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);
}
}
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;
}
}
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);
}
/* s16x8 mm mul */
void mw_neon_mm_mul_s16x8(short * A, int Row, int T, short * B, int Col, short * C)
{
int i, k, j;
int16x8_t neon_b, neon_c;
int16x8_t neon_a0, neon_a1, neon_a2, neon_a3, neon_a4, neon_a5, neon_a6, neon_a7;
int16x8_t neon_b0, neon_b1, neon_b2, neon_b3, neon_b4, neon_b5, neon_b6, neon_b7;
for (i = 0; i < Row; i+=8)
{
for (k = 0; k < Col; k+=1)
{
neon_c = vmovq_n_s16(0);
for (j = 0; j < T; j+=8)
{
int j_T = j * T + i;
int k_Row = k * Row;
neon_a0 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a1 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a2 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a3 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a4 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a5 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a6 = vld1q_s16(A + j_T);
j_T+=Row;
neon_a7 = vld1q_s16(A + j_T);
neon_b = vld1q_s16(B + k_Row + j);
neon_b0 = vdupq_n_s16(vgetq_lane_s16(neon_b, 0));
neon_b1 = vdupq_n_s16(vgetq_lane_s16(neon_b, 1));
neon_b2 = vdupq_n_s16(vgetq_lane_s16(neon_b, 2));
neon_b3 = vdupq_n_s16(vgetq_lane_s16(neon_b, 3));
neon_b4 = vdupq_n_s16(vgetq_lane_s16(neon_b, 4));
neon_b5 = vdupq_n_s16(vgetq_lane_s16(neon_b, 5));
neon_b6 = vdupq_n_s16(vgetq_lane_s16(neon_b, 6));
neon_b7 = vdupq_n_s16(vgetq_lane_s16(neon_b, 7));
neon_c = vaddq_s16(vmulq_s16(neon_a0, neon_b0), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a1, neon_b1), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a2, neon_b2), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a3, neon_b3), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a4, neon_b4), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a5, neon_b5), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a6, neon_b6), neon_c);
neon_c = vaddq_s16(vmulq_s16(neon_a7, neon_b7), neon_c);
vst1q_lane_s16(C + k_Row + i, neon_c, 0);
vst1q_lane_s16(C + k_Row + i + 1, neon_c, 1);
vst1q_lane_s16(C + k_Row + i + 2, neon_c, 2);
vst1q_lane_s16(C + k_Row + i + 3, neon_c, 3);
vst1q_lane_s16(C + k_Row + i + 4, neon_c, 4);
vst1q_lane_s16(C + k_Row + i + 5, neon_c, 5);
vst1q_lane_s16(C + k_Row + i + 6, neon_c, 6);
vst1q_lane_s16(C + k_Row + i + 7, neon_c, 7);
}
}
}
}
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 vaddq(const int16x8_t & v0, const int16x8_t & v1) { return vaddq_s16(v0, v1); }
static INLINE void IADST8X8_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, d4s16, d5s16, d6s16, d7s16;
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 q2s16, q4s16, q5s16, q6s16;
int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q7s32, q8s32;
int32x4_t q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
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);
d14s16 = vdup_n_s16(cospi_2_64);
d15s16 = vdup_n_s16(cospi_30_64);
q1s32 = vmull_s16(d30s16, d14s16);
q2s32 = vmull_s16(d31s16, d14s16);
q3s32 = vmull_s16(d30s16, d15s16);
q4s32 = vmull_s16(d31s16, d15s16);
d30s16 = vdup_n_s16(cospi_18_64);
