diff --git a/rpcs3/Emu/CMakeLists.txt b/rpcs3/Emu/CMakeLists.txt
index 62a91d1d34..0164376dff 100644
--- a/rpcs3/Emu/CMakeLists.txt
+++ b/rpcs3/Emu/CMakeLists.txt
@@ -511,6 +511,7 @@ target_sources(rpcs3_emu PRIVATE
     RSX/Overlays/overlay_video.cpp
     RSX/Overlays/Shaders/shader_loading_dialog.cpp
     RSX/Overlays/Shaders/shader_loading_dialog_native.cpp
+    RSX/Program/CgBinaryProgram.cpp
     RSX/Program/CgBinaryFragmentProgram.cpp
     RSX/Program/CgBinaryVertexProgram.cpp
     RSX/Program/FragmentProgramDecompiler.cpp
diff --git a/rpcs3/Emu/RSX/Program/CgBinaryFragmentProgram.cpp b/rpcs3/Emu/RSX/Program/CgBinaryFragmentProgram.cpp
index 877e55c6af..a06818de10 100644
--- a/rpcs3/Emu/RSX/Program/CgBinaryFragmentProgram.cpp
+++ b/rpcs3/Emu/RSX/Program/CgBinaryFragmentProgram.cpp
@@ -230,7 +230,7 @@ void CgBinaryDisasm::TaskFP()
 {
 	m_size = 0;
 	u32* data = reinterpret_cast<u32*>(&m_buffer[m_offset]);
-	ensure((m_buffer_size - m_offset) % sizeof(u32) == 0);
+	ensure((m_buffer.size() - m_offset) % sizeof(u32) == 0);
 
