template<bool align, bool compensation> void ReduceGray3x3(
const uint8_t* src, size_t srcWidth, size_t srcHeight, size_t srcStride,
uint8_t* dst, size_t dstWidth, size_t dstHeight, size_t dstStride)
{
assert(srcWidth >= A && (srcWidth + 1)/2 == dstWidth && (srcHeight + 1)/2 == dstHeight);
if(align)
assert(Aligned(src) && Aligned(srcStride));
size_t lastOddCol = srcWidth - AlignLo(srcWidth, 2);
size_t bodyWidth = AlignLo(srcWidth, A);
for(size_t row = 0; row < srcHeight; row += 2, dst += dstStride, src += 2*srcStride)
{
const uint8_t * s1 = src;
const uint8_t * s0 = s1 - (row ? srcStride : 0);
const uint8_t * s2 = s1 + (row != srcHeight - 1 ? srcStride : 0);
vst1_u8(dst, ReduceRow<compensation>(ReduceColNose<align>(s0),
ReduceColNose<align>(s1), ReduceColNose<align>(s2)));
for(size_t srcCol = A, dstCol = HA; srcCol < bodyWidth; srcCol += A, dstCol += HA)
vst1_u8(dst + dstCol, ReduceRow<compensation>(ReduceColBody<align>(s0 + srcCol),
ReduceColBody<align>(s1 + srcCol), ReduceColBody<align>(s2 + srcCol)));
if(bodyWidth != srcWidth)
{
size_t srcCol = srcWidth - A - lastOddCol;
size_t dstCol = dstWidth - HA - lastOddCol;
vst1_u8(dst + dstCol, ReduceRow<compensation>(ReduceColBody<false>(s0 + srcCol),
ReduceColBody<false>(s1 + srcCol), ReduceColBody<false>(s2 + srcCol)));
if(lastOddCol)
dst[dstWidth - 1] = Base::GaussianBlur3x3<compensation>(s0 + srcWidth, s1 + srcWidth, s2 + srcWidth, -2, -1, -1);
}
}
}
int main(void)
{
uint8_t v1_init[8] = {1, 1, 1, 1, 1, 1, 1, 1};
uint8_t v2_init[8] = {2, 2, 2, 2, 2, 2, 2, 2};
uint8x8_t v1 = vld1_u8 (v1_init);
uint8x8_t v2 = vld1_u8 (v2_init);
uint8x8x2_t vd1, vd2;
union {uint8x8_t v; uint8_t buf[8];} d1, d2, d3, d4;
int i;
uint8_t odd, even;
vd1 = vzip_u8(v1, vdup_n_u8(0));
vd2 = vzip_u8(v2, vdup_n_u8(0));
vst1_u8(d1.buf, vd1.val[0]);
vst1_u8(d2.buf, vd1.val[1]);
vst1_u8(d3.buf, vd2.val[0]);
vst1_u8(d4.buf, vd2.val[1]);
#ifdef __ARMEL__
odd = 1;
even = 0;
#else
odd = 0;
even = 1;
#endif
for (i = 0; i < 8; i++)
if ((i % 2 == even && d4.buf[i] != 2)
|| (i % 2 == odd && d4.buf[i] != 0))
abort ();
return 0;
}
static void ConvertARGBToUV_NEON(const uint32_t* argb, uint8_t* u, uint8_t* v,
int src_width, int do_store) {
int i;
for (i = 0; i + 16 <= src_width; i += 16, u += 8, v += 8) {
const uint8x16x4_t RGB = vld4q_u8((const uint8_t*)&argb[i]);
const uint16x8_t R = vpaddlq_u8(RGB.val[2]); // pair-wise adds
const uint16x8_t G = vpaddlq_u8(RGB.val[1]);
const uint16x8_t B = vpaddlq_u8(RGB.val[0]);
int16x8_t U_tmp, V_tmp;
CONVERT_RGB_TO_UV(R, G, B, 1, U_tmp, V_tmp);
{
const uint8x8_t U = vqrshrun_n_s16(U_tmp, 1);
const uint8x8_t V = vqrshrun_n_s16(V_tmp, 1);
if (do_store) {
vst1_u8(u, U);
vst1_u8(v, V);
} else {
const uint8x8_t prev_u = vld1_u8(u);
const uint8x8_t prev_v = vld1_u8(v);
vst1_u8(u, vrhadd_u8(U, prev_u));
vst1_u8(v, vrhadd_u8(V, prev_v));
}
}
}
if (i < src_width) { // left-over
WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store);
}
}
template <bool even> void ReduceGray4x4(const uint8_t *src, size_t srcWidth, size_t srcHeight, size_t srcStride,
uint8_t *dst, size_t dstWidth, size_t dstHeight, size_t dstStride)
{
assert((srcWidth + 1) / 2 == dstWidth && (srcHeight + 1) / 2 == dstHeight && srcWidth > A);
size_t alignedDstWidth = Simd::AlignLo(dstWidth, HA);
size_t srcTail = Simd::AlignHi(srcWidth - A, 2);
Buffer buffer(Simd::AlignHi(dstWidth, A));
uint16x8_t tmp = ReduceColNose(src);
Store<true>(buffer.src0, tmp);
Store<true>(buffer.src1, tmp);
size_t srcCol = A, dstCol = HA;
for (; srcCol < srcWidth - A; srcCol += A, dstCol += HA)
{
tmp = ReduceColBody(src + srcCol);
Store<true>(buffer.src0 + dstCol, tmp);
Store<true>(buffer.src1 + dstCol, tmp);
}
tmp = ReduceColTail<even>(src + srcTail);
Store<false>(buffer.src0 + dstWidth - HA, tmp);
Store<false>(buffer.src1 + dstWidth - HA, tmp);
for (size_t row = 0; row < srcHeight; row += 2, dst += dstStride)
{
const uint8_t *src2 = src + srcStride*(row + 1);
const uint8_t *src3 = src2 + srcStride;
if (row >= srcHeight - 2)
{
src2 = src + srcStride*(srcHeight - 1);
src3 = src2;
}
Store<true>(buffer.src2, ReduceColNose(src2));
Store<true>(buffer.src3, ReduceColNose(src3));
size_t srcCol = A, dstCol = HA;
for (; srcCol < srcWidth - A; srcCol += A, dstCol += HA)
{
Store<true>(buffer.src2 + dstCol, ReduceColBody(src2 + srcCol));
Store<true>(buffer.src3 + dstCol, ReduceColBody(src3 + srcCol));
}
Store<false>(buffer.src2 + dstWidth - HA, ReduceColTail<even>(src2 + srcTail));
Store<false>(buffer.src3 + dstWidth - HA, ReduceColTail<even>(src3 + srcTail));
for (size_t col = 0; col < alignedDstWidth; col += HA)
vst1_u8(dst + col, ReduceRow<true>(buffer, col));
if (alignedDstWidth != dstWidth)
vst1_u8(dst + dstWidth - HA, ReduceRow<false>(buffer, dstWidth - HA));
Swap(buffer.src0, buffer.src2);
Swap(buffer.src1, buffer.src3);
}
}
void vp9_lpf_horizontal_4_neon(
unsigned char *src,
int pitch,
unsigned char *blimit,
unsigned char *limit,
unsigned char *thresh,
int count) {
int i;
uint8_t *s, *psrc;
uint8x8_t dblimit, dlimit, dthresh;
uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
if (count == 0) // end_vp9_lf_h_edge
return;
dblimit = vld1_u8(blimit);
dlimit = vld1_u8(limit);
dthresh = vld1_u8(thresh);
psrc = src - (pitch << 2);
for (i = 0; i < count; i++) {
s = psrc + i * 8;
d3u8 = vld1_u8(s);
s += pitch;
d4u8 = vld1_u8(s);
s += pitch;
d5u8 = vld1_u8(s);
s += pitch;
d6u8 = vld1_u8(s);
s += pitch;
d7u8 = vld1_u8(s);
s += pitch;
d16u8 = vld1_u8(s);
s += pitch;
d17u8 = vld1_u8(s);
s += pitch;
d18u8 = vld1_u8(s);
vp9_loop_filter_neon(dblimit, dlimit, dthresh,
d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
&d4u8, &d5u8, &d6u8, &d7u8);
s -= (pitch * 5);
vst1_u8(s, d4u8);
s += pitch;
vst1_u8(s, d5u8);
s += pitch;
vst1_u8(s, d6u8);
s += pitch;
vst1_u8(s, d7u8);
}
return;
}
static INLINE void scaledconvolve_horiz_w8(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const x_filters,
const int x0_q4, const int x_step_q4, const int w, const int h) {
DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
int x, y, z;
src -= SUBPEL_TAPS / 2 - 1;
// This function processes 8x8 areas. The intermediate height is not always
// a multiple of 8, so force it to be a multiple of 8 here.
