static void ConvertBGRAToRGB_SSE2(const uint32_t* src, int num_pixels,
uint8_t* dst) {
const __m128i* in = (const __m128i*)src;
__m128i* out = (__m128i*)dst;
while (num_pixels >= 32) {
// Load the BGRA buffers.
__m128i in0 = _mm_loadu_si128(in + 0);
__m128i in1 = _mm_loadu_si128(in + 1);
__m128i in2 = _mm_loadu_si128(in + 2);
__m128i in3 = _mm_loadu_si128(in + 3);
__m128i in4 = _mm_loadu_si128(in + 4);
__m128i in5 = _mm_loadu_si128(in + 5);
__m128i in6 = _mm_loadu_si128(in + 6);
__m128i in7 = _mm_loadu_si128(in + 7);
VP8L32bToPlanar_SSE2(&in0, &in1, &in2, &in3);
VP8L32bToPlanar_SSE2(&in4, &in5, &in6, &in7);
// At this points, in1/in5 contains red only, in2/in6 green only ...
// Pack the colors in 24b RGB.
VP8PlanarTo24b_SSE2(&in1, &in5, &in2, &in6, &in3, &in7);
_mm_storeu_si128(out + 0, in1);
_mm_storeu_si128(out + 1, in5);
_mm_storeu_si128(out + 2, in2);
_mm_storeu_si128(out + 3, in6);
_mm_storeu_si128(out + 4, in3);
_mm_storeu_si128(out + 5, in7);
in += 8;
out += 6;
num_pixels -= 32;
}
// left-overs
if (num_pixels > 0) {
VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
}
}
static void ConvertBGRAToRGB(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const uint32_t* const end = src + (num_pixels & ~15);
for (; src < end; src += 16) {
const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
const uint8x16x3_t tmp = { { pixel.val[2], pixel.val[1], pixel.val[0] } };
vst3q_u8(dst, tmp);
dst += 48;
}
VP8LConvertBGRAToRGB_C(src, num_pixels & 15, dst); // left-overs
}
static void ConvertBGRAToRGB(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const uint32_t* const end = src + (num_pixels & ~7);
const uint8x8_t shuffle0 = vld1_u8(kRGBShuffle[0]);
const uint8x8_t shuffle1 = vld1_u8(kRGBShuffle[1]);
const uint8x8_t shuffle2 = vld1_u8(kRGBShuffle[2]);
for (; src < end; src += 8) {
uint8x8x4_t pixels;
INIT_VECTOR4(pixels,
vld1_u8((const uint8_t*)(src + 0)),
vld1_u8((const uint8_t*)(src + 2)),
vld1_u8((const uint8_t*)(src + 4)),
vld1_u8((const uint8_t*)(src + 6)));
vst1_u8(dst + 0, vtbl4_u8(pixels, shuffle0));
vst1_u8(dst + 8, vtbl4_u8(pixels, shuffle1));
vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2));
dst += 8 * 3;
}
VP8LConvertBGRAToRGB_C(src, num_pixels & 7, dst); // left-overs
}