12tap filter for av1_highbd_convolve_2d_sr_ssse3()
SpeedTest: 12x compared to C-code
Change-Id: If11b3f47bf0ee556c8b15c7d6328ac11c419b6de
diff --git a/av1/common/convolve.c b/av1/common/convolve.c
index 054c33e..63dda39 100644
--- a/av1/common/convolve.c
+++ b/av1/common/convolve.c
@@ -1025,11 +1025,6 @@
const int subpel_y_qn, ConvolveParams *conv_params, int bd) {
const bool need_x = subpel_x_qn != 0;
const bool need_y = subpel_y_qn != 0;
- // Filters with taps > 8 are only for encoder side use.
- const int filter_x_taps_gt8 =
- (filter_params_x == NULL) ? 0 : ((filter_params_x->taps > 8) ? 1 : 0);
- const int filter_y_taps_gt8 =
- (filter_params_y == NULL) ? 0 : ((filter_params_y->taps > 8) ? 1 : 0);
if (!need_x && !need_y) {
aom_highbd_convolve_copy(src, src_stride, dst, dst_stride, w, h);
@@ -1041,15 +1036,9 @@
filter_params_y, subpel_y_qn, bd);
} else {
assert(need_x && need_y);
- if (filter_x_taps_gt8 || filter_y_taps_gt8) {
- av1_highbd_convolve_2d_sr_c(src, src_stride, dst, dst_stride, w, h,
- filter_params_x, filter_params_y, subpel_x_qn,
- subpel_y_qn, conv_params, bd);
- } else {
- av1_highbd_convolve_2d_sr(src, src_stride, dst, dst_stride, w, h,
- filter_params_x, filter_params_y, subpel_x_qn,
- subpel_y_qn, conv_params, bd);
- }
+ av1_highbd_convolve_2d_sr(src, src_stride, dst, dst_stride, w, h,
+ filter_params_x, filter_params_y, subpel_x_qn,
+ subpel_y_qn, conv_params, bd);
}
}
diff --git a/av1/common/x86/highbd_convolve_2d_avx2.c b/av1/common/x86/highbd_convolve_2d_avx2.c
index 12d1962..12046e4 100644
--- a/av1/common/x86/highbd_convolve_2d_avx2.c
+++ b/av1/common/x86/highbd_convolve_2d_avx2.c
@@ -20,6 +20,12 @@
#include "aom_dsp/aom_filter.h"
#include "av1/common/convolve.h"
+void av1_highbd_convolve_2d_sr_ssse3(
+ const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
+ int h, const InterpFilterParams *filter_params_x,
+ const InterpFilterParams *filter_params_y, const int subpel_x_qn,
+ const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride,
uint16_t *dst, int dst_stride, int w, int h,
const InterpFilterParams *filter_params_x,
@@ -27,6 +33,13 @@
const int subpel_x_qn,
const int subpel_y_qn,
ConvolveParams *conv_params, int bd) {
+ if (filter_params_x->taps == 12) {
+ av1_highbd_convolve_2d_sr_ssse3(src, src_stride, dst, dst_stride, w, h,
+ filter_params_x, filter_params_y,
+ subpel_x_qn, subpel_y_qn, conv_params, bd);
+ return;
+ }
+
DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
int im_h = h + filter_params_y->taps - 1;
int im_stride = 8;
diff --git a/av1/common/x86/highbd_convolve_2d_ssse3.c b/av1/common/x86/highbd_convolve_2d_ssse3.c
index 5318fca..148543f 100644
--- a/av1/common/x86/highbd_convolve_2d_ssse3.c
+++ b/av1/common/x86/highbd_convolve_2d_ssse3.c
@@ -18,6 +18,7 @@
#include "aom_dsp/aom_filter.h"
#include "aom_dsp/x86/convolve_sse2.h"
#include "av1/common/convolve.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
void av1_highbd_convolve_2d_sr_ssse3(
const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
@@ -35,7 +36,6 @@
// Check that, even with 12-bit input, the intermediate values will fit
// into an unsigned 16-bit intermediate array.
assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
- __m128i coeffs_x[4], coeffs_y[4], s[16];
const __m128i round_const_x = _mm_set1_epi32(
((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
@@ -54,163 +54,360 @@
_mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
const __m128i zero = _mm_setzero_si128();
- prepare_coeffs(filter_params_x, subpel_x_qn, coeffs_x);
- prepare_coeffs(filter_params_y, subpel_y_qn, coeffs_y);
+ if (filter_params_x->taps == 12) {
+ __m128i coeffs_x[6], coeffs_y[6], s[24];
+ prepare_coeffs_12tap(filter_params_x, subpel_x_qn, coeffs_x);
+ prepare_coeffs_12tap(filter_params_y, subpel_y_qn, coeffs_y);
- for (j = 0; j < w; j += 8) {
- /* Horizontal filter */
- {
- for (i = 0; i < im_h; i += 1) {
- const __m128i row00 =
- _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
- const __m128i row01 =
- _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
+ for (j = 0; j < w; j += 8) {
+ /* Horizontal filter */
+ {
+ for (i = 0; i < im_h; i += 1) {
+ const __m128i row00 =
+ _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+ const __m128i row01 =
+ _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
+ const __m128i row02 =
+ _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 16)]);
- // even pixels
- s[0] = _mm_alignr_epi8(row01, row00, 0);
- s[1] = _mm_alignr_epi8(row01, row00, 4);
- s[2] = _mm_alignr_epi8(row01, row00, 8);
- s[3] = _mm_alignr_epi8(row01, row00, 12);
+ // even pixels
+ s[0] = _mm_alignr_epi8(row01, row00, 0);
+ s[1] = _mm_alignr_epi8(row01, row00, 4);
+ s[2] = _mm_alignr_epi8(row01, row00, 8);
+ s[3] = _mm_alignr_epi8(row01, row00, 12);
+ s[4] = _mm_alignr_epi8(row02, row01, 0);
+ s[5] = _mm_alignr_epi8(row02, row01, 4);
- __m128i res_even = convolve(s, coeffs_x);
- res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
- round_shift_x);
+ __m128i res_even = convolve_12tap(s, coeffs_x);
+ res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
+ round_shift_x);
- // odd pixels
- s[0] = _mm_alignr_epi8(row01, row00, 2);
- s[1] = _mm_alignr_epi8(row01, row00, 6);
- s[2] = _mm_alignr_epi8(row01, row00, 10);
- s[3] = _mm_alignr_epi8(row01, row00, 14);
+ // odd pixels
+ s[0] = _mm_alignr_epi8(row01, row00, 2);
+ s[1] = _mm_alignr_epi8(row01, row00, 6);
+ s[2] = _mm_alignr_epi8(row01, row00, 10);
+ s[3] = _mm_alignr_epi8(row01, row00, 14);
+ s[4] = _mm_alignr_epi8(row02, row01, 2);
+ s[5] = _mm_alignr_epi8(row02, row01, 6);
- __m128i res_odd = convolve(s, coeffs_x);
- res_odd =
- _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x), round_shift_x);
+ __m128i res_odd = convolve_12tap(s, coeffs_x);
+ res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x),
+ round_shift_x);
- __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
- __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
- __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
+ __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
+ __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
+ __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
- _mm_store_si128((__m128i *)&im_block[i * im_stride], res);
+ _mm_store_si128((__m128i *)&im_block[i * im_stride], res);
+ }
+ }
+
+ /* Vertical filter */
+ {
