/*

  Simple DirectMedia Layer

  Copyright (C) 1997-2018 Sam Lantinga <slouken@libsdl.org>

  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 "../../SDL_internal.h"

#if SDL_VIDEO_RENDER_METAL && !SDL_RENDER_DISABLED

#include "SDL_hints.h"

#include "SDL_log.h"

#include "SDL_assert.h"

#include "SDL_syswm.h"

#include "../SDL_sysrender.h"

#ifdef __MACOSX__

#include "../../video/cocoa/SDL_cocoametalview.h"

#else

#include "../../video/uikit/SDL_uikitmetalview.h"

#endif

#include <Availability.h>

#import <Metal/Metal.h>

#import <QuartzCore/CAMetalLayer.h>

/* Regenerate these with build-metal-shaders.sh */

#ifdef __MACOSX__

#include "SDL_shaders_metal_osx.h"

#else

#include "SDL_shaders_metal_ios.h"

#endif

/* Apple Metal renderer implementation */

/* macOS requires constants in a buffer to have a 256 byte alignment. */

#ifdef __MACOSX__

#define CONSTANT_ALIGN 256

#else

#define CONSTANT_ALIGN 4

#endif

#define ALIGN_CONSTANTS(size) ((size + CONSTANT_ALIGN - 1) & (~(CONSTANT_ALIGN - 1)))

static const size_t CONSTANTS_OFFSET_INVALID = 0xFFFFFFFF;

static const size_t CONSTANTS_OFFSET_IDENTITY = 0;

static const size_t CONSTANTS_OFFSET_HALF_PIXEL_TRANSFORM = ALIGN_CONSTANTS(CONSTANTS_OFFSET_IDENTITY + sizeof(float) * 16);

static const size_t CONSTANTS_OFFSET_DECODE_JPEG = ALIGN_CONSTANTS(CONSTANTS_OFFSET_HALF_PIXEL_TRANSFORM + sizeof(float) * 16);

static const size_t CONSTANTS_OFFSET_DECODE_BT601 = ALIGN_CONSTANTS(CONSTANTS_OFFSET_DECODE_JPEG + sizeof(float) * 4 * 4);

static const size_t CONSTANTS_OFFSET_DECODE_BT709 = ALIGN_CONSTANTS(CONSTANTS_OFFSET_DECODE_BT601 + sizeof(float) * 4 * 4);

static const size_t CONSTANTS_LENGTH = CONSTANTS_OFFSET_DECODE_BT709 + sizeof(float) * 4 * 4;

typedef enum SDL_MetalVertexFunction

{

    SDL_METAL_VERTEX_SOLID,

    SDL_METAL_VERTEX_COPY,

} SDL_MetalVertexFunction;

typedef enum SDL_MetalFragmentFunction

{

    SDL_METAL_FRAGMENT_SOLID = 0,

    SDL_METAL_FRAGMENT_COPY,

    SDL_METAL_FRAGMENT_YUV,

    SDL_METAL_FRAGMENT_NV12,

    SDL_METAL_FRAGMENT_NV21,

    SDL_METAL_FRAGMENT_COUNT,

} SDL_MetalFragmentFunction;

typedef struct METAL_PipelineState

{

    SDL_BlendMode blendMode;

    void *pipe;

} METAL_PipelineState;

typedef struct METAL_PipelineCache

{

    METAL_PipelineState *states;

    int count;

    SDL_MetalVertexFunction vertexFunction;

    SDL_MetalFragmentFunction fragmentFunction;

    MTLPixelFormat renderTargetFormat;

    const char *label;

} METAL_PipelineCache;

/* Each shader combination used by drawing functions has a separate pipeline

 * cache, and we have a separate list of caches for each render target pixel

 * format. This is more efficient than iterating over a global cache to find

 * the pipeline based on the specified shader combination and RT pixel format,

 * since we know what the RT pixel format is when we set the render target, and

 * we know what the shader combination is inside each drawing function's code. */

typedef struct METAL_ShaderPipelines

{

    MTLPixelFormat renderTargetFormat;

    METAL_PipelineCache caches[SDL_METAL_FRAGMENT_COUNT];

} METAL_ShaderPipelines;

@interface METAL_RenderData : NSObject

    @property (nonatomic, retain) id<MTLDevice> mtldevice;

    @property (nonatomic, retain) id<MTLCommandQueue> mtlcmdqueue;

    @property (nonatomic, retain) id<MTLCommandBuffer> mtlcmdbuffer;

    @property (nonatomic, retain) id<MTLRenderCommandEncoder> mtlcmdencoder;

    @property (nonatomic, retain) id<MTLLibrary> mtllibrary;

    @property (nonatomic, retain) id<CAMetalDrawable> mtlbackbuffer;

    @property (nonatomic, retain) id<MTLSamplerState> mtlsamplernearest;

    @property (nonatomic, retain) id<MTLSamplerState> mtlsamplerlinear;

    @property (nonatomic, retain) id<MTLBuffer> mtlbufconstants;

    @property (nonatomic, retain) CAMetalLayer *mtllayer;

    @property (nonatomic, retain) MTLRenderPassDescriptor *mtlpassdesc;

    @property (nonatomic, assign) METAL_ShaderPipelines *activepipelines;

    @property (nonatomic, assign) METAL_ShaderPipelines *allpipelines;

    @property (nonatomic, assign) int pipelinescount;