 	enum
 	{
diff --git a/rpcs3/Emu/RSX/Program/CgBinaryProgram.cpp b/rpcs3/Emu/RSX/Program/CgBinaryProgram.cpp
new file mode 100644
index 0000000000..d9c090addb
--- /dev/null
+++ b/rpcs3/Emu/RSX/Program/CgBinaryProgram.cpp
@@ -0,0 +1,246 @@
+#include "stdafx.h"
+#include "CgBinaryProgram.h"
+
+#ifndef WITHOUT_OPENGL
+#include "Emu/RSX/GL/GLVertexProgram.h"
+#include "Emu/RSX/GL/GLFragmentProgram.h"
+#endif
+
+CgBinaryDisasm::CgBinaryDisasm(const std::string& path)
+	: m_path(path)
+{
+	fs::file f(path);
+	if (!f)
+	{
+		return;
+	}
+
+	usz buffer_size = f.size();
+	m_buffer.resize(buffer_size);
+	f.read(m_buffer, buffer_size);
+	fmt::append(m_arb_shader, "Loading... [%s]\n", path.c_str());
+}
+
+std::string CgBinaryDisasm::GetCgParamType(u32 type)
+{
+	switch (type)
+	{
+	case 1045: return "float";
+	case 1046:
+	case 1047:
+	case 1048: return fmt::format("float%d", type - 1044);
+	case 1064: return "float4x4";
+	case 1066: return "sampler2D";
+	case 1069: return "samplerCUBE";
+	case 1091: return "float1";
+
+	default: return fmt::format("!UnkCgType(%d)", type);
+	}
+}
+
+std::string CgBinaryDisasm::GetCgParamName(u32 offset) const
+{
+	return std::string(&m_buffer[offset]);
+}
+
+std::string CgBinaryDisasm::GetCgParamRes(u32 /*offset*/) const
+{
+	// rsx_log.warning("GetCgParamRes offset 0x%x", offset);
+	// TODO
+	return "";
+}
+
+std::string CgBinaryDisasm::GetCgParamSemantic(u32 offset) const
+{
+	return std::string(&m_buffer[offset]);
+}
+
+std::string CgBinaryDisasm::GetCgParamValue(u32 offset, u32 end_offset) const
+{
+	std::string offsets = "offsets:";
+
+	u32 num = 0;
+	offset += 6;
+	while (offset < end_offset)
+	{
+		fmt::append(offsets, " %d,", m_buffer[offset] << 8 | m_buffer[offset + 1]);
+		offset += 4;
+		num++;
+	}
+
+	if (num > 4)
+	{
+		return "";
+	}
+
+	offsets.pop_back();
+	return fmt::format("num %d ", num) + offsets;
+}
+
+void CgBinaryDisasm::ConvertToLE(CgBinaryProgram& prog)
+{
+	// BE payload, requires that data be swapped
+	const auto be_profile = prog.profile;
+
+	auto swap_be32 = [&](u32 start_offset, size_t size_bytes)
+	{
+		auto start = reinterpret_cast<u32*>(m_buffer.data() + start_offset);
+		auto end = reinterpret_cast<u32*>(m_buffer.data() + start_offset + size_bytes);
+
+		for (auto data = start; data < end; ++data)
+		{
+			*data = std::bit_cast<be_t<u32>>(*data);
+		}
+	};
+
+	// 1. Swap the header
+	swap_be32(0, sizeof(CgBinaryProgram));
+
+	// 2. Swap parameters
+	swap_be32(prog.parameterArray, sizeof(CgBinaryParameter) * prog.parameterCount);
+
+	// 3. Swap the ucode
+	swap_be32(prog.ucode, m_buffer.size() - prog.ucode);
+
+	// 4. Swap the domain header
+	if (be_profile == 7004u)
+	{
+		// Need to swap each field individually
+		auto& fprog = GetCgRef<CgBinaryFragmentProgram>(prog.program);
+		fprog.instructionCount = std::bit_cast<be_t<u32>>(fprog.instructionCount);
+		fprog.attributeInputMask = std::bit_cast<be_t<u32>>(fprog.attributeInputMask);
+		fprog.partialTexType = std::bit_cast<be_t<u32>>(fprog.partialTexType);
+		fprog.texCoordsInputMask = std::bit_cast<be_t<u16>>(fprog.texCoordsInputMask);
+		fprog.texCoords2D = std::bit_cast<be_t<u16>>(fprog.texCoords2D);
+		fprog.texCoordsCentroid = std::bit_cast<be_t<u16>>(fprog.texCoordsCentroid);
+	}
+	else
+	{
+		// Swap entire header block as all fields are u32
+		swap_be32(prog.program, sizeof(CgBinaryVertexProgram));
+	}
+}
+
+void CgBinaryDisasm::BuildShaderBody(bool include_glsl)
+{
+	ParamArray param_array;
+
+	auto& prog = GetCgRef<CgBinaryProgram>(0);
+
+	if (const u32 be_profile = std::bit_cast<be_t<u32>>(prog.profile);
+		be_profile == 7003u || be_profile == 7004u)
+	{
+		ConvertToLE(prog);
+		ensure(be_profile == prog.profile);
+	}
+
+	if (prog.profile == 7004u)
+	{
+		auto& fprog = GetCgRef<CgBinaryFragmentProgram>(prog.program);
+		m_arb_shader += "\n";
+		fmt::append(m_arb_shader, "# binaryFormatRevision 0x%x\n", prog.binaryFormatRevision);
+		fmt::append(m_arb_shader, "# profile sce_fp_rsx\n");
+		fmt::append(m_arb_shader, "# parameterCount %d\n", prog.parameterCount);
+		fmt::append(m_arb_shader, "# instructionCount %d\n", fprog.instructionCount);
+		fmt::append(m_arb_shader, "# attributeInputMask 0x%x\n", fprog.attributeInputMask);
+		fmt::append(m_arb_shader, "# registerCount %d\n\n", fprog.registerCount);
+
+		CgBinaryParameterOffset offset = prog.parameterArray;
+		for (u32 i = 0; i < prog.parameterCount; i++)
+		{
+			auto& fparam = GetCgRef<CgBinaryParameter>(offset);
+
+			std::string param_type = GetCgParamType(fparam.type) + " ";
+			std::string param_name = GetCgParamName(fparam.name) + " ";
+			std::string param_res = GetCgParamRes(fparam.res) + " ";
+			std::string param_semantic = GetCgParamSemantic(fparam.semantic) + " ";
+			std::string param_const = GetCgParamValue(fparam.embeddedConst, fparam.name);
+
+			fmt::append(m_arb_shader, "#%d%s%s%s%s\n", i, param_type, param_name, param_semantic, param_const);
+
+			offset += u32{sizeof(CgBinaryParameter)};
+		}
+
+		m_arb_shader += "\n";
+		m_offset = prog.ucode;
+		TaskFP();
+
+		if (!include_glsl)
+		{
+			return;
+		}
+
+		u32 unused;
+		std::vector<u32> be_data;
+
+		// Swap bytes. FP decompiler expects input in BE
+		for (u32* ptr = reinterpret_cast<u32*>(m_buffer.data() + m_offset),
+			*end = reinterpret_cast<u32*>(m_buffer.data() + m_buffer.size());
+			ptr < end; ++ptr)
+		{
+			be_data.push_back(std::bit_cast<be_t<u32>>(*ptr));
+		}
+
+		RSXFragmentProgram rsx_prog;
+		auto metadata = program_hash_util::fragment_program_utils::analyse_fragment_program(be_data.data());
+		rsx_prog.ctrl = (fprog.outputFromH0 ? 0 : 0x40) | (fprog.depthReplace ? 0xe : 0);
+		rsx_prog.offset = metadata.program_start_offset;
+		rsx_prog.ucode_length = metadata.program_ucode_length;
+		rsx_prog.total_length = metadata.program_ucode_length + metadata.program_start_offset;
+		rsx_prog.data = reinterpret_cast<u8*>(be_data.data()) + metadata.program_start_offset;
+		for (u32 i = 0; i < 16; ++i) rsx_prog.texture_state.set_dimension(rsx::texture_dimension_extended::texture_dimension_2d, i);
+#ifndef WITHOUT_OPENGL
+		GLFragmentDecompilerThread(m_glsl_shader, param_array, rsx_prog, unused).Task();
+#endif
+	}
+
+	else
+	{
+		const auto& vprog = GetCgRef<CgBinaryVertexProgram>(prog.program);
+		m_arb_shader += "\n";
+		fmt::append(m_arb_shader, "# binaryFormatRevision 0x%x\n", prog.binaryFormatRevision);
+		fmt::append(m_arb_shader, "# profile sce_vp_rsx\n");
+		fmt::append(m_arb_shader, "# parameterCount %d\n", prog.parameterCount);
+		fmt::append(m_arb_shader, "# instructionCount %d\n", vprog.instructionCount);
+		fmt::append(m_arb_shader, "# registerCount %d\n", vprog.registerCount);
+		fmt::append(m_arb_shader, "# attributeInputMask 0x%x\n", vprog.attributeInputMask);
+		fmt::append(m_arb_shader, "# attributeOutputMask 0x%x\n\n", vprog.attributeOutputMask);
+
+		CgBinaryParameterOffset offset = prog.parameterArray;
+		for (u32 i = 0; i < prog.parameterCount; i++)
+		{
+			auto& vparam = GetCgRef<CgBinaryParameter>(offset);
+
+			std::string param_type = GetCgParamType(vparam.type) + " ";
+			std::string param_name = GetCgParamName(vparam.name) + " ";
+			std::string param_res = GetCgParamRes(vparam.res) + " ";
+			std::string param_semantic = GetCgParamSemantic(vparam.semantic) + " ";
+			std::string param_const = GetCgParamValue(vparam.embeddedConst, vparam.name);
+
+			fmt::append(m_arb_shader, "#%d%s%s%s%s\n", i, param_type, param_name, param_semantic, param_const);
+
+			offset += u32{sizeof(CgBinaryParameter)};
+		}
+
+		m_arb_shader += "\n";
+		m_offset = prog.ucode;
+		ensure((m_buffer.size() - m_offset) % sizeof(u32) == 0);
+
+		u32* vdata = reinterpret_cast<u32*>(&m_buffer[m_offset]);
+		m_data.resize(prog.ucodeSize / sizeof(u32));
+		std::memcpy(m_data.data(), vdata, prog.ucodeSize);
+		TaskVP();
+
+		if (!include_glsl)
+		{
+			return;
+		}
+
+		RSXVertexProgram rsx_prog;
+		program_hash_util::vertex_program_utils::analyse_vertex_program(vdata, 0, rsx_prog);
+		for (u32 i = 0; i < 4; ++i) rsx_prog.texture_state.set_dimension(rsx::texture_dimension_extended::texture_dimension_2d, i);
+#ifndef WITHOUT_OPENGL
+		GLVertexDecompilerThread(rsx_prog, m_glsl_shader, param_array).Task();
+#endif
+	}
+}
diff --git a/rpcs3/Emu/RSX/Program/CgBinaryProgram.h b/rpcs3/Emu/RSX/Program/CgBinaryProgram.h
index 310be35baa..56d8f4d70b 100644
--- a/rpcs3/Emu/RSX/Program/CgBinaryProgram.h
+++ b/rpcs3/Emu/RSX/Program/CgBinaryProgram.h
@@ -7,11 +7,6 @@
 #include "Emu/RSX/Program/ShaderParam.h"
 #include "Utilities/File.h"
 
