ref: 580d32240f9a2f7783ced9b0e409de62ba07f4e1
parent: 1fa3ec3023ef1b841d993bc21cba32b658a790ea
author: Linfeng Zhang <linfengz@google.com>
date: Wed Oct 11 07:59:04 EDT 2017
Add 4 to 3 scaling SSSE3 optimization Note this change will trigger the different C version on SSSE3 and generate different scaled output. Its speed is 2x compared with the version calling vpx_scaled_2d_ssse3(). Change-Id: I17fff122cd0a5ac8aa451d84daa606582da8e194
--- a/vp9/encoder/vp9_frame_scale.c
+++ b/vp9/encoder/vp9_frame_scale.c
@@ -28,7 +28,7 @@
const InterpKernel *const kernel = vp9_filter_kernels[filter_type];
int x, y, i;
-#if HAVE_NEON
+#if HAVE_SSSE3 || HAVE_NEON
// TODO(linfengz): The 4:3 specialized C code is disabled by default since
// it's much slower than the general version which calls vpx_scaled_2d() even
// if vpx_scaled_2d() is not optimized. It will only be enabled as a reference
--- a/vp9/encoder/x86/vp9_frame_scale_ssse3.c
+++ b/vp9/encoder/x86/vp9_frame_scale_ssse3.c
@@ -438,6 +438,202 @@
} while (x);
}
+typedef void (*shuffle_filter_funcs)(const int16_t *const filter,
+ __m128i *const f);
+
+typedef __m128i (*convolve8_funcs)(const __m128i *const s,
+ const __m128i *const f);
+
+static void scale_plane_4_to_3_general(const uint8_t *src, const int src_stride,
+ uint8_t *dst, const int dst_stride,
+ const int w, const int h,
+ const InterpKernel *const coef,
+ const int phase_scaler,
+ uint8_t *const temp_buffer) {+ static const int step_q4 = 16 * 4 / 3;
+ const int width_hor = (w + 5) - ((w + 5) % 6);
+ const int stride_hor = 2 * width_hor + 4; // store 4 extra pixels
+ const int width_ver = (w + 7) & ~7;
+ // We need (SUBPEL_TAPS - 1) extra rows: (SUBPEL_TAPS / 2 - 1) extra rows
+ // above and (SUBPEL_TAPS / 2) extra rows below.
+ const int height_hor = (4 * h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
+ const int height_ver = (h + 5) - ((h + 5) % 6);
+ int x, y = height_hor;
+ uint8_t *t = temp_buffer;
+ __m128i s[12], d[6], dd[4];
+ __m128i f0[4], f1[5], f2[5];
+ // The offset of the first row is always less than 1 pixel.
+ const int offset1_q4 = phase_scaler + 1 * step_q4;
+ const int offset2_q4 = phase_scaler + 2 * step_q4;
+ // offset_idxx indicates the pixel offset is even (0) or odd (1).
+ // It's used to choose the src offset and filter coefficient offset.
+ const int offset_idx1 = (offset1_q4 >> 4) & 1;
+ const int offset_idx2 = (offset2_q4 >> 4) & 1;
+ static const shuffle_filter_funcs shuffle_filter_funcs[2] = {+ shuffle_filter_ssse3, shuffle_filter_odd_ssse3
+ };
+ static const convolve8_funcs convolve8_funcs[2] = {+ convolve8_8_even_offset_ssse3, convolve8_8_odd_offset_ssse3
+ };
+
+ assert(w && h);
+
+ shuffle_filter_ssse3(coef[(phase_scaler + 0 * step_q4) & SUBPEL_MASK], f0);
+ shuffle_filter_funcs[offset_idx1](coef[offset1_q4 & SUBPEL_MASK], f1);
+ shuffle_filter_funcs[offset_idx2](coef[offset2_q4 & SUBPEL_MASK], f2);
+
+ // Sub 64 to avoid overflow.
+ // Coef 128 would be treated as -128 in PMADDUBSW. Sub 64 here.
+ // Coef 128 is in either fx[1] or fx[2] depending on the phase idx.
+ // When filter phase idx is 1, the two biggest coefficients are shuffled
+ // together, and the sum of them are always no less than 128. Sub 64 here.
+ // After the subtraction, when the sum of all positive coefficients are no
+ // larger than 128, and the sum of all negative coefficients are no
+ // less than -128, there will be no overflow in the convolve8 functions.