+ __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride));
+ __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride));
+ __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride));
+ __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride));
+ __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride));
+ __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride));
+ __m128i s6 = _mm_loadu_si128((__m128i *)(im_block + 6 * im_stride));
+ __m128i s7 = _mm_loadu_si128((__m128i *)(im_block + 7 * im_stride));
+ __m128i s8 = _mm_loadu_si128((__m128i *)(im_block + 8 * im_stride));
+ __m128i s9 = _mm_loadu_si128((__m128i *)(im_block + 9 * im_stride));
+ __m128i s10 = _mm_loadu_si128((__m128i *)(im_block + 10 * im_stride));
+
+ s[0] = _mm_unpacklo_epi16(s0, s1);
+ s[1] = _mm_unpacklo_epi16(s2, s3);
+ s[2] = _mm_unpacklo_epi16(s4, s5);
+ s[3] = _mm_unpacklo_epi16(s6, s7);
+ s[4] = _mm_unpacklo_epi16(s8, s9);
+
+ s[6] = _mm_unpackhi_epi16(s0, s1);
+ s[7] = _mm_unpackhi_epi16(s2, s3);
+ s[8] = _mm_unpackhi_epi16(s4, s5);
+ s[9] = _mm_unpackhi_epi16(s6, s7);
+ s[10] = _mm_unpackhi_epi16(s8, s9);
+
+ s[12] = _mm_unpacklo_epi16(s1, s2);
+ s[13] = _mm_unpacklo_epi16(s3, s4);
+ s[14] = _mm_unpacklo_epi16(s5, s6);
+ s[15] = _mm_unpacklo_epi16(s7, s8);
+ s[16] = _mm_unpacklo_epi16(s9, s10);
+
+ s[18] = _mm_unpackhi_epi16(s1, s2);
+ s[19] = _mm_unpackhi_epi16(s3, s4);
+ s[20] = _mm_unpackhi_epi16(s5, s6);
+ s[21] = _mm_unpackhi_epi16(s7, s8);
+ s[22] = _mm_unpackhi_epi16(s9, s10);
+
+ for (i = 0; i < h; i += 2) {
+ const int16_t *data = &im_block[i * im_stride];
+
+ __m128i s11 = _mm_loadu_si128((__m128i *)(data + 11 * im_stride));
+ __m128i s12 = _mm_loadu_si128((__m128i *)(data + 12 * im_stride));
+
+ s[5] = _mm_unpacklo_epi16(s10, s11);
+ s[11] = _mm_unpackhi_epi16(s10, s11);
+
+ s[17] = _mm_unpacklo_epi16(s11, s12);
+ s[23] = _mm_unpackhi_epi16(s11, s12);
+
+ const __m128i res_a0 = convolve_12tap(s, coeffs_y);
+ __m128i res_a_round0 = _mm_sra_epi32(
+ _mm_add_epi32(res_a0, round_const_y), round_shift_y);
+ res_a_round0 = _mm_sra_epi32(
+ _mm_add_epi32(res_a_round0, round_const_bits), round_shift_bits);
+
+ const __m128i res_a1 = convolve_12tap(s + 12, coeffs_y);
+ __m128i res_a_round1 = _mm_sra_epi32(
+ _mm_add_epi32(res_a1, round_const_y), round_shift_y);
+ res_a_round1 = _mm_sra_epi32(
+ _mm_add_epi32(res_a_round1, round_const_bits), round_shift_bits);
+
+ if (w - j > 4) {
+ const __m128i res_b0 = convolve_12tap(s + 6, coeffs_y);
+ __m128i res_b_round0 = _mm_sra_epi32(
+ _mm_add_epi32(res_b0, round_const_y), round_shift_y);
+ res_b_round0 =
+ _mm_sra_epi32(_mm_add_epi32(res_b_round0, round_const_bits),
+ round_shift_bits);
+
+ const __m128i res_b1 = convolve_12tap(s + 18, coeffs_y);
+ __m128i res_b_round1 = _mm_sra_epi32(
+ _mm_add_epi32(res_b1, round_const_y), round_shift_y);
+ res_b_round1 =
+ _mm_sra_epi32(_mm_add_epi32(res_b_round1, round_const_bits),
+ round_shift_bits);
+
+ __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
+ res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
+ res_16bit0 = _mm_max_epi16(res_16bit0, zero);
+
+ __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
+ res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
