love/src/modules/graphics/metal/Shader.mm

/**
 * Copyright (c) 2006-2020 LOVE Development Team
 *
 * This software is provided 'as-is', without any express or implied
 * warranty.  In no event will the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Permission is granted to anyone to use this software for any purpose,
 * including commercial applications, and to alter it and redistribute it
 * freely, subject to the following restrictions:
 *
 * 1. The origin of this software must not be misrepresented; you must not
 *    claim that you wrote the original software. If you use this software
 *    in a product, an acknowledgment in the product documentation would be
 *    appreciated but is not required.
 * 2. Altered source versions must be plainly marked as such, and must not be
 *    misrepresented as being the original software.
 * 3. This notice may not be removed or altered from any source distribution.
 **/

#include "Shader.h"
#include "Graphics.h"
#include "common/int.h"

// glslang
#include "libraries/glslang/glslang/Public/ShaderLang.h"
#include "libraries/glslang/SPIRV/GlslangToSpv.h"
#include "libraries/spirv_cross/spirv_msl.hpp"
#include "libraries/spirv_cross/spirv_reflect.hpp"

#include <algorithm>

namespace love
{
namespace graphics
{
namespace metal
{

static_assert(MAX_COLOR_RENDER_TARGETS <= 8, "Metal pipeline cache key only stores 8 render target pixel formats.");

// TODO: Use love.graphics to determine actual limits?
static const TBuiltInResource defaultTBuiltInResource = {
	/* .MaxLights = */ 32,
	/* .MaxClipPlanes = */ 6,
	/* .MaxTextureUnits = */ 32,
	/* .MaxTextureCoords = */ 32,
	/* .MaxVertexAttribs = */ 64,
	/* .MaxVertexUniformComponents = */ 16384,
	/* .MaxVaryingFloats = */ 128,
	/* .MaxVertexTextureImageUnits = */ 32,
	/* .MaxCombinedTextureImageUnits = */ 80,
	/* .MaxTextureImageUnits = */ 32,
	/* .MaxFragmentUniformComponents = */ 16384,
	/* .MaxDrawBuffers = */ 8,
	/* .MaxVertexUniformVectors = */ 4096,
	/* .MaxVaryingVectors = */ 32,
	/* .MaxFragmentUniformVectors = */ 4096,
	/* .MaxVertexOutputVectors = */ 32,
	/* .MaxFragmentInputVectors = */ 31,
	/* .MinProgramTexelOffset = */ -8,
	/* .MaxProgramTexelOffset = */ 7,
	/* .MaxClipDistances = */ 8,
	/* .MaxComputeWorkGroupCountX = */ 65535,
	/* .MaxComputeWorkGroupCountY = */ 65535,
	/* .MaxComputeWorkGroupCountZ = */ 65535,
	/* .MaxComputeWorkGroupSizeX = */ 1024,
	/* .MaxComputeWorkGroupSizeY = */ 1024,
	/* .MaxComputeWorkGroupSizeZ = */ 64,
	/* .MaxComputeUniformComponents = */ 1024,
	/* .MaxComputeTextureImageUnits = */ 32,
	/* .MaxComputeImageUniforms = */ 16,
	/* .MaxComputeAtomicCounters = */ 4096,
	/* .MaxComputeAtomicCounterBuffers = */ 8,
	/* .MaxVaryingComponents = */ 128,
	/* .MaxVertexOutputComponents = */ 128,
	/* .MaxGeometryInputComponents = */ 128,
	/* .MaxGeometryOutputComponents = */ 128,
	/* .MaxFragmentInputComponents = */ 128,
	/* .MaxImageUnits = */ 192,
	/* .MaxCombinedImageUnitsAndFragmentOutputs = */ 144,
	/* .MaxCombinedShaderOutputResources = */ 144,
	/* .MaxImageSamples = */ 32,
	/* .MaxVertexImageUniforms = */ 16,
	/* .MaxTessControlImageUniforms = */ 16,
	/* .MaxTessEvaluationImageUniforms = */ 16,
	/* .MaxGeometryImageUniforms = */ 16,
	/* .MaxFragmentImageUniforms = */ 16,
	/* .MaxCombinedImageUniforms = */ 80,
	/* .MaxGeometryTextureImageUnits = */ 16,
	/* .MaxGeometryOutputVertices = */ 256,
	/* .MaxGeometryTotalOutputComponents = */ 1024,
	/* .MaxGeometryUniformComponents = */ 1024,
	/* .MaxGeometryVaryingComponents = */ 64,
	/* .MaxTessControlInputComponents = */ 128,
	/* .MaxTessControlOutputComponents = */ 128,
	/* .MaxTessControlTextureImageUnits = */ 16,
	/* .MaxTessControlUniformComponents = */ 1024,
	/* .MaxTessControlTotalOutputComponents = */ 4096,
	/* .MaxTessEvaluationInputComponents = */ 128,
	/* .MaxTessEvaluationOutputComponents = */ 128,
	/* .MaxTessEvaluationTextureImageUnits = */ 16,
	/* .MaxTessEvaluationUniformComponents = */ 1024,
	/* .MaxTessPatchComponents = */ 120,
	/* .MaxPatchVertices = */ 32,
	/* .MaxTessGenLevel = */ 64,
	/* .MaxViewports = */ 16,
	/* .MaxVertexAtomicCounters = */ 4096,
	/* .MaxTessControlAtomicCounters = */ 4096,
	/* .MaxTessEvaluationAtomicCounters = */ 4096,
	/* .MaxGeometryAtomicCounters = */ 4096,
	/* .MaxFragmentAtomicCounters = */ 4096,
	/* .MaxCombinedAtomicCounters = */ 4096,
	/* .MaxAtomicCounterBindings = */ 8,
	/* .MaxVertexAtomicCounterBuffers = */ 8,
	/* .MaxTessControlAtomicCounterBuffers = */ 8,
	/* .MaxTessEvaluationAtomicCounterBuffers = */ 8,
	/* .MaxGeometryAtomicCounterBuffers = */ 8,
	/* .MaxFragmentAtomicCounterBuffers = */ 8,
	/* .MaxCombinedAtomicCounterBuffers = */ 8,
	/* .MaxAtomicCounterBufferSize = */ 16384,
	/* .MaxTransformFeedbackBuffers = */ 4,
	/* .MaxTransformFeedbackInterleavedComponents = */ 64,
	/* .MaxCullDistances = */ 8,
	/* .MaxCombinedClipAndCullDistances = */ 8,
	/* .MaxSamples = */ 32,
	/* .maxMeshOutputVerticesNV = */ 256,
	/* .maxMeshOutputPrimitivesNV = */ 512,
	/* .maxMeshWorkGroupSizeX_NV = */ 32,
	/* .maxMeshWorkGroupSizeY_NV = */ 1,
	/* .maxMeshWorkGroupSizeZ_NV = */ 1,
	/* .maxTaskWorkGroupSizeX_NV = */ 32,
	/* .maxTaskWorkGroupSizeY_NV = */ 1,
	/* .maxTaskWorkGroupSizeZ_NV = */ 1,
	/* .maxMeshViewCountNV = */ 4,
	/* .maxDualSourceDrawBuffersEXT = */ 1,
	/* .limits = */ {
		/* .nonInductiveForLoops = */ 1,
		/* .whileLoops = */ 1,
		/* .doWhileLoops = */ 1,
		/* .generalUniformIndexing = */ 1,
		/* .generalAttributeMatrixVectorIndexing = */ 1,
		/* .generalVaryingIndexing = */ 1,
		/* .generalSamplerIndexing = */ 1,
		/* .generalVariableIndexing = */ 1,
		/* .generalConstantMatrixVectorIndexing = */ 1,
	}
};

