ref: a9bbe53dbb8ce599100ceb7ce864d8d81dfcb858
parent: fb40b5d7a76978e6fcedf5997fd7a45571aa6b0d
author: Linfeng Zhang <linfengz@google.com>
date: Thu Sep 7 08:50:36 EDT 2017
Add 4 to 1 scaling NEON optimization BUG=webm:1419 Change-Id: If82a93935d2453e61b7647aae70983db1740bec7
--- a/vp9/encoder/arm/neon/vp9_frame_scale_neon.c
+++ b/vp9/encoder/arm/neon/vp9_frame_scale_neon.c
@@ -45,11 +45,61 @@
} while (--y);
}
+static INLINE void scale_plane_4_to_1_phase_0_neon(const uint8_t *src,
+ const int src_stride,
+ uint8_t *dst,
+ const int dst_stride,
+ const int w, const int h) {+ const int max_width = (w + 15) & ~15;
+ int y = h;
+
+ assert(w && h);
+
+ do {+ int x = max_width;
+ do {+ const uint8x16x4_t s = vld4q_u8(src);
+ vst1q_u8(dst, s.val[0]);
+ src += 64;
+ dst += 16;
+ x -= 16;
+ } while (x);
+ src += 4 * (src_stride - max_width);
+ dst += dst_stride - max_width;
+ } while (--y);
+}
+
+static INLINE void scale_plane_bilinear_phase_non_0_kernel(
+ const uint8x16_t in0, const uint8x16_t in1, const uint8x16_t in2,
+ const uint8x16_t in3, const uint8x8_t coef0, const uint8x8_t coef1,
+ uint8_t *const dst) {+ const uint16x8_t h0 = vmull_u8(vget_low_u8(in0), coef0);
+ const uint16x8_t h1 = vmull_u8(vget_high_u8(in0), coef0);
+ const uint16x8_t h2 = vmull_u8(vget_low_u8(in2), coef0);
+ const uint16x8_t h3 = vmull_u8(vget_high_u8(in2), coef0);
+ const uint16x8_t h4 = vmlal_u8(h0, vget_low_u8(in1), coef1);
+ const uint16x8_t h5 = vmlal_u8(h1, vget_high_u8(in1), coef1);
+ const uint16x8_t h6 = vmlal_u8(h2, vget_low_u8(in3), coef1);
+ const uint16x8_t h7 = vmlal_u8(h3, vget_high_u8(in3), coef1);
+
+ const uint8x8_t hor0 = vrshrn_n_u16(h4, 7); // temp: 00 01 02 03 04 05 06 07
+ const uint8x8_t hor1 = vrshrn_n_u16(h5, 7); // temp: 08 09 0A 0B 0C 0D 0E 0F
+ const uint8x8_t hor2 = vrshrn_n_u16(h6, 7); // temp: 10 11 12 13 14 15 16 17
+ const uint8x8_t hor3 = vrshrn_n_u16(h7, 7); // temp: 18 19 1A 1B 1C 1D 1E 1F
+ const uint16x8_t v0 = vmull_u8(hor0, coef0);
+ const uint16x8_t v1 = vmull_u8(hor1, coef0);
+ const uint16x8_t v2 = vmlal_u8(v0, hor2, coef1);
+ const uint16x8_t v3 = vmlal_u8(v1, hor3, coef1);
+ // dst: 0 1 2 3 4 5 6 7 8 9 A B C D E F
+ const uint8x16_t d = vcombine_u8(vrshrn_n_u16(v2, 7), vrshrn_n_u16(v3, 7));
+ vst1q_u8(dst, d);
+}
+
static INLINE void scale_plane_2_to_1_bilinear_phase_non_0_neon(
const uint8_t *const src, const int src_stride, uint8_t *dst,
const int dst_stride, const int w, const int h, const int16_t c0,
const int16_t c1) {- const int max_width = (w + 7) & ~7;
+ const int max_width = (w + 15) & ~15;
const uint8_t *src0 = src;
const uint8_t *src1 = src + src_stride;
const uint8x8_t coef0 = vdup_n_u8(c0);
@@ -61,25 +111,18 @@
do {int x = max_width;
do {- // 00 02 04 06 08 0A 0C 0E 01 03 05 07 09 0B 0D 0F
- const uint8x8x2_t s0 = vld2_u8(src0);
- // 10 12 14 16 18 1A 1C 1E 11 13 15 17 19 1B 1D 1F
- const uint8x8x2_t s1 = vld2_u8(src1);
- const uint16x8_t h0 = vmull_u8(s0.val[0], coef0);
- const uint16x8_t h1 = vmull_u8(s1.val[0], coef0);