+ res_16bit1 = _mm_max_epi16(res_16bit1, zero);
+
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ res_16bit1);
+ } else if (w == 4) {
+ res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+ res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+ res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+ res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+ res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+ res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ res_a_round1);
+ } else {
+ res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+ res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+ res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+ res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+ res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+ res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+ *((uint32_t *)(&dst[i * dst_stride + j])) =
+ _mm_cvtsi128_si32(res_a_round0);
+
+ *((uint32_t *)(&dst[i * dst_stride + j + dst_stride])) =
+ _mm_cvtsi128_si32(res_a_round1);
+ }
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+ s[3] = s[4];
+ s[4] = s[5];
+
+ s[6] = s[7];
+ s[7] = s[8];
+ s[8] = s[9];
+ s[9] = s[10];
+ s[10] = s[11];
+
+ s[12] = s[13];
+ s[13] = s[14];
+ s[14] = s[15];
+ s[15] = s[16];
+ s[16] = s[17];
+
+ s[18] = s[19];
+ s[19] = s[20];
+ s[20] = s[21];
+ s[21] = s[22];
+ s[22] = s[23];
+
+ s10 = s12;
+ }
}
}
- /* Vertical filter */
- {
- __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride));
- __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride));
- __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride));
- __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride));
- __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride));
- __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride));
- __m128i s6 = _mm_loadu_si128((__m128i *)(im_block + 6 * im_stride));
+ } else {
+ __m128i coeffs_x[4], coeffs_y[4], s[16];
+ prepare_coeffs(filter_params_x, subpel_x_qn, coeffs_x);
+ prepare_coeffs(filter_params_y, subpel_y_qn, coeffs_y);
- s[0] = _mm_unpacklo_epi16(s0, s1);
- s[1] = _mm_unpacklo_epi16(s2, s3);
- s[2] = _mm_unpacklo_epi16(s4, s5);
+ for (j = 0; j < w; j += 8) {
+ /* Horizontal filter */
+ {
+ for (i = 0; i < im_h; i += 1) {
+ const __m128i row00 =
+ _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+ const __m128i row01 =
+ _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
- s[4] = _mm_unpackhi_epi16(s0, s1);
- s[5] = _mm_unpackhi_epi16(s2, s3);
- s[6] = _mm_unpackhi_epi16(s4, s5);
+ // even pixels
+ s[0] = _mm_alignr_epi8(row01, row00, 0);
+ s[1] = _mm_alignr_epi8(row01, row00, 4);
+ s[2] = _mm_alignr_epi8(row01, row00, 8);
+ s[3] = _mm_alignr_epi8(row01, row00, 12);
- s[0 + 8] = _mm_unpacklo_epi16(s1, s2);
- s[1 + 8] = _mm_unpacklo_epi16(s3, s4);
- s[2 + 8] = _mm_unpacklo_epi16(s5, s6);
+ __m128i res_even = convolve(s, coeffs_x);
+ res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
+ round_shift_x);
- s[4 + 8] = _mm_unpackhi_epi16(s1, s2);
- s[5 + 8] = _mm_unpackhi_epi16(s3, s4);
- s[6 + 8] = _mm_unpackhi_epi16(s5, s6);
+ // odd pixels
+ s[0] = _mm_alignr_epi8(row01, row00, 2);
+ s[1] = _mm_alignr_epi8(row01, row00, 6);
+ s[2] = _mm_alignr_epi8(row01, row00, 10);