-#ifndef WITHOUT_OPENGL
-#include "Emu/RSX/GL/GLVertexProgram.h"
-#include "Emu/RSX/GL/GLFragmentProgram.h"
-#endif
-
 using CGprofile = u32;
 using CGbool = s32;
 using CGresource = u32;
@@ -133,8 +128,7 @@ class CgBinaryDisasm
 
 	std::string m_path; // used for FP decompiler thread, delete this later
 
-	u8* m_buffer = nullptr;
-	usz m_buffer_size = 0;
+	std::vector<char> m_buffer;
 	std::string m_arb_shader;
 	std::string m_glsl_shader;
 	std::string m_dst_reg_name;
@@ -190,76 +184,17 @@ public:
 	void SetDSTScaDisasm(const std::string& code);
 
 
-	CgBinaryDisasm(const std::string& path)
-		: m_path(path)
-	{
-		fs::file f(path);
-		if (!f) return;
+	CgBinaryDisasm(const std::string& path);
 
-		m_buffer_size = f.size();
-		m_buffer = new u8[m_buffer_size];
-		f.read(m_buffer, m_buffer_size);
-		fmt::append(m_arb_shader, "Loading... [%s]\n", path.c_str());
+	template <typename T>
+	CgBinaryDisasm(const std::span<T>& data)
+		: m_path("<raw>")
+	{
+		m_buffer.resize(data.size_bytes());
+		std::memcpy(m_buffer.data(), data.data(), data.size_bytes());
 	}
 