static MTLVertexFormat getMTLVertexFormat(DataFormat format)
{
	switch (format)
	{
		case DATAFORMAT_FLOAT: return MTLVertexFormatFloat;
		case DATAFORMAT_FLOAT_VEC2: return MTLVertexFormatFloat2;
		case DATAFORMAT_FLOAT_VEC3: return MTLVertexFormatFloat3;
		case DATAFORMAT_FLOAT_VEC4: return MTLVertexFormatFloat4;

		case DATAFORMAT_INT32: return MTLVertexFormatInt;
		case DATAFORMAT_INT32_VEC2: return MTLVertexFormatInt2;
		case DATAFORMAT_INT32_VEC3: return MTLVertexFormatInt3;
		case DATAFORMAT_INT32_VEC4: return MTLVertexFormatInt4;

		case DATAFORMAT_UINT32: return MTLVertexFormatUInt;
		case DATAFORMAT_UINT32_VEC2: return MTLVertexFormatUInt2;
		case DATAFORMAT_UINT32_VEC3: return MTLVertexFormatUInt3;
		case DATAFORMAT_UINT32_VEC4: return MTLVertexFormatUInt4;

		case DATAFORMAT_SNORM8_VEC4: return MTLVertexFormatChar4Normalized;
		case DATAFORMAT_UNORM8_VEC4: return MTLVertexFormatUChar4Normalized;
		case DATAFORMAT_INT8_VEC4: return MTLVertexFormatChar4;
		case DATAFORMAT_UINT8_VEC4: return MTLVertexFormatUChar4;

