void vp9_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
uint8x8_t d0u8 = vdup_n_u8(0);
uint64x1_t d1u64 = vdup_n_u64(0);
(void)above;
d1u64 = vld1_u64((const uint64_t *)left);
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
vst1_u8(dst, d0u8);
}
void test_vdup_laneu8 (void)
{
uint8x8_t out_uint8x8_t;
uint8x8_t arg0_uint8x8_t;
out_uint8x8_t = vdup_lane_u8 (arg0_uint8x8_t, 1);
}
void vp9_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
uint8x8_t d0u8 = vdup_n_u8(0);
uint32x2_t d1u32 = vdup_n_u32(0);
(void)above;
d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
}
// 'do_above' and 'do_left' facilitate branch removal when inlined.
static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left,
int do_above, int do_left) {
uint16x8_t sum_top;
uint16x8_t sum_left;
uint8x8_t dc0;
if (do_above) {
const uint8x8_t A = vld1_u8(above); // top row
const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top
const uint16x4_t p1 = vpadd_u16(p0, p0);
const uint16x4_t p2 = vpadd_u16(p1, p1);
sum_top = vcombine_u16(p2, p2);
}
if (do_left) {
const uint8x8_t L = vld1_u8(left); // left border
const uint16x4_t p0 = vpaddl_u8(L); // cascading summation of the left
const uint16x4_t p1 = vpadd_u16(p0, p0);
const uint16x4_t p2 = vpadd_u16(p1, p1);
sum_left = vcombine_u16(p2, p2);
}
if (do_above && do_left) {
const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
dc0 = vrshrn_n_u16(sum, 4);
} else if (do_above) {
dc0 = vrshrn_n_u16(sum_top, 3);
} else if (do_left) {
dc0 = vrshrn_n_u16(sum_left, 3);
} else {
dc0 = vdup_n_u8(0x80);
}
{
const uint8x8_t dc = vdup_lane_u8(dc0, 0);
int i;
for (i = 0; i < 8; ++i) {
vst1_u32((uint32_t*)(dst + i * stride), vreinterpret_u32_u8(dc));
}
}
}
uint8x8_t test_vdup_lane_u8(uint8x8_t v1) {
// CHECK: test_vdup_lane_u8
return vdup_lane_u8(v1, 5);
// CHECK: dup {{v[0-9]+}}.8b, {{v[0-9]+}}.b[5]
}
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;
}
}