-	~CgBinaryDisasm()
-	{
-		delete[] m_buffer;
-	}
-
-	static std::string GetCgParamType(u32 type)
-	{
-		switch (type)
-		{
-		case 1045: return "float";
-		case 1046:
-		case 1047:
-		case 1048: return fmt::format("float%d", type - 1044);
-		case 1064: return "float4x4";
-		case 1066: return "sampler2D";
-		case 1069: return "samplerCUBE";
-		case 1091: return "float1";
-
-		default: return fmt::format("!UnkCgType(%d)", type);
-		}
-	}
-
-	std::string GetCgParamName(u32 offset) const
-	{
-		return std::string(reinterpret_cast<char*>(&m_buffer[offset]));
-	}
-
-	std::string GetCgParamRes(u32 /*offset*/) const
-	{
-		// rsx_log.warning("GetCgParamRes offset 0x%x", offset);
-		// TODO
-		return "";
-	}
-
-	std::string GetCgParamSemantic(u32 offset) const
-	{
-		return std::string(reinterpret_cast<char*>(&m_buffer[offset]));
-	}
-
-	std::string GetCgParamValue(u32 offset, u32 end_offset) const
-	{
-		std::string offsets = "offsets:";
-
-		u32 num = 0;
-		offset += 6;
-		while (offset < end_offset)
-		{
-			fmt::append(offsets, " %d,", m_buffer[offset] << 8 | m_buffer[offset + 1]);
-			offset += 4;
-			num++;
-		}
-
-		if (num > 4)
-			return "";
-
-		offsets.pop_back();
-		return fmt::format("num %d ", num) + offsets;
-	}
+	~CgBinaryDisasm() = default;
 
 	template<typename T>
 	T& GetCgRef(const u32 offset)
@@ -267,163 +202,14 @@ public:
 		return reinterpret_cast<T&>(m_buffer[offset]);
 	}
 
-	void ConvertToLE(CgBinaryProgram& prog)
-	{
-		// BE payload, requires that data be swapped
-		const auto be_profile = prog.profile;
+	static std::string GetCgParamType(u32 type);
+	std::string GetCgParamName(u32 offset) const;
+	std::string GetCgParamRes(u32 /*offset*/) const;
+	std::string GetCgParamSemantic(u32 offset) const;
+	std::string GetCgParamValue(u32 offset, u32 end_offset) const;
 