@end

@implementation METAL_RenderData

#if !__has_feature(objc_arc)

- (void)dealloc

{

    [_mtldevice release];

    [_mtlcmdqueue release];

    [_mtlcmdbuffer release];

    [_mtlcmdencoder release];

    [_mtllibrary release];

    [_mtlbackbuffer release];

    [_mtlsamplernearest release];

    [_mtlsamplerlinear release];

    [_mtlbufconstants release];

    [_mtllayer release];

    [_mtlpassdesc release];

    [super dealloc];

}

#endif

@end

@interface METAL_TextureData : NSObject

    @property (nonatomic, retain) id<MTLTexture> mtltexture;

    @property (nonatomic, retain) id<MTLTexture> mtltexture_uv;

    @property (nonatomic, retain) id<MTLSamplerState> mtlsampler;

    @property (nonatomic, assign) SDL_MetalFragmentFunction fragmentFunction;

    @property (nonatomic, assign) BOOL yuv;

    @property (nonatomic, assign) BOOL nv12;

    @property (nonatomic, assign) size_t conversionBufferOffset;

@end

@implementation METAL_TextureData

#if !__has_feature(objc_arc)

- (void)dealloc

{

    [_mtltexture release];

    [_mtltexture_uv release];

    [_mtlsampler release];

    [super dealloc];

}

#endif

@end

static int

IsMetalAvailable(const SDL_SysWMinfo *syswm)

{

    if (syswm->subsystem != SDL_SYSWM_COCOA && syswm->subsystem != SDL_SYSWM_UIKIT) {

        return SDL_SetError("Metal render target only supports Cocoa and UIKit video targets at the moment.");

    }

    // this checks a weak symbol.

#if (defined(__MACOSX__) && (MAC_OS_X_VERSION_MIN_REQUIRED < 101100))

    if (MTLCreateSystemDefaultDevice == NULL) {  // probably on 10.10 or lower.

        return SDL_SetError("Metal framework not available on this system");

    }

#endif

    return 0;

}

static const MTLBlendOperation invalidBlendOperation = (MTLBlendOperation)0xFFFFFFFF;

static const MTLBlendFactor invalidBlendFactor = (MTLBlendFactor)0xFFFFFFFF;

static MTLBlendOperation

GetBlendOperation(SDL_BlendOperation operation)

{

    switch (operation) {

        case SDL_BLENDOPERATION_ADD: return MTLBlendOperationAdd;

        case SDL_BLENDOPERATION_SUBTRACT: return MTLBlendOperationSubtract;

        case SDL_BLENDOPERATION_REV_SUBTRACT: return MTLBlendOperationReverseSubtract;

        case SDL_BLENDOPERATION_MINIMUM: return MTLBlendOperationMin;

        case SDL_BLENDOPERATION_MAXIMUM: return MTLBlendOperationMax;

        default: return invalidBlendOperation;

    }

}

static MTLBlendFactor

GetBlendFactor(SDL_BlendFactor factor)

{

    switch (factor) {

        case SDL_BLENDFACTOR_ZERO: return MTLBlendFactorZero;

        case SDL_BLENDFACTOR_ONE: return MTLBlendFactorOne;

        case SDL_BLENDFACTOR_SRC_COLOR: return MTLBlendFactorSourceColor;

        case SDL_BLENDFACTOR_ONE_MINUS_SRC_COLOR: return MTLBlendFactorOneMinusSourceColor;

        case SDL_BLENDFACTOR_SRC_ALPHA: return MTLBlendFactorSourceAlpha;

        case SDL_BLENDFACTOR_ONE_MINUS_SRC_ALPHA: return MTLBlendFactorOneMinusSourceAlpha;

        case SDL_BLENDFACTOR_DST_COLOR: return MTLBlendFactorDestinationColor;

        case SDL_BLENDFACTOR_ONE_MINUS_DST_COLOR: return MTLBlendFactorOneMinusDestinationColor;

        case SDL_BLENDFACTOR_DST_ALPHA: return MTLBlendFactorDestinationAlpha;

        case SDL_BLENDFACTOR_ONE_MINUS_DST_ALPHA: return MTLBlendFactorOneMinusDestinationAlpha;

        default: return invalidBlendFactor;

    }

}

static NSString *

GetVertexFunctionName(SDL_MetalVertexFunction function)

{

    switch (function) {

        case SDL_METAL_VERTEX_SOLID: return @"SDL_Solid_vertex";

        case SDL_METAL_VERTEX_COPY: return @"SDL_Copy_vertex";

        default: return nil;

    }

}

static NSString *

GetFragmentFunctionName(SDL_MetalFragmentFunction function)

{

    switch (function) {

        case SDL_METAL_FRAGMENT_SOLID: return @"SDL_Solid_fragment";

        case SDL_METAL_FRAGMENT_COPY: return @"SDL_Copy_fragment";

        case SDL_METAL_FRAGMENT_YUV: return @"SDL_YUV_fragment";

        case SDL_METAL_FRAGMENT_NV12: return @"SDL_NV12_fragment";

        case SDL_METAL_FRAGMENT_NV21: return @"SDL_NV21_fragment";

        default: return nil;

    }

}

static id<MTLRenderPipelineState>

MakePipelineState(METAL_RenderData *data, METAL_PipelineCache *cache,

                  NSString *blendlabel, SDL_BlendMode blendmode)

