ref: 2030c49cf87681712ba200d2d1f06d09cc58adb1
parent: 4ef08dcec82085b69e528812c5cdd334410b7bd4
author: Scott LaVarnway <slavarnway@google.com>
date: Tue Aug 18 11:13:02 EDT 2015
VPX ssse3 scaled convolve Change-Id: I71d5994e21813554a927d35ebcc26bf7a68984fd
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -399,7 +399,7 @@
specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa/;
add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_scaled_2d/;
+specialize qw/vpx_scaled_2d ssse3/;
add_proto qw/void vpx_scaled_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
specialize qw/vpx_scaled_horiz/;
--- a/vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
+++ b/vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
@@ -15,7 +15,9 @@
#include <tmmintrin.h>
#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_filter.h"
#include "vpx_dsp/x86/convolve.h"
+#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/emmintrin_compat.h"
@@ -587,7 +589,576 @@
FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
ssse3);
-// void vpx_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, \
+ out0, out1, out2, out3, out4, out5, out6, out7) { \+ const __m128i tr0_0 = _mm_unpacklo_epi8(in0, in1); \
+ const __m128i tr0_1 = _mm_unpacklo_epi8(in2, in3); \
+ const __m128i tr0_2 = _mm_unpacklo_epi8(in4, in5); \
+ const __m128i tr0_3 = _mm_unpacklo_epi8(in6, in7); \
+ \
+ const __m128i tr1_0 = _mm_unpacklo_epi16(tr0_0, tr0_1); \
+ const __m128i tr1_1 = _mm_unpackhi_epi16(tr0_0, tr0_1); \
+ const __m128i tr1_2 = _mm_unpacklo_epi16(tr0_2, tr0_3); \
+ const __m128i tr1_3 = _mm_unpackhi_epi16(tr0_2, tr0_3); \
+ \
+ const __m128i tr2_0 = _mm_unpacklo_epi32(tr1_0, tr1_2); \
+ const __m128i tr2_1 = _mm_unpackhi_epi32(tr1_0, tr1_2); \
+ const __m128i tr2_2 = _mm_unpacklo_epi32(tr1_1, tr1_3); \
+ const __m128i tr2_3 = _mm_unpackhi_epi32(tr1_1, tr1_3); \
+ \
+ out0 = _mm_unpacklo_epi64(tr2_0, tr2_0); \
+ out1 = _mm_unpackhi_epi64(tr2_0, tr2_0); \
+ out2 = _mm_unpacklo_epi64(tr2_1, tr2_1); \
+ out3 = _mm_unpackhi_epi64(tr2_1, tr2_1); \
+ out4 = _mm_unpacklo_epi64(tr2_2, tr2_2); \
+ out5 = _mm_unpackhi_epi64(tr2_2, tr2_2); \
+ out6 = _mm_unpacklo_epi64(tr2_3, tr2_3); \
+ out7 = _mm_unpackhi_epi64(tr2_3, tr2_3); \
+}
+
+static void filter_horiz_w8_ssse3(const uint8_t *src_x, int src_pitch,
+ uint8_t *dst, const int16_t *x_filter) {+ const __m128i k_256 = _mm_set1_epi16(1 << 8);
+ const __m128i f_values = _mm_load_si128((const __m128i *)x_filter);
+ // pack and duplicate the filter values
+ const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+ const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+ const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+ const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+ const __m128i A = _mm_loadl_epi64((const __m128i *)src_x);
+ const __m128i B = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch));
+ const __m128i C = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 2));
+ const __m128i D = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 3));
+ const __m128i E = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 4));
+ const __m128i F = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 5));
+ const __m128i G = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 6));
+ const __m128i H = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 7));
+ // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17
+ const __m128i tr0_0 = _mm_unpacklo_epi16(A, B);
+ // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37
+ const __m128i tr0_1 = _mm_unpacklo_epi16(C, D);
+ // 40 41 50 51 42 43 52 53 44 45 54 55 46 47 56 57
+ const __m128i tr0_2 = _mm_unpacklo_epi16(E, F);
+ // 60 61 70 71 62 63 72 73 64 65 74 75 66 67 76 77
+ const __m128i tr0_3 = _mm_unpacklo_epi16(G, H);
+ // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33
+ const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+ // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37
+ const __m128i tr1_1 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+ // 40 41 50 51 60 61 70 71 42 43 52 53 62 63 72 73