-		auto swap_be32 = [&](u32 start_offset, size_t size_bytes)
-		{
-			auto start = reinterpret_cast<u32*>(m_buffer + start_offset);
-			auto end = reinterpret_cast<u32*>(m_buffer + start_offset + size_bytes);
-
-			for (auto data = start; data < end; ++data)
-			{
-				*data = std::bit_cast<be_t<u32>>(*data);
-			}
-		};
-
-		// 1. Swap the header
-		swap_be32(0, sizeof(CgBinaryProgram));
-
-		// 2. Swap parameters
-		swap_be32(prog.parameterArray, sizeof(CgBinaryParameter) * prog.parameterCount);
-
-		// 3. Swap the ucode
-		swap_be32(prog.ucode, m_buffer_size - prog.ucode);
-
-		// 4. Swap the domain header
-		if (be_profile == 7004u)
-		{
-			// Need to swap each field individually
-			auto& fprog = GetCgRef<CgBinaryFragmentProgram>(prog.program);
-			fprog.instructionCount = std::bit_cast<be_t<u32>>(fprog.instructionCount);
-			fprog.attributeInputMask = std::bit_cast<be_t<u32>>(fprog.attributeInputMask);
-			fprog.partialTexType = std::bit_cast<be_t<u32>>(fprog.partialTexType);
-			fprog.texCoordsInputMask = std::bit_cast<be_t<u16>>(fprog.texCoordsInputMask);
-			fprog.texCoords2D = std::bit_cast<be_t<u16>>(fprog.texCoords2D);
-			fprog.texCoordsCentroid = std::bit_cast<be_t<u16>>(fprog.texCoordsCentroid);
-		}
-		else
-		{
-			// Swap entire header block as all fields are u32
-			swap_be32(prog.program, sizeof(CgBinaryVertexProgram));
-		}
-	}
-
-	void BuildShaderBody()
-	{
-		ParamArray param_array;
-
-		auto& prog = GetCgRef<CgBinaryProgram>(0);
-
-		if (const u32 be_profile = std::bit_cast<be_t<u32>>(prog.profile);
-			be_profile == 7003u || be_profile == 7004u)
-		{
-			ConvertToLE(prog);
-			ensure(be_profile == prog.profile);
-		}
-
-		if (prog.profile == 7004u)
-		{
-			auto& fprog = GetCgRef<CgBinaryFragmentProgram>(prog.program);
-			m_arb_shader += "\n";
-			fmt::append(m_arb_shader, "# binaryFormatRevision 0x%x\n", prog.binaryFormatRevision);
-			fmt::append(m_arb_shader, "# profile sce_fp_rsx\n");
-			fmt::append(m_arb_shader, "# parameterCount %d\n", prog.parameterCount);
-			fmt::append(m_arb_shader, "# instructionCount %d\n", fprog.instructionCount);
-			fmt::append(m_arb_shader, "# attributeInputMask 0x%x\n", fprog.attributeInputMask);
-			fmt::append(m_arb_shader, "# registerCount %d\n\n", fprog.registerCount);
-
-			CgBinaryParameterOffset offset = prog.parameterArray;
-			for (u32 i = 0; i < prog.parameterCount; i++)
-			{
-				auto& fparam = GetCgRef<CgBinaryParameter>(offset);
-
-				std::string param_type = GetCgParamType(fparam.type) + " ";
-				std::string param_name = GetCgParamName(fparam.name) + " ";
-				std::string param_res = GetCgParamRes(fparam.res) + " ";
-				std::string param_semantic = GetCgParamSemantic(fparam.semantic) + " ";
-				std::string param_const = GetCgParamValue(fparam.embeddedConst, fparam.name);
-
-				fmt::append(m_arb_shader, "#%d%s%s%s%s\n", i, param_type, param_name, param_semantic, param_const);
-
-				offset += u32{sizeof(CgBinaryParameter)};
-			}
-
-			m_arb_shader += "\n";
-			m_offset = prog.ucode;
-			TaskFP();
-
-			u32 unused;
-			std::vector<u32> be_data;
-
-			// Swap bytes. FP decompiler expects input in BE
-			for (u32* ptr = reinterpret_cast<u32*>(m_buffer + m_offset),
-				*end = reinterpret_cast<u32*>(m_buffer + m_buffer_size);
-				ptr < end; ++ptr)
-			{
-				be_data.push_back(std::bit_cast<be_t<u32>>(*ptr));
-			}
-
-			RSXFragmentProgram prog;
-			auto metadata = program_hash_util::fragment_program_utils::analyse_fragment_program(be_data.data());
-			prog.ctrl = (fprog.outputFromH0 ? 0 : 0x40) | (fprog.depthReplace ? 0xe : 0);
-			prog.offset = metadata.program_start_offset;
-			prog.ucode_length = metadata.program_ucode_length;
-			prog.total_length = metadata.program_ucode_length + metadata.program_start_offset;
-			prog.data = reinterpret_cast<u8*>(be_data.data()) + metadata.program_start_offset;
-			for (u32 i = 0; i < 16; ++i) prog.texture_state.set_dimension(rsx::texture_dimension_extended::texture_dimension_2d, i);
-#ifndef WITHOUT_OPENGL
-			GLFragmentDecompilerThread(m_glsl_shader, param_array, prog, unused).Task();
-#endif
-		}
-
-		else
-		{
-			const auto& vprog = GetCgRef<CgBinaryVertexProgram>(prog.program);
-			m_arb_shader += "\n";
-			fmt::append(m_arb_shader, "# binaryFormatRevision 0x%x\n", prog.binaryFormatRevision);
-			fmt::append(m_arb_shader, "# profile sce_vp_rsx\n");
-			fmt::append(m_arb_shader, "# parameterCount %d\n", prog.parameterCount);
-			fmt::append(m_arb_shader, "# instructionCount %d\n", vprog.instructionCount);
-			fmt::append(m_arb_shader, "# registerCount %d\n", vprog.registerCount);
-			fmt::append(m_arb_shader, "# attributeInputMask 0x%x\n", vprog.attributeInputMask);
-			fmt::append(m_arb_shader, "# attributeOutputMask 0x%x\n\n", vprog.attributeOutputMask);
-
-			CgBinaryParameterOffset offset = prog.parameterArray;
-			for (u32 i = 0; i < prog.parameterCount; i++)
-			{
-				auto& vparam = GetCgRef<CgBinaryParameter>(offset);
-
-				std::string param_type = GetCgParamType(vparam.type) + " ";
-				std::string param_name = GetCgParamName(vparam.name) + " ";
-				std::string param_res = GetCgParamRes(vparam.res) + " ";
-				std::string param_semantic = GetCgParamSemantic(vparam.semantic) + " ";
-				std::string param_const = GetCgParamValue(vparam.embeddedConst, vparam.name);
-
-				fmt::append(m_arb_shader, "#%d%s%s%s%s\n", i, param_type, param_name, param_semantic, param_const);
-
-				offset += u32{sizeof(CgBinaryParameter)};
-			}
-
-			m_arb_shader += "\n";
-			m_offset = prog.ucode;
-			ensure((m_buffer_size - m_offset) % sizeof(u32) == 0);
-
-			u32* vdata = reinterpret_cast<u32*>(&m_buffer[m_offset]);
-			m_data.resize(prog.ucodeSize / sizeof(u32));
-			std::memcpy(m_data.data(), vdata, prog.ucodeSize);
-			TaskVP();
-
-			RSXVertexProgram prog;
-			program_hash_util::vertex_program_utils::analyse_vertex_program(vdata, 0, prog);
-			for (u32 i = 0; i < 4; ++i) prog.texture_state.set_dimension(rsx::texture_dimension_extended::texture_dimension_2d, i);
-#ifndef WITHOUT_OPENGL
-			GLVertexDecompilerThread(prog, m_glsl_shader, param_array).Task();
-#endif
-		}
-	}
+	void ConvertToLE(CgBinaryProgram& prog);
+	void BuildShaderBody(bool include_glsl = true);
 
 	static u32 GetData(const u32 d) { return d << 16 | d >> 16; }
 	void TaskFP();
diff --git a/rpcs3/Emu/RSX/VK/VKMemAlloc.cpp b/rpcs3/Emu/RSX/VK/VKMemAlloc.cpp
index 92c59a0773..d90fb7936c 100644
--- a/rpcs3/Emu/RSX/VK/VKMemAlloc.cpp
+++ b/rpcs3/Emu/RSX/VK/VKMemAlloc.cpp
@@ -1,4 +1,5 @@
 #define VMA_IMPLEMENTATION
+#define VMA_VULKAN_VERSION 1000000
 