+ s[3] = _mm_alignr_epi8(row01, row00, 14);
- for (i = 0; i < h; i += 2) {
- const int16_t *data = &im_block[i * im_stride];
+ __m128i res_odd = convolve(s, coeffs_x);
+ res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x),
+ round_shift_x);
- __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * im_stride));
- __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * im_stride));
+ __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
+ __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
+ __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
- s[3] = _mm_unpacklo_epi16(s6, s7);
- s[7] = _mm_unpackhi_epi16(s6, s7);
-
- s[3 + 8] = _mm_unpacklo_epi16(s7, s8);
- s[7 + 8] = _mm_unpackhi_epi16(s7, s8);
-
- const __m128i res_a0 = convolve(s, coeffs_y);
- __m128i res_a_round0 =
- _mm_sra_epi32(_mm_add_epi32(res_a0, round_const_y), round_shift_y);
- res_a_round0 = _mm_sra_epi32(
- _mm_add_epi32(res_a_round0, round_const_bits), round_shift_bits);
-
- const __m128i res_a1 = convolve(s + 8, coeffs_y);
- __m128i res_a_round1 =
- _mm_sra_epi32(_mm_add_epi32(res_a1, round_const_y), round_shift_y);
- res_a_round1 = _mm_sra_epi32(
- _mm_add_epi32(res_a_round1, round_const_bits), round_shift_bits);
-
- if (w - j > 4) {
- const __m128i res_b0 = convolve(s + 4, coeffs_y);
- __m128i res_b_round0 = _mm_sra_epi32(
- _mm_add_epi32(res_b0, round_const_y), round_shift_y);
- res_b_round0 = _mm_sra_epi32(
- _mm_add_epi32(res_b_round0, round_const_bits), round_shift_bits);
-
- const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y);
- __m128i res_b_round1 = _mm_sra_epi32(
- _mm_add_epi32(res_b1, round_const_y), round_shift_y);
- res_b_round1 = _mm_sra_epi32(
- _mm_add_epi32(res_b_round1, round_const_bits), round_shift_bits);
-
- __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
- res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
- res_16bit0 = _mm_max_epi16(res_16bit0, zero);
-
- __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
- res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
- res_16bit1 = _mm_max_epi16(res_16bit1, zero);
-
- _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
- _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
- res_16bit1);
- } else if (w == 4) {
- res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
- res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
- res_a_round0 = _mm_max_epi16(res_a_round0, zero);
-
- res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
- res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
- res_a_round1 = _mm_max_epi16(res_a_round1, zero);
-
- _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
- _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
- res_a_round1);
- } else {
- res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
- res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
- res_a_round0 = _mm_max_epi16(res_a_round0, zero);
-
- res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
- res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
- res_a_round1 = _mm_max_epi16(res_a_round1, zero);
-
- *((uint32_t *)(&dst[i * dst_stride + j])) =
- _mm_cvtsi128_si32(res_a_round0);
-
- *((uint32_t *)(&dst[i * dst_stride + j + dst_stride])) =
