#include <private/qblendfunctions_p.h>

#include <private/qmath_p.h>

#include <private/qpaintengine_raster_p.h>

extern "C" void

pixman_composite_over_8888_0565_asm_neon (int32_t   w,

                                          int32_t   h,

                                          uint16_t *dst,

                                          int32_t   dst_stride,

                                          uint32_t *src,

                                          int32_t   src_stride);

extern "C" void

pixman_composite_over_8888_8888_asm_neon (int32_t   w,

                                          int32_t   h,

                                          uint32_t *dst,

                                          int32_t   dst_stride,

                                          uint32_t *src,

                                          int32_t   src_stride);

extern "C" void

pixman_composite_src_0565_8888_asm_neon (int32_t   w,

                                         int32_t   h,

                                         uint32_t *dst,

                                         int32_t   dst_stride,

                                         uint16_t *src,

                                         int32_t   src_stride);

extern "C" void

pixman_composite_over_n_8_0565_asm_neon (int32_t    w,

                                         int32_t    h,

                                         uint16_t  *dst,

                                         int32_t    dst_stride,

                                         uint32_t   src,

                                         int32_t    unused,

                                         uint8_t   *mask,

                                         int32_t    mask_stride);

extern "C" void

pixman_composite_scanline_over_asm_neon (int32_t         w,

                                         const uint32_t *dst,

                                         const uint32_t *src);

// qblendfunctions.cpp

void qt_blend_argb32_on_rgb16_const_alpha(uchar *destPixels, int dbpl,

                                          const uchar *srcPixels, int sbpl,

                                          int w, int h,

                                          int const_alpha);

void qt_blend_rgb16_on_argb32_neon(uchar *destPixels, int dbpl,

                                   const uchar *srcPixels, int sbpl,

                                   int w, int h,

                                   int const_alpha)

{

    dbpl /= 4;

    sbpl /= 2;

    quint32 *dst = (quint32 *) destPixels;

    quint16 *src = (quint16 *) srcPixels;

    if (const_alpha != 256) {

        quint8 a = (255 * const_alpha) >> 8;

        quint8 ia = 255 - a;

        while (h--) {

            for (int x=0; x<w; ++x)

                dst[x] = INTERPOLATE_PIXEL_255(qt_colorConvert(src[x], dst[x]), a, dst[x], ia);

            dst += dbpl;

            src += sbpl;

        }

        return;

    }

    pixman_composite_src_0565_8888_asm_neon(w, h, dst, dbpl, src, sbpl);

}

extern "C" void blend_8_pixels_argb32_on_rgb16_neon(quint16 *dst, const quint32 *src, int const_alpha);

void qt_blend_argb32_on_rgb16_neon(uchar *destPixels, int dbpl,

                                   const uchar *srcPixels, int sbpl,

                                   int w, int h,

                                   int const_alpha)

{

    quint16 *dst = (quint16 *) destPixels;

    quint32 *src = (quint32 *) srcPixels;

    if (const_alpha != 256) {

        for (int y=0; y<h; ++y) {

            int i = 0;

            for (; i < w-7; i += 8)

                blend_8_pixels_argb32_on_rgb16_neon(&dst[i], &src[i], const_alpha);

            if (i < w) {

                int tail = w - i;

                quint16 dstBuffer[8];

                quint32 srcBuffer[8];

                for (int j = 0; j < tail; ++j) {

                    dstBuffer[j] = dst[i + j];

                    srcBuffer[j] = src[i + j];

                }

                blend_8_pixels_argb32_on_rgb16_neon(dstBuffer, srcBuffer, const_alpha);

                for (int j = 0; j < tail; ++j) {

                    dst[i + j] = dstBuffer[j];

                    src[i + j] = srcBuffer[j];

                }

            }

            dst = (quint16 *)(((uchar *) dst) + dbpl);

            src = (quint32 *)(((uchar *) src) + sbpl);

        }

        return;

    }

    pixman_composite_over_8888_0565_asm_neon(w, h, dst, dbpl / 2, src, sbpl / 4);

}

void qt_blend_argb32_on_argb32_scanline_neon(uint *dest, const uint *src, int length, uint const_alpha)

{

    if (const_alpha == 255) {

        pixman_composite_scanline_over_asm_neon(length, dest, src);

    } else {

        qt_blend_argb32_on_argb32_neon((uchar *)dest, 4 * length, (uchar *)src, 4 * length, length, 1, (const_alpha * 256) / 255);

    }

}

        pixman_composite_over_8888_8888_asm_neon(w, h, (uint32_t *)destPixels, dbpl / 4, (uint32_t *)srcPixels, sbpl / 4);

void qt_alphamapblit_quint16_neon(QRasterBuffer *rasterBuffer,

                                  int x, int y, quint32 color,

                                  const uchar *bitmap,

                                  int mapWidth, int mapHeight, int mapStride,

                                  const QClipData *)

{

    quint16 *dest = reinterpret_cast<quint16*>(rasterBuffer->scanLine(y)) + x;

    const int destStride = rasterBuffer->bytesPerLine() / sizeof(quint16);

    uchar *mask = const_cast<uchar *>(bitmap);

    pixman_composite_over_n_8_0565_asm_neon(mapWidth, mapHeight, dest, destStride, color, 0, mask, mapStride);

}

extern "C" void blend_8_pixels_rgb16_on_rgb16_neon(quint16 *dst, const quint16 *src, int const_alpha);

template <typename SRC, typename BlendFunc>

struct Blend_on_RGB16_SourceAndConstAlpha_Neon {

    Blend_on_RGB16_SourceAndConstAlpha_Neon(BlendFunc blender, int const_alpha)