+ const __m128i tr1_2 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+ // 44 45 54 55 64 65 74 75 46 47 56 57 66 67 76 77
+ const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+ // 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71
+ const __m128i s1s0 = _mm_unpacklo_epi64(tr1_0, tr1_2);
+ const __m128i s3s2 = _mm_unpackhi_epi64(tr1_0, tr1_2);
+ const __m128i s5s4 = _mm_unpacklo_epi64(tr1_1, tr1_3);
+ const __m128i s7s6 = _mm_unpackhi_epi64(tr1_1, tr1_3);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+ const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+ const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+ const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+ // add and saturate the results together
+ const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+ const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+ __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_mulhrs_epi16(temp, k_256);
+ // shrink to 8 bit each 16 bits
+ temp = _mm_packus_epi16(temp, temp);
+ // save only 8 bytes convolve result
+ _mm_storel_epi64((__m128i*)dst, temp);
+}
+
+static void transpose8x8_to_dst(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride) {+ __m128i A, B, C, D, E, F, G, H;
+
+ A = _mm_loadl_epi64((const __m128i *)src);
+ B = _mm_loadl_epi64((const __m128i *)(src + src_stride));
+ C = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+ D = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3));
+ E = _mm_loadl_epi64((const __m128i *)(src + src_stride * 4));
+ F = _mm_loadl_epi64((const __m128i *)(src + src_stride * 5));
+ G = _mm_loadl_epi64((const __m128i *)(src + src_stride * 6));
+ H = _mm_loadl_epi64((const __m128i *)(src + src_stride * 7));
+
+ TRANSPOSE_8X8(A, B, C, D, E, F, G, H,
+ A, B, C, D, E, F, G, H);
+
+ _mm_storel_epi64((__m128i*)dst, A);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 1), B);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 2), C);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 3), D);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 4), E);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 5), F);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 6), G);
+ _mm_storel_epi64((__m128i*)(dst + dst_stride * 7), H);
+}
+
+static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *x_filters,
+ int x0_q4, int x_step_q4, int w, int h) {+ DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
+ int x, y, z;
+ src -= SUBPEL_TAPS / 2 - 1;
+
+ // This function processes 8x8 areas. The intermediate height is not always
+ // a multiple of 8, so force it to be a multiple of 8 here.
+ y = h + (8 - (h & 0x7));
+
+ do {+ int x_q4 = x0_q4;
+ for (x = 0; x < w; x += 8) {+ // process 8 src_x steps
+ for (z = 0; z < 8; ++z) {+ const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+ const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+ if (x_q4 & SUBPEL_MASK) {+ filter_horiz_w8_ssse3(src_x, src_stride, temp + (z * 8), x_filter);
+ } else {+ int i;
+ for (i = 0; i < 8; ++i) {+ temp[z * 8 + i] = src_x[i * src_stride + 3];
+ }
+ }
+ x_q4 += x_step_q4;
+ }
+
+ // transpose the 8x8 filters values back to dst
+ transpose8x8_to_dst(temp, 8, dst + x, dst_stride);
+ }
+
+ src += src_stride * 8;
+ dst += dst_stride * 8;
+ } while (y -= 8);
+}
+
+static void filter_horiz_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+ uint8_t *dst, const int16_t *filter) {+ const __m64 k_256 = _mm_set1_pi16(1 << 8);
+ const __m64 f_values_lo = *((const __m64 *)filter);
+ const __m64 f_values_hi = *((const __m64 *)(filter + 4));
+ // pack and duplicate the filter values
+ const __m64 f1f0 = _mm_shuffle_pi8(f_values_lo, _mm_set1_pi16(0x0200u));
+ const __m64 f3f2 = _mm_shuffle_pi8(f_values_lo, _mm_set1_pi16(0x0604u));
+ const __m64 f5f4 = _mm_shuffle_pi8(f_values_hi, _mm_set1_pi16(0x0a08u));
+ const __m64 f7f6 = _mm_shuffle_pi8(f_values_hi, _mm_set1_pi16(0x0e0cu));
+ const __m64 A = *((const __m64 *)src_ptr);
+ const __m64 B = *((const __m64 *)(src_ptr + src_pitch));
+ const __m64 C = *((const __m64 *)(src_ptr + src_pitch * 2));
+ const __m64 D = *((const __m64 *)(src_ptr + src_pitch * 3));
+ // TRANSPOSE...