 #include "util/atomic.hpp"
 #include "Utilities/mutex.h"
diff --git a/rpcs3/Emu/RSX/VK/vkutils/memory.cpp b/rpcs3/Emu/RSX/VK/vkutils/memory.cpp
index 3bce6c3550..c1018389e7 100644
--- a/rpcs3/Emu/RSX/VK/vkutils/memory.cpp
+++ b/rpcs3/Emu/RSX/VK/vkutils/memory.cpp
@@ -167,6 +167,7 @@ namespace vk
 		allocatorInfo.physicalDevice = pdev;
 		allocatorInfo.device = dev;
 		allocatorInfo.instance = inst;
+		allocatorInfo.vulkanApiVersion = VK_API_VERSION_1_0;
 
 		std::vector<VkDeviceSize> heap_limits;
 		const auto vram_allocation_limit = g_cfg.video.vk.vram_allocation_limit * 0x100000ull;
diff --git a/rpcs3/Emu/RSX/VK/vkutils/memory.h b/rpcs3/Emu/RSX/VK/vkutils/memory.h
index 7a8ee99934..0b4186666d 100644
--- a/rpcs3/Emu/RSX/VK/vkutils/memory.h
+++ b/rpcs3/Emu/RSX/VK/vkutils/memory.h
@@ -8,6 +8,7 @@
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wnullability-completeness"
 #endif
+#define VMA_VULKAN_VERSION 1000000
 #include "3rdparty/GPUOpen/VulkanMemoryAllocator/include/vk_mem_alloc.h"
 #ifdef __clang__
 #pragma clang diagnostic pop
diff --git a/rpcs3/Emu/RSX/rsx_methods.h b/rpcs3/Emu/RSX/rsx_methods.h
index dbcb8bb15d..60be7ec14e 100644
--- a/rpcs3/Emu/RSX/rsx_methods.h
+++ b/rpcs3/Emu/RSX/rsx_methods.h
@@ -705,6 +705,11 @@ namespace rsx
 			return decode<NV4097_SET_USER_CLIP_PLANE_CONTROL>().clip_plane5();
 		}
 
+		u32 clip_planes_mask() const
+		{
+			return registers[NV4097_SET_USER_CLIP_PLANE_CONTROL];
+		}
+
 		front_face front_face_mode() const
 		{
 			return decode<NV4097_SET_FRONT_FACE>().front_face_mode();
diff --git a/rpcs3/emucore.vcxproj b/rpcs3/emucore.vcxproj
index 0f07b59a3f..3e623663d4 100644
--- a/rpcs3/emucore.vcxproj
+++ b/rpcs3/emucore.vcxproj
@@ -454,6 +454,7 @@
     <ClCompile Include="Emu\Io\usb_vfs.cpp" />
     <ClCompile Include="Emu\RSX\Capture\rsx_capture.cpp" />
     <ClCompile Include="Emu\RSX\Capture\rsx_replay.cpp" />
+    <ClCompile Include="Emu\RSX\Program\CgBinaryProgram.cpp" />
     <ClCompile Include="Emu\RSX\Program\CgBinaryFragmentProgram.cpp" />
     <ClCompile Include="Emu\RSX\Program\CgBinaryVertexProgram.cpp" />
     <ClCompile Include="Emu\RSX\Common\BufferUtils.cpp" />
diff --git a/rpcs3/emucore.vcxproj.filters b/rpcs3/emucore.vcxproj.filters
index ef03ca416f..94f0382bcc 100644
--- a/rpcs3/emucore.vcxproj.filters
+++ b/rpcs3/emucore.vcxproj.filters
@@ -1099,6 +1099,9 @@
     <ClCompile Include="Emu\RSX\Program\VertexProgramDecompiler.cpp">
       <Filter>Emu\GPU\RSX\Program</Filter>
     </ClCompile>
+    <ClCompile Include="Emu\RSX\Program\CgBinaryProgram.cpp">
+      <Filter>Emu\GPU\RSX\Program</Filter>
+    </ClCompile>
     <ClCompile Include="Emu\RSX\Program\CgBinaryFragmentProgram.cpp">
       <Filter>Emu\GPU\RSX\Program</Filter>
     </ClCompile>
diff --git a/rpcs3/rpcs3qt/rsx_debugger.cpp b/rpcs3/rpcs3qt/rsx_debugger.cpp
index 809ce37525..bdea183d2e 100644
--- a/rpcs3/rpcs3qt/rsx_debugger.cpp
+++ b/rpcs3/rpcs3qt/rsx_debugger.cpp
@@ -4,6 +4,8 @@
 #include "table_item_delegate.h"
 #include "Emu/RSX/RSXThread.h"
 #include "Emu/RSX/gcm_printing.h"
+#include "Emu/RSX/Common/BufferUtils.h"
+#include "Emu/RSX/Program/CgBinaryProgram.h"
 #include "Utilities/File.h"
 