y = (h + 7) & ~7;
do {
int x_q4 = x0_q4;
x = 0;
do {
uint8x8_t d[8];
// process 8 src_x steps
for (z = 0; z < 8; ++z) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
if (x_q4 & SUBPEL_MASK) {
const int16x8_t filters = vld1q_s16(x_filters[x_q4 & SUBPEL_MASK]);
uint8x8_t s[8];
load_u8_8x8(src_x, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4],
&s[5], &s[6], &s[7]);
transpose_u8_8x8(&s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
&s[7]);
d[0] = scale_filter_8(s, filters);
vst1_u8(&temp[8 * z], d[0]);
} else {
int i;
for (i = 0; i < 8; ++i) {
temp[z * 8 + i] = src_x[i * src_stride + 3];
}
}
x_q4 += x_step_q4;
}
// transpose the 8x8 filters values back to dst
load_u8_8x8(temp, 8, &d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6],
&d[7]);
transpose_u8_8x8(&d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
vst1_u8(&dst[x + 0 * dst_stride], d[0]);
vst1_u8(&dst[x + 1 * dst_stride], d[1]);
vst1_u8(&dst[x + 2 * dst_stride], d[2]);
vst1_u8(&dst[x + 3 * dst_stride], d[3]);
vst1_u8(&dst[x + 4 * dst_stride], d[4]);
vst1_u8(&dst[x + 5 * dst_stride], d[5]);
vst1_u8(&dst[x + 6 * dst_stride], d[6]);
vst1_u8(&dst[x + 7 * dst_stride], d[7]);
x += 8;
} while (x < w);
src += src_stride * 8;
dst += dst_stride * 8;
} while (y -= 8);
}
void vp9_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
uint8x8_t d0u8 = vdup_n_u8(0);
uint64x1_t d1u64 = vdup_n_u64(0);
(void)above;
d1u64 = vld1_u64((const uint64_t *)left);
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
vst1_u8(dst, d0u8);
}
static void ConvertRGBA32ToUV_NEON(const uint16_t* rgb,
uint8_t* u, uint8_t* v, int width) {
int i;
for (i = 0; i + 8 <= width; i += 8, rgb += 4 * 8) {
const uint16x8x4_t RGB = vld4q_u16((const uint16_t*)rgb);
int16x8_t U, V;
CONVERT_RGB_TO_UV(RGB.val[0], RGB.val[1], RGB.val[2], 2, U, V);
vst1_u8(u + i, vqrshrun_n_s16(U, 2));
vst1_u8(v + i, vqrshrun_n_s16(V, 2));
}
for (; i < width; i += 1, rgb += 4) {
const int r = rgb[0], g = rgb[1], b = rgb[2];
u[i] = VP8RGBToU(r, g, b, YUV_HALF << 2);
v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2);
}
}
static INLINE void scaledconvolve_vert_w8(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
const int y0_q4, const int y_step_q4, const int w, const int h) {
int y;
int y_q4 = y0_q4;
src -= src_stride * (SUBPEL_TAPS / 2 - 1);
y = h;
do {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
if (y_q4 & SUBPEL_MASK) {
const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
uint8x8_t s[8], d;
load_u8_8x8(src_y, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
&s[6], &s[7]);
d = scale_filter_8(s, filters);
vst1_u8(dst, d);
} else {
memcpy(dst, &src_y[3 * src_stride], w);
}
dst += dst_stride;
y_q4 += y_step_q4;
} while (--y);
}
void test_vst1u8 (void)
{
uint8_t *arg0_uint8_t;
uint8x8_t arg1_uint8x8_t;
vst1_u8 (arg0_uint8_t, arg1_uint8x8_t);
}
void
ComputesRGBLuminanceMask_NEON(const uint8_t *aSourceData,
int32_t aSourceStride,
uint8_t *aDestData,
int32_t aDestStride,
const IntSize &aSize,
float aOpacity)
{
int32_t redFactor = 55 * aOpacity; // 255 * 0.2125 * opacity
int32_t greenFactor = 183 * aOpacity; // 255 * 0.7154 * opacity
int32_t blueFactor = 18 * aOpacity; // 255 * 0.0721
const uint8_t *sourcePixel = aSourceData;
int32_t sourceOffset = aSourceStride - 4 * aSize.width;
uint8_t *destPixel = aDestData;
int32_t destOffset = aDestStride - aSize.width;
sourcePixel = aSourceData;
int32_t remainderWidth = aSize.width % 8;
int32_t roundedWidth = aSize.width - remainderWidth;
uint16x8_t temp;
uint8x8_t gray;
uint8x8_t redVector = vdup_n_u8(redFactor);
uint8x8_t greenVector = vdup_n_u8(greenFactor);
uint8x8_t blueVector = vdup_n_u8(blueFactor);
uint8x8_t fullBitVector = vdup_n_u8(255);
uint8x8_t oneVector = vdup_n_u8(1);
for (int32_t y = 0; y < aSize.height; y++) {
// Calculate luminance by neon with 8 pixels per loop
for (int32_t x = 0; x < roundedWidth; x += 8) {
uint8x8x4_t argb = vld4_u8(sourcePixel);
temp = vmull_u8(argb.val[GFX_ARGB32_OFFSET_R], redVector); // temp = red * redFactor
temp = vmlal_u8(temp, argb.val[GFX_ARGB32_OFFSET_G], greenVector); // temp += green * greenFactor
temp = vmlal_u8(temp, argb.val[GFX_ARGB32_OFFSET_B], blueVector); // temp += blue * blueFactor
gray = vshrn_n_u16(temp, 8); // gray = temp >> 8
// Check alpha value
uint8x8_t alphaVector = vtst_u8(argb.val[GFX_ARGB32_OFFSET_A], fullBitVector);
gray = vmul_u8(gray, vand_u8(alphaVector, oneVector));
// Put the result to the 8 pixels
vst1_u8(destPixel, gray);
sourcePixel += 8 * 4;
destPixel += 8;
}
// Calculate the rest pixels of the line by cpu
for (int32_t x = 0; x < remainderWidth; x++) {
if (sourcePixel[GFX_ARGB32_OFFSET_A] > 0) {
*destPixel = (redFactor * sourcePixel[GFX_ARGB32_OFFSET_R]+
greenFactor * sourcePixel[GFX_ARGB32_OFFSET_G] +
blueFactor * sourcePixel[GFX_ARGB32_OFFSET_B]) >> 8;
} else {
*destPixel = 0;
}
sourcePixel += 4;
destPixel++;
}
void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int i;
uint8x8_t d0u8 = vdup_n_u8(0);
(void)left;
d0u8 = vld1_u8(above);
for (i = 0; i < 8; i++, dst += stride) vst1_u8(dst, d0u8);
}
void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
