static INLINE void IDCT4x4_1D(int16x4_t *d0s16, int16x4_t *d1s16,
int16x4_t *d2s16, int16x8_t *q8s16,
int16x8_t *q9s16) {
int16x4_t d16s16, d17s16, d18s16, d19s16, d23s16, d24s16;
int16x4_t d26s16, d27s16, d28s16, d29s16;
int32x4_t q10s32, q13s32, q14s32, q15s32;
int16x8_t q13s16, q14s16;
d16s16 = vget_low_s16(*q8s16);
d17s16 = vget_high_s16(*q8s16);
d18s16 = vget_low_s16(*q9s16);
d19s16 = vget_high_s16(*q9s16);
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, *d2s16);
q10s32 = vmull_s16(d17s16, *d0s16);
q13s32 = vmull_s16(d23s16, *d1s16);
q14s32 = vmull_s16(d24s16, *d1s16);
q15s32 = vmlsl_s16(q15s32, d19s16, *d0s16);
q10s32 = vmlal_s16(q10s32, d19s16, *d2s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q10s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
*q8s16 = vaddq_s16(q13s16, q14s16);
*q9s16 = vsubq_s16(q13s16, q14s16);
*q9s16 = vcombine_s16(vget_high_s16(*q9s16), vget_low_s16(*q9s16)); // vswp
return;
}
int64_t av1_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff,
int block_size) {
int64x2_t error = vdupq_n_s64(0);
assert(block_size >= 8);
assert((block_size % 8) == 0);
do {
const int16x8_t c = vld1q_s16(coeff);
const int16x8_t d = vld1q_s16(dqcoeff);
const int16x8_t diff = vsubq_s16(c, d);
const int16x4_t diff_lo = vget_low_s16(diff);
const int16x4_t diff_hi = vget_high_s16(diff);
// diff is 15-bits, the squares 30, so we can store 2 in 31-bits before
// accumulating them in 64-bits.
const int32x4_t err0 = vmull_s16(diff_lo, diff_lo);
const int32x4_t err1 = vmlal_s16(err0, diff_hi, diff_hi);
const int64x2_t err2 = vaddl_s32(vget_low_s32(err1), vget_high_s32(err1));
error = vaddq_s64(error, err2);
coeff += 8;
dqcoeff += 8;
block_size -= 8;
} while (block_size != 0);
return vgetq_lane_s64(error, 0) + vgetq_lane_s64(error, 1);
}
static uint64_t SharpYUVUpdateY_NEON(const uint16_t* ref, const uint16_t* src,
uint16_t* dst, int len) {
int i;
const int16x8_t zero = vdupq_n_s16(0);
const int16x8_t max = vdupq_n_s16(MAX_Y);
uint64x2_t sum = vdupq_n_u64(0);
uint64_t diff;
for (i = 0; i + 8 <= len; i += 8) {
const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i));
const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i));
const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i));
const int16x8_t D = vsubq_s16(A, B); // diff_y
const int16x8_t F = vaddq_s16(C, D); // new_y
const uint16x8_t H =
vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero));
const int16x8_t I = vabsq_s16(D); // abs(diff_y)
vst1q_u16(dst + i, H);
sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I)));
}
diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1);
for (; i < len; ++i) {
const int diff_y = ref[i] - src[i];
const int new_y = (int)(dst[i]) + diff_y;
dst[i] = clip_y(new_y);
diff += (uint64_t)(abs(diff_y));
}
return diff;
}
static inline void char_to_float_vectors(const unsigned char * sourcep,
float32x4_t *mp0, float32x4_t * mp1)
{
uint8x8_t rawpixels; /* source pixels as {[YUYV]0 [YUYV]1} */
int16x8_t widerpixels; /* rawpixels promoted to shorts per component */
int16x4_t high16, low16;
int32x4_t high32, low32;
const int16x8_t uvbias = {0, 128, 0, 128, 0, 128, 0, 128};
rawpixels = vld1_u8(sourcep);
widerpixels = vreinterpretq_s16_u16(vmovl_u8(rawpixels));
/* subtract uvbias from widerpixels */
widerpixels = vsubq_s16(widerpixels, uvbias);
/* now take widerpixels apart into (low16, high16) and */
/* then expand those into (low32, high32) */
low16 = vget_low_s16(widerpixels);
high16 = vget_high_s16(widerpixels);
high32 = vmovl_s16(high16);
low32 = vmovl_s16(low16);
/* now convert low32 and high32 into floats and store them in */
/* *mp0, *mp1 */
*mp0 = vcvtq_f32_s32(low32);
*mp1 = vcvtq_f32_s32(high32);
}
void test_vsubQs16 (void)
{
int16x8_t out_int16x8_t;
int16x8_t arg0_int16x8_t;
int16x8_t arg1_int16x8_t;
out_int16x8_t = vsubq_s16 (arg0_int16x8_t, arg1_int16x8_t);
}
static inline int16x8_t qvsource_over_s16(int16x8_t src16, int16x8_t dst16, int16x8_t half, int16x8_t full)
{
const int16x4_t alpha16_high = vdup_lane_s16(vget_high_s16(src16), 3);
const int16x4_t alpha16_low = vdup_lane_s16(vget_low_s16(src16), 3);
const int16x8_t alpha16 = vsubq_s16(full, vcombine_s16(alpha16_low, alpha16_high));
return vaddq_s16(src16, qvbyte_mul_s16(dst16, alpha16, half));
}
void aom_hadamard_16x16_neon(const int16_t *src_diff, int src_stride,
int16_t *coeff) {
int i;
/* Rearrange 16x16 to 8x32 and remove stride.
* Top left first. */
aom_hadamard_8x8_neon(src_diff + 0 + 0 * src_stride, src_stride, coeff + 0);
/* Top right. */
aom_hadamard_8x8_neon(src_diff + 8 + 0 * src_stride, src_stride, coeff + 64);
/* Bottom left. */
aom_hadamard_8x8_neon(src_diff + 0 + 8 * src_stride, src_stride, coeff + 128);
/* Bottom right. */
aom_hadamard_8x8_neon(src_diff + 8 + 8 * src_stride, src_stride, coeff + 192);
for (i = 0; i < 64; i += 8) {
const int16x8_t a0 = vld1q_s16(coeff + 0);
const int16x8_t a1 = vld1q_s16(coeff + 64);
const int16x8_t a2 = vld1q_s16(coeff + 128);
const int16x8_t a3 = vld1q_s16(coeff + 192);
const int16x8_t b0 = vhaddq_s16(a0, a1);
const int16x8_t b1 = vhsubq_s16(a0, a1);
const int16x8_t b2 = vhaddq_s16(a2, a3);
const int16x8_t b3 = vhsubq_s16(a2, a3);
const int16x8_t c0 = vaddq_s16(b0, b2);
const int16x8_t c1 = vaddq_s16(b1, b3);
const int16x8_t c2 = vsubq_s16(b0, b2);
const int16x8_t c3 = vsubq_s16(b1, b3);
vst1q_s16(coeff + 0, c0);
vst1q_s16(coeff + 64, c1);
vst1q_s16(coeff + 128, c2);
vst1q_s16(coeff + 192, c3);
coeff += 8;
}
}
void rfx_decode_YCbCr_to_RGB_NEON(sint16 * y_r_buffer, sint16 * cb_g_buffer, sint16 * cr_b_buffer)
{
int16x8_t zero = vdupq_n_s16(0);
int16x8_t max = vdupq_n_s16(255);
int16x8_t y_add = vdupq_n_s16(128);
int16x8_t* y_r_buf = (int16x8_t*)y_r_buffer;
int16x8_t* cb_g_buf = (int16x8_t*)cb_g_buffer;
int16x8_t* cr_b_buf = (int16x8_t*)cr_b_buffer;
int i;
for (i = 0; i < 4096 / 8; i++)
{
int16x8_t y = vld1q_s16((sint16*)&y_r_buf[i]);
y = vaddq_s16(y, y_add);
int16x8_t cr = vld1q_s16((sint16*)&cr_b_buf[i]);
// r = between((y + cr + (cr >> 2) + (cr >> 3) + (cr >> 5)), 0, 255);
int16x8_t r = vaddq_s16(y, cr);
r = vaddq_s16(r, vshrq_n_s16(cr, 2));
r = vaddq_s16(r, vshrq_n_s16(cr, 3));
r = vaddq_s16(r, vshrq_n_s16(cr, 5));
r = vminq_s16(vmaxq_s16(r, zero), max);
vst1q_s16((sint16*)&y_r_buf[i], r);
// cb = cb_g_buf[i];
int16x8_t cb = vld1q_s16((sint16*)&cb_g_buf[i]);
// g = between(y - (cb >> 2) - (cb >> 4) - (cb >> 5) - (cr >> 1) - (cr >> 3) - (cr >> 4) - (cr >> 5), 0, 255);
int16x8_t g = vsubq_s16(y, vshrq_n_s16(cb, 2));
g = vsubq_s16(g, vshrq_n_s16(cb, 4));
g = vsubq_s16(g, vshrq_n_s16(cb, 5));
g = vsubq_s16(g, vshrq_n_s16(cr, 1));
g = vsubq_s16(g, vshrq_n_s16(cr, 3));
g = vsubq_s16(g, vshrq_n_s16(cr, 4));
g = vsubq_s16(g, vshrq_n_s16(cr, 5));
g = vminq_s16(vmaxq_s16(g, zero), max);
vst1q_s16((sint16*)&cb_g_buf[i], g);
// b = between((y + cb + (cb >> 1) + (cb >> 2) + (cb >> 6)), 0, 255);
int16x8_t b = vaddq_s16(y, cb);
b = vaddq_s16(b, vshrq_n_s16(cb, 1));
b = vaddq_s16(b, vshrq_n_s16(cb, 2));
b = vaddq_s16(b, vshrq_n_s16(cb, 6));
b = vminq_s16(vmaxq_s16(b, zero), max);
vst1q_s16((sint16*)&cr_b_buf[i], b);
}
}
static void SharpYUVUpdateRGB_NEON(const int16_t* ref, const int16_t* src,
int16_t* dst, int len) {
int i;
for (i = 0; i + 8 <= len; i += 8) {
const int16x8_t A = vld1q_s16(ref + i);
const int16x8_t B = vld1q_s16(src + i);
const int16x8_t C = vld1q_s16(dst + i);
const int16x8_t D = vsubq_s16(A, B); // diff_uv
const int16x8_t E = vaddq_s16(C, D); // new_uv
vst1q_s16(dst + i, E);
}
for (; i < len; ++i) {
const int diff_uv = ref[i] - src[i];
dst[i] += diff_uv;
}
}
// ref, src = [0, 510] - max diff = 16-bits
// bwl = {2, 3, 4}, width = {16, 32, 64}
int vp9_vector_var_neon(int16_t const *ref, int16_t const *src, const int bwl) {
int width = 4 << bwl;
int32x4_t sse = vdupq_n_s32(0);
int16x8_t total = vdupq_n_s16(0);
assert(width >= 8);
assert((width % 8) == 0);
do {
const int16x8_t r = vld1q_s16(ref);
const int16x8_t s = vld1q_s16(src);
const int16x8_t diff = vsubq_s16(r, s); // [-510, 510], 10 bits.