		case DATAFORMAT_SNORM16_VEC2: return MTLVertexFormatShort2Normalized;
		case DATAFORMAT_SNORM16_VEC4: return MTLVertexFormatShort4Normalized;

		case DATAFORMAT_UNORM16_VEC2: return MTLVertexFormatUShort2Normalized;
		case DATAFORMAT_UNORM16_VEC4: return MTLVertexFormatUShort4Normalized;

		case DATAFORMAT_INT16_VEC2: return MTLVertexFormatShort2;
		case DATAFORMAT_INT16_VEC4: return MTLVertexFormatShort4;

		case DATAFORMAT_UINT16: return MTLVertexFormatUShort;
		case DATAFORMAT_UINT16_VEC2: return MTLVertexFormatUShort2;
		case DATAFORMAT_UINT16_VEC4: return MTLVertexFormatUShort4;

		default: return MTLVertexFormatInvalid;
	}
}

static MTLBlendOperation getMTLBlendOperation(BlendOperation op)
{
	switch (op)
	{
		case BLENDOP_ADD: return MTLBlendOperationAdd;
		case BLENDOP_SUBTRACT: return MTLBlendOperationSubtract;
		case BLENDOP_REVERSE_SUBTRACT: return MTLBlendOperationReverseSubtract;
		case BLENDOP_MIN: return MTLBlendOperationMin;
		case BLENDOP_MAX: return MTLBlendOperationMax;
		case BLENDOP_MAX_ENUM: return MTLBlendOperationAdd;
	}
	return MTLBlendOperationAdd;
}

static MTLBlendFactor getMTLBlendFactor(BlendFactor factor)
{
	switch (factor)
	{
		case BLENDFACTOR_ZERO: return MTLBlendFactorZero;
		case BLENDFACTOR_ONE: return MTLBlendFactorOne;
		case BLENDFACTOR_SRC_COLOR: return MTLBlendFactorSourceColor;
		case BLENDFACTOR_ONE_MINUS_SRC_COLOR: return MTLBlendFactorOneMinusSourceColor;
		case BLENDFACTOR_SRC_ALPHA: return MTLBlendFactorSourceAlpha;
		case BLENDFACTOR_ONE_MINUS_SRC_ALPHA: return MTLBlendFactorOneMinusSourceAlpha;
		case BLENDFACTOR_DST_COLOR: return MTLBlendFactorDestinationColor;
		case BLENDFACTOR_ONE_MINUS_DST_COLOR: return MTLBlendFactorOneMinusDestinationColor;
		case BLENDFACTOR_DST_ALPHA: return MTLBlendFactorDestinationAlpha;
		case BLENDFACTOR_ONE_MINUS_DST_ALPHA: return MTLBlendFactorOneMinusDestinationAlpha;
		case BLENDFACTOR_SRC_ALPHA_SATURATED: return MTLBlendFactorSourceAlphaSaturated;
		case BLENDFACTOR_MAX_ENUM: return MTLBlendFactorZero;
	}
	return MTLBlendFactorZero;
}

static inline id<MTLTexture> getMTLTexture(love::graphics::Texture *tex)
{
	return tex ? (__bridge id<MTLTexture>)(void *) tex->getHandle() : nil;
}

static inline id<MTLTexture> getMTLTexture(love::graphics::Buffer *buffer)
{
	return buffer ? (__bridge id<MTLTexture>)(void *) buffer->getTexelBufferHandle() : nil;
}

static inline id<MTLSamplerState> getMTLSampler(love::graphics::Texture *tex)
{
	return tex ? (__bridge id<MTLSamplerState>)(void *) tex->getSamplerHandle() : nil;
}

static EShLanguage getGLSLangStage(ShaderStage::StageType stage)
{
	switch (stage)
	{
		case ShaderStage::STAGE_VERTEX: return EShLangVertex;
		case ShaderStage::STAGE_PIXEL: return EShLangFragment;
		case ShaderStage::STAGE_MAX_ENUM: return EShLangCount;
	}
	return EShLangCount;
}

Shader::Shader(id<MTLDevice> device, love::graphics::ShaderStage *vertex, love::graphics::ShaderStage *pixel)
	: love::graphics::Shader(vertex, pixel)
	, functions()
	, builtinUniformInfo()
	, localUniformBufferData(nullptr)
	, localUniformBufferSize(0)
	, builtinUniformDataOffset(0)
{ @autoreleasepool {
	using namespace glslang;

	TShader *glslangShaders[ShaderStage::STAGE_MAX_ENUM] = {};

