[PICA] Shader uploads

CMakeLists.txt

@@ -54,7 +54,7 @@ set(SERVICE_SOURCE_FILES src/core/services/service_manager.cpp src/core/services
                          src/core/services/fs.cpp src/core/services/gsp_gpu.cpp src/core/services/gsp_lcd.cpp
                          src/core/services/ndm.cpp
 )
-set(PICA_SOURCE_FILES src/core/PICA/gpu.cpp src/core/PICA/regs.cpp)
+set(PICA_SOURCE_FILES src/core/PICA/gpu.cpp src/core/PICA/regs.cpp src/core/PICA/shader_unit.cpp)
 
 set(HEADER_FILES include/emulator.hpp include/helpers.hpp include/opengl.hpp include/termcolor.hpp
                  include/cpu.hpp include/cpu_dynarmic.hpp include/memory.hpp include/kernel/kernel.hpp
@@ -62,7 +62,8 @@ set(HEADER_FILES include/emulator.hpp include/helpers.hpp include/opengl.hpp inc
                  include/kernel/config_mem.hpp include/services/service_manager.hpp include/services/apt.hpp
                  include/kernel/handles.hpp include/services/hid.hpp include/services/fs.hpp
                  include/services/gsp_gpu.hpp include/services/gsp_lcd.hpp include/arm_defs.hpp
-                 include/gpu.hpp include/services/ndm.hpp
+                 include/PICA/gpu.hpp include/PICA/regs.hpp include/services/ndm.hpp
+                 include/PICA/shader.hpp include/PICA/shader_unit.hpp
 )
 
 set(THIRD_PARTY_SOURCE_FILES third_party/imgui/imgui.cpp

include/PICA/gpu.hpp

@@ -2,10 +2,12 @@
 #include <array>
 #include "helpers.hpp"
 #include "memory.hpp"
+#include "PICA/shader_unit.hpp"
 
 class GPU {
 	Memory& mem;
 	static constexpr u32 regNum = 0x300;
+	ShaderUnit shaderUnit;
 	std::array<u32, regNum> regs; // GPU internal registers
 
 public:

include/PICA/regs.hpp

@@ -0,0 +1,16 @@
+#pragma once
+
+namespace PICAInternalRegs {
+	enum : u32 {
+		VertexShaderTransferEnd = 0x2BF,
+		VertexShaderTransferIndex = 0x2CB,
+		VertexShaderData0 = 0x2CC,
+		VertexShaderData1 = 0x2CD,
+		VertexShaderData2 = 0x2CE,
+		VertexShaderData3 = 0x2CF,
+		VertexShaderData4 = 0x2D0,
+		VertexShaderData5 = 0x2D1,
+		VertexShaderData6 = 0x2D2,
+		VertexShaderData7 = 0x2D3,
+	};
+}

include/PICA/shader.hpp

@@ -0,0 +1,41 @@
+#pragma once
+#include <algorithm>
+#include <array>
+#include <cstring>
+#include "helpers.hpp"
+
+class PICAShader {
+	int bufferIndex; // Index of the next instruction to overwrite
+
+public:
+	std::array<u32, 512> loadedShader; // Currently loaded & active shader
+	std::array<u32, 512> bufferedShader; // Shader to be transferred when the SH_CODETRANSFER_END reg gets written to
+	
+	u32 boolUniform;
+	std::array<u32, 4> intUniforms;
+	std::array<u32, 8> floatUniforms;
+
+	void reset() {
+		loadedShader.fill(0);
+		bufferedShader.fill(0);
+
+		intUniforms.fill(0);
+		floatUniforms.fill(0);
+		boolUniform = 0;
+		bufferIndex = 0;
+	}
+
+	void finalize() {
+		std::memcpy(&loadedShader[0], &bufferedShader[0], 512 * sizeof(u32));
+	}
+
+	void setBufferIndex(u32 offset) {
+		if (offset != 0) Helpers::panic("Is this register 9 or 11 bit?");
+		bufferIndex = (offset >> 2) & 511;
+	}
+
+	void uploadWord(u32 word) {
+		bufferedShader[bufferIndex++] = word;
+		bufferIndex &= 511;
+	}
+};

include/PICA/shader_unit.hpp

@@ -0,0 +1,11 @@
+#pragma once
+#include "PICA/shader.hpp"
+
+class ShaderUnit {
+
+public:
+	PICAShader vs; // Vertex shader
+	PICAShader gs; // Geometry shader
+
+	void reset();
+};