const int16x4_t diff_lo = vget_low_s16(diff);
const int16x4_t diff_hi = vget_high_s16(diff);
sse = vmlal_s16(sse, diff_lo, diff_lo); // dynamic range 26 bits.
sse = vmlal_s16(sse, diff_hi, diff_hi);
total = vaddq_s16(total, diff); // dynamic range 16 bits.
ref += 8;
src += 8;
width -= 8;
} while (width != 0);
{
// Note: 'total''s pairwise addition could be implemented similarly to
// horizontal_add_u16x8(), but one less vpaddl with 'total' when paired
// with the summation of 'sse' performed better on a Cortex-A15.
const int32x4_t t0 = vpaddlq_s16(total); // cascading summation of 'total'
const int32x2_t t1 = vadd_s32(vget_low_s32(t0), vget_high_s32(t0));
const int32x2_t t2 = vpadd_s32(t1, t1);
const int t = vget_lane_s32(t2, 0);
const int64x2_t s0 = vpaddlq_s32(sse); // cascading summation of 'sse'.
const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
vreinterpret_s32_s64(vget_high_s64(s0)));
const int s = vget_lane_s32(s1, 0);
const int shift_factor = bwl + 2;
return s - ((t * t) >> shift_factor);
}
}
/* s16x8 sub */
void mw_neon_mm_sub_s16x8(short * A, int Row, int Col, short * B, short * C)
{
int16x8_t neon_a, neon_b, neon_c;
int size = Row * Col;
int i = 0;
int k = 0;
for (i = 8; i <= size ; i+=8)
{
k = i - 8;
neon_a = vld1q_s16(A + k);
neon_b = vld1q_s16(B + k);
neon_c = vsubq_s16(neon_a, neon_b);
vst1q_s16(C + k, neon_c);
}
k = i - 8;
for (i = 0; i < size % 8; i++)
{
C[k + i] = A[k + i] - B[k + i];
}
}
rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband_width)
{
int x, n;
INT16 * l_ptr = l;
INT16 * h_ptr = h;
INT16 * dst_ptr = dst;
int total_width = subband_width + subband_width;
/* Even coefficients */
for (n = 0; n < subband_width; n++)
{
for (x = 0; x < total_width; x+=8)
{
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t tmp_n = vaddq_s16(h_n, vdupq_n_s16(1));;
if (n == 0)
tmp_n = vaddq_s16(tmp_n, h_n);
else
{
int16x8_t h_n_m = vld1q_s16((h_ptr - total_width));
tmp_n = vaddq_s16(tmp_n, h_n_m);
}
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vsubq_s16(l_n, tmp_n);
vst1q_s16(dst_ptr, dst_n);
l_ptr+=8;
h_ptr+=8;
dst_ptr+=8;
}
dst_ptr+=total_width;
}
h_ptr = h;
dst_ptr = dst + total_width;
/* Odd coefficients */
for (n = 0; n < subband_width; n++)
{
for (x = 0; x < total_width; x+=8)
{
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t dst_n_m = vld1q_s16(dst_ptr - total_width);
h_n = vshlq_n_s16(h_n, 1);
int16x8_t tmp_n = dst_n_m;
if (n == subband_width - 1)
tmp_n = vaddq_s16(tmp_n, dst_n_m);
else
{
int16x8_t dst_n_p = vld1q_s16((dst_ptr + total_width));
tmp_n = vaddq_s16(tmp_n, dst_n_p);
}
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vaddq_s16(tmp_n, h_n);
vst1q_s16(dst_ptr, dst_n);
h_ptr+=8;
dst_ptr+=8;
}
dst_ptr+=total_width;
}
}
rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband_width)
{
int y, n;
INT16 * l_ptr = l;
INT16 * h_ptr = h;
INT16 * dst_ptr = dst;
for (y = 0; y < subband_width; y++)
{
/* Even coefficients */
for (n = 0; n < subband_width; n+=8)
{
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t h_n_m = vld1q_s16(h_ptr - 1);
if (n == 0)
{
int16_t first = vgetq_lane_s16(h_n_m, 1);
h_n_m = vsetq_lane_s16(first, h_n_m, 0);
}
int16x8_t tmp_n = vaddq_s16(h_n, h_n_m);
tmp_n = vaddq_s16(tmp_n, vdupq_n_s16(1));
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vsubq_s16(l_n, tmp_n);
vst1q_s16(l_ptr, dst_n);
l_ptr+=8;
h_ptr+=8;
}
l_ptr -= subband_width;
h_ptr -= subband_width;
/* Odd coefficients */
for (n = 0; n < subband_width; n+=8)
{
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr);
h_n = vshlq_n_s16(h_n, 1);
int16x8x2_t dst_n;
dst_n.val[0] = vld1q_s16(l_ptr);
int16x8_t dst_n_p = vld1q_s16(l_ptr + 1);
if (n == subband_width - 8)
{
int16_t last = vgetq_lane_s16(dst_n_p, 6);
dst_n_p = vsetq_lane_s16(last, dst_n_p, 7);
}
dst_n.val[1] = vaddq_s16(dst_n_p, dst_n.val[0]);
dst_n.val[1] = vshrq_n_s16(dst_n.val[1], 1);
dst_n.val[1] = vaddq_s16(dst_n.val[1], h_n);
vst2q_s16(dst_ptr, dst_n);
l_ptr+=8;
h_ptr+=8;
dst_ptr+=16;
}
}
}
void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
uint8x8_t d26u8, d27u8;
uint32x2_t d26u32, d27u32;
uint16x8_t q8u16, q9u16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
int16x8_t q8s16, q9s16, q13s16, q14s16;
int32x4_t q1s32, q13s32, q14s32, q15s32;
int16x4x2_t d0x2s16, d1x2s16;
int32x4x2_t q0x2s32;
uint8_t *d;
d26u32 = d27u32 = vdup_n_u32(0);
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_low_s16(q9s16);
d19s16 = vget_high_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
d20s16 = vdup_n_s16((int16_t)cospi_8_64);
d21s16 = vdup_n_s16((int16_t)cospi_16_64);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d22s16 = vdup_n_s16((int16_t)cospi_24_64);
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_high_s16(q9s16); // vswp d18 d19
d19s16 = vget_low_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
// do the transform on columns
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
q8s16 = vrshrq_n_s16(q8s16, 4);
q9s16 = vrshrq_n_s16(q9s16, 4);
d = dest;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
d += dest_stride;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d = dest;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
return;
}
PRIM_STATIC pstatus_t neon_yCbCrToRGB_16s16s_P3P3(
const INT16 *pSrc[3],
int srcStep,
INT16 *pDst[3],
int dstStep,
const prim_size_t *roi) /* region of interest */
{
/* TODO: If necessary, check alignments and call the general version. */
int16x8_t zero = vdupq_n_s16(0);
int16x8_t max = vdupq_n_s16(255);
int16x8_t y_add = vdupq_n_s16(128);
int16x8_t* y_buf = (int16x8_t*) pSrc[0];
int16x8_t* cb_buf = (int16x8_t*) pSrc[1];
int16x8_t* cr_buf = (int16x8_t*) pSrc[2];
int16x8_t* r_buf = (int16x8_t*) pDst[0];
int16x8_t* g_buf = (int16x8_t*) pDst[1];
int16x8_t* b_buf = (int16x8_t*) pDst[2];
int srcbump = srcStep / sizeof(int16x8_t);
int dstbump = dstStep / sizeof(int16x8_t);
int yp;
int imax = roi->width * sizeof(INT16) / sizeof(int16x8_t);
for (yp=0; yp<roi->height; ++yp)
{
int i;
for (i=0; i<imax; i++)
{
int16x8_t y = vld1q_s16((INT16*) (y_buf+i));
y = vaddq_s16(y, y_add);
int16x8_t cr = vld1q_s16((INT16*) (cr_buf+i));
/* r = between((y + cr + (cr >> 2) + (cr >> 3) + (cr >> 5)),
* 0, 255);
*/
int16x8_t r = vaddq_s16(y, cr);
r = vaddq_s16(r, vshrq_n_s16(cr, 2));
r = vaddq_s16(r, vshrq_n_s16(cr, 3));
r = vaddq_s16(r, vshrq_n_s16(cr, 5));
r = vminq_s16(vmaxq_s16(r, zero), max);
vst1q_s16((INT16*) (r_buf+i), r);
/* cb = cb_g_buf[i]; */
int16x8_t cb = vld1q_s16((INT16*) (cb_buf+i));
/* g = between(y - (cb >> 2) - (cb >> 4) - (cb >> 5) - (cr >> 1)
* - (cr >> 3) - (cr >> 4) - (cr >> 5), 0, 255);
*/
int16x8_t g = vsubq_s16(y, vshrq_n_s16(cb, 2));
g = vsubq_s16(g, vshrq_n_s16(cb, 4));
g = vsubq_s16(g, vshrq_n_s16(cb, 5));
g = vsubq_s16(g, vshrq_n_s16(cr, 1));
g = vsubq_s16(g, vshrq_n_s16(cr, 3));
g = vsubq_s16(g, vshrq_n_s16(cr, 4));
