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 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;
}
void test_vmaxQs16 (void)
{
int16x8_t out_int16x8_t;
int16x8_t arg0_int16x8_t;
int16x8_t arg1_int16x8_t;
out_int16x8_t = vmaxq_s16 (arg0_int16x8_t, arg1_int16x8_t);
}
static void SharpYUVFilterRow_NEON(const int16_t* A, const int16_t* B, int len,
const uint16_t* best_y, uint16_t* out) {
int i;
const int16x8_t max = vdupq_n_s16(MAX_Y);
const int16x8_t zero = vdupq_n_s16(0);
for (i = 0; i + 8 <= len; i += 8) {
const int16x8_t a0 = vld1q_s16(A + i + 0);
const int16x8_t a1 = vld1q_s16(A + i + 1);
const int16x8_t b0 = vld1q_s16(B + i + 0);
const int16x8_t b1 = vld1q_s16(B + i + 1);
const int16x8_t a0b1 = vaddq_s16(a0, b1);
const int16x8_t a1b0 = vaddq_s16(a1, b0);
const int16x8_t a0a1b0b1 = vaddq_s16(a0b1, a1b0); // A0+A1+B0+B1
const int16x8_t a0b1_2 = vaddq_s16(a0b1, a0b1); // 2*(A0+B1)
const int16x8_t a1b0_2 = vaddq_s16(a1b0, a1b0); // 2*(A1+B0)
const int16x8_t c0 = vshrq_n_s16(vaddq_s16(a0b1_2, a0a1b0b1), 3);
const int16x8_t c1 = vshrq_n_s16(vaddq_s16(a1b0_2, a0a1b0b1), 3);
const int16x8_t d0 = vaddq_s16(c1, a0);
const int16x8_t d1 = vaddq_s16(c0, a1);
const int16x8_t e0 = vrshrq_n_s16(d0, 1);
const int16x8_t e1 = vrshrq_n_s16(d1, 1);
const int16x8x2_t f = vzipq_s16(e0, e1);
const int16x8_t g0 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 0));
const int16x8_t g1 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 8));
const int16x8_t h0 = vaddq_s16(g0, f.val[0]);
const int16x8_t h1 = vaddq_s16(g1, f.val[1]);
const int16x8_t i0 = vmaxq_s16(vminq_s16(h0, max), zero);
const int16x8_t i1 = vmaxq_s16(vminq_s16(h1, max), zero);
vst1q_u16(out + 2 * i + 0, vreinterpretq_u16_s16(i0));
vst1q_u16(out + 2 * i + 8, vreinterpretq_u16_s16(i1));
}
for (; i < len; ++i) {
const int a0b1 = A[i + 0] + B[i + 1];
const int a1b0 = A[i + 1] + B[i + 0];
const int a0a1b0b1 = a0b1 + a1b0 + 8;
const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
out[2 * i + 0] = clip_y(best_y[2 * i + 0] + v0);
out[2 * i + 1] = clip_y(best_y[2 * i + 1] + v1);
}
}
int16x8_t test_vmaxq_s16(int16x8_t a1, int16x8_t a2) {
// CHECK: test_vmaxq_s16
return vmaxq_s16(a1, a2);
// CHECK llvm.aarch64.neon.smax.v8i16
}
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;
}
}
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;
}
}
inline int16x8_t vmaxq(const int16x8_t & v0, const int16x8_t & v1) { return vmaxq_s16(v0, v1); }
int16x8_t
test_vmaxq_s16 (int16x8_t __a, int16x8_t __b)
{
return vmaxq_s16(__a, __b);
}
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 r_cr = vdupq_n_s16(22986); // 1.403 << 14
int16x8_t g_cb = vdupq_n_s16(-5636); // -0.344 << 14
int16x8_t g_cr = vdupq_n_s16(-11698); // -0.714 << 14
int16x8_t b_cb = vdupq_n_s16(28999); // 1.770 << 14
int16x8_t c4096 = vdupq_n_s16(4096);
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++)
{
/*
In order to use NEON signed 16-bit integer multiplication we need to convert
the floating point factors to signed int without loosing information.
The result of this multiplication is 32 bit and we have a NEON instruction
that returns the hi word of the saturated double.
Thus we will multiply the factors by the highest possible 2^n, take the
upper 16 bits of the signed 32-bit result (vqdmulhq_s16 followed by a right
shift by 1 to reverse the doubling) and correct this result by multiplying it
by 2^(16-n).
For the given factors in the conversion matrix the best possible n is 14.
Example for calculating r:
r = (y>>5) + 128 + (cr*1.403)>>5 // our base formula
r = (y>>5) + 128 + (HIWORD(cr*(1.403<<14)<<2))>>5 // see above
r = (y+4096)>>5 + (HIWORD(cr*22986)<<2)>>5 // simplification
r = ((y+4096)>>2 + HIWORD(cr*22986)) >> 3
*/
/* y = (y_buf[i] + 4096) >> 2 */
int16x8_t y = vld1q_s16((INT16*) &y_buf[i]);
y = vaddq_s16(y, c4096);
y = vshrq_n_s16(y, 2);
/* cb = cb_buf[i]; */
int16x8_t cb = vld1q_s16((INT16*)&cb_buf[i]);
/* cr = cr_buf[i]; */
int16x8_t cr = vld1q_s16((INT16*) &cr_buf[i]);
/* (y + HIWORD(cr*22986)) >> 3 */
int16x8_t r = vaddq_s16(y, vshrq_n_s16(vqdmulhq_s16(cr, r_cr), 1));
r = vshrq_n_s16(r, 3);
/* r_buf[i] = MINMAX(r, 0, 255); */
r = vminq_s16(vmaxq_s16(r, zero), max);
vst1q_s16((INT16*)&r_buf[i], r);
/* (y + HIWORD(cb*-5636) + HIWORD(cr*-11698)) >> 3 */
int16x8_t g = vaddq_s16(y, vshrq_n_s16(vqdmulhq_s16(cb, g_cb), 1));
g = vaddq_s16(g, vshrq_n_s16(vqdmulhq_s16(cr, g_cr), 1));
g = vshrq_n_s16(g, 3);
/* g_buf[i] = MINMAX(g, 0, 255); */
g = vminq_s16(vmaxq_s16(g, zero), max);
vst1q_s16((INT16*)&g_buf[i], g);
/* (y + HIWORD(cb*28999)) >> 3 */
int16x8_t b = vaddq_s16(y, vshrq_n_s16(vqdmulhq_s16(cb, b_cb), 1));
b = vshrq_n_s16(b, 3);
/* b_buf[i] = MINMAX(b, 0, 255); */
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;
}
return PRIMITIVES_SUCCESS;
}