- const uint16x8_t h2 = vmlal_u8(h0, s0.val[1], coef1);
- const uint16x8_t h3 = vmlal_u8(h1, s1.val[1], coef1);
-
- const uint8x8_t hor0 = vrshrn_n_u16(h2, 7); // 00 01 02 03 04 05 06 07
- const uint8x8_t hor1 = vrshrn_n_u16(h3, 7); // 10 11 12 13 14 15 16 17
- const uint16x8_t v0 = vmull_u8(hor0, coef0);
- const uint16x8_t v1 = vmlal_u8(v0, hor1, coef1);
- const uint8x8_t d = vrshrn_n_u16(v1, 7); // 00 01 02 03 04 05 06 07
- vst1_u8(dst, d);
- src0 += 16;
- src1 += 16;
- dst += 8;
- x -= 8;
+ // 000 002 004 006 008 00A 00C 00E 010 012 014 016 018 01A 01C 01E
+ // 001 003 005 007 009 00B 00D 00F 011 013 015 017 019 01B 01D 01F
+ const uint8x16x2_t s0 = vld2q_u8(src0);
+ // 100 102 104 106 108 10A 10C 10E 110 112 114 116 118 11A 11C 11E
+ // 101 103 105 107 109 10B 10D 10F 111 113 115 117 119 11B 11D 11F
+ const uint8x16x2_t s1 = vld2q_u8(src1);
+ scale_plane_bilinear_phase_non_0_kernel(s0.val[0], s0.val[1], s1.val[0],
+ s1.val[1], coef0, coef1, dst);
+ src0 += 32;
+ src1 += 32;
+ dst += 16;
+ x -= 16;
} while (x);
src0 += 2 * (src_stride - max_width);
src1 += 2 * (src_stride - max_width);
@@ -87,6 +130,46 @@
} while (--y);
}
+static INLINE void scale_plane_4_to_1_bilinear_phase_non_0_neon(
+ const uint8_t *const src, const int src_stride, uint8_t *dst,
+ const int dst_stride, const int w, const int h, const int16_t c0,
+ const int16_t c1) {+ const int max_width = (w + 15) & ~15;
+ const uint8_t *src0 = src;
+ const uint8_t *src1 = src + src_stride;
+ const uint8x8_t coef0 = vdup_n_u8(c0);
+ const uint8x8_t coef1 = vdup_n_u8(c1);
+ int y = h;
+
+ assert(w && h);
+
+ do {+ int x = max_width;
+ do {+ // (*) -- useless
+ // 000 004 008 00C 010 014 018 01C 020 024 028 02C 030 034 038 03C
+ // 001 005 009 00D 011 015 019 01D 021 025 029 02D 031 035 039 03D
+ // 002 006 00A 00E 012 016 01A 01E 022 026 02A 02E 032 036 03A 03E (*)
+ // 003 007 00B 00F 013 017 01B 01F 023 027 02B 02F 033 037 03B 03F (*)
+ const uint8x16x4_t s0 = vld4q_u8(src0);
+ // 100 104 108 10C 110 114 118 11C 120 124 128 12C 130 134 138 13C
+ // 101 105 109 10D 111 115 119 11D 121 125 129 12D 131 135 139 13D
+ // 102 106 10A 10E 112 116 11A 11E 122 126 12A 12E 132 136 13A 13E (*)
+ // 103 107 10B 10F 113 117 11B 11F 123 127 12B 12F 133 137 13B 13F (*)
+ const uint8x16x4_t s1 = vld4q_u8(src1);
+ scale_plane_bilinear_phase_non_0_kernel(s0.val[0], s0.val[1], s1.val[0],
+ s1.val[1], coef0, coef1, dst);
+ src0 += 64;
+ src1 += 64;
+ dst += 16;
+ x -= 16;
+ } while (x);
+ src0 += 4 * (src_stride - max_width);
+ src1 += 4 * (src_stride - max_width);
+ dst += dst_stride - max_width;
+ } while (--y);
+}
+
static INLINE uint8x8_t scale_filter(const uint8x8_t *const s,
const int16x8_t filters) {const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
@@ -126,7 +209,7 @@
src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 1;
// horizontal
- // Note: processing 8x8 is about 20% faster than processing row by row using
+ // Note: processing 4x8 is about 20% faster than processing row by row using
// vld4_u8().