g = vsubq_s16(g, vshrq_n_s16(cr, 5));
g = vminq_s16(vmaxq_s16(g, zero), max);
vst1q_s16((INT16*) (g_buf+i), g);
/* b = between((y + cb + (cb >> 1) + (cb >> 2) + (cb >> 6)),
* 0, 255);
*/
int16x8_t b = vaddq_s16(y, cb);
b = vaddq_s16(b, vshrq_n_s16(cb, 1));
b = vaddq_s16(b, vshrq_n_s16(cb, 2));
b = vaddq_s16(b, vshrq_n_s16(cb, 6));
b = vminq_s16(vmaxq_s16(b, zero), max);
vst1q_s16((INT16*) (b_buf+i), b);
}
y_buf += srcbump;
cb_buf += srcbump;
cr_buf += srcbump;
r_buf += dstbump;
g_buf += dstbump;
b_buf += dstbump;
}
}
// vp9/common/vp9_scan.c:vp9_default_iscan_32x32 arranges the first 34 non-zero
// coefficients as follows:
// 0 1 2 3 4 5 6 7
// 0 0 2 5 10 17 25
// 1 1 4 8 15 22 30
// 2 3 7 12 18 28
// 3 6 11 16 23 31
// 4 9 14 19 29
// 5 13 20 26
// 6 21 27 33
// 7 24 32
void vpx_idct32_6_neon(const tran_low_t *input, int16_t *output) {
int16x8_t in[8], s1[32], s2[32], s3[32];
in[0] = load_tran_low_to_s16q(input);
input += 32;
in[1] = load_tran_low_to_s16q(input);
input += 32;
in[2] = load_tran_low_to_s16q(input);
input += 32;
in[3] = load_tran_low_to_s16q(input);
input += 32;
in[4] = load_tran_low_to_s16q(input);
input += 32;
in[5] = load_tran_low_to_s16q(input);
input += 32;
in[6] = load_tran_low_to_s16q(input);
input += 32;
in[7] = load_tran_low_to_s16q(input);
transpose_s16_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6],
&in[7]);
// stage 1
// input[1] * cospi_31_64 - input[31] * cospi_1_64 (but input[31] == 0)
s1[16] = multiply_shift_and_narrow_s16(in[1], cospi_31_64);
// input[1] * cospi_1_64 + input[31] * cospi_31_64 (but input[31] == 0)
s1[31] = multiply_shift_and_narrow_s16(in[1], cospi_1_64);
s1[20] = multiply_shift_and_narrow_s16(in[5], cospi_27_64);
s1[27] = multiply_shift_and_narrow_s16(in[5], cospi_5_64);
s1[23] = multiply_shift_and_narrow_s16(in[3], -cospi_29_64);
s1[24] = multiply_shift_and_narrow_s16(in[3], cospi_3_64);
// stage 2
s2[8] = multiply_shift_and_narrow_s16(in[2], cospi_30_64);
s2[15] = multiply_shift_and_narrow_s16(in[2], cospi_2_64);
// stage 3
s1[4] = multiply_shift_and_narrow_s16(in[4], cospi_28_64);
s1[7] = multiply_shift_and_narrow_s16(in[4], cospi_4_64);
s1[17] = multiply_accumulate_shift_and_narrow_s16(s1[16], -cospi_4_64, s1[31],
cospi_28_64);
s1[30] = multiply_accumulate_shift_and_narrow_s16(s1[16], cospi_28_64, s1[31],
cospi_4_64);
s1[21] = multiply_accumulate_shift_and_narrow_s16(s1[20], -cospi_20_64,
s1[27], cospi_12_64);
s1[26] = multiply_accumulate_shift_and_narrow_s16(s1[20], cospi_12_64, s1[27],
cospi_20_64);
s1[22] = multiply_accumulate_shift_and_narrow_s16(s1[23], -cospi_12_64,
s1[24], -cospi_20_64);
s1[25] = multiply_accumulate_shift_and_narrow_s16(s1[23], -cospi_20_64,
s1[24], cospi_12_64);
// stage 4
s1[0] = multiply_shift_and_narrow_s16(in[0], cospi_16_64);
s2[9] = multiply_accumulate_shift_and_narrow_s16(s2[8], -cospi_8_64, s2[15],
cospi_24_64);
s2[14] = multiply_accumulate_shift_and_narrow_s16(s2[8], cospi_24_64, s2[15],
cospi_8_64);
s2[20] = vsubq_s16(s1[23], s1[20]);
s2[21] = vsubq_s16(s1[22], s1[21]);
s2[22] = vaddq_s16(s1[21], s1[22]);
s2[23] = vaddq_s16(s1[20], s1[23]);
s2[24] = vaddq_s16(s1[24], s1[27]);
s2[25] = vaddq_s16(s1[25], s1[26]);
s2[26] = vsubq_s16(s1[25], s1[26]);
s2[27] = vsubq_s16(s1[24], s1[27]);
// stage 5
s1[5] = sub_multiply_shift_and_narrow_s16(s1[7], s1[4], cospi_16_64);
s1[6] = add_multiply_shift_and_narrow_s16(s1[4], s1[7], cospi_16_64);
s1[18] = multiply_accumulate_shift_and_narrow_s16(s1[17], -cospi_8_64, s1[30],
cospi_24_64);
s1[29] = multiply_accumulate_shift_and_narrow_s16(s1[17], cospi_24_64, s1[30],
cospi_8_64);
s1[19] = multiply_accumulate_shift_and_narrow_s16(s1[16], -cospi_8_64, s1[31],
cospi_24_64);
s1[28] = multiply_accumulate_shift_and_narrow_s16(s1[16], cospi_24_64, s1[31],
cospi_8_64);
s1[20] = multiply_accumulate_shift_and_narrow_s16(s2[20], -cospi_24_64,
s2[27], -cospi_8_64);
s1[27] = multiply_accumulate_shift_and_narrow_s16(s2[20], -cospi_8_64, s2[27],
cospi_24_64);
s1[21] = multiply_accumulate_shift_and_narrow_s16(s2[21], -cospi_24_64,
s2[26], -cospi_8_64);
s1[26] = multiply_accumulate_shift_and_narrow_s16(s2[21], -cospi_8_64, s2[26],
cospi_24_64);
// stage 6
s2[0] = vaddq_s16(s1[0], s1[7]);
s2[1] = vaddq_s16(s1[0], s1[6]);
s2[2] = vaddq_s16(s1[0], s1[5]);
s2[3] = vaddq_s16(s1[0], s1[4]);
s2[4] = vsubq_s16(s1[0], s1[4]);
s2[5] = vsubq_s16(s1[0], s1[5]);
s2[6] = vsubq_s16(s1[0], s1[6]);
s2[7] = vsubq_s16(s1[0], s1[7]);
s2[10] = sub_multiply_shift_and_narrow_s16(s2[14], s2[9], cospi_16_64);
s2[13] = add_multiply_shift_and_narrow_s16(s2[9], s2[14], cospi_16_64);
s2[11] = sub_multiply_shift_and_narrow_s16(s2[15], s2[8], cospi_16_64);
s2[12] = add_multiply_shift_and_narrow_s16(s2[8], s2[15], cospi_16_64);
s2[16] = vaddq_s16(s1[16], s2[23]);
s2[17] = vaddq_s16(s1[17], s2[22]);
s2[18] = vaddq_s16(s1[18], s1[21]);
s2[19] = vaddq_s16(s1[19], s1[20]);
s2[20] = vsubq_s16(s1[19], s1[20]);
s2[21] = vsubq_s16(s1[18], s1[21]);
s2[22] = vsubq_s16(s1[17], s2[22]);
s2[23] = vsubq_s16(s1[16], s2[23]);
s3[24] = vsubq_s16(s1[31], s2[24]);
s3[25] = vsubq_s16(s1[30], s2[25]);
s3[26] = vsubq_s16(s1[29], s1[26]);
s3[27] = vsubq_s16(s1[28], s1[27]);
s2[28] = vaddq_s16(s1[27], s1[28]);
s2[29] = vaddq_s16(s1[26], s1[29]);
s2[30] = vaddq_s16(s2[25], s1[30]);
s2[31] = vaddq_s16(s2[24], s1[31]);
// stage 7
s1[0] = vaddq_s16(s2[0], s2[15]);
s1[1] = vaddq_s16(s2[1], s2[14]);
s1[2] = vaddq_s16(s2[2], s2[13]);
s1[3] = vaddq_s16(s2[3], s2[12]);
s1[4] = vaddq_s16(s2[4], s2[11]);
s1[5] = vaddq_s16(s2[5], s2[10]);
s1[6] = vaddq_s16(s2[6], s2[9]);
s1[7] = vaddq_s16(s2[7], s2[8]);
s1[8] = vsubq_s16(s2[7], s2[8]);
s1[9] = vsubq_s16(s2[6], s2[9]);
s1[10] = vsubq_s16(s2[5], s2[10]);
s1[11] = vsubq_s16(s2[4], s2[11]);
s1[12] = vsubq_s16(s2[3], s2[12]);
s1[13] = vsubq_s16(s2[2], s2[13]);
s1[14] = vsubq_s16(s2[1], s2[14]);
s1[15] = vsubq_s16(s2[0], s2[15]);
s1[20] = sub_multiply_shift_and_narrow_s16(s3[27], s2[20], cospi_16_64);
s1[27] = add_multiply_shift_and_narrow_s16(s2[20], s3[27], cospi_16_64);
s1[21] = sub_multiply_shift_and_narrow_s16(s3[26], s2[21], cospi_16_64);
s1[26] = add_multiply_shift_and_narrow_s16(s2[21], s3[26], cospi_16_64);
s1[22] = sub_multiply_shift_and_narrow_s16(s3[25], s2[22], cospi_16_64);
s1[25] = add_multiply_shift_and_narrow_s16(s2[22], s3[25], cospi_16_64);
s1[23] = sub_multiply_shift_and_narrow_s16(s3[24], s2[23], cospi_16_64);
s1[24] = add_multiply_shift_and_narrow_s16(s2[23], s3[24], cospi_16_64);
// final stage
