SIMD_INLINE uint16x8_t ReduceColBody(const uint8_t *src)
{
const uint8x8x2_t t01 = vld2_u8(src - 1);
const uint8x8x2_t t23 = vld2_u8(src + 1);
return vaddq_u16(vaddl_u8(t01.val[0], t23.val[1]), vmulq_u16(vaddl_u8(t01.val[1], t23.val[0]), K16_0003));
}
template <> SIMD_INLINE uint16x8_t ReduceColTail<false>(const uint8_t *src)
{
const uint8x8x2_t t01 = vld2_u8(src - 1);
const uint8x8x2_t t23 = Deinterleave(LoadAfterLast<1>(LoadAfterLast<1>(vld1q_u8(src - 1))));
return vaddq_u16(vaddl_u8(t01.val[0], t23.val[1]), vmulq_u16(vaddl_u8(t01.val[1], t23.val[0]), K16_0003));
}
SIMD_INLINE uint16x8_t ReduceColNose(const uint8_t * src)
{
const uint8x8x2_t t01 = Deinterleave(LoadBeforeFirst<1>(vld1q_u8(src)));
const uint8x8x2_t t23 = vld2_u8(src + 1);
return vaddq_u16(vaddl_u8(t01.val[0], t23.val[1]), vmulq_u16(vaddl_u8(t01.val[1], t23.val[0]), K16_0003));
}
void test_vld2u8 (void)
{
uint8x8x2_t out_uint8x8x2_t;
out_uint8x8x2_t = vld2_u8 (0);
}
inline uint8x8x2_t vld2(const u8 * ptr) { return vld2_u8(ptr); }
void combineUYVY(const Size2D &size,
const u8 * srcyBase, ptrdiff_t srcyStride,
const u8 * srcuBase, ptrdiff_t srcuStride,
const u8 * srcvBase, ptrdiff_t srcvStride,
u8 * dstBase, ptrdiff_t dstStride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
#ifndef __ANDROID__
size_t roiw32 = size.width >= 31 ? size.width - 31 : 0;
#endif
size_t roiw8 = size.width >= 7 ? size.width - 7 : 0;
for (size_t i = 0u; i < size.height; ++i)
{
const u8 * srcy = internal::getRowPtr(srcyBase, srcyStride, i);
const u8 * srcu = internal::getRowPtr(srcuBase, srcuStride, i);
const u8 * srcv = internal::getRowPtr(srcvBase, srcvStride, i);
u8 * dst = internal::getRowPtr(dstBase, dstStride, i);
size_t syj = 0u, sj = 0u, dj = 0u;
#ifndef __ANDROID__
for (; sj < roiw32; sj += 32, syj += 64, dj += 128)
{
internal::prefetch(srcy + syj);
internal::prefetch(srcu + sj);
internal::prefetch(srcv + sj);
uint8x16x2_t v_y = vld2q_u8(srcy + syj);
uint8x16x4_t v_dst;
v_dst.val[0] = vld1q_u8(srcu + sj);
v_dst.val[1] = v_y.val[0];
v_dst.val[2] = vld1q_u8(srcv + sj);
v_dst.val[3] = v_y.val[1];
vst4q_u8(dst + dj, v_dst);
v_y = vld2q_u8(srcy + syj + 32);
v_dst.val[0] = vld1q_u8(srcu + sj + 16);
v_dst.val[1] = v_y.val[0];
v_dst.val[2] = vld1q_u8(srcv + sj + 16);
v_dst.val[3] = v_y.val[1];
vst4q_u8(dst + dj + 64, v_dst);
}
#endif
for (; sj < roiw8; sj += 8, syj += 16, dj += 32)
{
uint8x8x2_t v_y = vld2_u8(srcy + syj);
uint8x8x4_t v_dst;
v_dst.val[0] = vld1_u8(srcu + sj);
v_dst.val[1] = v_y.val[0];
v_dst.val[2] = vld1_u8(srcv + sj);
v_dst.val[3] = v_y.val[1];
vst4_u8(dst + dj, v_dst);
}
for (; sj < size.width; ++sj, syj += 2, dj += 4)
{
dst[dj] = srcu[sj];
dst[dj + 1] = srcy[syj];
dst[dj + 2] = srcv[sj];
dst[dj + 3] = srcy[syj + 1];
}
}
#else
(void)size;
(void)srcyBase;
(void)srcyStride;
(void)srcuBase;
(void)srcuStride;
(void)srcvBase;
(void)srcvStride;
(void)dstBase;
(void)dstStride;
#endif
}
void rotate_down_scale_cbcr_to_cr_cb(int wDest, int hDest, int full_width, uint8_t* cbcr_src, uint8_t* cr_dst, uint8_t* cb_dst,bool_t clockWise,bool_t down_scale) {
#ifdef __ARM_NEON__
int hSrc = down_scale?wDest*2:wDest;
int wSrc = down_scale?hDest*2:hDest;
int src_stride = 2*full_width;
int signed_dst_stride;
int incr;
int y_step=down_scale?2:1;
if (clockWise) {
/* ms_warning("start writing destination buffer from top right");*/
cb_dst += wDest - 1;
cr_dst += wDest - 1;
incr = 1;
signed_dst_stride = wDest;
} else {
/* ms_warning("start writing destination buffer from top right");*/
cb_dst += wDest * (hDest - 1);
cr_dst += wDest * (hDest - 1);
incr = -1;
signed_dst_stride = -wDest;
}
int x,y;
for (y=0; y<hSrc; y+=y_step) {
uint8_t* cb_dst2 = cb_dst;
uint8_t* cr_dst2 = cr_dst;
for (x=0; x<2*wSrc; x+=16) {
uint8x8x2_t tmp = vld2_u8 (cbcr_src+x);
vst1_lane_u8 (cb_dst2, tmp.val[0], 0);
vst1_lane_u8 (cr_dst2, tmp.val[1], 0);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
if (!down_scale) {
vst1_lane_u8 (cb_dst2, tmp.val[0], 1);
vst1_lane_u8 (cr_dst2, tmp.val[1], 1);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
}
vst1_lane_u8 (cb_dst2, tmp.val[0], 2);
vst1_lane_u8 (cr_dst2, tmp.val[1], 2);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
if (!down_scale) {
vst1_lane_u8 (cb_dst2, tmp.val[0], 3);
vst1_lane_u8 (cr_dst2, tmp.val[1], 3);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
}
vst1_lane_u8 (cb_dst2, tmp.val[0], 4);
vst1_lane_u8 (cr_dst2, tmp.val[1], 4);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
if (!down_scale) {
vst1_lane_u8 (cb_dst2, tmp.val[0], 5);
vst1_lane_u8 (cr_dst2, tmp.val[1], 5);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
}
vst1_lane_u8 (cb_dst2, tmp.val[0], 6);
vst1_lane_u8 (cr_dst2, tmp.val[1], 6);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
if (!down_scale) {
vst1_lane_u8 (cb_dst2, tmp.val[0], 7);
vst1_lane_u8 (cr_dst2, tmp.val[1], 7);
cb_dst2+=signed_dst_stride;
cr_dst2+=signed_dst_stride;
}
}
cb_dst -= incr;
cr_dst -= incr;
cbcr_src += src_stride*y_step;
}
#else
ms_error("Neon function '%s' used without hw neon support", __FUNCTION__);
#endif
}