static const uint8_t shuffle1[8] = { 1, 2, 3, 4, 5, 6, 7, 7 };
static const uint8_t shuffle2[8] = { 2, 3, 4, 5, 6, 7, 7, 7 };
const uint8x8_t sh_12345677 = vld1_u8(shuffle1);
const uint8x8_t sh_23456777 = vld1_u8(shuffle2);
const uint8x8_t A0 = vld1_u8(above); // top row
const uint8x8_t A1 = vtbl1_u8(A0, sh_12345677);
const uint8x8_t A2 = vtbl1_u8(A0, sh_23456777);
const uint8x8_t avg1 = vhadd_u8(A0, A2);
uint8x8_t row = vrhadd_u8(avg1, A1);
int i;
(void)left;
for (i = 0; i < 7; ++i) {
vst1_u8(dst + i * stride, row);
row = vtbl1_u8(row, sh_12345677);
}
vst1_u8(dst + i * stride, row);
}
static void ConvertBGRAToBGR(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(kBGRShuffle[0]);
const uint8x8_t shuffle1 = vld1_u8(kBGRShuffle[1]);
const uint8x8_t shuffle2 = vld1_u8(kBGRShuffle[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;
}
VP8LConvertBGRAToBGR_C(src, num_pixels & 7, dst); // left-overs
}
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 ConvertBGRAToRGBA(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const uint32_t* const end = src + (num_pixels & ~1);
const uint8x8_t shuffle = vld1_u8(kRGBAShuffle);
for (; src < end; src += 2) {
const uint8x8_t pixels = vld1_u8((uint8_t*)src);
vst1_u8(dst, vtbl1_u8(pixels, shuffle));
dst += 8;
}
VP8LConvertBGRAToRGBA_C(src, num_pixels & 1, dst); // left-overs
}
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 vp8_copy_mem8x8_neon(unsigned char *src, int src_stride,
unsigned char *dst, int dst_stride) {
uint8x8_t vtmp;
int r;
for (r = 0; r < 8; ++r) {
vtmp = vld1_u8(src);
vst1_u8(dst, vtmp);
src += src_stride;
dst += dst_stride;
}
}
static void ConvertARGBToY_NEON(const uint32_t* argb, uint8_t* y, int width) {
int i;
for (i = 0; i + 8 <= width; i += 8) {
const uint8x8x4_t RGB = vld4_u8((const uint8_t*)&argb[i]);
const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[2], RGB.val[1], RGB.val[0]);
vst1_u8(y + i, Y);
}
for (; i < width; ++i) { // left-over
const uint32_t p = argb[i];
y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff,
YUV_HALF);
}
}
void vpx_lpf_horizontal_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
const uint8_t *limit, const uint8_t *thresh,
int count) {
int i;
uint8_t *s, *psrc;
uint8x8_t dblimit, dlimit, dthresh;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
uint8x8_t d16u8, d17u8, d18u8;
if (count == 0) // end_vpx_mblf_h_edge
return;
dblimit = vld1_u8(blimit);
dlimit = vld1_u8(limit);
dthresh = vld1_u8(thresh);
psrc = src - (pitch << 2);
for (i = 0; i < count; i++) {
s = psrc + i * 8;
d3u8 = vld1_u8(s);
s += pitch;
d4u8 = vld1_u8(s);
s += pitch;
d5u8 = vld1_u8(s);
s += pitch;
d6u8 = vld1_u8(s);
s += pitch;
d7u8 = vld1_u8(s);
s += pitch;
d16u8 = vld1_u8(s);
s += pitch;
d17u8 = vld1_u8(s);
s += pitch;
d18u8 = vld1_u8(s);
mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
&d5u8);
s -= (pitch * 6);
vst1_u8(s, d0u8);
s += pitch;
vst1_u8(s, d1u8);
s += pitch;
vst1_u8(s, d2u8);
s += pitch;
vst1_u8(s, d3u8);
s += pitch;
vst1_u8(s, d4u8);
s += pitch;
vst1_u8(s, d5u8);
}
return;
}
/**
* @brief Grayscale 변환을 수행하는 thread 함수. NEON 명령어를 사용한다.
*
* @param arg 계산해야할 정보가 담겨있는 _thread_data 형식의 구조체
*
* @return NULL
*/
static void *thread_calc(void *arg)
{
struct _thread_data *data = (struct _thread_data *)arg;
uint8x8_t rfac = vdup_n_u8 (76);
uint8x8_t gfac = vdup_n_u8 (151);
uint8x8_t bfac = vdup_n_u8 (29);
int n = data->size / 8;
int m = data->size % 8;
int iTemp;
unsigned char szTemp[8];
unsigned int *data_in = data->data_in;
unsigned int *data_out = data->data_out;
// 한 루프당 8픽셀씩 변환 (32 bytes)
while (n--) {
uint16x8_t temp;
uint8x8x4_t rgb = vld4_u8 ((unsigned char *)data_in);
uint8x8_t result;
temp = vmull_u8(rgb.val[0], bfac);
temp = vmlal_u8(temp, rgb.val[1], gfac);
temp = vmlal_u8(temp, rgb.val[2], rfac);
result = vshrn_n_u16 (temp, 8);
vst1_u8(szTemp, result);
iTemp = szTemp[0]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[1]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[2]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[3]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[4]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[5]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[6]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
iTemp = szTemp[7]; *data_out = iTemp | (iTemp << 8) | (iTemp << 16) | (*data_in & 0xff000000); data_in++; data_out++;
}
if (m) {
argb8888_to_gray(data_in, data_out, m);
}
return NULL;
}
void compare_neon_ge(float *psrc1, float src2, uchar *pdst, int size)
{
int remainder = size - 7;
float32x4_t vsrc2 = vdupq_n_f32(src2);
int i = 0;
for(; i < remainder; i += 8){
float32x4_t vsrc1_32x4 = vld1q_f32(psrc1 + i );
float32x4_t vsrc2_32x4 = vld1q_f32(psrc1 + i + 4);
uint32x4_t vdst1 = vcgeq_f32(vsrc1_32x4, vsrc2);
uint32x4_t vdst2 = vcgeq_f32(vsrc2_32x4, vsrc2);
uint16x4_t vdst1_16x4 = vmovn_u32(vdst1);
uint16x4_t vdst2_16x4 = vmovn_u32(vdst2);
uint16x8_t vdst_16x8 = vcombine_u16(vdst1_16x4, vdst2_16x4);
uint8x8_t vdst_8x8 = vmovn_u16(vdst_16x8);
vst1_u8(pdst + i, vdst_8x8);
}
for( ; i < size; i++){
pdst[i] = (psrc1[i] >= src2 ) ? 255 : 0;
}
}
// Process a block exactly 8 wide and any height.