vst1q_s16(output, vaddq_s16(s1[0], s2[31]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[1], s2[30]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[2], s2[29]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[3], s2[28]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[4], s1[27]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[5], s1[26]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[6], s1[25]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[7], s1[24]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[8], s1[23]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[9], s1[22]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[10], s1[21]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[11], s1[20]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[12], s2[19]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[13], s2[18]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[14], s2[17]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[15], s2[16]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[15], s2[16]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[14], s2[17]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[13], s2[18]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[12], s2[19]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[11], s1[20]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[10], s1[21]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[9], s1[22]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[8], s1[23]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[7], s1[24]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[6], s1[25]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[5], s1[26]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[4], s1[27]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[3], s2[28]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[2], s2[29]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[1], s2[30]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[0], s2[31]));
}
void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
int skip_block, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) {
// TODO(jingning) Decide the need of these arguments after the
// quantization process is completed.
(void)zbin_ptr;
(void)quant_shift_ptr;
(void)scan;
if (!skip_block) {
// Quantization pass: All coefficients with index >= zero_flag are
// skippable. Note: zero_flag can be zero.
int i;
const int16x8_t v_zero = vdupq_n_s16(0);
const int16x8_t v_one = vdupq_n_s16(1);
int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
// adjust for dc
v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
// process dc and the first seven ac coeffs
{
const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr);
const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
const int32x4_t v_tmp_lo =
vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
const int32x4_t v_tmp_hi =
vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
const int16x8_t v_tmp2 =
vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
store_s16q_to_tran_low(qcoeff_ptr, v_qcoeff);
store_s16q_to_tran_low(dqcoeff_ptr, v_dqcoeff);
v_round = vmovq_n_s16(round_ptr[1]);
v_quant = vmovq_n_s16(quant_ptr[1]);
v_dequant = vmovq_n_s16(dequant_ptr[1]);
}
// now process the rest of the ac coeffs
for (i = 8; i < count; i += 8) {
const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr + i);
const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
const int32x4_t v_tmp_lo =
vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
const int32x4_t v_tmp_hi =
vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
const int16x8_t v_tmp2 =
vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
store_s16q_to_tran_low(qcoeff_ptr + i, v_qcoeff);
store_s16q_to_tran_low(dqcoeff_ptr + i, v_dqcoeff);
}
{
const int16x4_t v_eobmax_3210 = vmax_s16(
vget_low_s16(v_eobmax_76543210), vget_high_s16(v_eobmax_76543210));
const int64x1_t v_eobmax_xx32 =
vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
const int16x4_t v_eobmax_tmp =
vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
const int64x1_t v_eobmax_xxx3 =
vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
const int16x4_t v_eobmax_final =
vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
*eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
}
} else {
memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
*eob_ptr = 0;
}
}
inline int16x8_t vsubq(const int16x8_t & v0, const int16x8_t & v1) { return vsubq_s16(v0, v1); }
void vpx_idct32_8_neon(const int16_t *input, void *const output, int stride,
const int highbd_flag) {
int16x8_t in[8], s1[32], s2[32], s3[32], out[32];
load_and_transpose_s16_8x8(input, 8, &in[0], &in[1], &in[2], &in[3], &in[4],
&in[5], &in[6], &in[7]);
// stage 1
s1[16] = multiply_shift_and_narrow_s16(in[1], cospi_31_64);
s1[31] = multiply_shift_and_narrow_s16(in[1], cospi_1_64);
// Different for _8_
s1[19] = multiply_shift_and_narrow_s16(in[7], -cospi_25_64);
s1[28] = multiply_shift_and_narrow_s16(in[7], cospi_7_64);
s1[20] = multiply_shift_and_narrow_s16(in[5], cospi_27_64);
s1[27] = multiply_shift_and_narrow_s16(in[5], cospi_5_64);
s1[23] = multiply_shift_and_narrow_s16(in[3], -cospi_29_64);
s1[24] = multiply_shift_and_narrow_s16(in[3], cospi_3_64);
// stage 2
s2[8] = multiply_shift_and_narrow_s16(in[2], cospi_30_64);
s2[15] = multiply_shift_and_narrow_s16(in[2], cospi_2_64);
s2[11] = multiply_shift_and_narrow_s16(in[6], -cospi_26_64);
s2[12] = multiply_shift_and_narrow_s16(in[6], cospi_6_64);
// stage 3
s1[4] = multiply_shift_and_narrow_s16(in[4], cospi_28_64);
s1[7] = multiply_shift_and_narrow_s16(in[4], cospi_4_64);
s1[17] = multiply_accumulate_shift_and_narrow_s16(s1[16], -cospi_4_64, s1[31],
cospi_28_64);
s1[30] = multiply_accumulate_shift_and_narrow_s16(s1[16], cospi_28_64, s1[31],
cospi_4_64);
// Different for _8_
s1[18] = multiply_accumulate_shift_and_narrow_s16(s1[19], -cospi_28_64,
s1[28], -cospi_4_64);
s1[29] = multiply_accumulate_shift_and_narrow_s16(s1[19], -cospi_4_64, s1[28],
cospi_28_64);
s1[21] = multiply_accumulate_shift_and_narrow_s16(s1[20], -cospi_20_64,
s1[27], cospi_12_64);
s1[26] = multiply_accumulate_shift_and_narrow_s16(s1[20], cospi_12_64, s1[27],
cospi_20_64);
s1[22] = multiply_accumulate_shift_and_narrow_s16(s1[23], -cospi_12_64,
s1[24], -cospi_20_64);
s1[25] = multiply_accumulate_shift_and_narrow_s16(s1[23], -cospi_20_64,
s1[24], cospi_12_64);
// stage 4
s1[0] = multiply_shift_and_narrow_s16(in[0], cospi_16_64);
s2[9] = multiply_accumulate_shift_and_narrow_s16(s2[8], -cospi_8_64, s2[15],
cospi_24_64);
s2[14] = multiply_accumulate_shift_and_narrow_s16(s2[8], cospi_24_64, s2[15],
cospi_8_64);
s2[10] = multiply_accumulate_shift_and_narrow_s16(s2[11], -cospi_24_64,
s2[12], -cospi_8_64);
s2[13] = multiply_accumulate_shift_and_narrow_s16(s2[11], -cospi_8_64, s2[12],
cospi_24_64);
s2[16] = vaddq_s16(s1[16], s1[19]);
s2[17] = vaddq_s16(s1[17], s1[18]);
s2[18] = vsubq_s16(s1[17], s1[18]);
s2[19] = vsubq_s16(s1[16], s1[19]);
s2[20] = vsubq_s16(s1[23], s1[20]);
s2[21] = vsubq_s16(s1[22], s1[21]);
s2[22] = vaddq_s16(s1[21], s1[22]);
s2[23] = vaddq_s16(s1[20], s1[23]);
s2[24] = vaddq_s16(s1[24], s1[27]);
s2[25] = vaddq_s16(s1[25], s1[26]);