+ // 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
+ //
+ // TO
+ //
+ // 00 10 20 30
+ // 01 11 21 31
+ // 02 12 22 32
+ // 03 13 23 33
+ // 04 14 24 34
+ // 05 15 25 35
+ // 06 16 26 36
+ // 07 17 27 37
+ //
+ // 00 01 10 11 02 03 12 13
+ const __m64 tr0_0 = _mm_unpacklo_pi16(A, B);
+ // 20 21 30 31 22 23 32 33
+ const __m64 tr0_1 = _mm_unpacklo_pi16(C, D);
+ // 04 05 14 15 06 07 16 17
+ const __m64 tr0_2 = _mm_unpackhi_pi16(A, B);
+ // 24 25 34 35 26 27 36 37
+ const __m64 tr0_3 = _mm_unpackhi_pi16(C, D);
+ // 00 01 10 11 20 21 30 31
+ const __m64 s1s0 = _mm_unpacklo_pi32(tr0_0, tr0_1);
+ // 02 03 12 13 22 23 32 33
+ const __m64 s3s2 = _mm_unpackhi_pi32(tr0_0, tr0_1);
+ // 04 05 14 15 24 25 34 35
+ const __m64 s5s4 = _mm_unpacklo_pi32(tr0_2, tr0_3);
+ // 06 07 16 17 26 27 36 37
+ const __m64 s7s6 = _mm_unpackhi_pi32(tr0_2, tr0_3);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m64 x0 = _mm_maddubs_pi16(s1s0, f1f0);
+ const __m64 x1 = _mm_maddubs_pi16(s3s2, f3f2);
+ const __m64 x2 = _mm_maddubs_pi16(s5s4, f5f4);
+ const __m64 x3 = _mm_maddubs_pi16(s7s6, f7f6);
+ // add and saturate the results together
+ const __m64 min_x2x1 = _mm_min_pi16(x2, x1);
+ const __m64 max_x2x1 = _mm_max_pi16(x2, x1);
+ __m64 temp = _mm_adds_pi16(x0, x3);
+ temp = _mm_adds_pi16(temp, min_x2x1);
+ temp = _mm_adds_pi16(temp, max_x2x1);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_mulhrs_pi16(temp, k_256);
+ // shrink to 8 bit each 16 bits
+ temp = _mm_packs_pu16(temp, temp);
+ // save only 4 bytes
+ *(int *)dst = _mm_cvtsi64_si32(temp);
+}
+
+static void transpose4x4_to_dst(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride) {+ __m64 A = _mm_cvtsi32_si64(*(const int *)src);
+ __m64 B = _mm_cvtsi32_si64(*(const int *)(src + src_stride));
+ __m64 C = _mm_cvtsi32_si64(*(const int *)(src + src_stride * 2));
+ __m64 D = _mm_cvtsi32_si64(*(const int *)(src + src_stride * 3));
+ // 00 10 01 11 02 12 03 13
+ const __m64 tr0_0 = _mm_unpacklo_pi8(A, B);
+ // 20 30 21 31 22 32 23 33
+ const __m64 tr0_1 = _mm_unpacklo_pi8(C, D);
+ // 00 10 20 30 01 11 21 31
+ A = _mm_unpacklo_pi16(tr0_0, tr0_1);
+ // 02 12 22 32 03 13 23 33
+ C = _mm_unpackhi_pi16(tr0_0, tr0_1);
+
+ B = _mm_unpackhi_pi32(A, A);
+ D = _mm_unpackhi_pi32(C, C);
+
+ *(int *)(dst) = _mm_cvtsi64_si32(A);
+ *(int *)(dst + dst_stride) = _mm_cvtsi64_si32(B);
+ *(int *)(dst + dst_stride * 2) = _mm_cvtsi64_si32(C);
+ *(int *)(dst + dst_stride * 3) = _mm_cvtsi64_si32(D);
+}
+
+static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *x_filters,
+ int x0_q4, int x_step_q4, int w, int h) {+ DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
+ int x, y, z;
+ src -= SUBPEL_TAPS / 2 - 1;
+
+ for (y = 0; y < h; y += 4) {+ int x_q4 = x0_q4;
+ for (x = 0; x < w; x += 4) {+ // process 4 src_x steps
+ for (z = 0; z < 4; ++z) {+ const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+ const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+ if (x_q4 & SUBPEL_MASK) {+ filter_horiz_w4_ssse3(src_x, src_stride, temp + (z * 4), x_filter);
+ } else {+ int i;
+ for (i = 0; i < 4; ++i) {+ temp[z * 4 + i] = src_x[i * src_stride + 3];
+ }
+ }
+ x_q4 += x_step_q4;
+ }
+
+ // transpose the 4x4 filters values back to dst
+ transpose4x4_to_dst(temp, 4, dst + x, dst_stride);
+ }
+
+ src += src_stride * 4;
+ dst += dst_stride * 4;
+ }
+}
+
+static void filter_vert_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+ uint8_t *dst, const int16_t *filter) {+ const __m64 k_256 = _mm_set1_pi16(1 << 8);
+ const __m64 f_values_lo = *((const __m64 *)filter);
+ const __m64 f_values_hi = *((const __m64 *)(filter + 4));
+ // pack and duplicate the filter values
+ const __m64 f1f0 = _mm_shuffle_pi8(f_values_lo, _mm_set1_pi16(0x0200u));
+ const __m64 f3f2 = _mm_shuffle_pi8(f_values_lo, _mm_set1_pi16(0x0604u));