{

    id<MTLFunction> mtlvertfn = [data.mtllibrary newFunctionWithName:GetVertexFunctionName(cache->vertexFunction)];

    id<MTLFunction> mtlfragfn = [data.mtllibrary newFunctionWithName:GetFragmentFunctionName(cache->fragmentFunction)];

    SDL_assert(mtlvertfn != nil);

    SDL_assert(mtlfragfn != nil);

    MTLRenderPipelineDescriptor *mtlpipedesc = [[MTLRenderPipelineDescriptor alloc] init];

    mtlpipedesc.vertexFunction = mtlvertfn;

    mtlpipedesc.fragmentFunction = mtlfragfn;

    MTLRenderPipelineColorAttachmentDescriptor *rtdesc = mtlpipedesc.colorAttachments[0];

    rtdesc.pixelFormat = cache->renderTargetFormat;

    if (blendmode != SDL_BLENDMODE_NONE) {

        rtdesc.blendingEnabled = YES;

        rtdesc.sourceRGBBlendFactor = GetBlendFactor(SDL_GetBlendModeSrcColorFactor(blendmode));

        rtdesc.destinationRGBBlendFactor = GetBlendFactor(SDL_GetBlendModeDstColorFactor(blendmode));

        rtdesc.rgbBlendOperation = GetBlendOperation(SDL_GetBlendModeColorOperation(blendmode));

        rtdesc.sourceAlphaBlendFactor = GetBlendFactor(SDL_GetBlendModeSrcAlphaFactor(blendmode));

        rtdesc.destinationAlphaBlendFactor = GetBlendFactor(SDL_GetBlendModeDstAlphaFactor(blendmode));

        rtdesc.alphaBlendOperation = GetBlendOperation(SDL_GetBlendModeAlphaOperation(blendmode));

    } else {

        rtdesc.blendingEnabled = NO;

    }

    mtlpipedesc.label = [@(cache->label) stringByAppendingString:blendlabel];

    NSError *err = nil;

    id<MTLRenderPipelineState> state = [data.mtldevice newRenderPipelineStateWithDescriptor:mtlpipedesc error:&err];

    SDL_assert(err == nil);

    METAL_PipelineState pipeline;

    pipeline.blendMode = blendmode;

    pipeline.pipe = (void *)CFBridgingRetain(state);

    METAL_PipelineState *states = SDL_realloc(cache->states, (cache->count + 1) * sizeof(pipeline));

#if !__has_feature(objc_arc)

    [mtlpipedesc release];  // !!! FIXME: can these be reused for each creation, or does the pipeline obtain it?

    [mtlvertfn release];

    [mtlfragfn release];

    [state release];

#endif

    if (states) {

        states[cache->count++] = pipeline;

        cache->states = states;

        return (__bridge id<MTLRenderPipelineState>)pipeline.pipe;

    } else {

        CFBridgingRelease(pipeline.pipe);

        SDL_OutOfMemory();

        return NULL;

    }

}

static void

MakePipelineCache(METAL_RenderData *data, METAL_PipelineCache *cache, const char *label,

                  MTLPixelFormat rtformat, SDL_MetalVertexFunction vertfn, SDL_MetalFragmentFunction fragfn)

{

    SDL_zerop(cache);

    cache->vertexFunction = vertfn;

    cache->fragmentFunction = fragfn;

    cache->renderTargetFormat = rtformat;

    cache->label = label;

    /* Create pipeline states for the default blend modes. Custom blend modes

     * will be added to the cache on-demand. */

    MakePipelineState(data, cache, @" (blend=none)", SDL_BLENDMODE_NONE);

    MakePipelineState(data, cache, @" (blend=blend)", SDL_BLENDMODE_BLEND);

    MakePipelineState(data, cache, @" (blend=add)", SDL_BLENDMODE_ADD);

    MakePipelineState(data, cache, @" (blend=mod)", SDL_BLENDMODE_MOD);

}

static void

DestroyPipelineCache(METAL_PipelineCache *cache)

{

    if (cache != NULL) {

        for (int i = 0; i < cache->count; i++) {

            CFBridgingRelease(cache->states[i].pipe);

        }

        SDL_free(cache->states);

    }

}

void

MakeShaderPipelines(METAL_RenderData *data, METAL_ShaderPipelines *pipelines, MTLPixelFormat rtformat)

{

    SDL_zerop(pipelines);

    pipelines->renderTargetFormat = rtformat;

    MakePipelineCache(data, &pipelines->caches[SDL_METAL_FRAGMENT_SOLID], "SDL primitives pipeline", rtformat, SDL_METAL_VERTEX_SOLID, SDL_METAL_FRAGMENT_SOLID);

    MakePipelineCache(data, &pipelines->caches[SDL_METAL_FRAGMENT_COPY], "SDL copy pipeline", rtformat, SDL_METAL_VERTEX_COPY, SDL_METAL_FRAGMENT_COPY);

    MakePipelineCache(data, &pipelines->caches[SDL_METAL_FRAGMENT_YUV], "SDL YUV pipeline", rtformat, SDL_METAL_VERTEX_COPY, SDL_METAL_FRAGMENT_YUV);

    MakePipelineCache(data, &pipelines->caches[SDL_METAL_FRAGMENT_NV12], "SDL NV12 pipeline", rtformat, SDL_METAL_VERTEX_COPY, SDL_METAL_FRAGMENT_NV12);