	TProgram *program = new TProgram();

	auto cleanup = [&]()
	{
		delete program;
		for (int i = 0; i < ShaderStage::STAGE_MAX_ENUM; i++)
			delete glslangShaders[i];
	};

	// We can't do this in ShaderStage because the mapIO call modifies the
	// TShader internals in a manner that prevents it from being shared.
	for (int i = 0; i < ShaderStage::STAGE_MAX_ENUM; i++)
	{
		if (!stages[i])
			continue;

		auto stage = (ShaderStage::StageType) i;
		auto glslangstage = getGLSLangStage(stage);
		auto tshader = new TShader(glslangstage);

		glslangShaders[i] = tshader;

		tshader->setEnvInput(EShSourceGlsl, glslangstage, EShClientVulkan, 450);
		tshader->setEnvClient(EShClientVulkan, EShTargetVulkan_1_2);
		tshader->setEnvTarget(EShTargetSpv, EShTargetSpv_1_5);
		tshader->setAutoMapLocations(true);
		tshader->setAutoMapBindings(true);

		// Needed for local uniforms to work (they will be converted into a
		// uniform block).
		// https://github.com/KhronosGroup/GLSL/blob/master/extensions/ext/GL_EXT_vulkan_glsl_relaxed.txt
		tshader->setEnvInputVulkanRulesRelaxed();
		tshader->setGlobalUniformBinding(0);
		tshader->setGlobalUniformSet(0);

		const std::string &source = stages[i]->getSource();
		const char *csrc = source.c_str();
		int srclen = (int) source.length();
		tshader->setStringsWithLengths(&csrc, &srclen, 1);

		int defaultversion = 450;
		EProfile defaultprofile = ECoreProfile;
		bool forcedefault = false;
		bool forwardcompat = true;

		if (!tshader->parse(&defaultTBuiltInResource, defaultversion, defaultprofile, forcedefault, forwardcompat, EShMsgSuppressWarnings))
		{
			const char *stagename = "unknown";
			ShaderStage::getConstant(stage, stagename);

			std::string err = "Error parsing " + std::string(stagename) + " shader:\n\n"
				+ std::string(tshader->getInfoLog()) + "\n"
				+ std::string(tshader->getInfoDebugLog());

			cleanup();
			throw love::Exception("%s", err.c_str());
		}

		program->addShader(tshader);
	}

	if (!program->link(EShMsgDefault))
	{
		//err = "Cannot compile shader:\n\n" + std::string(program->getInfoLog()) + "\n" + std::string(program->getInfoDebugLog());
		cleanup();
		throw love::Exception("link failed! %s\n", program->getInfoLog());
	}

	if (!program->mapIO())
	{
		cleanup();
		throw love::Exception("mapIO failed");
	}

	try
	{
		compileFromGLSLang(device, *program);
	}
	catch (love::Exception &)
	{
		cleanup();
		throw;
	}

	cleanup();
}}

void Shader::compileFromGLSLang(id<MTLDevice> device, const glslang::TProgram &program)
{
	using namespace glslang;
	using namespace spirv_cross;

	auto gfx = Graphics::getInstance();

	std::map<std::string, int> varyings;
	int nextVaryingLocation = 0;

	for (int stageindex = 0; stageindex < ShaderStage::STAGE_MAX_ENUM; stageindex++)
	{
		auto glslangstage = getGLSLangStage((ShaderStage::StageType) stageindex);
		auto intermediate = program.getIntermediate(glslangstage);
		if (intermediate == nullptr)
			continue;

		spv::SpvBuildLogger logger;
		glslang::SpvOptions opt;
		opt.validate = true;

		std::vector<unsigned int> spirv;
		GlslangToSpv(*intermediate, spirv, &logger, &opt);

		std::string msgs = logger.getAllMessages();
//		printf("spirv length: %ld, messages:\n%s\n", spirv.size(), msgs.c_str());

		try
		{
//			printf("GLSL INPUT SOURCE:\n\n%s\n\n", pixel->getSource().c_str());

			CompilerMSL msl(std::move(spirv));

			auto interfacevars = msl.get_active_interface_variables();

			msl.set_enabled_interface_variables(interfacevars);

			ShaderResources resources = msl.get_shader_resources();

			for (const auto &resource : resources.sampled_images)
			{
				const SPIRType &basetype = msl.get_type(resource.base_type_id);
				const SPIRType &type = msl.get_type(resource.type_id);
				const SPIRType &imagetype = msl.get_type(basetype.image.type);