- _mm_cvtsi128_si32(res_a_round1);
+ _mm_store_si128((__m128i *)&im_block[i * im_stride], res);
}
- s[0] = s[1];
- s[1] = s[2];
- s[2] = s[3];
+ }
- s[4] = s[5];
- s[5] = s[6];
- s[6] = s[7];
+ /* Vertical filter */
+ {
+ __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride));
+ __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride));
+ __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride));
+ __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride));
+ __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride));
+ __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride));
+ __m128i s6 = _mm_loadu_si128((__m128i *)(im_block + 6 * im_stride));
- s[0 + 8] = s[1 + 8];
- s[1 + 8] = s[2 + 8];
- s[2 + 8] = s[3 + 8];
+ s[0] = _mm_unpacklo_epi16(s0, s1);
+ s[1] = _mm_unpacklo_epi16(s2, s3);
+ s[2] = _mm_unpacklo_epi16(s4, s5);
- s[4 + 8] = s[5 + 8];
- s[5 + 8] = s[6 + 8];
- s[6 + 8] = s[7 + 8];
+ s[4] = _mm_unpackhi_epi16(s0, s1);
+ s[5] = _mm_unpackhi_epi16(s2, s3);
+ s[6] = _mm_unpackhi_epi16(s4, s5);
- s6 = s8;
+ s[0 + 8] = _mm_unpacklo_epi16(s1, s2);
+ s[1 + 8] = _mm_unpacklo_epi16(s3, s4);
+ s[2 + 8] = _mm_unpacklo_epi16(s5, s6);
+
+ s[4 + 8] = _mm_unpackhi_epi16(s1, s2);
+ s[5 + 8] = _mm_unpackhi_epi16(s3, s4);
+ s[6 + 8] = _mm_unpackhi_epi16(s5, s6);
+
+ for (i = 0; i < h; i += 2) {
+ const int16_t *data = &im_block[i * im_stride];
+
+ __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * im_stride));
+ __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * im_stride));
+
+ s[3] = _mm_unpacklo_epi16(s6, s7);
+ s[7] = _mm_unpackhi_epi16(s6, s7);
+
+ s[3 + 8] = _mm_unpacklo_epi16(s7, s8);
+ s[7 + 8] = _mm_unpackhi_epi16(s7, s8);
+
+ const __m128i res_a0 = convolve(s, coeffs_y);
+ __m128i res_a_round0 = _mm_sra_epi32(
+ _mm_add_epi32(res_a0, round_const_y), round_shift_y);
+ res_a_round0 = _mm_sra_epi32(
+ _mm_add_epi32(res_a_round0, round_const_bits), round_shift_bits);
+
+ const __m128i res_a1 = convolve(s + 8, coeffs_y);
+ __m128i res_a_round1 = _mm_sra_epi32(
+ _mm_add_epi32(res_a1, round_const_y), round_shift_y);
+ res_a_round1 = _mm_sra_epi32(
+ _mm_add_epi32(res_a_round1, round_const_bits), round_shift_bits);
+
+ if (w - j > 4) {
+ const __m128i res_b0 = convolve(s + 4, coeffs_y);
+ __m128i res_b_round0 = _mm_sra_epi32(
+ _mm_add_epi32(res_b0, round_const_y), round_shift_y);
+ res_b_round0 =
+ _mm_sra_epi32(_mm_add_epi32(res_b_round0, round_const_bits),
+ round_shift_bits);
+
+ const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y);
+ __m128i res_b_round1 = _mm_sra_epi32(
+ _mm_add_epi32(res_b1, round_const_y), round_shift_y);
+ res_b_round1 =
+ _mm_sra_epi32(_mm_add_epi32(res_b_round1, round_const_bits),
+ round_shift_bits);
+
+ __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
+ res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
+ res_16bit0 = _mm_max_epi16(res_16bit0, zero);
+
+ __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
+ res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
+ res_16bit1 = _mm_max_epi16(res_16bit1, zero);
+
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ res_16bit1);
+ } else if (w == 4) {