    MakePipelineCache(data, &pipelines->caches[SDL_METAL_FRAGMENT_NV21], "SDL NV21 pipeline", rtformat, SDL_METAL_VERTEX_COPY, SDL_METAL_FRAGMENT_NV21);

}

static METAL_ShaderPipelines *

ChooseShaderPipelines(METAL_RenderData *data, MTLPixelFormat rtformat)

{

    METAL_ShaderPipelines *allpipelines = data.allpipelines;

    int count = data.pipelinescount;

    for (int i = 0; i < count; i++) {

        if (allpipelines[i].renderTargetFormat == rtformat) {

            return &allpipelines[i];

        }

    }

    allpipelines = SDL_realloc(allpipelines, (count + 1) * sizeof(METAL_ShaderPipelines));

    if (allpipelines == NULL) {

        SDL_OutOfMemory();

        return NULL;

    }

    MakeShaderPipelines(data, &allpipelines[count], rtformat);

    data.allpipelines = allpipelines;

    data.pipelinescount = count + 1;

    return &data.allpipelines[count];

}

static void

DestroyAllPipelines(METAL_ShaderPipelines *allpipelines, int count)

{

    if (allpipelines != NULL) {

        for (int i = 0; i < count; i++) {

            for (int cache = 0; cache < SDL_METAL_FRAGMENT_COUNT; cache++) {

                DestroyPipelineCache(&allpipelines[i].caches[cache]);

            }

        }

        SDL_free(allpipelines);

    }

}

static inline id<MTLRenderPipelineState>

ChoosePipelineState(METAL_RenderData *data, METAL_ShaderPipelines *pipelines, SDL_MetalFragmentFunction fragfn, SDL_BlendMode blendmode)

{

    METAL_PipelineCache *cache = &pipelines->caches[fragfn];

    for (int i = 0; i < cache->count; i++) {

        if (cache->states[i].blendMode == blendmode) {

            return (__bridge id<MTLRenderPipelineState>)cache->states[i].pipe;

        }

    }

    return MakePipelineState(data, cache, [NSString stringWithFormat:@" (blend=custom 0x%x)", blendmode], blendmode);

}

static void

METAL_ActivateRenderCommandEncoder(SDL_Renderer * renderer, MTLLoadAction load, MTLClearColor *clear_color)

{

    METAL_RenderData *data = (__bridge METAL_RenderData *) renderer->driverdata;

    /* Our SetRenderTarget just signals that the next render operation should

     * set up a new render pass. This is where that work happens. */

    if (data.mtlcmdencoder == nil) {

        id<MTLTexture> mtltexture = nil;

        if (renderer->target != NULL) {

            METAL_TextureData *texdata = (__bridge METAL_TextureData *)renderer->target->driverdata;

            mtltexture = texdata.mtltexture;

        } else {

            if (data.mtlbackbuffer == nil) {

                /* The backbuffer's contents aren't guaranteed to persist after

                 * presenting, so we can leave it undefined when loading it. */

                data.mtlbackbuffer = [data.mtllayer nextDrawable];

                if (load == MTLLoadActionLoad) {

                    load = MTLLoadActionDontCare;

                }

            }

            mtltexture = data.mtlbackbuffer.texture;

        }

        SDL_assert(mtltexture);

        if (load == MTLLoadActionClear) {

            SDL_assert(clear_color != NULL);

            data.mtlpassdesc.colorAttachments[0].clearColor = *clear_color;

        }

        data.mtlpassdesc.colorAttachments[0].loadAction = load;

        data.mtlpassdesc.colorAttachments[0].texture = mtltexture;

        data.mtlcmdbuffer = [data.mtlcmdqueue commandBuffer];

        data.mtlcmdencoder = [data.mtlcmdbuffer renderCommandEncoderWithDescriptor:data.mtlpassdesc];

        if (data.mtlbackbuffer != nil && mtltexture == data.mtlbackbuffer.texture) {

            data.mtlcmdencoder.label = @"SDL metal renderer backbuffer";

        } else {

            data.mtlcmdencoder.label = @"SDL metal renderer render target";

        }

        data.activepipelines = ChooseShaderPipelines(data, mtltexture.pixelFormat);

        // make sure this has a definite place in the queue. This way it will

        //  execute reliably whether the app tries to make its own command buffers

        //  or whatever. This means we can _always_ batch rendering commands!

        [data.mtlcmdbuffer enqueue];

    }

}

static void

METAL_WindowEvent(SDL_Renderer * renderer, const SDL_WindowEvent *event)

{

    if (event->event == SDL_WINDOWEVENT_SHOWN ||

        event->event == SDL_WINDOWEVENT_HIDDEN) {

        // !!! FIXME: write me

    }

}

static int

METAL_GetOutputSize(SDL_Renderer * renderer, int *w, int *h)

{ @autoreleasepool {

    METAL_RenderData *data = (__bridge METAL_RenderData *) renderer->driverdata;

    if (w) {

        *w = (int)data.mtllayer.drawableSize.width;

    }

    if (h) {

        *h = (int)data.mtllayer.drawableSize.height;

    }

    return 0;

}}

static SDL_bool

METAL_SupportsBlendMode(SDL_Renderer * renderer, SDL_BlendMode blendMode)

{

    SDL_BlendFactor srcColorFactor = SDL_GetBlendModeSrcColorFactor(blendMode);