+ const __m64 f5f4 = _mm_shuffle_pi8(f_values_hi, _mm_set1_pi16(0x0a08u));
+ const __m64 f7f6 = _mm_shuffle_pi8(f_values_hi, _mm_set1_pi16(0x0e0cu));
+ const __m64 A = _mm_cvtsi32_si64(*(const int *)src_ptr);
+ const __m64 B = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch));
+ const __m64 C = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch * 2));
+ const __m64 D = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch * 3));
+ const __m64 E = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch * 4));
+ const __m64 F = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch * 5));
+ const __m64 G = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch * 6));
+ const __m64 H = _mm_cvtsi32_si64(*(const int *)(src_ptr + src_pitch * 7));
+ const __m64 s1s0 = _mm_unpacklo_pi8(A, B);
+ // 02 03 12 13 22 23 32 33
+ const __m64 s3s2 = _mm_unpacklo_pi8(C, D);
+ // 04 05 14 15 24 25 34 35
+ const __m64 s5s4 = _mm_unpacklo_pi8(E, F);
+ // 06 07 16 17 26 27 36 37
+ const __m64 s7s6 = _mm_unpacklo_pi8(G, H);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m64 x0 = _mm_maddubs_pi16(s1s0, f1f0);
+ const __m64 x1 = _mm_maddubs_pi16(s3s2, f3f2);
+ const __m64 x2 = _mm_maddubs_pi16(s5s4, f5f4);
+ const __m64 x3 = _mm_maddubs_pi16(s7s6, f7f6);
+ // add and saturate the results together
+ const __m64 min_x2x1 = _mm_min_pi16(x2, x1);
+ const __m64 max_x2x1 = _mm_max_pi16(x2, x1);
+ __m64 temp = _mm_adds_pi16(x0, x3);
+ temp = _mm_adds_pi16(temp, min_x2x1);
+ temp = _mm_adds_pi16(temp, max_x2x1);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_mulhrs_pi16(temp, k_256);
+ // shrink to 8 bit each 16 bits
+ temp = _mm_packs_pu16(temp, temp);
+ // save only 4 bytes
+ *(int *)dst = _mm_cvtsi64_si32(temp);
+}
+
+static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *y_filters,
+ int y0_q4, int y_step_q4, int w, int h) {+ int y;
+ int y_q4 = y0_q4;
+
+ src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+ for (y = 0; y < h; ++y) {+ const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+ const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+
+ if (y_q4 & SUBPEL_MASK) {+ filter_vert_w4_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+ } else {+ memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+ }
+
+ y_q4 += y_step_q4;
+ }
+}
+
+static void filter_vert_w8_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+ uint8_t *dst, const int16_t *filter) {+ const __m128i k_256 = _mm_set1_epi16(1 << 8);
+ const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+ // pack and duplicate the filter values
+ const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+ const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+ const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+ const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+ const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr);
+ const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+ const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+ const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+ const __m128i E = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+ const __m128i F = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+ const __m128i G = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+ const __m128i H = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+ const __m128i s1s0 = _mm_unpacklo_epi8(A, B);
+ const __m128i s3s2 = _mm_unpacklo_epi8(C, D);
+ const __m128i s5s4 = _mm_unpacklo_epi8(E, F);
+ const __m128i s7s6 = _mm_unpacklo_epi8(G, H);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+ const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+ const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+ const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+ // add and saturate the results together