 #include <QHBoxLayout>
@@ -123,17 +125,6 @@ rsx_debugger::rsx_debugger(std::shared_ptr<gui_settings> gui_settings, QWidget*
 	vbox_tools->addLayout(hbox_controls);
 	vbox_tools->addWidget(m_tw_rsx);
 
-	// State explorer
-	m_text_transform_program = new QLabel();
-	m_text_transform_program->setTextInteractionFlags(Qt::TextSelectableByMouse | Qt::TextSelectableByKeyboard);
-	m_text_transform_program->setFont(mono);
-	m_text_transform_program->setText("");
-
-	m_text_shader_program = new QLabel();
-	m_text_shader_program->setTextInteractionFlags(Qt::TextSelectableByMouse | Qt::TextSelectableByKeyboard);
-	m_text_shader_program->setFont(mono);
-	m_text_shader_program->setText("");
-
 	m_list_index_buffer = new QListWidget();
 	m_list_index_buffer->setFont(mono);
 
@@ -208,11 +199,31 @@ rsx_debugger::rsx_debugger(std::shared_ptr<gui_settings> gui_settings, QWidget*
 	QWidget* buffers = new QWidget();
 	buffers->setLayout(buffer_layout);
 
+	auto xp_layout = new QHBoxLayout();
+	m_transform_disasm = new QTextEdit(this);
+	m_transform_disasm->setReadOnly(true);
+	m_transform_disasm->setWordWrapMode(QTextOption::NoWrap);
+	m_transform_disasm->setFont(QFontDatabase::systemFont(QFontDatabase::FixedFont));
+	xp_layout->addWidget(m_transform_disasm);
+
+	auto xp_tab = new QWidget();
+	xp_tab->setLayout(xp_layout);
+
+	auto fp_layout = new QHBoxLayout();
+	m_fragment_disasm = new QTextEdit(this);
+	m_fragment_disasm->setReadOnly(true);
+	m_fragment_disasm->setWordWrapMode(QTextOption::NoWrap);
+	m_fragment_disasm->setFont(QFontDatabase::systemFont(QFontDatabase::FixedFont));
+	fp_layout->addWidget(m_fragment_disasm);
+
+	auto fp_tab = new QWidget();
+	fp_tab->setLayout(fp_layout);
+
 	QTabWidget* state_rsx = new QTabWidget();
 	state_rsx->addTab(buffers, tr("RTTs and DS"));
-	state_rsx->addTab(m_text_transform_program, tr("Transform program"));
-	state_rsx->addTab(m_text_shader_program, tr("Shader program"));
-	state_rsx->addTab(m_list_index_buffer, tr("Index buffer"));
+	state_rsx->addTab(xp_tab, tr("Vertex Program"));
+	state_rsx->addTab(fp_tab, tr("Fragment Program"));
+	state_rsx->addTab(m_list_index_buffer, tr("Index Buffer"));
 
 	QHBoxLayout* main_layout = new QHBoxLayout();
 	main_layout->addLayout(vbox_tools, 1);
@@ -592,10 +603,10 @@ void rsx_debugger::OnClickDrawCalls()
 	}
 
 	// Programs
-	m_text_transform_program->clear();
-	m_text_transform_program->setText(qstr(frame_debug.draw_calls[draw_id].programs.first));
-	m_text_shader_program->clear();
-	m_text_shader_program->setText(qstr(frame_debug.draw_calls[draw_id].programs.second));
+	m_transform_disasm->clear();
+	m_transform_disasm->setText(qstr(frame_debug.draw_calls[draw_id].programs.first));
+	m_fragment_disasm->clear();
+	m_fragment_disasm->setText(qstr(frame_debug.draw_calls[draw_id].programs.second));
 
 	m_list_index_buffer->clear();
 	//m_list_index_buffer->insertColumn(0, "Index", 0, 700);
@@ -621,6 +632,8 @@ void rsx_debugger::UpdateInformation() const
 {
 	GetMemory();
 	GetBuffers();
+	GetVertexProgram();
+	GetFragmentProgram();
 }
 
 void rsx_debugger::GetMemory() const
@@ -1202,3 +1215,111 @@ void rsx_debugger::GetBuffers() const
 