+ f0[1] = _mm_sub_epi8(f0[1], _mm_set1_epi8(64));
+ f1[1 + offset_idx1] = _mm_sub_epi8(f1[1 + offset_idx1], _mm_set1_epi8(64));
+ f2[1 + offset_idx2] = _mm_sub_epi8(f2[1 + offset_idx2], _mm_set1_epi8(64));
+
+ src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 - 1;
+
+ // horizontal 6x8
+ do {+ load_8bit_8x8(src, src_stride, s);
+ // 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71
+ // 02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73
+ // 04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75
+ // 06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77
+ transpose_16bit_4x8(s, s);
+ x = width_hor;
+
+ do {+ src += 8;
+ load_8bit_8x8(src, src_stride, &s[4]);
+ // 08 09 18 19 28 29 38 39 48 49 58 59 68 69 78 79
+ // 0A 0B 1A 1B 2A 2B 3A 3B 4A 4B 5A 5B 6A 6B 7A 7B
+ // OC 0D 1C 1D 2C 2D 3C 3D 4C 4D 5C 5D 6C 6D 7C 7D
+ // 0E 0F 1E 1F 2E 2F 3E 3F 4E 4F 5E 5F 6E 6F 7E 7F
+ transpose_16bit_4x8(&s[4], &s[4]);
+
+ // 00 10 20 30 40 50 60 70
+ // 01 11 21 31 41 51 61 71
+ // 02 12 22 32 42 52 62 72
+ // 03 13 23 33 43 53 63 73
+ // 04 14 24 34 44 54 64 74
+ // 05 15 25 35 45 55 65 75
+ d[0] = convolve8_8_even_offset_ssse3(&s[0], f0);
+ d[1] = convolve8_funcs[offset_idx1](&s[offset1_q4 >> 5], f1);
+ d[2] = convolve8_funcs[offset_idx2](&s[offset2_q4 >> 5], f2);
+ d[3] = convolve8_8_even_offset_ssse3(&s[2], f0);
+ d[4] = convolve8_funcs[offset_idx1](&s[2 + (offset1_q4 >> 5)], f1);
+ d[5] = convolve8_funcs[offset_idx2](&s[2 + (offset2_q4 >> 5)], f2);
+
+ // 00 10 20 30 40 50 60 70 02 12 22 32 42 52 62 72
+ // 01 11 21 31 41 51 61 71 03 13 23 33 43 53 63 73
+ // 04 14 24 34 44 54 64 74 xx xx xx xx xx xx xx xx
+ // 05 15 25 35 45 55 65 75 xx xx xx xx xx xx xx xx
+ dd[0] = _mm_packus_epi16(d[0], d[2]);
+ dd[1] = _mm_packus_epi16(d[1], d[3]);
+ dd[2] = _mm_packus_epi16(d[4], d[4]);
+ dd[3] = _mm_packus_epi16(d[5], d[5]);
+
+ // 00 10 01 11 20 30 21 31 40 50 41 51 60 70 61 71
+ // 02 12 03 13 22 32 23 33 42 52 43 53 62 72 63 73
+ // 04 14 05 15 24 34 25 35 44 54 45 55 64 74 65 75
+ d[0] = _mm_unpacklo_epi16(dd[0], dd[1]);
+ d[1] = _mm_unpackhi_epi16(dd[0], dd[1]);
+ d[2] = _mm_unpacklo_epi16(dd[2], dd[3]);
+
+ // 00 10 01 11 02 12 03 13 20 30 21 31 22 32 23 33
+ // 40 50 41 51 42 52 43 53 60 70 61 71 62 72 63 73
+ // 04 14 05 15 xx xx xx xx 24 34 25 35 xx xx xx xx
+ // 44 54 45 55 xx xx xx xx 64 74 65 75 xx xx xx xx
+ dd[0] = _mm_unpacklo_epi32(d[0], d[1]);
+ dd[1] = _mm_unpackhi_epi32(d[0], d[1]);
+ dd[2] = _mm_unpacklo_epi32(d[2], d[2]);
+ dd[3] = _mm_unpackhi_epi32(d[2], d[2]);
+
+ // 00 10 01 11 02 12 03 13 04 14 05 15 xx xx xx xx
+ // 20 30 21 31 22 32 23 33 24 34 25 35 xx xx xx xx
+ // 40 50 41 51 42 52 43 53 44 54 45 55 xx xx xx xx
+ // 60 70 61 71 62 72 63 73 64 74 65 75 xx xx xx xx
+ d[0] = _mm_unpacklo_epi64(dd[0], dd[2]);
+ d[1] = _mm_unpackhi_epi64(dd[0], dd[2]);
+ d[2] = _mm_unpacklo_epi64(dd[1], dd[3]);
+ d[3] = _mm_unpackhi_epi64(dd[1], dd[3]);
+
+ // store 4 extra pixels