    SDL_BlendFactor srcAlphaFactor = SDL_GetBlendModeSrcAlphaFactor(blendMode);

    SDL_BlendOperation colorOperation = SDL_GetBlendModeColorOperation(blendMode);

    SDL_BlendFactor dstColorFactor = SDL_GetBlendModeDstColorFactor(blendMode);

    SDL_BlendFactor dstAlphaFactor = SDL_GetBlendModeDstAlphaFactor(blendMode);

    SDL_BlendOperation alphaOperation = SDL_GetBlendModeAlphaOperation(blendMode);

    if (GetBlendFactor(srcColorFactor) == invalidBlendFactor ||

        GetBlendFactor(srcAlphaFactor) == invalidBlendFactor ||

        GetBlendOperation(colorOperation) == invalidBlendOperation ||

        GetBlendFactor(dstColorFactor) == invalidBlendFactor ||

        GetBlendFactor(dstAlphaFactor) == invalidBlendFactor ||

        GetBlendOperation(alphaOperation) == invalidBlendOperation) {

        return SDL_FALSE;

    }

    return SDL_TRUE;

}

static int

METAL_CreateTexture(SDL_Renderer * renderer, SDL_Texture * texture)

{ @autoreleasepool {

    METAL_RenderData *data = (__bridge METAL_RenderData *) renderer->driverdata;

    MTLPixelFormat pixfmt;

    switch (texture->format) {

        case SDL_PIXELFORMAT_ABGR8888:

            pixfmt = MTLPixelFormatRGBA8Unorm;

            break;

        case SDL_PIXELFORMAT_ARGB8888:

            pixfmt = MTLPixelFormatBGRA8Unorm;

            break;

        case SDL_PIXELFORMAT_IYUV:

        case SDL_PIXELFORMAT_YV12:

        case SDL_PIXELFORMAT_NV12:

        case SDL_PIXELFORMAT_NV21:

            pixfmt = MTLPixelFormatR8Unorm;

            break;

        default:

            return SDL_SetError("Texture format %s not supported by Metal", SDL_GetPixelFormatName(texture->format));

    }

    MTLTextureDescriptor *mtltexdesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:pixfmt

                                            width:(NSUInteger)texture->w height:(NSUInteger)texture->h mipmapped:NO];

    /* Not available in iOS 8. */

    if ([mtltexdesc respondsToSelector:@selector(usage)]) {

        if (texture->access == SDL_TEXTUREACCESS_TARGET) {

            mtltexdesc.usage = MTLTextureUsageShaderRead | MTLTextureUsageRenderTarget;

        } else {

            mtltexdesc.usage = MTLTextureUsageShaderRead;

        }

    }

    id<MTLTexture> mtltexture = [data.mtldevice newTextureWithDescriptor:mtltexdesc];

    if (mtltexture == nil) {

        return SDL_SetError("Texture allocation failed");

    }

    id<MTLTexture> mtltexture_uv = nil;

    BOOL yuv = (texture->format == SDL_PIXELFORMAT_IYUV) || (texture->format == SDL_PIXELFORMAT_YV12);

    BOOL nv12 = (texture->format == SDL_PIXELFORMAT_NV12) || (texture->format == SDL_PIXELFORMAT_NV21);

    if (yuv) {

        mtltexdesc.pixelFormat = MTLPixelFormatR8Unorm;

        mtltexdesc.width = (texture->w + 1) / 2;

        mtltexdesc.height = (texture->h + 1) / 2;

        mtltexdesc.textureType = MTLTextureType2DArray;

        mtltexdesc.arrayLength = 2;

    } else if (nv12) {

        mtltexdesc.pixelFormat = MTLPixelFormatRG8Unorm;

        mtltexdesc.width = (texture->w + 1) / 2;

        mtltexdesc.height = (texture->h + 1) / 2;

    }

    if (yuv || nv12) {

        mtltexture_uv = [data.mtldevice newTextureWithDescriptor:mtltexdesc];

        if (mtltexture_uv == nil) {

#if !__has_feature(objc_arc)

            [mtltexture release];

#endif

            return SDL_SetError("Texture allocation failed");

        }

    }

    METAL_TextureData *texturedata = [[METAL_TextureData alloc] init];

    if (texture->scaleMode == SDL_ScaleModeNearest) {

        texturedata.mtlsampler = data.mtlsamplernearest;

    } else {

        texturedata.mtlsampler = data.mtlsamplerlinear;

    }

    texturedata.mtltexture = mtltexture;

    texturedata.mtltexture_uv = mtltexture_uv;

    texturedata.yuv = yuv;

    texturedata.nv12 = nv12;

    if (yuv) {

        texturedata.fragmentFunction = SDL_METAL_FRAGMENT_YUV;

    } else if (texture->format == SDL_PIXELFORMAT_NV12) {

        texturedata.fragmentFunction = SDL_METAL_FRAGMENT_NV12;

    } else if (texture->format == SDL_PIXELFORMAT_NV21) {

        texturedata.fragmentFunction = SDL_METAL_FRAGMENT_NV21;

    } else {

        texturedata.fragmentFunction = SDL_METAL_FRAGMENT_COPY;

    }

    if (yuv || nv12) {

        size_t offset = 0;