d31s16 = vdup_n_s16(cospi_14_64);
q1s32 = vmlal_s16(q1s32, d16s16, d15s16);
q2s32 = vmlal_s16(q2s32, d17s16, d15s16);
q3s32 = vmlsl_s16(q3s32, d16s16, d14s16);
q4s32 = vmlsl_s16(q4s32, d17s16, d14s16);
q5s32 = vmull_s16(d22s16, d30s16);
q6s32 = vmull_s16(d23s16, d30s16);
q7s32 = vmull_s16(d22s16, d31s16);
q8s32 = vmull_s16(d23s16, d31s16);
q5s32 = vmlal_s16(q5s32, d24s16, d31s16);
q6s32 = vmlal_s16(q6s32, d25s16, d31s16);
q7s32 = vmlsl_s16(q7s32, d24s16, d30s16);
q8s32 = vmlsl_s16(q8s32, d25s16, d30s16);
q11s32 = vaddq_s32(q1s32, q5s32);
q12s32 = vaddq_s32(q2s32, q6s32);
q1s32 = vsubq_s32(q1s32, q5s32);
q2s32 = vsubq_s32(q2s32, q6s32);
d22s16 = vqrshrn_n_s32(q11s32, 14);
d23s16 = vqrshrn_n_s32(q12s32, 14);
*q11s16 = vcombine_s16(d22s16, d23s16);
q12s32 = vaddq_s32(q3s32, q7s32);
q15s32 = vaddq_s32(q4s32, q8s32);
q3s32 = vsubq_s32(q3s32, q7s32);
q4s32 = vsubq_s32(q4s32, q8s32);
d2s16 = vqrshrn_n_s32(q1s32, 14);
d3s16 = vqrshrn_n_s32(q2s32, 14);
d24s16 = vqrshrn_n_s32(q12s32, 14);
d25s16 = vqrshrn_n_s32(q15s32, 14);
d6s16 = vqrshrn_n_s32(q3s32, 14);
d7s16 = vqrshrn_n_s32(q4s32, 14);
*q12s16 = vcombine_s16(d24s16, d25s16);
d0s16 = vdup_n_s16(cospi_10_64);
d1s16 = vdup_n_s16(cospi_22_64);
q4s32 = vmull_s16(d26s16, d0s16);
q5s32 = vmull_s16(d27s16, d0s16);
q2s32 = vmull_s16(d26s16, d1s16);
q6s32 = vmull_s16(d27s16, d1s16);
d30s16 = vdup_n_s16(cospi_26_64);
d31s16 = vdup_n_s16(cospi_6_64);
q4s32 = vmlal_s16(q4s32, d20s16, d1s16);
q5s32 = vmlal_s16(q5s32, d21s16, d1s16);
q2s32 = vmlsl_s16(q2s32, d20s16, d0s16);
q6s32 = vmlsl_s16(q6s32, d21s16, d0s16);
q0s32 = vmull_s16(d18s16, d30s16);
q13s32 = vmull_s16(d19s16, d30s16);
q0s32 = vmlal_s16(q0s32, d28s16, d31s16);
q13s32 = vmlal_s16(q13s32, d29s16, d31s16);
q10s32 = vmull_s16(d18s16, d31s16);
q9s32 = vmull_s16(d19s16, d31s16);
q10s32 = vmlsl_s16(q10s32, d28s16, d30s16);
q9s32 = vmlsl_s16(q9s32, d29s16, d30s16);
q14s32 = vaddq_s32(q2s32, q10s32);
q15s32 = vaddq_s32(q6s32, q9s32);
q2s32 = vsubq_s32(q2s32, q10s32);
q6s32 = vsubq_s32(q6s32, q9s32);
d28s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d4s16 = vqrshrn_n_s32(q2s32, 14);
d5s16 = vqrshrn_n_s32(q6s32, 14);
*q14s16 = vcombine_s16(d28s16, d29s16);
q9s32 = vaddq_s32(q4s32, q0s32);
q10s32 = vaddq_s32(q5s32, q13s32);
q4s32 = vsubq_s32(q4s32, q0s32);
q5s32 = vsubq_s32(q5s32, q13s32);
d30s16 = vdup_n_s16(cospi_8_64);
d31s16 = vdup_n_s16(cospi_24_64);
d18s16 = vqrshrn_n_s32(q9s32, 14);
d19s16 = vqrshrn_n_s32(q10s32, 14);
d8s16 = vqrshrn_n_s32(q4s32, 14);
d9s16 = vqrshrn_n_s32(q5s32, 14);
*q9s16 = vcombine_s16(d18s16, d19s16);
q5s32 = vmull_s16(d2s16, d30s16);
q6s32 = vmull_s16(d3s16, d30s16);
q7s32 = vmull_s16(d2s16, d31s16);
q0s32 = vmull_s16(d3s16, d31s16);
q5s32 = vmlal_s16(q5s32, d6s16, d31s16);
q6s32 = vmlal_s16(q6s32, d7s16, d31s16);
q7s32 = vmlsl_s16(q7s32, d6s16, d30s16);
q0s32 = vmlsl_s16(q0s32, d7s16, d30s16);
q1s32 = vmull_s16(d4s16, d30s16);
q3s32 = vmull_s16(d5s16, d30s16);
q10s32 = vmull_s16(d4s16, d31s16);
q2s32 = vmull_s16(d5s16, d31s16);
q1s32 = vmlsl_s16(q1s32, d8s16, d31s16);
q3s32 = vmlsl_s16(q3s32, d9s16, d31s16);