 	m_buffer_tex->showImage(QImage(m_buffer_tex->cache.data(), width, height, QImage::Format_RGB32));
 }
+
+void rsx_debugger::GetVertexProgram() const
+{
+	const auto render = rsx::get_current_renderer();
+	if (!render || !render->is_initialized || !render->local_mem_size || !render->is_paused())
+	{
+		return;
+	}
+
+	RSXVertexProgram vp;
+	vp.data.reserve(512 * 4);
+
+	const u32 vp_entry = rsx::method_registers.transform_program_start();
+	program_hash_util::vertex_program_utils::analyse_vertex_program
+	(
+		rsx::method_registers.transform_program.data(),  // Input raw block
+		vp_entry,                                       // Address of entry point
+		vp                                          // [out] Program object
+	);
+
+	const u32 ucode_len = static_cast<u32>(vp.data.size() * sizeof(u32));
+	std::vector<u32> vp_blob = {
+		7003u,             // Type
+		6u,                // Revision
+		56u + ucode_len,   // Total size
+		0,                 // paramCount
+		0,                 // paramArray
+		32u,               // Program header offset
+		ucode_len,         // Ucode length
+		56u,               // Ucode start
+
+		::size32(vp.data) / 4, // Instruction count
+		0,                     // Slot
+		16u,                   // Registers used
+		rsx::method_registers.vertex_attrib_input_mask(),
+		rsx::method_registers.vertex_attrib_output_mask(),
+		rsx::method_registers.clip_planes_mask()
+	};
+
+	vp_blob.resize(vp_blob.size() + vp.data.size());
+	std::copy(vp.data.begin(), vp.data.end(), vp_blob.begin() + 14);
+
+	std::span<u32> vp_binary(vp_blob);
+	CgBinaryDisasm vp_disasm(vp_binary);
+	vp_disasm.BuildShaderBody(false);
+
+	m_transform_disasm->clear();
+	m_transform_disasm->setText(QString::fromStdString(vp_disasm.GetArbShader()));
+}
+
+void rsx_debugger::GetFragmentProgram() const
+{
+	const auto render = rsx::get_current_renderer();
+	if (!render || !render->is_initialized || !render->local_mem_size || !render->is_paused())
+	{
+		return;
+	}
+
+	const auto [program_offset, program_location] = rsx::method_registers.shader_program_address();
+	const auto address = rsx::get_address(program_offset, program_location, 4);
+	if (!address)
+	{
+		m_fragment_disasm->clear();
+		return;
+	}
+
+	// NOTE: Reading through super ptr while crash-safe means we're probably reading incorrect bytes, but should be fine in 99% of cases
+	auto data_ptr = vm::get_super_ptr(address);
+	const auto fp_metadata = program_hash_util::fragment_program_utils::analyse_fragment_program(data_ptr);
+
+	const bool output_h0 = rsx::method_registers.shader_control() & CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS ? false : true;
+	const bool depth_replace = rsx::method_registers.shader_control() & CELL_GCM_SHADER_CONTROL_DEPTH_EXPORT ? true : false;
+	const u32 flags = 16u /* Reg count */ | ((output_h0 /* 16-bit exports? */ ? 1u : 0u) << 8u) | ((depth_replace /* Export depth? */ ? 1u : 0u) << 16u) | (0u /* Uses KILL? */ << 24u);
+	const auto ucode_len = fp_metadata.program_ucode_length;
+
+	std::vector<u32> blob = {
+		7004u,             // Type
+		6u,                // Revision
+		56u + ucode_len,   // Total size
+		0,                 // paramCount
+		0,                 // paramArray
+		32u,               // Program header offset
+		ucode_len,         // Ucode length
+		56u,               // Ucode start
+
+		ucode_len / 16,                                               // Instruction count
+		rsx::method_registers.vertex_attrib_output_mask(),            // Slot
+		0u,                                                           // Partial load
+		0u | (rsx::method_registers.texcoord_control_mask() << 16u),  // Texcoord input mask | tex2d control
+		0u | (flags << 16u),                                          // Centroid inputs (xor tex2d control) | flags (regs, 16-bit, fragDepth, KILL)
+
+		0u,                                                           // Padding
+	};
+
+	// Copy in the program bytes, swapped
+	const u32 start_offset_in_words = fp_metadata.program_start_offset / 4;
+	const u32 ucode_length_in_words = fp_metadata.program_ucode_length / 4;
+	blob.resize(blob.size() + ucode_length_in_words);
+	copy_data_swap_u32(blob.data() + 14, utils::bless<u32>(data_ptr) + start_offset_in_words, ucode_length_in_words);
+	//std::memcpy(blob.data() + 14, utils::bless<char>(data_ptr) + fp_metadata.program_start_offset, fp_metadata.program_ucode_length);
+
+	std::span<u32> fp_binary(blob);
+	CgBinaryDisasm fp_disasm(fp_binary);
+	fp_disasm.BuildShaderBody(false);
+
+	m_fragment_disasm->clear();
+	m_fragment_disasm->setText(QString::fromStdString(fp_disasm.GetArbShader()));
+}
diff --git a/rpcs3/rpcs3qt/rsx_debugger.h b/rpcs3/rpcs3qt/rsx_debugger.h
index 95ac67ddcd..2c189c3e8a 100644
--- a/rpcs3/rpcs3qt/rsx_debugger.h
+++ b/rpcs3/rpcs3qt/rsx_debugger.h
@@ -11,6 +11,7 @@
 #include <QListWidget>
 #include <QTableWidget>
 #include <QTabWidget>
+#include <QTextEdit>
 
 #include <memory>
 
@@ -61,8 +62,8 @@ class rsx_debugger : public QDialog
 	Buffer* m_buffer_tex;
 	QLabel* m_enabled_textures_label;
 
-	QLabel* m_text_transform_program;
-	QLabel* m_text_shader_program;
+	QTextEdit* m_transform_disasm;
+	QTextEdit* m_fragment_disasm;
 
 	u32 m_cur_texture = 0;
 	u32 m_texture_format_override = 0;
@@ -87,4 +88,7 @@ protected:
 
 private:
 	void PerformJump(u32 address);
+
+	void GetVertexProgram() const;
+	void GetFragmentProgram() const;
 };