				UniformInfo u = {};
				u.baseType = UNIFORM_SAMPLER;
				u.name = resource.name;
				u.count = type.array.empty() ? 1 : type.array[0];
				u.isDepthSampler = type.image.depth;

				switch (imagetype.basetype)
				{
				case SPIRType::Float:
					u.dataBaseType = DATA_BASETYPE_FLOAT;
					break;
				case SPIRType::Int:
					u.dataBaseType = DATA_BASETYPE_INT;
					break;
				case SPIRType::UInt:
					u.dataBaseType = DATA_BASETYPE_UINT;
					break;
				default:
					break;
				}

				switch (basetype.image.dim)
				{
				case spv::Dim2D:
					u.textureType = basetype.image.arrayed ? TEXTURE_2D_ARRAY : TEXTURE_2D;
					u.textures = new love::graphics::Texture*[u.count];
					break;
				case spv::Dim3D:
					u.textureType = TEXTURE_VOLUME;
					u.textures = new love::graphics::Texture*[u.count];
					break;
				case spv::DimCube:
					if (basetype.image.arrayed)
						throw love::Exception("Cubemap Arrays are not currently supported.");
					u.textureType = TEXTURE_CUBE;
					u.textures = new love::graphics::Texture*[u.count];
					break;
				case spv::DimBuffer:
					u.baseType = UNIFORM_TEXELBUFFER;
					u.buffers = new love::graphics::Buffer*[u.count];
					break;
				default:
					// TODO: error? continue?
					break;
				}

				u.data = malloc(sizeof(int) * u.count);
				for (int i = 0; i < u.count; i++)
					u.ints[i] = -1; // Initialized below, after compiling.

				if (u.baseType == UNIFORM_SAMPLER)
				{
					auto tex = gfx->getDefaultTexture(u.textureType);
					for (int i = 0; i < u.count; i++)
					{
						tex->retain();
						u.textures[i] = tex;
					}
				}
				else if (u.baseType == UNIFORM_TEXELBUFFER)
				{
					for (int i = 0; i < u.count; i++)
						u.buffers[i] = nullptr; // TODO
				}

				uniforms[u.name] = u;

				BuiltinUniform builtin;
				if (getConstant(resource.name.c_str(), builtin))
					builtinUniformInfo[builtin] = &uniforms[u.name];
			}

			for (const auto &resource : resources.uniform_buffers)
			{
				if (resource.name == "gl_DefaultUniformBlock")
				{
					MSLResourceBinding binding;
					binding.stage = msl.get_execution_model();
					binding.binding = msl.get_decoration(resource.id, spv::DecorationBinding);
					binding.desc_set = msl.get_decoration(resource.id, spv::DecorationDescriptorSet);
					binding.msl_buffer = getUniformBufferBinding();
					msl.add_msl_resource_binding(binding);

					const SPIRType &type = msl.get_type(resource.base_type_id);
					const auto &membertypes = type.member_types;

					size_t size = msl.get_declared_struct_size(type);

					if (localUniformBufferSize != 0)
					{
						if (localUniformBufferSize != size)
							throw love::Exception("Local uniform buffer size mismatch");
						continue;
					}

					localUniformBufferData = new uint8[size];
					localUniformBufferSize = size;

					memset(localUniformBufferData, 0, size);

					for (size_t uindex = 0; uindex < membertypes.size(); uindex++)
					{
						const auto &membertype = msl.get_type(membertypes[uindex]);
						size_t membersize = msl.get_declared_struct_member_size(type, uindex);
						size_t offset = msl.type_struct_member_offset(type, uindex);

						UniformInfo u = {};
						u.name = msl.get_member_name(type.self, uindex);
						u.dataSize = membersize;
						u.count = membertype.array.empty() ? 1 : membertype.array[0];

						switch (membertype.basetype)
						{
						case SPIRType::Int:
						case SPIRType::UInt:
						case SPIRType::Float:
							u.data = localUniformBufferData + offset;
							if (membertype.columns == 1)
							{
								if (membertype.basetype == SPIRType::Int)
									u.baseType = UNIFORM_INT;
								else if (membertype.basetype == SPIRType::UInt)
									u.baseType = UNIFORM_UINT;
								else
									u.baseType = UNIFORM_FLOAT;
								u.components = membertype.vecsize;
							}
							else
							{
								u.baseType = UNIFORM_MATRIX;
								u.matrix.rows = membertype.vecsize;
								u.matrix.columns = membertype.columns;
							}
							break;
						case SPIRType::Struct:
							// TODO
							break;
						default:
							break;
						}

						uniforms[u.name] = u;

						BuiltinUniform builtin = BUILTIN_MAX_ENUM;
						if (getConstant(u.name.c_str(), builtin))
						{
							if (builtin == BUILTIN_UNIFORMS_PER_DRAW)
								builtinUniformDataOffset = offset;
							builtinUniformInfo[builtin] = &uniforms[u.name];
						}
					}