        SDL_YUV_CONVERSION_MODE mode = SDL_GetYUVConversionModeForResolution(texture->w, texture->h);

        switch (mode) {

            case SDL_YUV_CONVERSION_JPEG: offset = CONSTANTS_OFFSET_DECODE_JPEG; break;

            case SDL_YUV_CONVERSION_BT601: offset = CONSTANTS_OFFSET_DECODE_BT601; break;

            case SDL_YUV_CONVERSION_BT709: offset = CONSTANTS_OFFSET_DECODE_BT709; break;

            default: offset = 0; break;

        }

        texturedata.conversionBufferOffset = offset;

    }

    texture->driverdata = (void*)CFBridgingRetain(texturedata);

#if !__has_feature(objc_arc)

    [texturedata release];

    [mtltexture release];

    [mtltexture_uv release];

#endif

    return 0;

}}

static int

METAL_UpdateTexture(SDL_Renderer * renderer, SDL_Texture * texture,

                 const SDL_Rect * rect, const void *pixels, int pitch)

{ @autoreleasepool {

    METAL_TextureData *texturedata = (__bridge METAL_TextureData *)texture->driverdata;

    /* !!! FIXME: replaceRegion does not do any synchronization, so it might

     * !!! FIXME: stomp on a previous frame's data that's currently being read

     * !!! FIXME: by the GPU. */

    [texturedata.mtltexture replaceRegion:MTLRegionMake2D(rect->x, rect->y, rect->w, rect->h)

                              mipmapLevel:0

                                withBytes:pixels

                              bytesPerRow:pitch];

    if (texturedata.yuv) {

        int Uslice = texture->format == SDL_PIXELFORMAT_YV12 ? 1 : 0;

        int Vslice = texture->format == SDL_PIXELFORMAT_YV12 ? 0 : 1;

        /* Skip to the correct offset into the next texture */

        pixels = (const void*)((const Uint8*)pixels + rect->h * pitch);

        [texturedata.mtltexture_uv replaceRegion:MTLRegionMake2D(rect->x / 2, rect->y / 2, (rect->w + 1) / 2, (rect->h + 1) / 2)

                                     mipmapLevel:0

                                           slice:Uslice

                                       withBytes:pixels

                                     bytesPerRow:(pitch + 1) / 2

                                   bytesPerImage:0];

        /* Skip to the correct offset into the next texture */

        pixels = (const void*)((const Uint8*)pixels + ((rect->h + 1) / 2) * ((pitch + 1)/2));

        [texturedata.mtltexture_uv replaceRegion:MTLRegionMake2D(rect->x / 2, rect->y / 2, (rect->w + 1) / 2, (rect->h + 1) / 2)

                                     mipmapLevel:0

                                           slice:Vslice

                                       withBytes:pixels

                                     bytesPerRow:(pitch + 1) / 2

                                   bytesPerImage:0];

    }

    if (texturedata.nv12) {

        /* Skip to the correct offset into the next texture */

        pixels = (const void*)((const Uint8*)pixels + rect->h * pitch);

        [texturedata.mtltexture_uv replaceRegion:MTLRegionMake2D(rect->x / 2, rect->y / 2, (rect->w + 1) / 2, (rect->h + 1) / 2)

                                     mipmapLevel:0

                                           slice:0

                                       withBytes:pixels

                                     bytesPerRow:2 * ((pitch + 1) / 2)

                                   bytesPerImage:0];

    }

    return 0;

}}

static int

METAL_UpdateTextureYUV(SDL_Renderer * renderer, SDL_Texture * texture,

                    const SDL_Rect * rect,

                    const Uint8 *Yplane, int Ypitch,

                    const Uint8 *Uplane, int Upitch,

                    const Uint8 *Vplane, int Vpitch)

{ @autoreleasepool {

    METAL_TextureData *texturedata = (__bridge METAL_TextureData *)texture->driverdata;

    const int Uslice = 0;

    const int Vslice = 1;

    /* Bail out if we're supposed to update an empty rectangle */

    if (rect->w <= 0 || rect->h <= 0) {

        return 0;

    }

    [texturedata.mtltexture replaceRegion:MTLRegionMake2D(rect->x, rect->y, rect->w, rect->h)

                              mipmapLevel:0

                                withBytes:Yplane

                              bytesPerRow:Ypitch];

    [texturedata.mtltexture_uv replaceRegion:MTLRegionMake2D(rect->x / 2, rect->y / 2, (rect->w + 1) / 2, (rect->h + 1) / 2)

                                 mipmapLevel:0

                                       slice:Uslice

                                   withBytes:Uplane

                                 bytesPerRow:Upitch

                               bytesPerImage:0];

    [texturedata.mtltexture_uv replaceRegion:MTLRegionMake2D(rect->x / 2, rect->y / 2, (rect->w + 1) / 2, (rect->h + 1) / 2)

                                 mipmapLevel:0

                                       slice:Vslice

                                   withBytes:Vplane

                                 bytesPerRow:Vpitch

                               bytesPerImage:0];

    return 0;

}}

static int

METAL_LockTexture(SDL_Renderer * renderer, SDL_Texture * texture,

               const SDL_Rect * rect, void **pixels, int *pitch)

{

    return SDL_Unsupported();   // !!! FIXME: write me

}

static void

METAL_UnlockTexture(SDL_Renderer * renderer, SDL_Texture * texture)

{

    // !!! FIXME: write me

}

static int

METAL_SetRenderTarget(SDL_Renderer * renderer, SDL_Texture * texture)