+ res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+ res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+ res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+ res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+ res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+ res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ res_a_round1);
+ } else {
+ res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+ res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+ res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+ res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+ res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+ res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+ *((uint32_t *)(&dst[i * dst_stride + j])) =
+ _mm_cvtsi128_si32(res_a_round0);
+
+ *((uint32_t *)(&dst[i * dst_stride + j + dst_stride])) =
+ _mm_cvtsi128_si32(res_a_round1);
+ }
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+
+ s[4] = s[5];
+ s[5] = s[6];
+ s[6] = s[7];
+
+ s[0 + 8] = s[1 + 8];
+ s[1 + 8] = s[2 + 8];
+ s[2 + 8] = s[3 + 8];
+
+ s[4 + 8] = s[5 + 8];
+ s[5 + 8] = s[6 + 8];
+ s[6 + 8] = s[7 + 8];
+
+ s6 = s8;
+ }
}
}
}
diff --git a/test/av1_convolve_test.cc b/test/av1_convolve_test.cc
index dfcdaab..0c90280 100644
--- a/test/av1_convolve_test.cc
+++ b/test/av1_convolve_test.cc
@@ -842,8 +842,11 @@
void RunTest() {
for (int sub_x = 0; sub_x < 16; ++sub_x) {
for (int sub_y = 0; sub_y < 16; ++sub_y) {
- for (int h_f = EIGHTTAP_REGULAR; h_f < INTERP_FILTERS_ALL; ++h_f) {
- for (int v_f = EIGHTTAP_REGULAR; v_f < INTERP_FILTERS_ALL; ++v_f) {
+ for (int h_f = EIGHTTAP_REGULAR; h_f <= INTERP_FILTERS_ALL; ++h_f) {
+ for (int v_f = EIGHTTAP_REGULAR; v_f <= INTERP_FILTERS_ALL; ++v_f) {
+ if (((h_f == MULTITAP_SHARP2) && (v_f < MULTITAP_SHARP2)) ||
+ ((h_f < MULTITAP_SHARP2) && (v_f == MULTITAP_SHARP2)))
+ continue;
TestConvolve(static_cast<InterpFilter>(h_f),
static_cast<InterpFilter>(v_f), sub_x, sub_y);
}
@@ -854,30 +857,17 @@
public:
void SpeedTest() {
- for (int h_f = EIGHTTAP_REGULAR; h_f < INTERP_FILTERS_ALL; ++h_f) {
- for (int v_f = EIGHTTAP_REGULAR; v_f < INTERP_FILTERS_ALL; ++v_f) {
+ for (int h_f = EIGHTTAP_REGULAR; h_f <= INTERP_FILTERS_ALL; ++h_f) {
+ for (int v_f = EIGHTTAP_REGULAR; v_f <= INTERP_FILTERS_ALL; ++v_f) {
+ if (((h_f == MULTITAP_SHARP2) && (v_f < MULTITAP_SHARP2)) ||
+ ((h_f < MULTITAP_SHARP2) && (v_f == MULTITAP_SHARP2)))
+ continue;
TestConvolveSpeed(static_cast<InterpFilter>(h_f),
static_cast<InterpFilter>(v_f), 10000);
}
}
}
- public:
- void RunTest12Tap() {
- for (int sub_x = 0; sub_x < 16; ++sub_x) {
- for (int sub_y = 0; sub_y < 16; ++sub_y) {
- TestConvolve(static_cast<InterpFilter>(MULTITAP_SHARP2),
- static_cast<InterpFilter>(MULTITAP_SHARP2), sub_x, sub_y);
- }
- }
- }
-
- public:
- void SpeedTest12Tap() {
- TestConvolveSpeed(static_cast<InterpFilter>(MULTITAP_SHARP2),
- static_cast<InterpFilter>(MULTITAP_SHARP2), 10000);
- }
-
private:
void TestConvolve(const InterpFilter h_f, const InterpFilter v_f,
const int sub_x, const int sub_y) {
@@ -939,12 +929,8 @@
TEST_P(AV1Convolve2DTest, RunTest) { RunTest(); }
-TEST_P(AV1Convolve2DTest, RunTest12Tap) { RunTest12Tap(); }
-
TEST_P(AV1Convolve2DTest, DISABLED_SpeedTest) { SpeedTest(); }
-TEST_P(AV1Convolve2DTest, DISABLED_SpeedTest12Tap) { SpeedTest12Tap(); }