				}
			}

			for (const auto &resource : resources.storage_buffers)
			{
				auto it = uniforms.find(resource.name);
				if (it != uniforms.end())
				{
					continue;
				}

				// TODO
			}

			if (stageindex == ShaderStage::STAGE_VERTEX)
			{
				int nextattributeindex = ATTRIB_MAX_ENUM;

				for (const auto &var : interfacevars)
				{
					spv::StorageClass storage = msl.get_storage_class(var);
					const std::string &name = msl.get_name(var);

					if (storage == spv::StorageClassInput)
					{
						int index = 0;

						BuiltinVertexAttribute builtinattribute;
						if (graphics::getConstant(name.c_str(), builtinattribute))
							index = (int) builtinattribute;
						else
							index = nextattributeindex++;

						msl.set_decoration(var, spv::DecorationLocation, index);
						attributes[name] = msl.get_decoration(var, spv::DecorationLocation);
					}
				}

				for (const auto &varying : resources.stage_outputs)
				{
//					printf("vertex shader output %s: %d\n", inp.name.c_str(), msl.get_decoration(inp.id, spv::DecorationLocation));
					varyings[varying.name] = nextVaryingLocation;
					msl.set_decoration(varying.id, spv::DecorationLocation, nextVaryingLocation++);
				}
			}
			else if (stageindex == ShaderStage::STAGE_PIXEL)
			{
				for (const auto &varying : resources.stage_inputs)
				{
					const auto it = varyings.find(varying.name);
					if (it != varyings.end())
						msl.set_decoration(varying.id, spv::DecorationLocation, it->second);
				}
			}

//			printf("// ubos: %ld, storage: %ld, inputs: %ld, outputs: %ld, images: %ld, samplers: %ld, push: %ld\n", resources.uniform_buffers.size(), resources.storage_buffers.size(), resources.stage_inputs.size(), resources.stage_outputs.size(), resources.storage_images.size(), resources.sampled_images.size(), resources.push_constant_buffers.size());

			CompilerMSL::Options options;

			options.set_msl_version(2, 1);
			options.texture_buffer_native = true;

#ifdef LOVE_IOS
			options.platform = CompilerMSL::Options::iOS;
#else
			options.platform = CompilerMSL::Options::macOS;
#endif

			msl.set_msl_options(options);

			std::string source = msl.compile();
//			printf("// MSL SOURCE for stage %d:\n\n%s\n\n", stageindex, source.c_str());

			NSString *nssource = [[NSString alloc] initWithBytes:source.c_str()
														  length:source.length()
														encoding:NSUTF8StringEncoding];

			NSError *err = nil;
			id<MTLLibrary> library = [device newLibraryWithSource:nssource options:nil error:&err];
			if (library == nil && err != nil)
			{
				NSLog(@"errors: %@", err);
				throw love::Exception("Error compiling converted Metal shader code");
			}

			functions[stageindex] = [library newFunctionWithName:library.functionNames[0]];

			for (const auto &resource : resources.sampled_images)
			{
				auto it = uniforms.find(resource.name);
				if (it == uniforms.end())
					continue;

				UniformInfo &u = it->second;

				uint32 texturebinding = msl.get_automatic_msl_resource_binding(resource.id);
				uint32 samplerbinding = msl.get_automatic_msl_resource_binding_secondary(resource.id);

				if (texturebinding == (uint32)-1)
					continue;

				for (int i = 0; i < u.count; i++)
				{
					if (u.ints[i] == -1)
					{
						u.ints[i] = (int)textureBindings.size();
						TextureBinding b = {};

						if (u.baseType == UNIFORM_TEXELBUFFER)
						{
							// TODO
						}
						else
						{
							b.texture = getMTLTexture(u.textures[i]);
							b.sampler = getMTLSampler(u.textures[i]);

							BuiltinUniform builtin = BUILTIN_MAX_ENUM;
							if (getConstant(u.name.c_str(), builtin) && builtin == BUILTIN_TEXTURE_MAIN)
								b.isMainTexture = true;
						}

						for (uint8 &stagebinding : b.texturestages)
							stagebinding = LOVE_UINT8_MAX;
						for (uint8 &stagebinding : b.samplerstages)
							stagebinding = LOVE_UINT8_MAX;

						textureBindings.push_back(b);
					}

					auto &b = textureBindings[u.ints[i]];
					b.texturestages[stageindex] = (uint8) texturebinding;
					b.samplerstages[stageindex] = (uint8) samplerbinding;
				}
			}
		}
		catch (std::exception &e)
		{
			printf("Error parsing SPIR-V shader source: %s\n", e.what());
		}
	}
}