static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
uint8_t *output_ptr,
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
const uint8_t *filter) {
const uint8x8_t f0 = vdup_n_u8(filter[0]);
const uint8x8_t f1 = vdup_n_u8(filter[1]);
unsigned int i;
for (i = 0; i < output_height; ++i) {
const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
const uint16x8_t a = vmull_u8(src_0, f0);
const uint16x8_t b = vmlal_u8(a, src_1, f1);
const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
vst1_u8(output_ptr, out);
src_ptr += src_pixels_per_line;
output_ptr += 8;
}
}
static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
uint8_t *output_ptr,
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
unsigned int output_width,
const uint16_t *vpx_filter) {
const uint8x8_t f0 = vmov_n_u8((uint8_t)vpx_filter[0]);
const uint8x8_t f1 = vmov_n_u8((uint8_t)vpx_filter[1]);
unsigned int i;
for (i = 0; i < output_height; ++i) {
const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
const uint16x8_t a = vmull_u8(src_0, f0);
const uint16x8_t b = vmlal_u8(a, src_1, f1);
const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
vst1_u8(&output_ptr[0], out);
// Next row...
src_ptr += src_pixels_per_line;
output_ptr += output_width;
}
}
inline void vst1(u8 * ptr, const uint8x8_t & v) { return vst1_u8(ptr, v); }
void ne10_img_vresize_linear_neon (const int** src, unsigned char* dst, const short* beta, int width)
{
const int *S0 = src[0], *S1 = src[1];
int32x4_t qS0_0123, qS0_4567, qS1_0123, qS1_4567;
int32x4_t qT_0123, qT_4567;
int16x4_t dT_0123, dT_4567;
uint16x8_t qT_01234567;
uint8x8_t dT_01234567, dDst_01234567;
int32x2_t dBeta;
dBeta = vset_lane_s32 ( (int) (beta[0]), dBeta, 0);
dBeta = vset_lane_s32 ( (int) (beta[1]), dBeta, 1);
int32x4_t qDelta, qMin, qMax;
qDelta = vdupq_n_s32 (DELTA);
qMin = vdupq_n_s32 (0);
qMax = vdupq_n_s32 (255);
int x = 0;
for (; x <= width - 8; x += 8)
{
qS0_0123 = vld1q_s32 (&S0[x]);
qS0_4567 = vld1q_s32 (&S0[x + 4]);
qS1_0123 = vld1q_s32 (&S1[x]);
qS1_4567 = vld1q_s32 (&S1[x + 4]);
qT_0123 = vmulq_lane_s32 (qS0_0123, dBeta, 0);
qT_4567 = vmulq_lane_s32 (qS0_4567, dBeta, 0);
qT_0123 = vmlaq_lane_s32 (qT_0123, qS1_0123, dBeta, 1);
qT_4567 = vmlaq_lane_s32 (qT_4567, qS1_4567, dBeta, 1);
qT_0123 = vaddq_s32 (qT_0123, qDelta);
qT_4567 = vaddq_s32 (qT_4567, qDelta);
qT_0123 = vshrq_n_s32 (qT_0123, BITS);
qT_4567 = vshrq_n_s32 (qT_4567, BITS);
qT_0123 = vmaxq_s32 (qT_0123, qMin);
qT_4567 = vmaxq_s32 (qT_4567, qMin);
qT_0123 = vminq_s32 (qT_0123, qMax);
qT_4567 = vminq_s32 (qT_4567, qMax);
dT_0123 = vmovn_s32 (qT_0123);
dT_4567 = vmovn_s32 (qT_4567);
qT_01234567 = vreinterpretq_u16_s16 (vcombine_s16 (dT_0123, dT_4567));
dT_01234567 = vmovn_u16 (qT_01234567);
vst1_u8 (&dst[x], dT_01234567);
}
if (x < width)
{
uint8x8_t dMask;
dMask = vld1_u8 ( (uint8_t *) (&ne10_img_vresize_linear_mask_residual_table[ (width - x - 1)]));
dDst_01234567 = vld1_u8 (&dst[x]);
qS0_0123 = vld1q_s32 (&S0[x]);
qS0_4567 = vld1q_s32 (&S0[x + 4]);
qS1_0123 = vld1q_s32 (&S1[x]);
qS1_4567 = vld1q_s32 (&S1[x + 4]);
qT_0123 = vmulq_lane_s32 (qS0_0123, dBeta, 0);
qT_4567 = vmulq_lane_s32 (qS0_4567, dBeta, 0);
qT_0123 = vmlaq_lane_s32 (qT_0123, qS1_0123, dBeta, 1);
qT_4567 = vmlaq_lane_s32 (qT_4567, qS1_4567, dBeta, 1);
qT_0123 = vaddq_s32 (qT_0123, qDelta);
qT_4567 = vaddq_s32 (qT_4567, qDelta);
qT_0123 = vshrq_n_s32 (qT_0123, BITS);
qT_4567 = vshrq_n_s32 (qT_4567, BITS);
qT_0123 = vmaxq_s32 (qT_0123, qMin);
qT_4567 = vmaxq_s32 (qT_4567, qMin);
qT_0123 = vminq_s32 (qT_0123, qMax);
qT_4567 = vminq_s32 (qT_4567, qMax);
dT_0123 = vmovn_s32 (qT_0123);
dT_4567 = vmovn_s32 (qT_4567);
qT_01234567 = vreinterpretq_u16_s16 (vcombine_s16 (dT_0123, dT_4567));
dT_01234567 = vmovn_u16 (qT_01234567);
dMask = vbsl_u8 (dMask, dT_01234567, dDst_01234567);
vst1_u8 (&dst[x], dMask);
}
}
void vp8_sixtap_predict8x8_neon(
unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
unsigned char *dst_ptr,
int dst_pitch) {
unsigned char *src, *tmpp;
unsigned char tmp[64];
int i;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
uint8x8_t d18u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8, d25u8;
uint8x8_t d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q9u8, q10u8, q11u8, q12u8;
if (xoffset == 0) { // secondpass_filter8x8_only
// load second_pass filter
dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
// load src data
src = src_ptr - src_pixels_per_line * 2;
d18u8 = vld1_u8(src);
src += src_pixels_per_line;
d19u8 = vld1_u8(src);
src += src_pixels_per_line;
d20u8 = vld1_u8(src);
src += src_pixels_per_line;
d21u8 = vld1_u8(src);
src += src_pixels_per_line;
d22u8 = vld1_u8(src);
src += src_pixels_per_line;
d23u8 = vld1_u8(src);
src += src_pixels_per_line;
d24u8 = vld1_u8(src);
src += src_pixels_per_line;
d25u8 = vld1_u8(src);
src += src_pixels_per_line;
d26u8 = vld1_u8(src);
src += src_pixels_per_line;
d27u8 = vld1_u8(src);