q10s32 = vmlal_s16(q10s32, d8s16, d30s16);
q2s32 = vmlal_s16(q2s32, d9s16, d30s16);
*q8s16 = vaddq_s16(*q11s16, *q9s16);
*q11s16 = vsubq_s16(*q11s16, *q9s16);
q4s16 = vaddq_s16(*q12s16, *q14s16);
*q12s16 = vsubq_s16(*q12s16, *q14s16);
q14s32 = vaddq_s32(q5s32, q1s32);
q15s32 = vaddq_s32(q6s32, q3s32);
q5s32 = vsubq_s32(q5s32, q1s32);
q6s32 = vsubq_s32(q6s32, q3s32);
d18s16 = vqrshrn_n_s32(q14s32, 14);
d19s16 = vqrshrn_n_s32(q15s32, 14);
d10s16 = vqrshrn_n_s32(q5s32, 14);
d11s16 = vqrshrn_n_s32(q6s32, 14);
*q9s16 = vcombine_s16(d18s16, d19s16);
q1s32 = vaddq_s32(q7s32, q10s32);
q3s32 = vaddq_s32(q0s32, q2s32);
q7s32 = vsubq_s32(q7s32, q10s32);
q0s32 = vsubq_s32(q0s32, q2s32);
d28s16 = vqrshrn_n_s32(q1s32, 14);
d29s16 = vqrshrn_n_s32(q3s32, 14);
d14s16 = vqrshrn_n_s32(q7s32, 14);
d15s16 = vqrshrn_n_s32(q0s32, 14);
*q14s16 = vcombine_s16(d28s16, d29s16);
d30s16 = vdup_n_s16(cospi_16_64);
d22s16 = vget_low_s16(*q11s16);
d23s16 = vget_high_s16(*q11s16);
q2s32 = vmull_s16(d22s16, d30s16);
q3s32 = vmull_s16(d23s16, d30s16);
q13s32 = vmull_s16(d22s16, d30s16);
q1s32 = vmull_s16(d23s16, d30s16);
d24s16 = vget_low_s16(*q12s16);
d25s16 = vget_high_s16(*q12s16);
q2s32 = vmlal_s16(q2s32, d24s16, d30s16);
q3s32 = vmlal_s16(q3s32, d25s16, d30s16);
q13s32 = vmlsl_s16(q13s32, d24s16, d30s16);
q1s32 = vmlsl_s16(q1s32, d25s16, d30s16);
d4s16 = vqrshrn_n_s32(q2s32, 14);
d5s16 = vqrshrn_n_s32(q3s32, 14);
d24s16 = vqrshrn_n_s32(q13s32, 14);
d25s16 = vqrshrn_n_s32(q1s32, 14);
q2s16 = vcombine_s16(d4s16, d5s16);
*q12s16 = vcombine_s16(d24s16, d25s16);
q13s32 = vmull_s16(d10s16, d30s16);
q1s32 = vmull_s16(d11s16, d30s16);
q11s32 = vmull_s16(d10s16, d30s16);
q0s32 = vmull_s16(d11s16, d30s16);
q13s32 = vmlal_s16(q13s32, d14s16, d30s16);
q1s32 = vmlal_s16(q1s32, d15s16, d30s16);
q11s32 = vmlsl_s16(q11s32, d14s16, d30s16);
q0s32 = vmlsl_s16(q0s32, d15s16, d30s16);
d20s16 = vqrshrn_n_s32(q13s32, 14);
d21s16 = vqrshrn_n_s32(q1s32, 14);
d12s16 = vqrshrn_n_s32(q11s32, 14);
d13s16 = vqrshrn_n_s32(q0s32, 14);
*q10s16 = vcombine_s16(d20s16, d21s16);
q6s16 = vcombine_s16(d12s16, d13s16);
q5s16 = vdupq_n_s16(0);
*q9s16 = vsubq_s16(q5s16, *q9s16);
*q11s16 = vsubq_s16(q5s16, q2s16);
*q13s16 = vsubq_s16(q5s16, q6s16);
*q15s16 = vsubq_s16(q5s16, q4s16);
return;
}
rfx_dwt_2d_decode_block_vert_NEON(sint16 * l, sint16 * h, sint16 * dst, int subband_width)
{
int x, n;
sint16 * l_ptr = l;
sint16 * h_ptr = h;
sint16 * 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;
}
}
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);
}
}
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;
}
rfx_dwt_2d_decode_block_horiz_NEON(sint16 * l, sint16 * h, sint16 * dst, int subband_width)
{
int y, n;
sint16 * l_ptr = l;
sint16 * h_ptr = h;
sint16 * 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 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_idct4x4_16_add_neon(const tran_low_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 = load_tran_low_to_s16(input);
q9s16 = load_tran_low_to_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);
}
void vp9_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