Shader::~Shader()
{ @autoreleasepool {
	for (int i = 0; i < ShaderStage::STAGE_MAX_ENUM; i++)
		functions[i] = nil;

	for (const auto &kvp : cachedRenderPipelines)
		CFBridgingRelease(kvp.second);

	cachedRenderPipelines.clear();

	for (const auto &it : uniforms)
	{
		const auto &u = it.second;
		if (u.baseType == UNIFORM_SAMPLER)
		{
			free(u.data);
			for (int i = 0; i < u.count; i++)
			{
				if (u.textures[i] != nullptr)
					u.textures[i]->release();
			}
			delete[] u.textures;
		}
		else if (u.baseType == UNIFORM_TEXELBUFFER || u.baseType == UNIFORM_STORAGEBUFFER)
		{
			free(u.data);
			for (int i = 0; i < u.count; i++)
			{
				if (u.buffers[i] != nullptr)
					u.buffers[i]->release();
			}
			delete[] u.buffers;
		}
	}

	delete[] localUniformBufferData;
}}

void Shader::attach()
{
	if (current != this)
	{
		Graphics *gfx = Graphics::getInstance();
		gfx->flushBatchedDraws();
		gfx->attachShader(this);
		current = this;
	}
}

int Shader::getVertexAttributeIndex(const std::string &name)
{
	const auto it = attributes.find(name);
	return it != attributes.end() ? it->second : -1;
}

const Shader::UniformInfo *Shader::getUniformInfo(const std::string &name) const
{
	const auto it = uniforms.find(name);
	return it != uniforms.end() ? &(it->second) : nullptr;
}

const Shader::UniformInfo *Shader::getUniformInfo(BuiltinUniform builtin) const
{
	return builtinUniformInfo[(int)builtin];
}

void Shader::updateUniform(const UniformInfo * /*info*/, int /*count*/)
{
	// Nothing needed here, All uniform data will be memcpy'd to the main
	// uniform buffer before drawing.
	// FIXME: do we need to copy to a second buffer here, in order for batch
	// flushing to work (and possibly padding)?
}

void Shader::sendTextures(const UniformInfo *info, love::graphics::Texture **textures, int count)
{ @autoreleasepool {
	if (info->baseType != UNIFORM_SAMPLER)
		return;

	if (current == this)
		Graphics::flushBatchedDrawsGlobal();

	count = std::min(count, info->count);

	for (int i = 0; i < count; i++)
	{
		love::graphics::Texture *tex = textures[i];

		if (tex != nullptr)
		{
			if (!validateTexture(info, tex, false))
				continue;
		}
		else
		{
			auto gfx = Graphics::getInstance();
			tex = gfx->getDefaultTexture(info->textureType);
		}

		tex->retain();

		if (info->textures[i] != nullptr)
			info->textures[i]->release();

		info->textures[i] = tex;

		// TODO: handle changing a sampler after Shader:send(texture)...
		textureBindings[info->ints[i]].texture = getMTLTexture(tex);
		textureBindings[info->ints[i]].sampler = getMTLSampler(tex);
	}
}}

void Shader::sendBuffers(const UniformInfo *info, love::graphics::Buffer **buffers, int count)
{
	bool texelbinding = info->baseType == UNIFORM_TEXELBUFFER;
	bool storagebinding = info->baseType == UNIFORM_STORAGEBUFFER;

	if (!texelbinding && !storagebinding)
		return;

	if (current == this)
		Graphics::flushBatchedDrawsGlobal();

	count = std::min(count, info->count);

	// Bind the textures to the texture units.
	for (int i = 0; i < count; i++)
	{
		love::graphics::Buffer *buffer = buffers[i];

		if (buffer != nullptr)
		{
			if (!validateBuffer(info, buffer, false))
				continue;
			buffer->retain();
		}

		if (info->buffers[i] != nullptr)
			info->buffers[i]->release();

		info->buffers[i] = buffer;

		if (texelbinding)
		{
			textureBindings[info->ints[i]].texture = getMTLTexture(buffer);
		}
		else if (storagebinding)
		{
			// TODO
		}
	}
}

void Shader::setVideoTextures(love::graphics::Texture *ytexture, love::graphics::Texture *cbtexture, love::graphics::Texture *crtexture)
{
	// TODO
}

bool Shader::hasUniform(const std::string &name) const
{
	return uniforms.find(name) != uniforms.end();
}

id<MTLRenderPipelineState> Shader::getCachedRenderPipeline(const RenderPipelineKey &key)
{
	auto it = cachedRenderPipelines.find(key);

	if (it != cachedRenderPipelines.end())
		return (__bridge id<MTLRenderPipelineState>) it->second;

	id<MTLDevice> device = Graphics::getInstance()->device;