{ @autoreleasepool {

    METAL_RenderData *data = (__bridge METAL_RenderData *) renderer->driverdata;

    if (data.mtlcmdencoder) {

        /* End encoding for the previous render target so we can set up a new

         * render pass for this one. */

        [data.mtlcmdencoder endEncoding];

        [data.mtlcmdbuffer commit];

        data.mtlcmdencoder = nil;

        data.mtlcmdbuffer = nil;

    }

    /* We don't begin a new render pass right away - we delay it until an actual

     * draw or clear happens. That way we can use hardware clears when possible,

     * which are only available when beginning a new render pass. */

    return 0;

}}

// normalize a value from 0.0f to len into 0.0f to 1.0f.

static inline float

normtex(const float _val, const float len)

{

    return _val / len;

}

static int

METAL_QueueSetViewport(SDL_Renderer * renderer, SDL_RenderCommand *cmd)

{

    float projection[4][4];    /* Prepare an orthographic projection */

    const int w = cmd->data.viewport.rect.w;

    const int h = cmd->data.viewport.rect.h;

    const size_t matrixlen = sizeof (projection);

    float *matrix = (float *) SDL_AllocateRenderVertices(renderer, matrixlen, CONSTANT_ALIGN, &cmd->data.viewport.first);

    if (!matrix) {

        return -1;

    }

    SDL_memset(projection, '\0', matrixlen);

    if (w && h) {

        projection[0][0] = 2.0f / w;

        projection[1][1] = -2.0f / h;

        projection[3][0] = -1.0f;

        projection[3][1] = 1.0f;

        projection[3][3] = 1.0f;

    }

    SDL_memcpy(matrix, projection, matrixlen);

    return 0;

}

static int

METAL_QueueSetDrawColor(SDL_Renderer *renderer, SDL_RenderCommand *cmd)

{

    const size_t vertlen = sizeof (float) * 4;

    float *verts = (float *) SDL_AllocateRenderVertices(renderer, vertlen, CONSTANT_ALIGN, &cmd->data.color.first);

    if (!verts) {

        return -1;

    }

    *(verts++) = ((float)cmd->data.color.r) / 255.0f;

    *(verts++) = ((float)cmd->data.color.g) / 255.0f;

    *(verts++) = ((float)cmd->data.color.b) / 255.0f;

    *(verts++) = ((float)cmd->data.color.a) / 255.0f;

    return 0;

}

static int

METAL_QueueDrawPoints(SDL_Renderer * renderer, SDL_RenderCommand *cmd, const SDL_FPoint * points, int count)

{

    const size_t vertlen = (sizeof (float) * 2) * count;

    float *verts = (float *) SDL_AllocateRenderVertices(renderer, vertlen, 0, &cmd->data.draw.first);

    if (!verts) {

        return -1;

    }

    cmd->data.draw.count = count;

    SDL_memcpy(verts, points, vertlen);

    return 0;

}

static int

METAL_QueueFillRects(SDL_Renderer * renderer, SDL_RenderCommand *cmd, const SDL_FRect * rects, int count)

{

    // !!! FIXME: use an index buffer

    const size_t vertlen = (sizeof (float) * 8) * count;

    float *verts = (float *) SDL_AllocateRenderVertices(renderer, vertlen, 0, &cmd->data.draw.first);

    if (!verts) {

        return -1;

    }

    cmd->data.draw.count = count;

    for (int i = 0; i < count; i++, rects++) {

        if ((rects->w <= 0.0f) || (rects->h <= 0.0f)) {

            cmd->data.draw.count--;

        } else {

            *(verts++) = rects->x;

            *(verts++) = rects->y + rects->h;

            *(verts++) = rects->x;

            *(verts++) = rects->y;

            *(verts++) = rects->x + rects->w;

            *(verts++) = rects->y + rects->h;

            *(verts++) = rects->x + rects->w;

            *(verts++) = rects->y;

        }

    }

    if (cmd->data.draw.count == 0) {

        cmd->command = SDL_RENDERCMD_NO_OP;  // nothing to do, just skip this one later.

    }

    return 0;

}

static int

METAL_QueueCopy(SDL_Renderer * renderer, SDL_RenderCommand *cmd, SDL_Texture * texture,

                const SDL_Rect * srcrect, const SDL_FRect * dstrect)

{

    METAL_TextureData *texturedata = (__bridge METAL_TextureData *)texture->driverdata;

    const float texw = (float) texturedata.mtltexture.width;

    const float texh = (float) texturedata.mtltexture.height;

    // !!! FIXME: use an index buffer

    const size_t vertlen = (sizeof (float) * 16);

    float *verts = (float *) SDL_AllocateRenderVertices(renderer, vertlen, 0, &cmd->data.draw.first);

    if (!verts) {

        return -1;

    }

    cmd->data.draw.count = 1;

    *(verts++) = dstrect->x;

    *(verts++) = dstrect->y + dstrect->h;

    *(verts++) = dstrect->x;

    *(verts++) = dstrect->y;

    *(verts++) = dstrect->x + dstrect->w;

    *(verts++) = dstrect->y + dstrect->h;

    *(verts++) = dstrect->x + dstrect->w;

    *(verts++) = dstrect->y;

    *(verts++) = normtex(srcrect->x, texw);

    *(verts++) = normtex(srcrect->y + srcrect->h, texh);

    *(verts++) = normtex(srcrect->x, texw);