src += src_pixels_per_line;
d28u8 = vld1_u8(src);
src += src_pixels_per_line;
d29u8 = vld1_u8(src);
src += src_pixels_per_line;
d30u8 = vld1_u8(src);
for (i = 2; i > 0; i--) {
q3u16 = vmull_u8(d18u8, d0u8);
q4u16 = vmull_u8(d19u8, d0u8);
q5u16 = vmull_u8(d20u8, d0u8);
q6u16 = vmull_u8(d21u8, d0u8);
q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
q7u16 = vmull_u8(d21u8, d3u8);
q8u16 = vmull_u8(d22u8, d3u8);
q9u16 = vmull_u8(d23u8, d3u8);
q10u16 = vmull_u8(d24u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d6u8 = vqrshrun_n_s16(q7s16, 7);
d7u8 = vqrshrun_n_s16(q8s16, 7);
d8u8 = vqrshrun_n_s16(q9s16, 7);
d9u8 = vqrshrun_n_s16(q10s16, 7);
d18u8 = d22u8;
d19u8 = d23u8;
d20u8 = d24u8;
d21u8 = d25u8;
d22u8 = d26u8;
d23u8 = d27u8;
d24u8 = d28u8;
d25u8 = d29u8;
d26u8 = d30u8;
vst1_u8(dst_ptr, d6u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d7u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d8u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d9u8);
dst_ptr += dst_pitch;
}
return;
}
// load first_pass filter
dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
// First pass: output_height lines x output_width columns (9x4)
if (yoffset == 0) // firstpass_filter4x4_only
src = src_ptr - 2;
else
src = src_ptr - 2 - (src_pixels_per_line * 2);
tmpp = tmp;
for (i = 2; i > 0; i--) {
q3u8 = vld1q_u8(src);
src += src_pixels_per_line;
q4u8 = vld1q_u8(src);
src += src_pixels_per_line;
q5u8 = vld1q_u8(src);
src += src_pixels_per_line;
q6u8 = vld1q_u8(src);
src += src_pixels_per_line;
__builtin_prefetch(src);
__builtin_prefetch(src + src_pixels_per_line);
__builtin_prefetch(src + src_pixels_per_line * 2);
q7u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
q8u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
q9u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
q10u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
q7u16 = vmlsl_u8(q7u16, d28u8, d1u8);
q8u16 = vmlsl_u8(q8u16, d29u8, d1u8);
q9u16 = vmlsl_u8(q9u16, d30u8, d1u8);
q10u16 = vmlsl_u8(q10u16, d31u8, d1u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
q7u16 = vmlsl_u8(q7u16, d28u8, d4u8);
q8u16 = vmlsl_u8(q8u16, d29u8, d4u8);
q9u16 = vmlsl_u8(q9u16, d30u8, d4u8);
q10u16 = vmlsl_u8(q10u16, d31u8, d4u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
q7u16 = vmlal_u8(q7u16, d28u8, d2u8);
q8u16 = vmlal_u8(q8u16, d29u8, d2u8);
q9u16 = vmlal_u8(q9u16, d30u8, d2u8);
q10u16 = vmlal_u8(q10u16, d31u8, d2u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
q7u16 = vmlal_u8(q7u16, d28u8, d5u8);
q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
q9u16 = vmlal_u8(q9u16, d30u8, d5u8);
q10u16 = vmlal_u8(q10u16, d31u8, d5u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
q3u16 = vmull_u8(d28u8, d3u8);
q4u16 = vmull_u8(d29u8, d3u8);
q5u16 = vmull_u8(d30u8, d3u8);
q6u16 = vmull_u8(d31u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d22u8 = vqrshrun_n_s16(q7s16, 7);
d23u8 = vqrshrun_n_s16(q8s16, 7);
d24u8 = vqrshrun_n_s16(q9s16, 7);
d25u8 = vqrshrun_n_s16(q10s16, 7);
if (yoffset == 0) { // firstpass_filter8x4_only
vst1_u8(dst_ptr, d22u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d23u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d24u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d25u8);
dst_ptr += dst_pitch;
} else {
vst1_u8(tmpp, d22u8);
tmpp += 8;
vst1_u8(tmpp, d23u8);
tmpp += 8;
vst1_u8(tmpp, d24u8);
tmpp += 8;
vst1_u8(tmpp, d25u8);
tmpp += 8;
}
}
if (yoffset == 0)
return;
// First Pass on rest 5-line data
q3u8 = vld1q_u8(src);
src += src_pixels_per_line;
q4u8 = vld1q_u8(src);
src += src_pixels_per_line;
q5u8 = vld1q_u8(src);
src += src_pixels_per_line;
q6u8 = vld1q_u8(src);
src += src_pixels_per_line;
q7u8 = vld1q_u8(src);
q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
q11u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
q12u16 = vmull_u8(vget_low_u8(q7u8), d0u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 1);
q8u16 = vmlsl_u8(q8u16, d27u8, d1u8);
q9u16 = vmlsl_u8(q9u16, d28u8, d1u8);
q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
q11u16 = vmlsl_u8(q11u16, d30u8, d1u8);
q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 4);
q8u16 = vmlsl_u8(q8u16, d27u8, d4u8);
q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
q11u16 = vmlsl_u8(q11u16, d30u8, d4u8);
q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 2);
q8u16 = vmlal_u8(q8u16, d27u8, d2u8);
q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
q11u16 = vmlal_u8(q11u16, d30u8, d2u8);
q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 5);
q8u16 = vmlal_u8(q8u16, d27u8, d5u8);
q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
q11u16 = vmlal_u8(q11u16, d30u8, d5u8);
q12u16 = vmlal_u8(q12u16, d31u8, d5u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 3);
q3u16 = vmull_u8(d27u8, d3u8);
q4u16 = vmull_u8(d28u8, d3u8);
q5u16 = vmull_u8(d29u8, d3u8);
q6u16 = vmull_u8(d30u8, d3u8);