do {load_u8_8x8(src + 2, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
@@ -141,10 +224,14 @@
transpose_u8_8x8(&s[6], &s[7], &s[8], &s[9], &s[10], &s[11], &s[12],
&s[13]);
- d[0] = scale_filter(&s[0], filters);
- d[1] = scale_filter(&s[2], filters);
- d[2] = scale_filter(&s[4], filters);
- d[3] = scale_filter(&s[6], filters);
+ d[0] = scale_filter(&s[0], filters); // 00 10 20 30 40 50 60 70
+ d[1] = scale_filter(&s[2], filters); // 01 11 21 31 41 51 61 71
+ d[2] = scale_filter(&s[4], filters); // 02 12 22 32 42 52 62 72
+ d[3] = scale_filter(&s[6], filters); // 03 13 23 33 43 53 63 73
+ // 00 01 02 03 40 41 42 43
+ // 10 11 12 13 50 51 52 53
+ // 20 21 22 23 60 61 62 63
+ // 30 31 32 33 70 71 72 73
transpose_u8_8x4(&d[0], &d[1], &d[2], &d[3]);
vst1_lane_u32((uint32_t *)(t + 0 * max_width_hor),
vreinterpret_u32_u8(d[0]), 0);
@@ -219,6 +306,111 @@
} while (x);
}
+static void scale_plane_4_to_1_general_neon(const uint8_t *src,
+ const int src_stride, uint8_t *dst,
+ const int dst_stride, const int w,
+ const int h,
+ const int16_t *const coef,
+ uint8_t *const temp_buffer) {+ const int max_width_hor = (w + 1) & ~1;
+ const int max_width_ver = (w + 7) & ~7;
+ const int max_height_hor = (4 * h + SUBPEL_TAPS - 2 + 7) & ~7;
+ const int max_height_ver = (h + 1) & ~1;
+ const int16x8_t filters = vld1q_s16(coef);
+ int x, y = max_height_hor;
+ uint8_t *t = temp_buffer;
+ uint8x8_t s[12], d[2];
+
+ assert(w && h);
+
+ src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 3;
+
+ // horizontal
+ // Note: processing 2x8 is about 20% faster than processing row by row using
+ // vld4_u8().
+ do {+ load_u8_8x8(src + 4, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
+ &s[6], &s[7]);
+ transpose_u8_4x8(&s[0], &s[1], &s[2], &s[3], s[4], s[5], s[6], s[7]);
+ x = max_width_hor;
+
+ do {+ uint8x8x2_t dd;
+ src += 8;
+ load_u8_8x8(src, src_stride, &s[4], &s[5], &s[6], &s[7], &s[8], &s[9],
+ &s[10], &s[11]);
+ transpose_u8_8x8(&s[4], &s[5], &s[6], &s[7], &s[8], &s[9], &s[10],
+ &s[11]);
+
+ d[0] = scale_filter(&s[0], filters); // 00 10 20 30 40 50 60 70
+ d[1] = scale_filter(&s[4], filters); // 01 11 21 31 41 51 61 71
+ // dd.val[0]: 00 01 20 21 40 41 60 61
+ // dd.val[1]: 10 11 30 31 50 51 70 71
+ dd = vtrn_u8(d[0], d[1]);
+ vst1_lane_u16((uint16_t *)(t + 0 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[0]), 0);
+ vst1_lane_u16((uint16_t *)(t + 1 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[1]), 0);
+ vst1_lane_u16((uint16_t *)(t + 2 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[0]), 1);
+ vst1_lane_u16((uint16_t *)(t + 3 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[1]), 1);
+ vst1_lane_u16((uint16_t *)(t + 4 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[0]), 2);
+ vst1_lane_u16((uint16_t *)(t + 5 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[1]), 2);
+ vst1_lane_u16((uint16_t *)(t + 6 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[0]), 3);
+ vst1_lane_u16((uint16_t *)(t + 7 * max_width_hor),
+ vreinterpret_u16_u8(dd.val[1]), 3);
+
+ s[0] = s[8];
+ s[1] = s[9];
+ s[2] = s[10];
+ s[3] = s[11];
+
+ t += 2;
+ x -= 2;
+ } while (x);
+ src += 8 * src_stride - 4 * max_width_hor;
+ t += 7 * max_width_hor;
+ y -= 8;
+ } while (y);
+
+ // vertical
+ x = max_width_ver;
+ t = temp_buffer;
+ do {+ load_u8_8x4(t, max_width_hor, &s[0], &s[1], &s[2], &s[3]);
+ t += 4 * max_width_hor;
+ y = max_height_ver;
+
+ do {+ load_u8_8x8(t, max_width_hor, &s[4], &s[5], &s[6], &s[7], &s[8], &s[9],
+ &s[10], &s[11]);
+ t += 8 * max_width_hor;
+
+ d[0] = scale_filter(&s[0], filters);
+ d[1] = scale_filter(&s[4], filters);
+ vst1_u8(dst + 0 * dst_stride, d[0]);