+ storeu_8bit_16x4(d, t, stride_hor);
+
+ s[0] = s[4];
+ s[1] = s[5];
+ s[2] = s[6];
+ s[3] = s[7];
+
+ t += 12;
+ x -= 6;
+ } while (x);
+ src += 8 * src_stride - 4 * width_hor / 3;
+ t += 3 * stride_hor + 4;
+ y -= 8;
+ } while (y);
+
+ // vertical 8x6
+ x = width_ver;
+ t = temp_buffer;
+ do {+ // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+ // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+ // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+ // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+ loadu_8bit_16x4(t, stride_hor, s);
+ y = height_ver;
+
+ do {+ // 80 90 81 91 82 92 83 93 84 94 85 95 86 96 87 97
+ // A0 B0 A1 B1 A2 B2 A3 B3 A4 B4 A5 B5 A6 B6 A7 B7
+ // C0 D0 C1 D1 C2 D2 C3 D3 C4 D4 C5 D5 C6 D6 C7 D7
+ // E0 F0 E1 F1 E2 F2 E3 F3 E4 F4 E5 F5 E6 F6 E7 F7
+ t += 4 * stride_hor;
+ loadu_8bit_16x4(t, stride_hor, &s[4]);
+
+ d[0] = convolve8_8_even_offset_ssse3(&s[0], f0);
+ d[1] = convolve8_funcs[offset_idx1](&s[offset1_q4 >> 5], f1);
+ d[2] = convolve8_funcs[offset_idx2](&s[offset2_q4 >> 5], f2);
+ d[3] = convolve8_8_even_offset_ssse3(&s[2], f0);
+ d[4] = convolve8_funcs[offset_idx1](&s[2 + (offset1_q4 >> 5)], f1);
+ d[5] = convolve8_funcs[offset_idx2](&s[2 + (offset2_q4 >> 5)], f2);
+
+ // 00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17
+ // 20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37
+ // 40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57
+ d[0] = _mm_packus_epi16(d[0], d[1]);
+ d[2] = _mm_packus_epi16(d[2], d[3]);
+ d[4] = _mm_packus_epi16(d[4], d[5]);
+
+ _mm_storel_epi64((__m128i *)(dst + 0 * dst_stride), d[0]);
+ _mm_storeh_epi64((__m128i *)(dst + 1 * dst_stride), d[0]);
+ _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), d[2]);
+ _mm_storeh_epi64((__m128i *)(dst + 3 * dst_stride), d[2]);
+ _mm_storel_epi64((__m128i *)(dst + 4 * dst_stride), d[4]);
+ _mm_storeh_epi64((__m128i *)(dst + 5 * dst_stride), d[4]);
+
+ s[0] = s[4];
+ s[1] = s[5];
+ s[2] = s[6];
+ s[3] = s[7];
+
+ dst += 6 * dst_stride;
+ y -= 6;
+ } while (y);
+ t -= stride_hor * 2 * height_ver / 3;
+ t += 16;
+ dst -= height_ver * dst_stride;
+ dst += 8;
+ x -= 8;
+ } while (x);
+}
+
static INLINE __m128i scale_1_to_2_phase_0_kernel(const __m128i *const s,
const __m128i *const f) {__m128i ss[4], temp;
@@ -651,6 +847,36 @@
} else {scaled = 0;
}
+ }
+ } else if (4 * dst_w == 3 * src_w && 4 * dst_h == 3 * src_h) {+ // 4 to 3
+ const int buffer_stride_hor = (dst_w + 5) - ((dst_w + 5) % 6) + 2;
+ const int buffer_stride_ver = (dst_w + 7) & ~7;
+ const int buffer_height = (4 * dst_h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
+ // When the vertical filter reads more pixels than the horizontal filter
+ // generated in each row, we need extra padding to avoid heap read overflow.
+ // For example, the horizontal filter generates 18 pixels but the vertical
+ // filter reads 24 pixels in a row. The difference is multiplied by 2 since
+ // two rows are interlaced together in the optimization.