include/emulator.hpp

@@ -4,9 +4,9 @@
 #include <fstream>
 
 #include "cpu.hpp"
-#include "gpu.hpp"
 #include "memory.hpp"
 #include "opengl.hpp"
+#include "PICA/gpu.hpp"
 #include "SFML/Window.hpp"
 #include "SFML/Graphics.hpp"
 

include/services/gsp_gpu.hpp

@@ -1,6 +1,6 @@
 #pragma once
 #include <cstring>
-#include "gpu.hpp"
+#include "PICA/gpu.hpp"
 #include "helpers.hpp"
 #include "kernel_types.hpp"
 #include "memory.hpp"

src/core/PICA/gpu.cpp

@@ -1,8 +1,9 @@
-#include "gpu.hpp"
+#include "PICA/gpu.hpp"
 #include <cstdio>
 
 void GPU::reset() {
 	regs.fill(0);
+	shaderUnit.reset();
 	// TODO: Reset blending, texturing, etc here
 }
 

src/core/PICA/regs.cpp

@@ -1,4 +1,5 @@
-#include "gpu.hpp"
+#include "PICA/gpu.hpp"
+#include "PICA/regs.hpp"
 
 u32 GPU::readReg(u32 address) {
 	printf("Ignoring read from GPU register %08X\n", address);
@@ -10,7 +11,7 @@ void GPU::writeReg(u32 address, u32 value) {
 		const u32 index = (address - 0x1EF01000) / sizeof(u32);
 		writeInternalReg(index, value, 0xffffffff);
 	} else {
-		Helpers::panic("Ignoring write to GPU register %08X. Value: %08X\n", address, value);
+		printf("Ignoring write to external GPU register %08X. Value: %08X\n", address, value);
 	}
 }
 
@@ -24,6 +25,8 @@ u32 GPU::readInternalReg(u32 index) {
 }
 
 void GPU::writeInternalReg(u32 index, u32 value, u32 mask) {
+	using namespace PICAInternalRegs;
+
 	if (index > regNum) {
 		Helpers::panic("Tried to write to invalid GPU register. Index: %X, value: %08X\n", index, value);
 		return;
@@ -32,7 +35,22 @@ void GPU::writeInternalReg(u32 index, u32 value, u32 mask) {
 	u32 currentValue = regs[index];
 	u32 newValue = (currentValue & ~mask) | (value & mask); // Only overwrite the bits specified by "mask"
 
+	// TODO: Figure out if things like the shader index use the unmasked value or the masked one
+	// We currently use the unmasked value like Citra does
 	switch (index) {
+		case VertexShaderTransferEnd:
+			if (value != 0) shaderUnit.vs.finalize();
+			break;
+
+		case VertexShaderTransferIndex:
+			shaderUnit.vs.setBufferIndex(value);
+			break;
+
+		case VertexShaderData0: case VertexShaderData1: case VertexShaderData2: case VertexShaderData3:
+		case VertexShaderData4: case VertexShaderData5: case VertexShaderData6: case VertexShaderData7:
+			shaderUnit.vs.uploadWord(value);
+			break;
+
 		default:
 			printf("GPU: Wrote to unimplemented internal reg: %X, value: %08X\n", index, newValue);
 			break;

src/core/PICA/shader_unit.cpp

@@ -0,0 +1,6 @@
+#include "PICA/shader_unit.hpp"
+
+void ShaderUnit::reset() {
+	vs.reset();
+	gs.reset();
+}

src/core/services/gsp_gpu.cpp

@@ -252,7 +252,7 @@ void GPUService::memoryFill(u32* cmd) {
 // Actually send command list (aka display list) to GPU
 void GPUService::processCommandList(u32* cmd) {
 	u32 address = cmd[1] & ~7; // Buffer address
-	u32 size = cmd[2] * 8; // Buffer size in bytes
+	u32 size = cmd[2] & ~3; // Buffer size in bytes
 	bool updateGas = cmd[3] == 1; // Update gas additive blend results (0 = don't update, 1 = update)
 	bool flushBuffer = cmd[7] == 1; // Flush buffer (0 = don't flush, 1 = flush)
 
@@ -305,5 +305,5 @@ void GPUService::processCommandList(u32* cmd) {
 		}
 	}
 
-	Helpers::panic("GPU::GSP::processCommandList. Address: %08X, size in bytes: %08X", address, size);
+	printf("GPU::GSP::processCommandList. Address: %08X, size in bytes: %08X\n", address, size);
 }