q7u16 = vmull_u8(d31u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q11s16 = vreinterpretq_s16_u16(q11u16);
q12s16 = vreinterpretq_s16_u16(q12u16);
q8s16 = vqaddq_s16(q8s16, q3s16);
q9s16 = vqaddq_s16(q9s16, q4s16);
q10s16 = vqaddq_s16(q10s16, q5s16);
q11s16 = vqaddq_s16(q11s16, q6s16);
q12s16 = vqaddq_s16(q12s16, q7s16);
d26u8 = vqrshrun_n_s16(q8s16, 7);
d27u8 = vqrshrun_n_s16(q9s16, 7);
d28u8 = vqrshrun_n_s16(q10s16, 7);
d29u8 = vqrshrun_n_s16(q11s16, 7);
d30u8 = vqrshrun_n_s16(q12s16, 7);
// Second pass: 8x8
dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
tmpp = tmp;
q9u8 = vld1q_u8(tmpp);
tmpp += 16;
q10u8 = vld1q_u8(tmpp);
tmpp += 16;
q11u8 = vld1q_u8(tmpp);
tmpp += 16;
q12u8 = vld1q_u8(tmpp);
d18u8 = vget_low_u8(q9u8);
d19u8 = vget_high_u8(q9u8);
d20u8 = vget_low_u8(q10u8);
d21u8 = vget_high_u8(q10u8);
d22u8 = vget_low_u8(q11u8);
d23u8 = vget_high_u8(q11u8);
d24u8 = vget_low_u8(q12u8);
d25u8 = vget_high_u8(q12u8);
for (i = 2; i > 0; i--) {
q3u16 = vmull_u8(d18u8, d0u8);
q4u16 = vmull_u8(d19u8, d0u8);
q5u16 = vmull_u8(d20u8, d0u8);
q6u16 = vmull_u8(d21u8, d0u8);
q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
q7u16 = vmull_u8(d21u8, d3u8);
q8u16 = vmull_u8(d22u8, d3u8);
q9u16 = vmull_u8(d23u8, d3u8);
q10u16 = vmull_u8(d24u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d6u8 = vqrshrun_n_s16(q7s16, 7);
d7u8 = vqrshrun_n_s16(q8s16, 7);
d8u8 = vqrshrun_n_s16(q9s16, 7);
d9u8 = vqrshrun_n_s16(q10s16, 7);
d18u8 = d22u8;
d19u8 = d23u8;
d20u8 = d24u8;
d21u8 = d25u8;
d22u8 = d26u8;
d23u8 = d27u8;
d24u8 = d28u8;
d25u8 = d29u8;
d26u8 = d30u8;
vst1_u8(dst_ptr, d6u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d7u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d8u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d9u8);
dst_ptr += dst_pitch;
}
return;
}
void vp8_sixtap_predict16x16_neon(
unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
unsigned char *dst_ptr,
int dst_pitch) {
unsigned char *src, *src_tmp, *dst, *tmpp;
unsigned char tmp[336];
int i, j;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d18u8, d19u8;
uint8x8_t d20u8, d21u8, d22u8, d23u8, d24u8, d25u8, d26u8, d27u8;
uint8x8_t d28u8, d29u8, d30u8, d31u8;
int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
uint8x16_t q3u8, q4u8;
uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16, q8u16, q9u16, q10u16;
uint16x8_t q11u16, q12u16, q13u16, q15u16;
int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16, q8s16, q9s16, q10s16;
int16x8_t q11s16, q12s16, q13s16, q15s16;
if (xoffset == 0) { // secondpass_filter8x8_only
// load second_pass filter
dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
// load src data
src_tmp = src_ptr - src_pixels_per_line * 2;
for (i = 0; i < 2; i++) {
src = src_tmp + i * 8;
dst = dst_ptr + i * 8;
d18u8 = vld1_u8(src);
src += src_pixels_per_line;
d19u8 = vld1_u8(src);
src += src_pixels_per_line;
d20u8 = vld1_u8(src);
src += src_pixels_per_line;
d21u8 = vld1_u8(src);
src += src_pixels_per_line;
d22u8 = vld1_u8(src);
src += src_pixels_per_line;
for (j = 0; j < 4; j++) {
d23u8 = vld1_u8(src);
src += src_pixels_per_line;
d24u8 = vld1_u8(src);
src += src_pixels_per_line;
d25u8 = vld1_u8(src);
src += src_pixels_per_line;
d26u8 = vld1_u8(src);
src += src_pixels_per_line;
q3u16 = vmull_u8(d18u8, d0u8);
q4u16 = vmull_u8(d19u8, d0u8);
q5u16 = vmull_u8(d20u8, d0u8);
q6u16 = vmull_u8(d21u8, d0u8);
q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
q7u16 = vmull_u8(d21u8, d3u8);
q8u16 = vmull_u8(d22u8, d3u8);
q9u16 = vmull_u8(d23u8, d3u8);
q10u16 = vmull_u8(d24u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d6u8 = vqrshrun_n_s16(q7s16, 7);
d7u8 = vqrshrun_n_s16(q8s16, 7);
d8u8 = vqrshrun_n_s16(q9s16, 7);
d9u8 = vqrshrun_n_s16(q10s16, 7);
d18u8 = d22u8;
d19u8 = d23u8;
d20u8 = d24u8;
d21u8 = d25u8;
d22u8 = d26u8;
vst1_u8(dst, d6u8);
dst += dst_pitch;
vst1_u8(dst, d7u8);
dst += dst_pitch;
vst1_u8(dst, d8u8);
dst += dst_pitch;
vst1_u8(dst, d9u8);
dst += dst_pitch;
}
}
return;
}
// load first_pass filter
dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
// First pass: output_height lines x output_width columns (9x4)
if (yoffset == 0) { // firstpass_filter4x4_only
src = src_ptr - 2;
dst = dst_ptr;
for (i = 0; i < 8; i++) {
d6u8 = vld1_u8(src);
d7u8 = vld1_u8(src + 8);
d8u8 = vld1_u8(src + 16);
src += src_pixels_per_line;
d9u8 = vld1_u8(src);
d10u8 = vld1_u8(src + 8);
d11u8 = vld1_u8(src + 16);
src += src_pixels_per_line;
__builtin_prefetch(src);
__builtin_prefetch(src + src_pixels_per_line);
q6u16 = vmull_u8(d6u8, d0u8);
q7u16 = vmull_u8(d7u8, d0u8);
q8u16 = vmull_u8(d9u8, d0u8);
q9u16 = vmull_u8(d10u8, d0u8);
d20u8 = vext_u8(d6u8, d7u8, 1);
d21u8 = vext_u8(d9u8, d10u8, 1);
d22u8 = vext_u8(d7u8, d8u8, 1);
d23u8 = vext_u8(d10u8, d11u8, 1);
d24u8 = vext_u8(d6u8, d7u8, 4);
d25u8 = vext_u8(d9u8, d10u8, 4);
d26u8 = vext_u8(d7u8, d8u8, 4);
d27u8 = vext_u8(d10u8, d11u8, 4);
d28u8 = vext_u8(d6u8, d7u8, 5);