s2[26] = vsubq_s16(s1[25], s1[26]);
s2[27] = vsubq_s16(s1[24], s1[27]);
s2[28] = vsubq_s16(s1[31], s1[28]);
s2[29] = vsubq_s16(s1[30], s1[29]);
s2[30] = vaddq_s16(s1[29], s1[30]);
s2[31] = vaddq_s16(s1[28], s1[31]);
// stage 5
s1[5] = sub_multiply_shift_and_narrow_s16(s1[7], s1[4], cospi_16_64);
s1[6] = add_multiply_shift_and_narrow_s16(s1[4], s1[7], cospi_16_64);
s1[8] = vaddq_s16(s2[8], s2[11]);
s1[9] = vaddq_s16(s2[9], s2[10]);
s1[10] = vsubq_s16(s2[9], s2[10]);
s1[11] = vsubq_s16(s2[8], s2[11]);
s1[12] = vsubq_s16(s2[15], s2[12]);
s1[13] = vsubq_s16(s2[14], s2[13]);
s1[14] = vaddq_s16(s2[13], s2[14]);
s1[15] = vaddq_s16(s2[12], s2[15]);
s1[18] = multiply_accumulate_shift_and_narrow_s16(s2[18], -cospi_8_64, s2[29],
cospi_24_64);
s1[29] = multiply_accumulate_shift_and_narrow_s16(s2[18], cospi_24_64, s2[29],
cospi_8_64);
s1[19] = multiply_accumulate_shift_and_narrow_s16(s2[19], -cospi_8_64, s2[28],
cospi_24_64);
s1[28] = multiply_accumulate_shift_and_narrow_s16(s2[19], cospi_24_64, s2[28],
cospi_8_64);
s1[20] = multiply_accumulate_shift_and_narrow_s16(s2[20], -cospi_24_64,
s2[27], -cospi_8_64);
s1[27] = multiply_accumulate_shift_and_narrow_s16(s2[20], -cospi_8_64, s2[27],
cospi_24_64);
s1[21] = multiply_accumulate_shift_and_narrow_s16(s2[21], -cospi_24_64,
s2[26], -cospi_8_64);
s1[26] = multiply_accumulate_shift_and_narrow_s16(s2[21], -cospi_8_64, s2[26],
cospi_24_64);
// stage 6
s2[0] = vaddq_s16(s1[0], s1[7]);
s2[1] = vaddq_s16(s1[0], s1[6]);
s2[2] = vaddq_s16(s1[0], s1[5]);
s2[3] = vaddq_s16(s1[0], s1[4]);
s2[4] = vsubq_s16(s1[0], s1[4]);
s2[5] = vsubq_s16(s1[0], s1[5]);
s2[6] = vsubq_s16(s1[0], s1[6]);
s2[7] = vsubq_s16(s1[0], s1[7]);
s2[10] = sub_multiply_shift_and_narrow_s16(s1[13], s1[10], cospi_16_64);
s2[13] = add_multiply_shift_and_narrow_s16(s1[10], s1[13], cospi_16_64);
s2[11] = sub_multiply_shift_and_narrow_s16(s1[12], s1[11], cospi_16_64);
s2[12] = add_multiply_shift_and_narrow_s16(s1[11], s1[12], cospi_16_64);
s1[16] = vaddq_s16(s2[16], s2[23]);
s1[17] = vaddq_s16(s2[17], s2[22]);
s2[18] = vaddq_s16(s1[18], s1[21]);
s2[19] = vaddq_s16(s1[19], s1[20]);
s2[20] = vsubq_s16(s1[19], s1[20]);
s2[21] = vsubq_s16(s1[18], s1[21]);
s1[22] = vsubq_s16(s2[17], s2[22]);
s1[23] = vsubq_s16(s2[16], s2[23]);
s3[24] = vsubq_s16(s2[31], s2[24]);
s3[25] = vsubq_s16(s2[30], s2[25]);
s3[26] = vsubq_s16(s1[29], s1[26]);
s3[27] = vsubq_s16(s1[28], s1[27]);
s2[28] = vaddq_s16(s1[27], s1[28]);
s2[29] = vaddq_s16(s1[26], s1[29]);
s2[30] = vaddq_s16(s2[25], s2[30]);
s2[31] = vaddq_s16(s2[24], s2[31]);
// stage 7
s1[0] = vaddq_s16(s2[0], s1[15]);
s1[1] = vaddq_s16(s2[1], s1[14]);
s1[2] = vaddq_s16(s2[2], s2[13]);
s1[3] = vaddq_s16(s2[3], s2[12]);
s1[4] = vaddq_s16(s2[4], s2[11]);
s1[5] = vaddq_s16(s2[5], s2[10]);
s1[6] = vaddq_s16(s2[6], s1[9]);
s1[7] = vaddq_s16(s2[7], s1[8]);
s1[8] = vsubq_s16(s2[7], s1[8]);
s1[9] = vsubq_s16(s2[6], s1[9]);
s1[10] = vsubq_s16(s2[5], s2[10]);
s1[11] = vsubq_s16(s2[4], s2[11]);
s1[12] = vsubq_s16(s2[3], s2[12]);
s1[13] = vsubq_s16(s2[2], s2[13]);
s1[14] = vsubq_s16(s2[1], s1[14]);
s1[15] = vsubq_s16(s2[0], s1[15]);
s1[20] = sub_multiply_shift_and_narrow_s16(s3[27], s2[20], cospi_16_64);
s1[27] = add_multiply_shift_and_narrow_s16(s2[20], s3[27], cospi_16_64);
s1[21] = sub_multiply_shift_and_narrow_s16(s3[26], s2[21], cospi_16_64);
s1[26] = add_multiply_shift_and_narrow_s16(s2[21], s3[26], cospi_16_64);
s2[22] = sub_multiply_shift_and_narrow_s16(s3[25], s1[22], cospi_16_64);
s1[25] = add_multiply_shift_and_narrow_s16(s1[22], s3[25], cospi_16_64);
s2[23] = sub_multiply_shift_and_narrow_s16(s3[24], s1[23], cospi_16_64);
s1[24] = add_multiply_shift_and_narrow_s16(s1[23], s3[24], cospi_16_64);
// final stage
out[0] = final_add(s1[0], s2[31]);
out[1] = final_add(s1[1], s2[30]);
out[2] = final_add(s1[2], s2[29]);
out[3] = final_add(s1[3], s2[28]);
out[4] = final_add(s1[4], s1[27]);
out[5] = final_add(s1[5], s1[26]);
out[6] = final_add(s1[6], s1[25]);
out[7] = final_add(s1[7], s1[24]);
out[8] = final_add(s1[8], s2[23]);
out[9] = final_add(s1[9], s2[22]);
out[10] = final_add(s1[10], s1[21]);
out[11] = final_add(s1[11], s1[20]);
out[12] = final_add(s1[12], s2[19]);
out[13] = final_add(s1[13], s2[18]);
out[14] = final_add(s1[14], s1[17]);
out[15] = final_add(s1[15], s1[16]);
out[16] = final_sub(s1[15], s1[16]);
out[17] = final_sub(s1[14], s1[17]);
out[18] = final_sub(s1[13], s2[18]);
out[19] = final_sub(s1[12], s2[19]);
out[20] = final_sub(s1[11], s1[20]);
out[21] = final_sub(s1[10], s1[21]);
out[22] = final_sub(s1[9], s2[22]);
out[23] = final_sub(s1[8], s2[23]);
out[24] = final_sub(s1[7], s1[24]);
out[25] = final_sub(s1[6], s1[25]);
out[26] = final_sub(s1[5], s1[26]);
out[27] = final_sub(s1[4], s1[27]);
out[28] = final_sub(s1[3], s2[28]);
out[29] = final_sub(s1[2], s2[29]);
out[30] = final_sub(s1[1], s2[30]);
out[31] = final_sub(s1[0], s2[31]);
if (highbd_flag) {
highbd_add_and_store_bd8(out, output, stride);
} else {
uint8_t *const outputT = (uint8_t *)output;
add_and_store_u8_s16(out[0], out[1], out[2], out[3], out[4], out[5], out[6],
out[7], outputT, stride);
add_and_store_u8_s16(out[8], out[9], out[10], out[11], out[12], out[13],
out[14], out[15], outputT + (8 * stride), stride);
add_and_store_u8_s16(out[16], out[17], out[18], out[19], out[20], out[21],
out[22], out[23], outputT + (16 * stride), stride);
add_and_store_u8_s16(out[24], out[25], out[26], out[27], out[28], out[29],
out[30], out[31], outputT + (24 * stride), stride);
}
}
void vp8_short_fdct8x4_neon(
int16_t *input,
int16_t *output,
int pitch) {
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16;
uint16x4_t d28u16, d29u16;
uint16x8_t q14u16;
int16x8_t q0s16, q1s16, q2s16, q3s16;
int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
int16x8x2_t v2tmp0, v2tmp1;
int32x4x2_t v2tmp2, v2tmp3;
d16s16 = vdup_n_s16(5352);
d17s16 = vdup_n_s16(2217);
q9s32 = vdupq_n_s32(14500);
q10s32 = vdupq_n_s32(7500);
// Part one
pitch >>= 1;
q0s16 = vld1q_s16(input);
input += pitch;
q1s16 = vld1q_s16(input);
input += pitch;
q2s16 = vld1q_s16(input);
input += pitch;
q3s16 = vld1q_s16(input);
v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
vreinterpretq_s32_s16(q2s16));
v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
vreinterpretq_s32_s16(q3s16));
v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0
vreinterpretq_s16_s32(v2tmp3.val[0])); // q1
v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2
vreinterpretq_s16_s32(v2tmp3.val[1])); // q3
q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q11s16 = vshlq_n_s16(q11s16, 3);
q12s16 = vshlq_n_s16(q12s16, 3);
q13s16 = vshlq_n_s16(q13s16, 3);
q14s16 = vshlq_n_s16(q14s16, 3);
q0s16 = vaddq_s16(q11s16, q12s16);
q2s16 = vsubq_s16(q11s16, q12s16);
q11s32 = q9s32;
q12s32 = q10s32;
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
d2s16 = vshrn_n_s32(q9s32, 12);