-
INSTANTIATE_TEST_SUITE_P(C, AV1Convolve2DTest,
BuildLowbdParams(av1_convolve_2d_sr_c));
@@ -980,8 +966,11 @@
void RunTest() {
for (int sub_x = 0; sub_x < 16; ++sub_x) {
for (int sub_y = 0; sub_y < 16; ++sub_y) {
- for (int h_f = EIGHTTAP_REGULAR; h_f < INTERP_FILTERS_ALL; ++h_f) {
- for (int v_f = EIGHTTAP_REGULAR; v_f < INTERP_FILTERS_ALL; ++v_f) {
+ for (int h_f = EIGHTTAP_REGULAR; h_f <= INTERP_FILTERS_ALL; ++h_f) {
+ for (int v_f = EIGHTTAP_REGULAR; v_f <= INTERP_FILTERS_ALL; ++v_f) {
+ if (((h_f == MULTITAP_SHARP2) && (v_f < MULTITAP_SHARP2)) ||
+ ((h_f < MULTITAP_SHARP2) && (v_f == MULTITAP_SHARP2)))
+ continue;
TestConvolve(static_cast<InterpFilter>(h_f),
static_cast<InterpFilter>(v_f), sub_x, sub_y);
}
@@ -990,6 +979,19 @@
}
}
+ public:
+ void SpeedTest() {
+ for (int h_f = EIGHTTAP_REGULAR; h_f <= INTERP_FILTERS_ALL; ++h_f) {
+ for (int v_f = EIGHTTAP_REGULAR; v_f <= INTERP_FILTERS_ALL; ++v_f) {
+ if (((h_f == MULTITAP_SHARP2) && (v_f < MULTITAP_SHARP2)) ||
+ ((h_f < MULTITAP_SHARP2) && (v_f == MULTITAP_SHARP2)))
+ continue;
+ TestConvolveSpeed(static_cast<InterpFilter>(h_f),
+ static_cast<InterpFilter>(v_f), 10000);
+ }
+ }
+ }
+
private:
void TestConvolve(const InterpFilter h_f, const InterpFilter v_f,
const int sub_x, const int sub_y) {
@@ -1015,10 +1017,47 @@
&conv_params2, bit_depth);
AssertOutputBufferEq(reference, test, width, height);
}
+
+ void TestConvolveSpeed(const InterpFilter h_f, const InterpFilter v_f,
+ int num_iters) {
+ const int width = GetParam().Block().Width();
+ const int height = GetParam().Block().Height();
+ const int bit_depth = GetParam().BitDepth();
+ const InterpFilterParams *filter_params_x =
+ av1_get_interp_filter_params_with_block_size(h_f, width);
+ const InterpFilterParams *filter_params_y =
+ av1_get_interp_filter_params_with_block_size(v_f, height);
+ const uint16_t *input = FirstRandomInput16(GetParam());
+ DECLARE_ALIGNED(32, uint16_t, reference[MAX_SB_SQUARE]);
+ ConvolveParams conv_params1 = get_conv_params_no_round(0, 0, NULL, 0, 0, 8);
+ aom_usec_timer timer;
+ aom_usec_timer_start(&timer);
+ for (int i = 0; i < num_iters; ++i) {
+ av1_highbd_convolve_2d_sr_c(input, width, reference, kOutputStride, width,
+ height, filter_params_x, filter_params_y, 0,
+ 0, &conv_params1, bit_depth);
+ }
+ aom_usec_timer_mark(&timer);
+ const double time1 = static_cast<double>(aom_usec_timer_elapsed(&timer));
+ DECLARE_ALIGNED(32, uint16_t, test[MAX_SB_SQUARE]);
+ ConvolveParams conv_params2 = get_conv_params_no_round(0, 0, NULL, 0, 0, 8);
+ aom_usec_timer_start(&timer);
+ for (int i = 0; i < num_iters; ++i) {
+ GetParam().TestFunction()(input, width, test, kOutputStride, width,
+ height, filter_params_x, filter_params_y, 0, 0,
+ &conv_params2, bit_depth);
+ }
+ aom_usec_timer_mark(&timer);
+ const double time2 = static_cast<double>(aom_usec_timer_elapsed(&timer));
+ printf("%d - %d %3dx%-3d:%7.2f/%7.2fns (%3.2f)\n", h_f, v_f, width, height,
+ time1, time2, time1 / time2);
+ }
};
TEST_P(AV1Convolve2DHighbdTest, RunTest) { RunTest(); }
+TEST_P(AV1Convolve2DHighbdTest, DISABLED_SpeedTest) { SpeedTest(); }
+
INSTANTIATE_TEST_SUITE_P(C, AV1Convolve2DHighbdTest,
BuildHighbdParams(av1_highbd_convolve_2d_sr_c));