        : m_index(0)

        , m_blender(blender)

        , m_const_alpha(const_alpha)

    {

    }

    inline void write(quint16 *dst, quint32 src)

    {

        srcBuffer[m_index++] = src;

        if (m_index == 8) {

            m_blender(dst - 7, srcBuffer, m_const_alpha);

            m_index = 0;

        }

    }

    inline void flush(quint16 *dst)

    {

        if (m_index > 0) {

            quint16 dstBuffer[8];

            for (int i = 0; i < m_index; ++i)

                dstBuffer[i] = dst[i - m_index];

            m_blender(dstBuffer, srcBuffer, m_const_alpha);

            for (int i = 0; i < m_index; ++i)

                dst[i - m_index] = dstBuffer[i];

            m_index = 0;

        }

    }

    SRC srcBuffer[8];

    int m_index;

    BlendFunc m_blender;

    int m_const_alpha;

};

template <typename SRC, typename BlendFunc>

Blend_on_RGB16_SourceAndConstAlpha_Neon<SRC, BlendFunc>

Blend_on_RGB16_SourceAndConstAlpha_Neon_create(BlendFunc blender, int const_alpha)

{

    return Blend_on_RGB16_SourceAndConstAlpha_Neon<SRC, BlendFunc>(blender, const_alpha);

}

void qt_scale_image_argb32_on_rgb16_neon(uchar *destPixels, int dbpl,

                                         const uchar *srcPixels, int sbpl,

                                         const QRectF &targetRect,

                                         const QRectF &sourceRect,

                                         const QRect &clip,

                                         int const_alpha)

{

    if (const_alpha == 0)

        return;

    qt_scale_image_16bit<quint32>(destPixels, dbpl, srcPixels, sbpl, targetRect, sourceRect, clip,

        Blend_on_RGB16_SourceAndConstAlpha_Neon_create<quint32>(blend_8_pixels_argb32_on_rgb16_neon, const_alpha));

}

void qt_scale_image_rgb16_on_rgb16(uchar *destPixels, int dbpl,

                                   const uchar *srcPixels, int sbpl,

                                   const QRectF &targetRect,

                                   const QRectF &sourceRect,

                                   const QRect &clip,

                                   int const_alpha);

void qt_scale_image_rgb16_on_rgb16_neon(uchar *destPixels, int dbpl,

                                        const uchar *srcPixels, int sbpl,

                                        const QRectF &targetRect,

                                        const QRectF &sourceRect,

                                        const QRect &clip,

                                        int const_alpha)

{

    if (const_alpha == 0)

        return;

    if (const_alpha == 256) {

        qt_scale_image_rgb16_on_rgb16(destPixels, dbpl, srcPixels, sbpl, targetRect, sourceRect, clip, const_alpha);

        return;

    }

    qt_scale_image_16bit<quint16>(destPixels, dbpl, srcPixels, sbpl, targetRect, sourceRect, clip,

        Blend_on_RGB16_SourceAndConstAlpha_Neon_create<quint16>(blend_8_pixels_rgb16_on_rgb16_neon, const_alpha));

}

extern void qt_transform_image_rgb16_on_rgb16(uchar *destPixels, int dbpl,

                                              const uchar *srcPixels, int sbpl,

                                              const QRectF &targetRect,

                                              const QRectF &sourceRect,

                                              const QRect &clip,

                                              const QTransform &targetRectTransform,

                                              int const_alpha);

void qt_transform_image_rgb16_on_rgb16_neon(uchar *destPixels, int dbpl,

                                            const uchar *srcPixels, int sbpl,

                                            const QRectF &targetRect,

                                            const QRectF &sourceRect,

                                            const QRect &clip,

                                            const QTransform &targetRectTransform,

                                            int const_alpha)

{

    if (const_alpha == 0)

        return;

    if (const_alpha == 256) {

        qt_transform_image_rgb16_on_rgb16(destPixels, dbpl, srcPixels, sbpl, targetRect, sourceRect, clip, targetRectTransform, const_alpha);

        return;

    }

    qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,

                       reinterpret_cast<const quint16 *>(srcPixels), sbpl, targetRect, sourceRect, clip, targetRectTransform,

        Blend_on_RGB16_SourceAndConstAlpha_Neon_create<quint16>(blend_8_pixels_rgb16_on_rgb16_neon, const_alpha));

}

void qt_transform_image_argb32_on_rgb16_neon(uchar *destPixels, int dbpl,

                                             const uchar *srcPixels, int sbpl,

                                             const QRectF &targetRect,

                                             const QRectF &sourceRect,

                                             const QRect &clip,

                                             const QTransform &targetRectTransform,

                                             int const_alpha)

{

    if (const_alpha == 0)

        return;

    qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,

                       reinterpret_cast<const quint32 *>(srcPixels), sbpl, targetRect, sourceRect, clip, targetRectTransform,

        Blend_on_RGB16_SourceAndConstAlpha_Neon_create<quint32>(blend_8_pixels_argb32_on_rgb16_neon, const_alpha));

}

static inline void convert_8_pixels_rgb16_to_argb32(quint32 *dst, const quint16 *src)

{

    asm volatile (

        "vld1.16     { d0, d1 }, [%[SRC]]\n\t"

        /* convert 8 r5g6b5 pixel data from {d0, d1} to planar 8-bit format

           and put data into d4 - red, d3 - green, d2 - blue */

        "vshrn.u16   d4,  q0,  #8\n\t"

        "vshrn.u16   d3,  q0,  #3\n\t"

        "vsli.u16    q0,  q0,  #5\n\t"