+ const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+ const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+ __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_mulhrs_epi16(temp, k_256);
+ // shrink to 8 bit each 16 bits
+ temp = _mm_packus_epi16(temp, temp);
+ // save only 8 bytes convolve result
+ _mm_storel_epi64((__m128i*)dst, temp);
+}
+
+static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *y_filters,
+ int y0_q4, int y_step_q4, int w, int h) {+ int y;
+ int y_q4 = y0_q4;
+
+ src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+ for (y = 0; y < h; ++y) {+ const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+ const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+ if (y_q4 & SUBPEL_MASK) {+ filter_vert_w8_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+ } else {+ memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+ }
+ y_q4 += y_step_q4;
+ }
+}
+
+static void filter_vert_w16_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+ uint8_t *dst, const int16_t *filter, int w) {+ const __m128i k_256 = _mm_set1_epi16(1 << 8);
+ const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+ // pack and duplicate the filter values
+ const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+ const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+ const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+ const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+ int i;
+
+ for (i = 0; i < w; i += 16) {+ const __m128i A = _mm_loadu_si128((const __m128i *)src_ptr);
+ const __m128i B = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
+ const __m128i C =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+ const __m128i D =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+ const __m128i E =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+ const __m128i F =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+ const __m128i G =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+ const __m128i H =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+ // merge the result together
+ const __m128i s1s0_lo = _mm_unpacklo_epi8(A, B);
+ const __m128i s7s6_lo = _mm_unpacklo_epi8(G, H);
+ const __m128i s1s0_hi = _mm_unpackhi_epi8(A, B);
+ const __m128i s7s6_hi = _mm_unpackhi_epi8(G, H);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i x0_lo = _mm_maddubs_epi16(s1s0_lo, f1f0);
+ const __m128i x3_lo = _mm_maddubs_epi16(s7s6_lo, f7f6);
+ const __m128i x0_hi = _mm_maddubs_epi16(s1s0_hi, f1f0);
+ const __m128i x3_hi = _mm_maddubs_epi16(s7s6_hi, f7f6);
+ // add and saturate the results together
+ const __m128i x3x0_lo = _mm_adds_epi16(x0_lo, x3_lo);
+ const __m128i x3x0_hi = _mm_adds_epi16(x0_hi, x3_hi);
+ // merge the result together
+ const __m128i s3s2_lo = _mm_unpacklo_epi8(C, D);
+ const __m128i s3s2_hi = _mm_unpackhi_epi8(C, D);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i x1_lo = _mm_maddubs_epi16(s3s2_lo, f3f2);
+ const __m128i x1_hi = _mm_maddubs_epi16(s3s2_hi, f3f2);
+ // merge the result together
+ const __m128i s5s4_lo = _mm_unpacklo_epi8(E, F);
+ const __m128i s5s4_hi = _mm_unpackhi_epi8(E, F);
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i x2_lo = _mm_maddubs_epi16(s5s4_lo, f5f4);
+ const __m128i x2_hi = _mm_maddubs_epi16(s5s4_hi, f5f4);
+ // add and saturate the results together
+ __m128i temp_lo = _mm_adds_epi16(x3x0_lo, _mm_min_epi16(x1_lo, x2_lo));
+ __m128i temp_hi = _mm_adds_epi16(x3x0_hi, _mm_min_epi16(x1_hi, x2_hi));
+
+ // add and saturate the results together
+ temp_lo = _mm_adds_epi16(temp_lo, _mm_max_epi16(x1_lo, x2_lo));
+ temp_hi = _mm_adds_epi16(temp_hi, _mm_max_epi16(x1_hi, x2_hi));
+ // round and shift by 7 bit each 16 bit
+ temp_lo = _mm_mulhrs_epi16(temp_lo, k_256);