    *(verts++) = normtex(srcrect->y, texh);

    *(verts++) = normtex(srcrect->x + srcrect->w, texw);

    *(verts++) = normtex(srcrect->y + srcrect->h, texh);

    *(verts++) = normtex(srcrect->x + srcrect->w, texw);

    *(verts++) = normtex(srcrect->y, texh);

    return 0;

}

static int

METAL_QueueCopyEx(SDL_Renderer * renderer, SDL_RenderCommand *cmd, SDL_Texture * texture,

                  const SDL_Rect * srcquad, const SDL_FRect * dstrect,

                  const double angle, const SDL_FPoint *center, const SDL_RendererFlip flip)

{

    METAL_TextureData *texturedata = (__bridge METAL_TextureData *)texture->driverdata;

    const float texw = (float) texturedata.mtltexture.width;

    const float texh = (float) texturedata.mtltexture.height;

    const float rads = (float)(M_PI * (float) angle / 180.0f);

    const float c = cosf(rads), s = sinf(rads);

    float minu, maxu, minv, maxv;

    const size_t vertlen = (sizeof (float) * 32);

    float *verts;

    // cheat and store this offset in (count) because it needs to be aligned in ways other fields don't and we aren't using count otherwise.

    verts = (float *) SDL_AllocateRenderVertices(renderer, vertlen, CONSTANT_ALIGN, &cmd->data.draw.count);

    if (!verts) {

        return -1;

    }

    // transform matrix

    SDL_memset(verts, '\0', sizeof (*verts) * 16);

    verts[10] = verts[15] = 1.0f;

    // rotation

    verts[0] = c;

    verts[1] = s;

    verts[4] = -s;

    verts[5] = c;

    // translation

    verts[12] = dstrect->x + center->x;

    verts[13] = dstrect->y + center->y;

    // rest of the vertices don't need the aggressive alignment. Pack them in.

    verts = (float *) SDL_AllocateRenderVertices(renderer, vertlen, 0, &cmd->data.draw.first);

    if (!verts) {

        return -1;

    }

    minu = normtex(srcquad->x, texw);

    maxu = normtex(srcquad->x + srcquad->w, texw);

    minv = normtex(srcquad->y, texh);

    maxv = normtex(srcquad->y + srcquad->h, texh);

    if (flip & SDL_FLIP_HORIZONTAL) {

        float tmp = maxu;

        maxu = minu;

        minu = tmp;

    }

    if (flip & SDL_FLIP_VERTICAL) {

        float tmp = maxv;

        maxv = minv;

        minv = tmp;

    }

    // vertices

    *(verts++) = -center->x;

    *(verts++) = dstrect->h - center->y;

    *(verts++) = -center->x;

    *(verts++) = -center->y;

    *(verts++) = dstrect->w - center->x;

    *(verts++) = dstrect->h - center->y;

    *(verts++) = dstrect->w - center->x;

    *(verts++) = -center->y;

    // texcoords

    *(verts++) = minu;

    *(verts++) = maxv;

    *(verts++) = minu;

    *(verts++) = minv;

    *(verts++) = maxu;

    *(verts++) = maxv;

    *(verts++) = maxu;

    *(verts++) = minv;

    return 0;

}

typedef struct

{

    #if __has_feature(objc_arc)

    __unsafe_unretained id<MTLRenderPipelineState> pipeline;

    #else

    id<MTLRenderPipelineState> pipeline;

    #endif

    size_t constants_offset;

    SDL_Texture *texture;

    SDL_bool cliprect_dirty;

    SDL_bool cliprect_enabled;

    SDL_Rect cliprect;

    SDL_bool viewport_dirty;

    SDL_Rect viewport;

    size_t projection_offset;

    SDL_bool color_dirty;

    size_t color_offset;

} METAL_DrawStateCache;

static void

SetDrawState(SDL_Renderer *renderer, const SDL_RenderCommand *cmd, const SDL_MetalFragmentFunction shader,

             const size_t constants_offset, id<MTLBuffer> mtlbufvertex, METAL_DrawStateCache *statecache)

{

    METAL_RenderData *data = (__bridge METAL_RenderData *) renderer->driverdata;

    const SDL_BlendMode blend = cmd->data.draw.blend;

    size_t first = cmd->data.draw.first;

    id<MTLRenderPipelineState> newpipeline;

    METAL_ActivateRenderCommandEncoder(renderer, MTLLoadActionLoad, NULL);

    if (statecache->viewport_dirty) {

        MTLViewport viewport;

        viewport.originX = statecache->viewport.x;

        viewport.originY = statecache->viewport.y;

        viewport.width = statecache->viewport.w;

        viewport.height = statecache->viewport.h;

        viewport.znear = 0.0;

        viewport.zfar = 1.0;

        [data.mtlcmdencoder setViewport:viewport];

        [data.mtlcmdencoder setVertexBuffer:mtlbufvertex offset:statecache->projection_offset atIndex:2];  // projection

        statecache->viewport_dirty = SDL_FALSE;

    }

    if (statecache->cliprect_dirty) {

        MTLScissorRect mtlrect;

        if (statecache->cliprect_enabled) {

            const SDL_Rect *rect = &statecache->cliprect;

            mtlrect.x = statecache->viewport.x + rect->x;

            mtlrect.y = statecache->viewport.y + rect->y;

            mtlrect.width = rect->w;