	MTLRenderPipelineDescriptor *desc = [MTLRenderPipelineDescriptor new];

	desc.vertexFunction = functions[ShaderStage::STAGE_VERTEX];
	desc.fragmentFunction = functions[ShaderStage::STAGE_PIXEL];

	desc.rasterSampleCount = std::max((int) key.msaa, 1);

	for (int i = 0; i < MAX_COLOR_RENDER_TARGETS; i++)
	{
		PixelFormat format = (PixelFormat)((key.colorRenderTargetFormats >> (i * 8)) & 0xFF);
		if (format == PIXELFORMAT_UNKNOWN)
			continue;

		MTLRenderPipelineColorAttachmentDescriptor *attachment = desc.colorAttachments[i];

		bool isSRGB = false;
		auto formatdesc = Metal::convertPixelFormat(format, isSRGB);
		attachment.pixelFormat = formatdesc.format;

		if (key.blend.enable)
		{
			attachment.blendingEnabled = YES;
			attachment.sourceRGBBlendFactor = getMTLBlendFactor(key.blend.srcFactorRGB);
			attachment.destinationRGBBlendFactor = getMTLBlendFactor(key.blend.dstFactorRGB);
			attachment.rgbBlendOperation = getMTLBlendOperation(key.blend.operationRGB);
			attachment.sourceAlphaBlendFactor = getMTLBlendFactor(key.blend.srcFactorA);
			attachment.destinationAlphaBlendFactor = getMTLBlendFactor(key.blend.dstFactorA);
			attachment.alphaBlendOperation = getMTLBlendOperation(key.blend.operationA);
		}

		MTLColorWriteMask writeMask = MTLColorWriteMaskNone;
		if (key.colorChannelMask.r)
			writeMask |= MTLColorWriteMaskRed;
		if (key.colorChannelMask.g)
			writeMask |= MTLColorWriteMaskGreen;
		if (key.colorChannelMask.b)
			writeMask |= MTLColorWriteMaskBlue;
		if (key.colorChannelMask.a)
			writeMask |= MTLColorWriteMaskAlpha;

		attachment.writeMask = writeMask;

		desc.colorAttachments[i] = attachment;
	}

	// TODO: depth/stencil attachment formats
	

	{
		MTLVertexDescriptor *vertdesc = [MTLVertexDescriptor vertexDescriptor];
		const auto &attributes = key.vertexAttributes;

		for (const auto &pair : this->attributes)
		{
			int i = pair.second;
			uint32 bit = 1u << i;

			if (attributes.enableBits & bit)
			{
				const auto &attrib = attributes.attribs[i];
				int metalBufferIndex = attrib.bufferIndex + VERTEX_BUFFER_BINDING_START;

				vertdesc.attributes[i].format = getMTLVertexFormat(attrib.format);
				vertdesc.attributes[i].offset = attrib.offsetFromVertex;
				vertdesc.attributes[i].bufferIndex = metalBufferIndex;

				const auto &layout = attributes.bufferLayouts[attrib.bufferIndex];

				bool instanced = attributes.instanceBits & (1u << attrib.bufferIndex);
				auto step = instanced ? MTLVertexStepFunctionPerInstance : MTLVertexStepFunctionPerVertex;

				vertdesc.layouts[metalBufferIndex].stride = layout.stride;
				vertdesc.layouts[metalBufferIndex].stepFunction = step;
			}
			else
			{
				vertdesc.attributes[i].format = MTLVertexFormatFloat4;
				vertdesc.attributes[i].offset = 0;
				vertdesc.attributes[i].bufferIndex = DEFAULT_VERTEX_BUFFER_BINDING;

				vertdesc.layouts[DEFAULT_VERTEX_BUFFER_BINDING].stride = sizeof(float) * 4;
				vertdesc.layouts[DEFAULT_VERTEX_BUFFER_BINDING].stepFunction = MTLVertexStepFunctionConstant;
				vertdesc.layouts[DEFAULT_VERTEX_BUFFER_BINDING].stepRate = 0;
			}
		}

		desc.vertexDescriptor = vertdesc;
	}

	NSError *err = nil;
	id<MTLRenderPipelineState> pipeline = [device newRenderPipelineStateWithDescriptor:desc error:&err];

	if (err != nil)
	{
		NSLog(@"Error creating render pipeline: %@", err);
		return nil;
	}

	cachedRenderPipelines[key] = CFBridgingRetain(pipeline);

	return pipeline;
}

int Shader::getUniformBufferBinding()
{
	return spirv_cross::ResourceBindingPushConstantBinding;
}

} // metal
} // graphics
} // love