+ temp_hi = _mm_mulhrs_epi16(temp_hi, k_256);
+ // shrink to 8 bit each 16 bits, the first lane contain the first
+ // convolve result and the second lane contain the second convolve
+ // result
+ temp_hi = _mm_packus_epi16(temp_lo, temp_hi);
+ src_ptr += 16;
+ // save 16 bytes convolve result
+ _mm_store_si128((__m128i*)&dst[i], temp_hi);
+ }
+}
+
+static void scaledconvolve_vert_w16(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *y_filters,
+ int y0_q4, int y_step_q4, int w, int h) {+ int y;
+ int y_q4 = y0_q4;
+
+ src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+ for (y = 0; y < h; ++y) {+ const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+ const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+ if (y_q4 & SUBPEL_MASK) {+ filter_vert_w16_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter,
+ w);
+ } else {+ memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+ }
+ y_q4 += y_step_q4;
+ }
+}
+
+static void scaledconvolve2d(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *const x_filters,
+ int x0_q4, int x_step_q4,
+ const InterpKernel *const y_filters,
+ int y0_q4, int y_step_q4,
+ int w, int h) {+ // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+ // 2d filtering proceeds in 2 steps:
+ // (1) Interpolate horizontally into an intermediate buffer, temp.
+ // (2) Interpolate temp vertically to derive the sub-pixel result.
+ // Deriving the maximum number of rows in the temp buffer (135):
+ // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+ // --Largest block size is 64x64 pixels.
+ // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+ // original frame (in 1/16th pixel units).
+ // --Must round-up because block may be located at sub-pixel position.
+ // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+ // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+ // --Require an additional 8 rows for the horiz_w8 transpose tail.
+ DECLARE_ALIGNED(16, uint8_t, temp[(135 + 8) * 64]);
+ const int intermediate_height =
+ (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+ assert(w <= 64);
+ assert(h <= 64);
+ assert(y_step_q4 <= 32);
+ assert(x_step_q4 <= 32);
+
+ if (w >= 8) {+ scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+ src_stride, temp, 64, x_filters, x0_q4, x_step_q4,
+ w, intermediate_height);
+ } else {+ scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+ src_stride, temp, 64, x_filters, x0_q4, x_step_q4,
+ w, intermediate_height);
+ }
+
+ if (w >= 16) {+ scaledconvolve_vert_w16(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+ dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+ } else if (w == 8) {+ scaledconvolve_vert_w8(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+ dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+ } else {+ scaledconvolve_vert_w4(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+ dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+ }
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {+ // NOTE: This assumes that the filter table is 256-byte aligned.
+ // TODO(agrange) Modify to make independent of table alignment.
+ return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {+ return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h) {+ const InterpKernel *const filters_x = get_filter_base(filter_x);
+ const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+ const InterpKernel *const filters_y = get_filter_base(filter_y);
+ const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+ scaledconvolve2d(src, src_stride, dst, dst_stride,
+ filters_x, x0_q4, x_step_q4,
+ filters_y, y0_q4, y_step_q4, w, h);
+}
+
+// void vp9_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride,
// uint8_t *dst, ptrdiff_t dst_stride,
// const int16_t *filter_x, int x_step_q4,
// const int16_t *filter_y, int y_step_q4,