uint8x16_t q0u8, q1u8, q2u8;
int16x8_t q12s16, q13s16, q14s16, q15s16;
uint16x4_t d6u16;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
q0u8 = vld1q_dup_u8(above - 1);
q1u8 = vld1q_u8(above);
q2u8 = vld1q_u8(above + 16);
q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
for (k = 0; k < 4; k++, left += 8) {
d26u8 = vld1_u8(left);
q3u16 = vmovl_u8(d26u8);
d6u16 = vget_low_u16(q3u16);
for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
q0u16 = vdupq_lane_u16(d6u16, 0);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 1);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 2);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 3);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
}
}
}
void vp9_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
uint8x16_t q0u8, q1u8;
int16x8_t q0s16, q1s16, q8s16, q11s16;
uint16x4_t d20u16;
uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
q0u8 = vld1q_dup_u8(above - 1);
q1u8 = vld1q_u8(above);
q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
for (k = 0; k < 2; k++, left += 8) {
d18u8 = vld1_u8(left);
q10u16 = vmovl_u8(d18u8);
d20u16 = vget_low_u16(q10u16);
for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
q0u16 = vdupq_lane_u16(d20u16, 0);
q8u16 = vdupq_lane_u16(d20u16, 1);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q2u16));
q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q3u16));
q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q2u16));
q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q3u16));
d2u8 = vqmovun_s16(q1s16);
d3u8 = vqmovun_s16(q0s16);
d22u8 = vqmovun_s16(q11s16);
d23u8 = vqmovun_s16(q8s16);
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
dst += stride;
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
dst += stride;
q0u16 = vdupq_lane_u16(d20u16, 2);
q8u16 = vdupq_lane_u16(d20u16, 3);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q2u16));
q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q3u16));
q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q2u16));
q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q3u16));
d2u8 = vqmovun_s16(q1s16);
d3u8 = vqmovun_s16(q0s16);
d22u8 = vqmovun_s16(q11s16);
d23u8 = vqmovun_s16(q8s16);
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
dst += stride;
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
dst += stride;
}
}
}
void vp9_quantize_fp_neon(const int16_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, int16_t *qcoeff_ptr,
int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, 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)zbin_oq_value;
(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 = vld1q_s16(&coeff_ptr[0]);
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);
vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
vst1q_s16(&dqcoeff_ptr[0], 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 = vld1q_s16(&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);
vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
vst1q_s16(&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 {
vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
*eob_ptr = 0;
}
}