d29u8 = vext_u8(d9u8, d10u8, 5);
q6u16 = vmlsl_u8(q6u16, d20u8, d1u8);
q8u16 = vmlsl_u8(q8u16, d21u8, d1u8);
q7u16 = vmlsl_u8(q7u16, d22u8, d1u8);
q9u16 = vmlsl_u8(q9u16, d23u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d24u8, d4u8);
q8u16 = vmlsl_u8(q8u16, d25u8, d4u8);
q7u16 = vmlsl_u8(q7u16, d26u8, d4u8);
q9u16 = vmlsl_u8(q9u16, d27u8, d4u8);
q6u16 = vmlal_u8(q6u16, d28u8, d5u8);
q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
d20u8 = vext_u8(d7u8, d8u8, 5);
d21u8 = vext_u8(d10u8, d11u8, 5);
d22u8 = vext_u8(d6u8, d7u8, 2);
d23u8 = vext_u8(d9u8, d10u8, 2);
d24u8 = vext_u8(d7u8, d8u8, 2);
d25u8 = vext_u8(d10u8, d11u8, 2);
d26u8 = vext_u8(d6u8, d7u8, 3);
d27u8 = vext_u8(d9u8, d10u8, 3);
d28u8 = vext_u8(d7u8, d8u8, 3);
d29u8 = vext_u8(d10u8, d11u8, 3);
q7u16 = vmlal_u8(q7u16, d20u8, d5u8);
q9u16 = vmlal_u8(q9u16, d21u8, d5u8);
q6u16 = vmlal_u8(q6u16, d22u8, d2u8);
q8u16 = vmlal_u8(q8u16, d23u8, d2u8);
q7u16 = vmlal_u8(q7u16, d24u8, d2u8);
q9u16 = vmlal_u8(q9u16, d25u8, d2u8);
q10u16 = vmull_u8(d26u8, d3u8);
q11u16 = vmull_u8(d27u8, d3u8);
q12u16 = vmull_u8(d28u8, d3u8);
q15u16 = vmull_u8(d29u8, d3u8);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q11s16 = vreinterpretq_s16_u16(q11u16);
q12s16 = vreinterpretq_s16_u16(q12u16);
q15s16 = vreinterpretq_s16_u16(q15u16);
q6s16 = vqaddq_s16(q6s16, q10s16);
q8s16 = vqaddq_s16(q8s16, q11s16);
q7s16 = vqaddq_s16(q7s16, q12s16);
q9s16 = vqaddq_s16(q9s16, q15s16);
d6u8 = vqrshrun_n_s16(q6s16, 7);
d7u8 = vqrshrun_n_s16(q7s16, 7);
d8u8 = vqrshrun_n_s16(q8s16, 7);
d9u8 = vqrshrun_n_s16(q9s16, 7);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
vst1q_u8(dst, q3u8);
dst += dst_pitch;
vst1q_u8(dst, q4u8);
dst += dst_pitch;
}
return;
}
src = src_ptr - 2 - src_pixels_per_line * 2;
tmpp = tmp;
for (i = 0; i < 7; i++) {
d6u8 = vld1_u8(src);
d7u8 = vld1_u8(src + 8);
d8u8 = vld1_u8(src + 16);
src += src_pixels_per_line;
d9u8 = vld1_u8(src);
d10u8 = vld1_u8(src + 8);
d11u8 = vld1_u8(src + 16);
src += src_pixels_per_line;
d12u8 = vld1_u8(src);
d13u8 = vld1_u8(src + 8);
d14u8 = vld1_u8(src + 16);
src += src_pixels_per_line;
__builtin_prefetch(src);
__builtin_prefetch(src + src_pixels_per_line);
__builtin_prefetch(src + src_pixels_per_line * 2);
q8u16 = vmull_u8(d6u8, d0u8);
q9u16 = vmull_u8(d7u8, d0u8);
q10u16 = vmull_u8(d9u8, d0u8);
q11u16 = vmull_u8(d10u8, d0u8);
q12u16 = vmull_u8(d12u8, d0u8);
q13u16 = vmull_u8(d13u8, d0u8);
d28u8 = vext_u8(d6u8, d7u8, 1);
d29u8 = vext_u8(d9u8, d10u8, 1);
d30u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlsl_u8(q8u16, d28u8, d1u8);
q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
q12u16 = vmlsl_u8(q12u16, d30u8, d1u8);
d28u8 = vext_u8(d7u8, d8u8, 1);
d29u8 = vext_u8(d10u8, d11u8, 1);
d30u8 = vext_u8(d13u8, d14u8, 1);
q9u16 = vmlsl_u8(q9u16, d28u8, d1u8);
q11u16 = vmlsl_u8(q11u16, d29u8, d1u8);
q13u16 = vmlsl_u8(q13u16, d30u8, d1u8);
d28u8 = vext_u8(d6u8, d7u8, 4);
d29u8 = vext_u8(d9u8, d10u8, 4);
d30u8 = vext_u8(d12u8, d13u8, 4);
q8u16 = vmlsl_u8(q8u16, d28u8, d4u8);
q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
q12u16 = vmlsl_u8(q12u16, d30u8, d4u8);
d28u8 = vext_u8(d7u8, d8u8, 4);
d29u8 = vext_u8(d10u8, d11u8, 4);
d30u8 = vext_u8(d13u8, d14u8, 4);
q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
q11u16 = vmlsl_u8(q11u16, d29u8, d4u8);
q13u16 = vmlsl_u8(q13u16, d30u8, d4u8);
d28u8 = vext_u8(d6u8, d7u8, 5);
d29u8 = vext_u8(d9u8, d10u8, 5);
d30u8 = vext_u8(d12u8, d13u8, 5);
q8u16 = vmlal_u8(q8u16, d28u8, d5u8);
q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
q12u16 = vmlal_u8(q12u16, d30u8, d5u8);
d28u8 = vext_u8(d7u8, d8u8, 5);
d29u8 = vext_u8(d10u8, d11u8, 5);
d30u8 = vext_u8(d13u8, d14u8, 5);
q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
q11u16 = vmlal_u8(q11u16, d29u8, d5u8);
q13u16 = vmlal_u8(q13u16, d30u8, d5u8);
d28u8 = vext_u8(d6u8, d7u8, 2);
d29u8 = vext_u8(d9u8, d10u8, 2);
d30u8 = vext_u8(d12u8, d13u8, 2);
q8u16 = vmlal_u8(q8u16, d28u8, d2u8);
q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
q12u16 = vmlal_u8(q12u16, d30u8, d2u8);
d28u8 = vext_u8(d7u8, d8u8, 2);
d29u8 = vext_u8(d10u8, d11u8, 2);
d30u8 = vext_u8(d13u8, d14u8, 2);
q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
q11u16 = vmlal_u8(q11u16, d29u8, d2u8);
q13u16 = vmlal_u8(q13u16, d30u8, d2u8);
d28u8 = vext_u8(d6u8, d7u8, 3);
d29u8 = vext_u8(d9u8, d10u8, 3);
d30u8 = vext_u8(d12u8, d13u8, 3);
d15u8 = vext_u8(d7u8, d8u8, 3);
d31u8 = vext_u8(d10u8, d11u8, 3);
d6u8 = vext_u8(d13u8, d14u8, 3);
q4u16 = vmull_u8(d28u8, d3u8);
q5u16 = vmull_u8(d29u8, d3u8);
q6u16 = vmull_u8(d30u8, d3u8);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q12s16 = vreinterpretq_s16_u16(q12u16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q10s16 = vqaddq_s16(q10s16, q5s16);
q12s16 = vqaddq_s16(q12s16, q6s16);
q6u16 = vmull_u8(d15u8, d3u8);
q7u16 = vmull_u8(d31u8, d3u8);
