

10
src/core/renderer_mtl/renderer_mtl.cpp
View file

@ -154,11 +154,10 @@ void RendererMTL::initGraphicsContext(SDL_Window* window) {
MTL::Function* vertexBlitFunction = library->newFunction(NS::String::string("vertexBlit", NS::ASCIIStringEncoding));
MTL::Function* fragmentBlitFunction = library->newFunction(NS::String::string("fragmentBlit", NS::ASCIIStringEncoding));
blitPipelineCache.set(device, vertexBlitFunction, fragmentBlitFunction, nullptr);
blitPipelineCache.set(device, vertexBlitFunction, fragmentBlitFunction);
// Draw
MTL::Function* vertexDrawFunction = library->newFunction(NS::String::string("vertexDraw", NS::ASCIIStringEncoding));
MTL::Function* fragmentDrawFunction = library->newFunction(NS::String::string("fragmentDraw", NS::ASCIIStringEncoding));
// -------- Vertex descriptor --------
MTL::VertexDescriptor* vertexDescriptor = MTL::VertexDescriptor::alloc()->init();
@ -216,7 +215,7 @@ void RendererMTL::initGraphicsContext(SDL_Window* window) {
vertexBufferLayout->setStepFunction(MTL::VertexStepFunctionPerVertex);
vertexBufferLayout->setStepRate(1);
drawPipelineCache.set(device, vertexDrawFunction, fragmentDrawFunction, vertexDescriptor);
drawPipelineCache.set(device, library, vertexDrawFunction, vertexDescriptor);
// Depth stencil cache
depthStencilCache.set(device);
@ -321,7 +320,7 @@ void RendererMTL::displayTransfer(u32 inputAddr, u32 outputAddr, u32 inputSize,
colorAttachment->setStoreAction(MTL::StoreActionStore);
// Pipeline
Metal::PipelineHash hash{destFramebuffer->format, DepthFmt::Unknown1};
Metal::BlitPipelineHash hash{destFramebuffer->format, DepthFmt::Unknown1};
auto blitPipeline = blitPipelineCache.get(hash);
beginRenderPassIfNeeded(renderPassDescriptor, destFramebuffer->texture);
@ -381,10 +380,11 @@ void RendererMTL::drawVertices(PICA::PrimType primType, std::span<const PICA::Ve
depthUniforms.depthMapEnable = regs[PICA::InternalRegs::DepthmapEnable] & 1;
// -------- Pipeline --------
Metal::PipelineHash pipelineHash{colorRenderTarget->format, DepthFmt::Unknown1};
Metal::DrawPipelineHash pipelineHash{colorRenderTarget->format, DepthFmt::Unknown1};
if (depthStencilRenderTarget) {
pipelineHash.depthFmt = depthStencilRenderTarget->format;
}
pipelineHash.lightingEnabled = regs[0x008F] & 1;
// Blending and logic op
pipelineHash.blendEnabled = (regs[PICA::InternalRegs::ColourOperation] & (1 << 8)) != 0;

16
src/host_shaders/metal_shaders.metal
View file

@ -393,12 +393,6 @@ void calcLighting(thread DrawVertexOut& in, constant PicaRegs& picaRegs, texture
float3 bitangent = normalize(in.bitangent);
float3 view = normalize(in.view);
uint GPUREG_LIGHTING_ENABLE = picaRegs.read(0x008Fu);
if (extract_bits(GPUREG_LIGHTING_ENABLE, 0, 1) == 0u) {
primaryColor = secondaryColor = float4(1.0);
return;
}
uint GPUREG_LIGHTING_AMBIENT = picaRegs.read(0x01C0u);
uint GPUREG_LIGHTING_NUM_LIGHTS = (picaRegs.read(0x01C2u) & 0x7u) + 1u;
uint GPUREG_LIGHTING_LIGHT_PERMUTATION = picaRegs.read(0x01D9u);
@ -541,12 +535,20 @@ float4 performLogicOp(LogicOp logicOp, float4 s, float4 d) {
return as_type<float4>(performLogicOpU(logicOp, as_type<uint4>(s), as_type<uint4>(d)));
}
constant int LIGHTING_ENABLED_INDEX = 0;
constant bool lightingEnabled [[function_constant(LIGHTING_ENABLED_INDEX)]];
fragment float4 fragmentDraw(DrawVertexOut in [[stage_in]], float4 prevColor [[color(0)]], constant PicaRegs& picaRegs [[buffer(0)]], constant FragTEV& tev [[buffer(1)]], constant LogicOp& logicOp [[buffer(2)]],
texture2d<float> tex0 [[texture(0)]], texture2d<float> tex1 [[texture(1)]], texture2d<float> tex2 [[texture(2)]], texture1d_array<float> texLightingLut [[texture(3)]],
sampler samplr0 [[sampler(0)]], sampler samplr1 [[sampler(1)]], sampler samplr2 [[sampler(2)]], sampler linearSampler [[sampler(3)]]) {
Globals globals;
globals.tevSources[0] = in.color;
calcLighting(in, picaRegs, texLightingLut, linearSampler, globals.tevSources[1], globals.tevSources[2]);
if (lightingEnabled) {
calcLighting(in, picaRegs, texLightingLut, linearSampler, globals.tevSources[1], globals.tevSources[2]);
} else {
globals.tevSources[1] = float4(0.0);
globals.tevSources[2] = float4(0.0);
}
uint textureConfig = picaRegs.read(0x80u);
float2 texCoord2 = (textureConfig & (1u << 13)) != 0u ? in.texCoord1 : in.texCoord2;