static void ConvertBGR24ToY_SSE41(const uint8_t* bgr, uint8_t* y, int width) {
const int max_width = width & ~31;
int i;
for (i = 0; i < max_width; bgr += 3 * 16 * 2) {
__m128i bgr_plane[6];
int j;
RGB24PackedToPlanar_SSE41(bgr, bgr_plane);
for (j = 0; j < 2; ++j, i += 16) {
const __m128i zero = _mm_setzero_si128();
__m128i r, g, b, Y0, Y1;
// Convert to 16-bit Y.
b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero);
g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero);
r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero);
ConvertRGBToY_SSE41(&r, &g, &b, &Y0);
// Convert to 16-bit Y.
b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero);
g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero);
r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero);
ConvertRGBToY_SSE41(&r, &g, &b, &Y1);
// Cast to 8-bit and store.
STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
}
}
for (; i < width; ++i, bgr += 3) { // left-over
y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
}
}
static void ConvertBGR24ToY_NEON(const uint8_t* bgr, uint8_t* y, int width) {
int i;
for (i = 0; i + 8 <= width; i += 8, bgr += 3 * 8) {
const uint8x8x3_t BGR = vld3_u8(bgr);
const uint8x8_t Y = ConvertRGBToY_NEON(BGR.val[2], BGR.val[1], BGR.val[0]);
vst1_u8(y + i, Y);
}
for (; i < width; ++i, bgr += 3) { // left-over
y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
}
}
static void ConvertRGB24ToY_NEON(const uint8_t* rgb, uint8_t* y, int width) {
int i;
for (i = 0; i + 8 <= width; i += 8, rgb += 3 * 8) {
const uint8x8x3_t RGB = vld3_u8(rgb);
const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[0], RGB.val[1], RGB.val[2]);
vst1_u8(y + i, Y);
}
for (; i < width; ++i, rgb += 3) { // left-over
y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
}
}
void WebPBlendAlpha(WebPPicture* pic, uint32_t background_rgb) {
const int red = (background_rgb >> 16) & 0xff;
const int green = (background_rgb >> 8) & 0xff;
const int blue = (background_rgb >> 0) & 0xff;
int x, y;
if (pic == NULL) return;
if (!pic->use_argb) {
const int uv_width = (pic->width >> 1); // omit last pixel during u/v loop
const int Y0 = VP8RGBToY(red, green, blue, YUV_HALF);
// VP8RGBToU/V expects the u/v values summed over four pixels
const int U0 = VP8RGBToU(4 * red, 4 * green, 4 * blue, 4 * YUV_HALF);
const int V0 = VP8RGBToV(4 * red, 4 * green, 4 * blue, 4 * YUV_HALF);
const int has_alpha = pic->colorspace & WEBP_CSP_ALPHA_BIT;
if (!has_alpha || pic->a == NULL) return; // nothing to do
for (y = 0; y < pic->height; ++y) {
// Luma blending
uint8_t* const y_ptr = pic->y + y * pic->y_stride;
uint8_t* const a_ptr = pic->a + y * pic->a_stride;
for (x = 0; x < pic->width; ++x) {
const int alpha = a_ptr[x];
if (alpha < 0xff) {
y_ptr[x] = BLEND(Y0, y_ptr[x], a_ptr[x]);
}
}
// Chroma blending every even line
if ((y & 1) == 0) {
uint8_t* const u = pic->u + (y >> 1) * pic->uv_stride;
uint8_t* const v = pic->v + (y >> 1) * pic->uv_stride;
uint8_t* const a_ptr2 =
(y + 1 == pic->height) ? a_ptr : a_ptr + pic->a_stride;
for (x = 0; x < uv_width; ++x) {
// Average four alpha values into a single blending weight.
// TODO(skal): might lead to visible contouring. Can we do better?
const int alpha =
a_ptr[2 * x + 0] + a_ptr[2 * x + 1] +
a_ptr2[2 * x + 0] + a_ptr2[2 * x + 1];
u[x] = BLEND_10BIT(U0, u[x], alpha);
v[x] = BLEND_10BIT(V0, v[x], alpha);
}
if (pic->width & 1) { // rightmost pixel
const int alpha = 2 * (a_ptr[2 * x + 0] + a_ptr2[2 * x + 0]);
u[x] = BLEND_10BIT(U0, u[x], alpha);
v[x] = BLEND_10BIT(V0, v[x], alpha);
}
}
memset(a_ptr, 0xff, pic->width);
}
} else {