d6s16 = vshrn_n_s32(q10s32, 12);
d3s16 = vshrn_n_s32(q11s32, 12);
d7s16 = vshrn_n_s32(q12s32, 12);
q1s16 = vcombine_s16(d2s16, d3s16);
q3s16 = vcombine_s16(d6s16, d7s16);
// Part two
q9s32 = vdupq_n_s32(12000);
q10s32 = vdupq_n_s32(51000);
v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
vreinterpretq_s32_s16(q2s16));
v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
vreinterpretq_s32_s16(q3s16));
v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0
vreinterpretq_s16_s32(v2tmp3.val[0])); // q1
v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2
vreinterpretq_s16_s32(v2tmp3.val[1])); // q3
q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q15s16 = vdupq_n_s16(7);
q11s16 = vaddq_s16(q11s16, q15s16);
q0s16 = vaddq_s16(q11s16, q12s16);
q1s16 = vsubq_s16(q11s16, q12s16);
q11s32 = q9s32;
q12s32 = q10s32;
d0s16 = vget_low_s16(q0s16);
d1s16 = vget_high_s16(q0s16);
d2s16 = vget_low_s16(q1s16);
d3s16 = vget_high_s16(q1s16);
d0s16 = vshr_n_s16(d0s16, 4);
d4s16 = vshr_n_s16(d1s16, 4);
d2s16 = vshr_n_s16(d2s16, 4);
d6s16 = vshr_n_s16(d3s16, 4);
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
d1s16 = vshrn_n_s32(q9s32, 16);
d3s16 = vshrn_n_s32(q10s32, 16);
d5s16 = vshrn_n_s32(q11s32, 16);
d7s16 = vshrn_n_s32(q12s32, 16);
qEmptys16 = vdupq_n_s16(0);
q14u16 = vceqq_s16(q14s16, qEmptys16);
q14u16 = vmvnq_u16(q14u16);
d28u16 = vget_low_u16(q14u16);
d29u16 = vget_high_u16(q14u16);
d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16));
d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16));
q0s16 = vcombine_s16(d0s16, d1s16);
q1s16 = vcombine_s16(d2s16, d3s16);
q2s16 = vcombine_s16(d4s16, d5s16);
q3s16 = vcombine_s16(d6s16, d7s16);
vst1q_s16(output, q0s16);
vst1q_s16(output + 8, q1s16);
vst1q_s16(output + 16, q2s16);
vst1q_s16(output + 24, q3s16);
return;
}
void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *final_output,
int stride) {
int i;
// input[M * stride] * 16
int16x4_t input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
int16x4_t input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
int16x4_t input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
int16x4_t input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);
// If the very first value != 0, then add 1.
if (input[0] != 0) {
const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1));
input_0 = vadd_s16(input_0, one);
}
for (i = 0; i < 2; ++i) {
const int16x8_t input_01 = vcombine_s16(input_0, input_1);
const int16x8_t input_32 = vcombine_s16(input_3, input_2);
// in_0 +/- in_3, in_1 +/- in_2
const int16x8_t s_01 = vaddq_s16(input_01, input_32);
const int16x8_t s_32 = vsubq_s16(input_01, input_32);
// step_0 +/- step_1, step_2 +/- step_3
const int16x4_t s_0 = vget_low_s16(s_01);
const int16x4_t s_1 = vget_high_s16(s_01);
const int16x4_t s_2 = vget_high_s16(s_32);
const int16x4_t s_3 = vget_low_s16(s_32);
// (s_0 +/- s_1) * cospi_16_64
// Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1);
const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1);
const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, cospi_16_64);
const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, cospi_16_64);
// fdct_round_shift
int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS);
int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS);
// s_3 * cospi_8_64 + s_2 * cospi_24_64
// s_3 * cospi_24_64 - s_2 * cospi_8_64
const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, cospi_8_64);
const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, cospi_24_64);
const int32x4_t temp3 = vmlal_n_s16(s_3_cospi_8_64, s_2, cospi_24_64);
const int32x4_t temp4 = vmlsl_n_s16(s_3_cospi_24_64, s_2, cospi_8_64);
// fdct_round_shift
int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS);
int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS);
transpose_s16_4x4d(&out_0, &out_1, &out_2, &out_3);
input_0 = out_0;
input_1 = out_1;
input_2 = out_2;
input_3 = out_3;
}
{
// Not quite a rounding shift. Only add 1 despite shifting by 2.
const int16x8_t one = vdupq_n_s16(1);
int16x8_t out_01 = vcombine_s16(input_0, input_1);
int16x8_t out_23 = vcombine_s16(input_2, input_3);
out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
store_s16q_to_tran_low(final_output + 0 * 8, out_01);
store_s16q_to_tran_low(final_output + 1 * 8, out_23);
}
}
void vpx_idct8x8_12_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
int16x4_t d26s16, d27s16, d28s16, d29s16;
uint64x1_t d0u64, d1u64, d2u64, d3u64;
int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
uint16x8_t q8u16, q9u16, q10u16, q11u16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
q10s16 = vld1q_s16(input + 16);
q11s16 = vld1q_s16(input + 24);
q12s16 = vld1q_s16(input + 32);
q13s16 = vld1q_s16(input + 40);
q14s16 = vld1q_s16(input + 48);
q15s16 = vld1q_s16(input + 56);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
// First transform rows
// stage 1
q0s16 = vdupq_n_s16(cospi_28_64 * 2);
q1s16 = vdupq_n_s16(cospi_4_64 * 2);
q4s16 = vqrdmulhq_s16(q9s16, q0s16);
q0s16 = vdupq_n_s16(-cospi_20_64 * 2);
q7s16 = vqrdmulhq_s16(q9s16, q1s16);
q1s16 = vdupq_n_s16(cospi_12_64 * 2);
q5s16 = vqrdmulhq_s16(q11s16, q0s16);
q0s16 = vdupq_n_s16(cospi_16_64 * 2);
q6s16 = vqrdmulhq_s16(q11s16, q1s16);
// stage 2 & stage 3 - even half
q1s16 = vdupq_n_s16(cospi_24_64 * 2);
q9s16 = vqrdmulhq_s16(q8s16, q0s16);
q0s16 = vdupq_n_s16(cospi_8_64 * 2);
q13s16 = vqrdmulhq_s16(q10s16, q1s16);
q15s16 = vqrdmulhq_s16(q10s16, q0s16);
// stage 3 -odd half
q0s16 = vaddq_s16(q9s16, q15s16);
q1s16 = vaddq_s16(q9s16, q13s16);
q2s16 = vsubq_s16(q9s16, q13s16);
q3s16 = vsubq_s16(q9s16, q15s16);
// stage 2 - odd half
q13s16 = vsubq_s16(q4s16, q5s16);
q4s16 = vaddq_s16(q4s16, q5s16);
q14s16 = vsubq_s16(q7s16, q6s16);
q7s16 = vaddq_s16(q7s16, q6s16);
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
d16s16 = vdup_n_s16(cospi_16_64);
q9s32 = vmull_s16(d28s16, d16s16);
q10s32 = vmull_s16(d29s16, d16s16);
q11s32 = vmull_s16(d28s16, d16s16);
q12s32 = vmull_s16(d29s16, d16s16);
q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
d10s16 = vqrshrn_n_s32(q9s32, 14);
d11s16 = vqrshrn_n_s32(q10s32, 14);
d12s16 = vqrshrn_n_s32(q11s32, 14);
d13s16 = vqrshrn_n_s32(q12s32, 14);
q5s16 = vcombine_s16(d10s16, d11s16);
q6s16 = vcombine_s16(d12s16, d13s16);
// stage 4
q8s16 = vaddq_s16(q0s16, q7s16);
q9s16 = vaddq_s16(q1s16, q6s16);
q10s16 = vaddq_s16(q2s16, q5s16);
q11s16 = vaddq_s16(q3s16, q4s16);
q12s16 = vsubq_s16(q3s16, q4s16);
q13s16 = vsubq_s16(q2s16, q5s16);
q14s16 = vsubq_s16(q1s16, q6s16);
q15s16 = vsubq_s16(q0s16, q7s16);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
q8s16 = vrshrq_n_s16(q8s16, 5);
q9s16 = vrshrq_n_s16(q9s16, 5);
q10s16 = vrshrq_n_s16(q10s16, 5);
q11s16 = vrshrq_n_s16(q11s16, 5);
q12s16 = vrshrq_n_s16(q12s16, 5);
q13s16 = vrshrq_n_s16(q13s16, 5);
