static void var_filter_block2d_bil_w16(const uint8_t *src_ptr,
uint8_t *output_ptr,
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
unsigned int output_width,
const uint8_t *filter) {
const uint8x8_t f0 = vmov_n_u8(filter[0]);
const uint8x8_t f1 = vmov_n_u8(filter[1]);
unsigned int i, j;
for (i = 0; i < output_height; ++i) {
for (j = 0; j < output_width; j += 16) {
const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]);
const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]);
const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0);
const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1);
const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS);
const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
}
// Next row...
src_ptr += src_pixels_per_line;
output_ptr += output_width;
}
}
// 'do_above' and 'do_left' facilitate branch removal when inlined.
static INLINE void dc_32x32(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 uint8x16_t A0 = vld1q_u8(above); // top row
const uint8x16_t A1 = vld1q_u8(above + 16);
const uint16x8_t p0 = vpaddlq_u8(A0); // cascading summation of the top
const uint16x8_t p1 = vpaddlq_u8(A1);
const uint16x8_t p2 = vaddq_u16(p0, p1);
const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
const uint16x4_t p4 = vpadd_u16(p3, p3);
const uint16x4_t p5 = vpadd_u16(p4, p4);
sum_top = vcombine_u16(p5, p5);
}
if (do_left) {
const uint8x16_t L0 = vld1q_u8(left); // left row
const uint8x16_t L1 = vld1q_u8(left + 16);
const uint16x8_t p0 = vpaddlq_u8(L0); // cascading summation of the left
const uint16x8_t p1 = vpaddlq_u8(L1);
const uint16x8_t p2 = vaddq_u16(p0, p1);
const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
const uint16x4_t p4 = vpadd_u16(p3, p3);
const uint16x4_t p5 = vpadd_u16(p4, p4);
sum_left = vcombine_u16(p5, p5);
}
if (do_above && do_left) {
const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
dc0 = vrshrn_n_u16(sum, 6);
} else if (do_above) {
dc0 = vrshrn_n_u16(sum_top, 5);
} else if (do_left) {
dc0 = vrshrn_n_u16(sum_left, 5);
} else {
dc0 = vdup_n_u8(0x80);
}
{
const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
int i;
for (i = 0; i < 32; ++i) {
vst1q_u8(dst + i * stride, dc);
vst1q_u8(dst + i * stride + 16, dc);
}
}
}
// 'do_above' and 'do_left' facilitate branch removal when inlined.
static INLINE void dc_16x16(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 uint8x16_t A = vld1q_u8(above); // top row
const uint16x8_t p0 = vpaddlq_u8(A); // cascading summation of the top
const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
const uint16x4_t p2 = vpadd_u16(p1, p1);
const uint16x4_t p3 = vpadd_u16(p2, p2);
sum_top = vcombine_u16(p3, p3);
}
if (do_left) {
const uint8x16_t L = vld1q_u8(left); // left row
const uint16x8_t p0 = vpaddlq_u8(L); // cascading summation of the left
const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
const uint16x4_t p2 = vpadd_u16(p1, p1);
const uint16x4_t p3 = vpadd_u16(p2, p2);
sum_left = vcombine_u16(p3, p3);
}
if (do_above && do_left) {
const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
dc0 = vrshrn_n_u16(sum, 5);
} else if (do_above) {
dc0 = vrshrn_n_u16(sum_top, 4);
} else if (do_left) {
dc0 = vrshrn_n_u16(sum_left, 4);
} else {
dc0 = vdup_n_u8(0x80);
}
{
const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
int i;
for (i = 0; i < 16; ++i) {
vst1q_u8(dst + i * stride, dc);
}
}
}
// '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));
}
}
}
// Process a block exactly 8 wide and any height.
static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
uint8_t *output_ptr,
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
const uint8_t *filter) {
const uint8x8_t f0 = vdup_n_u8(filter[0]);
const uint8x8_t f1 = vdup_n_u8(filter[1]);
unsigned int i;
for (i = 0; i < output_height; ++i) {
const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
const uint16x8_t a = vmull_u8(src_0, f0);
const uint16x8_t b = vmlal_u8(a, src_1, f1);
const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
vst1_u8(output_ptr, out);
src_ptr += src_pixels_per_line;
output_ptr += 8;
}
}
static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
uint8_t *output_ptr,
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
unsigned int output_width,
const uint16_t *vpx_filter) {
const uint8x8_t f0 = vmov_n_u8((uint8_t)vpx_filter[0]);
const uint8x8_t f1 = vmov_n_u8((uint8_t)vpx_filter[1]);
unsigned int i;
for (i = 0; i < output_height; ++i) {
const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
const uint16x8_t a = vmull_u8(src_0, f0);
const uint16x8_t b = vmlal_u8(a, src_1, f1);
const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
vst1_u8(&output_ptr[0], out);
// Next row...
src_ptr += src_pixels_per_line;
output_ptr += output_width;
}
}