+ vst1_u8(dst + 1 * dst_stride, d[1]);
+
+ s[0] = s[8];
+ s[1] = s[9];
+ s[2] = s[10];
+ s[3] = s[11];
+
+ dst += 2 * dst_stride;
+ y -= 2;
+ } while (y);
+ t -= max_width_hor * (4 * max_height_ver + 4);
+ t += 8;
+ dst -= max_height_ver * dst_stride;
+ dst += 8;
+ x -= 8;
+ } while (x);
+}
+
void vp9_scale_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
YV12_BUFFER_CONFIG *dst,
INTERP_FILTER filter_type,
@@ -269,6 +461,52 @@
dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
temp_buffer);
scale_plane_2_to_1_general_neon(
+ src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
+ dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
+ temp_buffer);
+ free(temp_buffer);
+ } else {+ scaled = 0;
+ }
+ }
+ } else if (4 * dst_w == src_w && 4 * dst_h == src_h) {+ scaled = 1;
+ if (phase_scaler == 0) {+ scale_plane_4_to_1_phase_0_neon(src->y_buffer, src->y_stride,
+ dst->y_buffer, dst->y_stride, dst_w,
+ dst_h);
+ scale_plane_4_to_1_phase_0_neon(src->u_buffer, src->uv_stride,
+ dst->u_buffer, dst->uv_stride, dst_uv_w,
+ dst_uv_h);
+ scale_plane_4_to_1_phase_0_neon(src->v_buffer, src->uv_stride,
+ dst->v_buffer, dst->uv_stride, dst_uv_w,
+ dst_uv_h);
+ } else if (filter_type == BILINEAR) {+ const int16_t c0 = vp9_filter_kernels[BILINEAR][phase_scaler][3];
+ const int16_t c1 = vp9_filter_kernels[BILINEAR][phase_scaler][4];
+ scale_plane_4_to_1_bilinear_phase_non_0_neon(src->y_buffer, src->y_stride,
+ dst->y_buffer, dst->y_stride,
+ dst_w, dst_h, c0, c1);
+ scale_plane_4_to_1_bilinear_phase_non_0_neon(
+ src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
+ dst_uv_w, dst_uv_h, c0, c1);
+ scale_plane_4_to_1_bilinear_phase_non_0_neon(
+ src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
+ dst_uv_w, dst_uv_h, c0, c1);
+ } else {+ const int buffer_max_width = (dst_w + 1) & ~1;
+ const int buffer_max_height = (4 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
+ uint8_t *const temp_buffer =
+ (uint8_t *)malloc(buffer_max_width * buffer_max_height);
+ if (temp_buffer) {+ scale_plane_4_to_1_general_neon(
+ src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
+ dst_h, vp9_filter_kernels[filter_type][phase_scaler], temp_buffer);
+ scale_plane_4_to_1_general_neon(
+ src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
+ dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
+ temp_buffer);
+ scale_plane_4_to_1_general_neon(
src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
temp_buffer);
--- a/vpx_dsp/arm/vpx_convolve8_neon.c
+++ b/vpx_dsp/arm/vpx_convolve8_neon.c
@@ -30,18 +30,6 @@
// instructions. This optimization is much faster in speed unit test, but slowed
// down the whole decoder by 5%.
-static INLINE void load_u8_8x4(const uint8_t *s, const ptrdiff_t p,
- uint8x8_t *const s0, uint8x8_t *const s1,
- uint8x8_t *const s2, uint8x8_t *const s3) {- *s0 = vld1_u8(s);
- s += p;
- *s1 = vld1_u8(s);
- s += p;
- *s2 = vld1_u8(s);
- s += p;
- *s3 = vld1_u8(s);
-}
-
static INLINE void store_u8_8x8(uint8_t *s, const ptrdiff_t p,
const uint8x8_t s0, const uint8x8_t s1,
const uint8x8_t s2, const uint8x8_t s3,
--- a/vpx_dsp/arm/vpx_convolve8_neon.h
+++ b/vpx_dsp/arm/vpx_convolve8_neon.h
@@ -13,6 +13,18 @@
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
+static INLINE void load_u8_8x4(const uint8_t *s, const ptrdiff_t p,
+ uint8x8_t *const s0, uint8x8_t *const s1,
+ uint8x8_t *const s2, uint8x8_t *const s3) {+ *s0 = vld1_u8(s);
+ s += p;
+ *s1 = vld1_u8(s);
+ s += p;
+ *s2 = vld1_u8(s);
+ s += p;
+ *s3 = vld1_u8(s);
+}
+
static INLINE void load_u8_8x8(const uint8_t *s, const ptrdiff_t p,
uint8x8_t *const s0, uint8x8_t *const s1,
uint8x8_t *const s2, uint8x8_t *const s3,
--
⑨