+ const int extra_padding = (buffer_stride_ver > buffer_stride_hor)
+ ? 2 * (buffer_stride_ver - buffer_stride_hor)
+ : 0;
+ const int buffer_size = buffer_stride_hor * buffer_height + extra_padding;
+ uint8_t *const temp_buffer = (uint8_t *)malloc(buffer_size);
+ if (temp_buffer) {+ scaled = 1;
+ scale_plane_4_to_3_general(
+ 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_3_general(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_3_general(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 if (dst_w == src_w * 2 && dst_h == src_h * 2 && phase_scaler == 0) {// 1 to 2
--- a/vpx_dsp/x86/convolve_ssse3.h
+++ b/vpx_dsp/x86/convolve_ssse3.h
@@ -11,6 +11,7 @@
#ifndef VPX_DSP_X86_CONVOLVE_SSSE3_H_
#define VPX_DSP_X86_CONVOLVE_SSSE3_H_
+#include <assert.h>
#include <tmmintrin.h> // SSSE3
#include "./vpx_config.h"
@@ -25,6 +26,20 @@
f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
}
+static INLINE void shuffle_filter_odd_ssse3(const int16_t *const filter,
+ __m128i *const f) {+ const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+ // pack and duplicate the filter values
+ // It utilizes the fact that the high byte of filter[3] is always 0 to clean
+ // half of f[0] and f[4].
+ assert(filter[3] >= 0 && filter[3] < 256);
+ f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0007u));
+ f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0402u));
+ f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0806u));
+ f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0c0au));
+ f[4] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x070eu));
+}
+
static INLINE __m128i convolve8_8_ssse3(const __m128i *const s,
const __m128i *const f) {// multiply 2 adjacent elements with the filter and add the result
@@ -39,6 +54,52 @@
__m128i temp = _mm_adds_epi16(x0, x3);
temp = _mm_adds_epi16(temp, min_x2x1);
temp = _mm_adds_epi16(temp, max_x2x1);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_adds_epi16(temp, k_64);
+ temp = _mm_srai_epi16(temp, 7);
+ return temp;
+}
+
+static INLINE __m128i convolve8_8_even_offset_ssse3(const __m128i *const s,
+ const __m128i *const f) {+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i k_64 = _mm_set1_epi16(1 << 6);
+ const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
+ const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
+ const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
+ const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
+ // compensate the subtracted 64 in f[1]. x4 is always non negative.
+ const __m128i x4 = _mm_maddubs_epi16(s[1], _mm_set1_epi8(64));
+ // add and saturate the results together
+ __m128i temp = _mm_adds_epi16(x0, x3);
+ temp = _mm_adds_epi16(temp, x1);
+ temp = _mm_adds_epi16(temp, x2);
+ temp = _mm_adds_epi16(temp, x4);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_adds_epi16(temp, k_64);
+ temp = _mm_srai_epi16(temp, 7);
+ return temp;
+}
+
+static INLINE __m128i convolve8_8_odd_offset_ssse3(const __m128i *const s,
+ const __m128i *const f) {+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i k_64 = _mm_set1_epi16(1 << 6);
+ const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
+ const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
+ const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
+ const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
+ const __m128i x4 = _mm_maddubs_epi16(s[4], f[4]);
+ // compensate the subtracted 64 in f[2]. x5 is always non negative.
+ const __m128i x5 = _mm_maddubs_epi16(s[2], _mm_set1_epi8(64));
+ __m128i temp;
+
+ // add and saturate the results together
+ temp = _mm_adds_epi16(x0, x1);
+ temp = _mm_adds_epi16(temp, x2);
+ temp = _mm_adds_epi16(temp, x3);
+ temp = _mm_adds_epi16(temp, x4);
+ temp = _mm_adds_epi16(temp, x5);
// round and shift by 7 bit each 16 bit
temp = _mm_adds_epi16(temp, k_64);
temp = _mm_srai_epi16(temp, 7);
--- a/vpx_dsp/x86/mem_sse2.h
+++ b/vpx_dsp/x86/mem_sse2.h
@@ -113,4 +113,12 @@
_mm_storel_epi64((__m128i *)(d + 7 * stride), s[7]);
}
+static INLINE void storeu_8bit_16x4(const __m128i *const s, uint8_t *const d,
+ const ptrdiff_t stride) {+ _mm_storeu_si128((__m128i *)(d + 0 * stride), s[0]);
+ _mm_storeu_si128((__m128i *)(d + 1 * stride), s[1]);
+ _mm_storeu_si128((__m128i *)(d + 2 * stride), s[2]);
+ _mm_storeu_si128((__m128i *)(d + 3 * stride), s[3]);
+}
+
#endif // VPX_DSP_X86_MEM_SSE2_H_