q14s16 = vrshrq_n_s16(q14s16, 5);
q15s16 = vrshrq_n_s16(q15s16, 5);
d1 = d2 = dest;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
q8s16 = q12s16;
q9s16 = q13s16;
q10s16 = q14s16;
q11s16 = q15s16;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
return;
}
void yuv422rgb_neon_int(const unsigned char * sourcep, int source_byte_count,
unsigned char * destp)
{
const unsigned char *source_endp;
const unsigned char *vector_endp;
int remainder;
const int16x8_t u_coeff = {0, -22, 113, 0, 0, -22, 113, 0};
const int16x8_t v_coeff = {90, -46, 0, 0, 90, -46, 0, 0};
const uint8x8_t zeroalpha = {0x0, 0x0, 0x0, 0xFF, 0x0, 0x0, 0x0, 0xFF};
const int16x8_t uvbias = {0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
int16x8_t mp0_rgba; /* macropixel 0's resulting RGBA RGBA pixels */
int16x8_t mp1_rgba; /* macropixel 1's resulting RGBA RGBA pixels */
uint8x8_t rawpixels; /* source pixels as {[YUYV]0 [YUYV]1} */
uint8x8_t rgba0, rgba1; /* rgba values as bytes */
uint8x16_t bothrgba;
uint8_t * destinationp; /* pointer into output buffer destp */
int16x8_t widerpixels; /* rawpixels promoted to shorts per component */
const uint8x8_t yselect = {0xff, 0xff, 0xff, 0xff,
0x00, 0x00, 0x00, 0x00};
/* we're working with things in 4-byte macropixels */
remainder = source_byte_count % 4;
source_endp = sourcep + source_byte_count;
vector_endp = source_endp - remainder;
destinationp = (uint8_t *)destp;
while (sourcep < vector_endp)
{
/* pull YUYV from 2 four byte macropixels starting at sourcep. */
/* we'll increment sourcep as we go to save the array dereference */
/* and separate increment instruction at the end of the loop */
/* load rawpixels with {[YUYV]0 [YUYV]1 } with byte components */
rawpixels = vld1_u8(sourcep);
sourcep += sizeof(rawpixels);
widerpixels = vreinterpretq_s16_u16(vmovl_u8(rawpixels));
/* ---------- process macropixel 0 --------------- */
/* take macropixel zero ([YUYV]0) from rawpixels and */
/* compute the two RGBA pixels that come from it. store */
/* those two pixels in mp0_rgba */
{
int16x8_t wider_yalpha;
int16x8_t u_vec, v_vec, uv_vec;
uint8x8_t narrow_yalpha;
uint8x8_t y0_vec, y1_vec;
int16x4_t yuyv;
/* narrow_yalpha is drawn from [YUYV]0 and formed into */
/* {Y0, Y0, Y0, alpha, Y1, Y1, Y1, alpha} */
/* this would have been a nice place for vtbx1_u8, but i */
/* can't get it to work. so i'll have to use vbsl_u8 instead. */
y0_vec = vdup_lane_u8(rawpixels, MP0_Y0);
y1_vec = vdup_lane_u8(rawpixels, MP0_Y1);
narrow_yalpha = vbsl_u8(yselect, y0_vec, y1_vec);
/* store ALPHA in elements 3 and 7 (after the RGB components) */
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 3);
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 7);
/* use vmovl_u8 to go from being unsigned 8-bit to */
/* unsigned 16-bit, the use vreinterpretq_s16_u16 to */
/* change interpretation from unsigned 16-bit to signed */
/* 16-bit. */
wider_yalpha = vreinterpretq_s16_u16(vmovl_u8(narrow_yalpha));
yuyv = vget_low_s16(widerpixels);
/* form a vector of the U component from MP0 */
u_vec = vdupq_lane_s16(yuyv, MP0_U);
/* subtract uvbias from u_vec */
u_vec = vsubq_s16(u_vec, uvbias);
/* form a vector of the V component from MP0 */
v_vec = vdupq_lane_s16(yuyv, MP0_V);
/* subtract uvbias from v_vec */
v_vec = vsubq_s16(v_vec, uvbias);
/* Multiply eight 16-bit values in u_vec by eight 16-bit */
/* values in u_coeff and store the results in u_vec. */
u_vec = vmulq_s16(u_vec, u_coeff);
/* likewise multiply eight 16-bit values in v_vec by */
/* v_coeff and store the results in v_vec */
v_vec = vmulq_s16(v_vec, v_coeff);
/* form uv_vec as the sum of u_vec & v_vec, then shift 6 places */
/* (dividing by 64) */
uv_vec = vaddq_s16(u_vec, v_vec);
uv_vec = vshrq_n_s16(uv_vec, 6);
/* now mp0_rgba = y_vec + u_vec + v_vec */
mp0_rgba = vaddq_s16(wider_yalpha, uv_vec);
}
/* ---------- process macropixel 1 --------------- */
/* take macropixel one ([YUYV]1) from rawpixels and */
/* compute the two RGBA pixels that come from it. store */
/* those two pixels in mp1_rgba */
{
int16x8_t wider_yalpha;
int16x8_t u_vec, v_vec, uv_vec;
uint8x8_t narrow_yalpha;
uint8x8_t y0_vec, y1_vec;
int16x4_t yuyv;
/* narrow_yalpha is drawn from [YUYV]1 and formed into */
/* {Y0, Y0, Y0, alpha, Y1, Y1, Y1, alpha} */
/* this would have been a nice place for vtbx1_u8, but i */
/* can't get it to work. so i'll have to use vbsl_u8 instead. */
y0_vec = vdup_lane_u8(rawpixels, MP1_Y0);
y1_vec = vdup_lane_u8(rawpixels, MP1_Y1);
narrow_yalpha = vbsl_u8(yselect, y0_vec, y1_vec);
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 3);
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 7);
/* use vmovl_u8 to go from being unsigned 8-bit to */
/* unsigned 16-bit, the use vreinterpretq_s16_u16 to */
wider_yalpha = vreinterpretq_s16_u16(vmovl_u8(narrow_yalpha));
yuyv = vget_high_s16(widerpixels);
u_vec = vdupq_lane_s16(yuyv, 1);
u_vec = vsubq_s16(u_vec, uvbias);
v_vec = vdupq_lane_s16(yuyv, 3);
v_vec = vsubq_s16(v_vec, uvbias);
/* Multiply eight 16-bit values in u_vec by eight 16-bit */
/* values in u_coeff and store the results in u_vec. */
u_vec = vmulq_s16(u_vec, u_coeff);
/* likewise multiply eight 16-bit values in v_vec by */
/* v_coeff and store the results in v_vec */
v_vec = vmulq_s16(v_vec, v_coeff);
/* form uv_vec as the sum of u_vec & v_vec, then shift 6 places */
/* (dividing by 64) */
uv_vec = vaddq_s16(u_vec, v_vec);
uv_vec = vshrq_n_s16(uv_vec, 6);
/* now mp1_rgba = y_vec + u_vec + v_vec */
mp1_rgba = vaddq_s16(wider_yalpha, uv_vec);
}
/* turn mp0_rgba from a vector of shorts to a vector of */
/* unsigned unsigned chars. this will saturate: clipping */
/* the values between 0 and 255. */
rgba0 = vqmovun_s16(mp0_rgba);
rgba1 = vqmovun_s16(mp1_rgba);
/* make it faster to copy these back out of vector registers into */
/* memory by combining rgba0 and rgba1 into the larger bothrgba. */
/* then store that back into memory at destinationp. */
bothrgba = vcombine_u8(rgba0, rgba1);
vst1q_u8(destinationp, bothrgba);
destinationp += 16;
}
}
bool decode_yuv_neon(unsigned char* out, unsigned char const* y, unsigned char const* uv, int width, int height, unsigned char fill_alpha=0xff)
{
// pre-condition : width, height must be even
if (0!=(width&1) || width<2 || 0!=(height&1) || height<2 || !out || !y || !uv)
return false;
// in & out pointers
unsigned char* dst = out;
// constants
int const stride = width*trait::bytes_per_pixel;
int const itHeight = height>>1;
int const itWidth = width>>3;
uint8x8_t const Yshift = vdup_n_u8(16);
int16x8_t const half = vdupq_n_u16(128);
int32x4_t const rounding = vdupq_n_s32(128);
// tmp variable
uint16x8_t t;
// pixel block to temporary store 8 pixels
typename trait::PixelBlock pblock = trait::init_pixelblock(fill_alpha);
for (int j=0; j<itHeight; ++j, y+=width, dst+=stride) {
for (int i=0; i<itWidth; ++i, y+=8, uv+=8, dst+=(8*trait::bytes_per_pixel)) {