q3u16 = vmull_u8(d6u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q11s16 = vreinterpretq_s16_u16(q11u16);
q13s16 = vreinterpretq_s16_u16(q13u16);
q9s16 = vqaddq_s16(q9s16, q6s16);
q11s16 = vqaddq_s16(q11s16, q7s16);
q13s16 = vqaddq_s16(q13s16, q3s16);
d6u8 = vqrshrun_n_s16(q8s16, 7);
d7u8 = vqrshrun_n_s16(q9s16, 7);
d8u8 = vqrshrun_n_s16(q10s16, 7);
d9u8 = vqrshrun_n_s16(q11s16, 7);
d10u8 = vqrshrun_n_s16(q12s16, 7);
d11u8 = vqrshrun_n_s16(q13s16, 7);
vst1_u8(tmpp, d6u8);
tmpp += 8;
vst1_u8(tmpp, d7u8);
tmpp += 8;
vst1_u8(tmpp, d8u8);
tmpp += 8;
vst1_u8(tmpp, d9u8);
tmpp += 8;
vst1_u8(tmpp, d10u8);
tmpp += 8;
vst1_u8(tmpp, d11u8);
tmpp += 8;
}
// Second pass: 16x16
dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
for (i = 0; i < 2; i++) {
dst = dst_ptr + 8 * i;
tmpp = tmp + 8 * i;
d18u8 = vld1_u8(tmpp);
tmpp += 16;
d19u8 = vld1_u8(tmpp);
tmpp += 16;
d20u8 = vld1_u8(tmpp);
tmpp += 16;
d21u8 = vld1_u8(tmpp);
tmpp += 16;
d22u8 = vld1_u8(tmpp);
tmpp += 16;
for (j = 0; j < 4; j++) {
d23u8 = vld1_u8(tmpp);
tmpp += 16;
d24u8 = vld1_u8(tmpp);
tmpp += 16;
d25u8 = vld1_u8(tmpp);
tmpp += 16;
d26u8 = vld1_u8(tmpp);
tmpp += 16;
q3u16 = vmull_u8(d18u8, d0u8);
q4u16 = vmull_u8(d19u8, d0u8);
q5u16 = vmull_u8(d20u8, d0u8);
q6u16 = vmull_u8(d21u8, d0u8);
q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
q7u16 = vmull_u8(d21u8, d3u8);
q8u16 = vmull_u8(d22u8, d3u8);
q9u16 = vmull_u8(d23u8, d3u8);
q10u16 = vmull_u8(d24u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d6u8 = vqrshrun_n_s16(q7s16, 7);
d7u8 = vqrshrun_n_s16(q8s16, 7);
d8u8 = vqrshrun_n_s16(q9s16, 7);
d9u8 = vqrshrun_n_s16(q10s16, 7);
d18u8 = d22u8;
d19u8 = d23u8;
d20u8 = d24u8;
d21u8 = d25u8;
d22u8 = d26u8;
vst1_u8(dst, d6u8);
dst += dst_pitch;
vst1_u8(dst, d7u8);
dst += dst_pitch;
vst1_u8(dst, d8u8);
dst += dst_pitch;
vst1_u8(dst, d9u8);
dst += dst_pitch;
}
}
return;
}
inline void vnst(u8* dst, uint32x4_t v1, uint32x4_t v2) { vst1_u8(dst, vmovn_u16(vcombine_u16(vmovn_u32(v1), vmovn_u32(v2)))); }
void phase(const Size2D &size,
const s16 * src0Base, ptrdiff_t src0Stride,
const s16 * src1Base, ptrdiff_t src1Stride,
u8 * dstBase, ptrdiff_t dstStride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
FASTATAN2CONST(256.0f / 360.0f)
size_t roiw16 = size.width >= 15 ? size.width - 15 : 0;
size_t roiw8 = size.width >= 7 ? size.width - 7 : 0;
float32x4_t v_05 = vdupq_n_f32(0.5f);
for (size_t i = 0; i < size.height; ++i)
{
const s16 * src0 = internal::getRowPtr(src0Base, src0Stride, i);
const s16 * src1 = internal::getRowPtr(src1Base, src1Stride, i);
u8 * dst = internal::getRowPtr(dstBase, dstStride, i);
size_t j = 0;
for (; j < roiw16; j += 16)
{
internal::prefetch(src0 + j);
internal::prefetch(src1 + j);
int16x8_t v_src00 = vld1q_s16(src0 + j), v_src01 = vld1q_s16(src0 + j + 8);
int16x8_t v_src10 = vld1q_s16(src1 + j), v_src11 = vld1q_s16(src1 + j + 8);
// 0
float32x4_t v_src0_p = vcvtq_f32_s32(vmovl_s16(vget_low_s16(v_src00)));
float32x4_t v_src1_p = vcvtq_f32_s32(vmovl_s16(vget_low_s16(v_src10)));
float32x4_t v_dst32f0;
FASTATAN2VECTOR(v_src1_p, v_src0_p, v_dst32f0)
v_src0_p = vcvtq_f32_s32(vmovl_s16(vget_high_s16(v_src00)));
v_src1_p = vcvtq_f32_s32(vmovl_s16(vget_high_s16(v_src10)));
float32x4_t v_dst32f1;
FASTATAN2VECTOR(v_src1_p, v_src0_p, v_dst32f1)
uint16x8_t v_dst16s0 = vcombine_u16(vmovn_u32(vcvtq_u32_f32(vaddq_f32(v_dst32f0, v_05))),
vmovn_u32(vcvtq_u32_f32(vaddq_f32(v_dst32f1, v_05))));
// 1
v_src0_p = vcvtq_f32_s32(vmovl_s16(vget_low_s16(v_src01)));
v_src1_p = vcvtq_f32_s32(vmovl_s16(vget_low_s16(v_src11)));
FASTATAN2VECTOR(v_src1_p, v_src0_p, v_dst32f0)
v_src0_p = vcvtq_f32_s32(vmovl_s16(vget_high_s16(v_src01)));
v_src1_p = vcvtq_f32_s32(vmovl_s16(vget_high_s16(v_src11)));
FASTATAN2VECTOR(v_src1_p, v_src0_p, v_dst32f1)
uint16x8_t v_dst16s1 = vcombine_u16(vmovn_u32(vcvtq_u32_f32(vaddq_f32(v_dst32f0, v_05))),
vmovn_u32(vcvtq_u32_f32(vaddq_f32(v_dst32f1, v_05))));
vst1q_u8(dst + j, vcombine_u8(vmovn_u16(v_dst16s0),
vmovn_u16(v_dst16s1)));
}
for (; j < roiw8; j += 8)
{
int16x8_t v_src0 = vld1q_s16(src0 + j);
int16x8_t v_src1 = vld1q_s16(src1 + j);
float32x4_t v_src0_p = vcvtq_f32_s32(vmovl_s16(vget_low_s16(v_src0)));
float32x4_t v_src1_p = vcvtq_f32_s32(vmovl_s16(vget_low_s16(v_src1)));
float32x4_t v_dst32f0;
FASTATAN2VECTOR(v_src1_p, v_src0_p, v_dst32f0)
v_src0_p = vcvtq_f32_s32(vmovl_s16(vget_high_s16(v_src0)));
v_src1_p = vcvtq_f32_s32(vmovl_s16(vget_high_s16(v_src1)));
float32x4_t v_dst32f1;
FASTATAN2VECTOR(v_src1_p, v_src0_p, v_dst32f1)
uint16x8_t v_dst = vcombine_u16(vmovn_u32(vcvtq_u32_f32(vaddq_f32(v_dst32f0, v_05))),
vmovn_u32(vcvtq_u32_f32(vaddq_f32(v_dst32f1, v_05))));
vst1_u8(dst + j, vmovn_u16(v_dst));
}
for (; j < size.width; j++)
{
f32 x = src0[j], y = src1[j];
f32 a;
FASTATAN2SCALAR(y, x, a)
dst[j] = (u8)(s32)floor(a + 0.5f);
}
}
#else
(void)size;
(void)src0Base;
(void)src0Stride;
(void)src1Base;
(void)src1Stride;
(void)dstBase;
(void)dstStride;
#endif
}