t = vmovl_u8(vqsub_u8(vld1_u8(y), Yshift));
int32x4_t const Y00 = vmulq_n_u32(vmovl_u16(vget_low_u16(t)), 298);
int32x4_t const Y01 = vmulq_n_u32(vmovl_u16(vget_high_u16(t)), 298);
t = vmovl_u8(vqsub_u8(vld1_u8(y+width), Yshift));
int32x4_t const Y10 = vmulq_n_u32(vmovl_u16(vget_low_u16(t)), 298);
int32x4_t const Y11 = vmulq_n_u32(vmovl_u16(vget_high_u16(t)), 298);
// trait::loadvu pack 4 sets of uv into a uint8x8_t, layout : { v0,u0, v1,u1, v2,u2, v3,u3 }
t = vsubq_s16((int16x8_t)vmovl_u8(trait::loadvu(uv)), half);
// UV.val[0] : v0, v1, v2, v3
// UV.val[1] : u0, u1, u2, u3
int16x4x2_t const UV = vuzp_s16(vget_low_s16(t), vget_high_s16(t));
// tR : 128+409V
// tG : 128-100U-208V
// tB : 128+516U
int32x4_t const tR = vmlal_n_s16(rounding, UV.val[0], 409);
int32x4_t const tG = vmlal_n_s16(vmlal_n_s16(rounding, UV.val[0], -208), UV.val[1], -100);
int32x4_t const tB = vmlal_n_s16(rounding, UV.val[1], 516);
int32x4x2_t const R = vzipq_s32(tR, tR); // [tR0, tR0, tR1, tR1] [ tR2, tR2, tR3, tR3]
int32x4x2_t const G = vzipq_s32(tG, tG); // [tG0, tG0, tG1, tG1] [ tG2, tG2, tG3, tG3]
int32x4x2_t const B = vzipq_s32(tB, tB); // [tB0, tB0, tB1, tB1] [ tB2, tB2, tB3, tB3]
// upper 8 pixels
trait::store_pixel_block(dst, pblock,
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(R.val[0], Y00)), vqmovun_s32(vaddq_s32(R.val[1], Y01))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(G.val[0], Y00)), vqmovun_s32(vaddq_s32(G.val[1], Y01))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(B.val[0], Y00)), vqmovun_s32(vaddq_s32(B.val[1], Y01))), 8));
// lower 8 pixels
trait::store_pixel_block(dst+stride, pblock,
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(R.val[0], Y10)), vqmovun_s32(vaddq_s32(R.val[1], Y11))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(G.val[0], Y10)), vqmovun_s32(vaddq_s32(G.val[1], Y11))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(B.val[0], Y10)), vqmovun_s32(vaddq_s32(B.val[1], Y11))), 8));
}
}
return true;
}
static INLINE void IDCT8x8_1D(
int16x8_t *q8s16,
int16x8_t *q9s16,
int16x8_t *q10s16,
int16x8_t *q11s16,
int16x8_t *q12s16,
int16x8_t *q13s16,
int16x8_t *q14s16,
int16x8_t *q15s16) {
int16x4_t d0s16, d1s16, d2s16, d3s16;
int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
d0s16 = vdup_n_s16(cospi_28_64);
d1s16 = vdup_n_s16(cospi_4_64);
d2s16 = vdup_n_s16(cospi_12_64);
d3s16 = vdup_n_s16(cospi_20_64);
d16s16 = vget_low_s16(*q8s16);
d17s16 = vget_high_s16(*q8s16);
d18s16 = vget_low_s16(*q9s16);
d19s16 = vget_high_s16(*q9s16);
d20s16 = vget_low_s16(*q10s16);
d21s16 = vget_high_s16(*q10s16);
d22s16 = vget_low_s16(*q11s16);
d23s16 = vget_high_s16(*q11s16);
d24s16 = vget_low_s16(*q12s16);
d25s16 = vget_high_s16(*q12s16);
d26s16 = vget_low_s16(*q13s16);
d27s16 = vget_high_s16(*q13s16);
d28s16 = vget_low_s16(*q14s16);
d29s16 = vget_high_s16(*q14s16);
d30s16 = vget_low_s16(*q15s16);
d31s16 = vget_high_s16(*q15s16);
q2s32 = vmull_s16(d18s16, d0s16);
q3s32 = vmull_s16(d19s16, d0s16);
q5s32 = vmull_s16(d26s16, d2s16);
q6s32 = vmull_s16(d27s16, d2s16);
q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
d8s16 = vqrshrn_n_s32(q2s32, 14);
d9s16 = vqrshrn_n_s32(q3s32, 14);
d10s16 = vqrshrn_n_s32(q5s32, 14);
d11s16 = vqrshrn_n_s32(q6s32, 14);
q4s16 = vcombine_s16(d8s16, d9s16);
q5s16 = vcombine_s16(d10s16, d11s16);
q2s32 = vmull_s16(d18s16, d1s16);
q3s32 = vmull_s16(d19s16, d1s16);
q9s32 = vmull_s16(d26s16, d3s16);
q13s32 = vmull_s16(d27s16, d3s16);
q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
d14s16 = vqrshrn_n_s32(q2s32, 14);
d15s16 = vqrshrn_n_s32(q3s32, 14);
d12s16 = vqrshrn_n_s32(q9s32, 14);
d13s16 = vqrshrn_n_s32(q13s32, 14);
q6s16 = vcombine_s16(d12s16, d13s16);
q7s16 = vcombine_s16(d14s16, d15s16);
d0s16 = vdup_n_s16(cospi_16_64);
q2s32 = vmull_s16(d16s16, d0s16);
q3s32 = vmull_s16(d17s16, d0s16);
q13s32 = vmull_s16(d16s16, d0s16);
q15s32 = vmull_s16(d17s16, d0s16);
q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
d0s16 = vdup_n_s16(cospi_24_64);
d1s16 = vdup_n_s16(cospi_8_64);
d18s16 = vqrshrn_n_s32(q2s32, 14);
d19s16 = vqrshrn_n_s32(q3s32, 14);
d22s16 = vqrshrn_n_s32(q13s32, 14);
d23s16 = vqrshrn_n_s32(q15s32, 14);
*q9s16 = vcombine_s16(d18s16, d19s16);
*q11s16 = vcombine_s16(d22s16, d23s16);
q2s32 = vmull_s16(d20s16, d0s16);
q3s32 = vmull_s16(d21s16, d0s16);
q8s32 = vmull_s16(d20s16, d1s16);
q12s32 = vmull_s16(d21s16, d1s16);
q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
d26s16 = vqrshrn_n_s32(q2s32, 14);
d27s16 = vqrshrn_n_s32(q3s32, 14);
d30s16 = vqrshrn_n_s32(q8s32, 14);
d31s16 = vqrshrn_n_s32(q12s32, 14);
*q13s16 = vcombine_s16(d26s16, d27s16);
*q15s16 = vcombine_s16(d30s16, d31s16);
q0s16 = vaddq_s16(*q9s16, *q15s16);
q1s16 = vaddq_s16(*q11s16, *q13s16);
q2s16 = vsubq_s16(*q11s16, *q13s16);
q3s16 = vsubq_s16(*q9s16, *q15s16);
*q13s16 = vsubq_s16(q4s16, q5s16);
q4s16 = vaddq_s16(q4s16, q5s16);
*q14s16 = vsubq_s16(q7s16, q6s16);
q7s16 = vaddq_s16(q7s16, q6s16);
d26s16 = vget_low_s16(*q13s16);
d27s16 = vget_high_s16(*q13s16);
d28s16 = vget_low_s16(*q14s16);
d29s16 = vget_high_s16(*q14s16);
d16s16 = vdup_n_s16(cospi_16_64);
q9s32 = vmull_s16(d28s16, d16s16);
q10s32 = vmull_s16(d29s16, d16s16);
q11s32 = vmull_s16(d28s16, d16s16);
q12s32 = vmull_s16(d29s16, d16s16);
q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
d10s16 = vqrshrn_n_s32(q9s32, 14);
d11s16 = vqrshrn_n_s32(q10s32, 14);
d12s16 = vqrshrn_n_s32(q11s32, 14);
d13s16 = vqrshrn_n_s32(q12s32, 14);
q5s16 = vcombine_s16(d10s16, d11s16);
q6s16 = vcombine_s16(d12s16, d13s16);
*q8s16 = vaddq_s16(q0s16, q7s16);
*q9s16 = vaddq_s16(q1s16, q6s16);
*q10s16 = vaddq_s16(q2s16, q5s16);
*q11s16 = vaddq_s16(q3s16, q4s16);
*q12s16 = vsubq_s16(q3s16, q4s16);
*q13s16 = vsubq_s16(q2s16, q5s16);
*q14s16 = vsubq_s16(q1s16, q6s16);
*q15s16 = vsubq_s16(q0s16, q7s16);
return;
}
static void hadamard8x8_one_pass(int16x8_t *a0, int16x8_t *a1,
int16x8_t *a2, int16x8_t *a3,
int16x8_t *a4, int16x8_t *a5,
int16x8_t *a6, int16x8_t *a7) {
const int16x8_t b0 = vaddq_s16(*a0, *a1);
const int16x8_t b1 = vsubq_s16(*a0, *a1);
const int16x8_t b2 = vaddq_s16(*a2, *a3);
const int16x8_t b3 = vsubq_s16(*a2, *a3);
const int16x8_t b4 = vaddq_s16(*a4, *a5);
const int16x8_t b5 = vsubq_s16(*a4, *a5);
const int16x8_t b6 = vaddq_s16(*a6, *a7);
const int16x8_t b7 = vsubq_s16(*a6, *a7);
const int16x8_t c0 = vaddq_s16(b0, b2);
const int16x8_t c1 = vaddq_s16(b1, b3);
const int16x8_t c2 = vsubq_s16(b0, b2);
const int16x8_t c3 = vsubq_s16(b1, b3);
const int16x8_t c4 = vaddq_s16(b4, b6);
const int16x8_t c5 = vaddq_s16(b5, b7);
const int16x8_t c6 = vsubq_s16(b4, b6);
const int16x8_t c7 = vsubq_s16(b5, b7);
*a0 = vaddq_s16(c0, c4);
*a1 = vsubq_s16(c2, c6);
*a2 = vsubq_s16(c0, c4);
*a3 = vaddq_s16(c2, c6);
*a4 = vaddq_s16(c3, c7);
*a5 = vsubq_s16(c3, c7);
*a6 = vsubq_s16(c1, c5);
*a7 = vaddq_s16(c1, c5);
}