Faster AVX2 implementation of motion compensation modules
Improvements have been made to av1_convolve_y_avx2 (~1.5x faster),
av1_convolve_y_sr_avx2 (~1.8x faster) and av1_convolve_2d_sr_avx2 (~1.3x faster).
Change-Id: Iaed764a7c4d069a4180c3edb0b1ac57ad36dad21
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index f61af74..6c1d89e 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -81,6 +81,7 @@
"${AOM_ROOT}/aom_dsp/x86/intrapred_avx2.c"
"${AOM_ROOT}/aom_dsp/x86/inv_txfm_avx2.c"
"${AOM_ROOT}/aom_dsp/x86/common_avx2.h"
+ "${AOM_ROOT}/aom_dsp/x86/convolve_avx2.h"
"${AOM_ROOT}/aom_dsp/x86/inv_txfm_common_avx2.h"
"${AOM_ROOT}/aom_dsp/x86/txfm_common_avx2.h")
diff --git a/aom_dsp/x86/convolve_avx2.h b/aom_dsp/x86/convolve_avx2.h
index f4ea91e..80a6067 100644
--- a/aom_dsp/x86/convolve_avx2.h
+++ b/aom_dsp/x86/convolve_avx2.h
@@ -33,4 +33,101 @@
6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
};
+static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params,
+ const int subpel_q4,
+ __m256i *const coeffs /* [4] */) {
+ const int16_t *const filter = av1_get_interp_filter_subpel_kernel(
+ *filter_params, subpel_q4 & SUBPEL_MASK);
+ const __m128i coeffs_8 = _mm_loadu_si128((__m128i *)filter);
+ const __m256i filter_coeffs = _mm256_broadcastsi128_si256(coeffs_8);
+
+ // right shift all filter co-efficients by 1 to reduce the bits required.
+ // This extra right shift will be taken care of at the end while rounding
+ // the result.
+ // Since all filter co-efficients are even, this change will not affect the
+ // end result
+ assert(_mm_test_all_zeros(_mm_and_si128(coeffs_8, _mm_set1_epi16(1)),
+ _mm_set1_epi16(0xffff)));
+
+ const __m256i coeffs_1 = _mm256_srai_epi16(filter_coeffs, 1);
+
+ // coeffs 0 1 0 1 0 1 0 1
+ coeffs[0] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0200u));
+ // coeffs 2 3 2 3 2 3 2 3
+ coeffs[1] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0604u));
+ // coeffs 4 5 4 5 4 5 4 5
+ coeffs[2] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0a08u));
+ // coeffs 6 7 6 7 6 7 6 7
+ coeffs[3] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0e0cu));
+}
+
+static INLINE void prepare_coeffs_y_2d(
+ const InterpFilterParams *const filter_params_y, const int subpel_y_q4,
+ __m256i *const coeffs /* [4] */) {
+ const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+ *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+
+ const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter);
+ const __m256i coeffs_y = _mm256_broadcastsi128_si256(coeffs_y8);
+
+ // coeffs 0 1 0 1 0 1 0 1
+ coeffs[0] = _mm256_shuffle_epi32(coeffs_y, 0x00);
+ // coeffs 2 3 2 3 2 3 2 3
+ coeffs[1] = _mm256_shuffle_epi32(coeffs_y, 0x55);
+ // coeffs 4 5 4 5 4 5 4 5
+ coeffs[2] = _mm256_shuffle_epi32(coeffs_y, 0xaa);
+ // coeffs 6 7 6 7 6 7 6 7
+ coeffs[3] = _mm256_shuffle_epi32(coeffs_y, 0xff);
+}
+
+static INLINE __m256i convolve(const __m256i *const s,
+ const __m256i *const coeffs) {
+ const __m256i res_01 = _mm256_maddubs_epi16(s[0], coeffs[0]);
+ const __m256i res_23 = _mm256_maddubs_epi16(s[1], coeffs[1]);
+ const __m256i res_45 = _mm256_maddubs_epi16(s[2], coeffs[2]);
+ const __m256i res_67 = _mm256_maddubs_epi16(s[3], coeffs[3]);
+
+ // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+ const __m256i res = _mm256_add_epi16(_mm256_add_epi16(res_01, res_45),
+ _mm256_add_epi16(res_23, res_67));
+
+ return res;
+}
+
+static INLINE __m256i convolve_y_2d(const __m256i *const s,
+ const __m256i *const coeffs) {
+ const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]);
+ const __m256i res_1 = _mm256_madd_epi16(s[1], coeffs[1]);
+ const __m256i res_2 = _mm256_madd_epi16(s[2], coeffs[2]);
+ const __m256i res_3 = _mm256_madd_epi16(s[3], coeffs[3]);
+
+ const __m256i res = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1),
+ _mm256_add_epi32(res_2, res_3));
+
+ return res;
+}
+
+static INLINE __m256i convolve_x(const __m256i data,
+ const __m256i *const coeffs,
+ const __m256i *const filt) {
+ __m256i s[4];
+
+ s[0] = _mm256_shuffle_epi8(data, filt[0]);
+ s[1] = _mm256_shuffle_epi8(data, filt[1]);
+ s[2] = _mm256_shuffle_epi8(data, filt[2]);
+ s[3] = _mm256_shuffle_epi8(data, filt[3]);
+
+ return convolve(s, coeffs);
+}
+
+static INLINE void add_store_aligned(CONV_BUF_TYPE *const dst,
+ const __m256i *const res,
+ const __m256i *const avg_mask) {
+ __m256i d;
+ d = _mm256_load_si256((__m256i *)dst);
+ d = _mm256_and_si256(d, *avg_mask);
+ d = _mm256_add_epi32(d, *res);
+ _mm256_store_si256((__m256i *)dst, d);
+}
+
#endif
diff --git a/av1/common/convolve.c b/av1/common/convolve.c
index 9f4ec26..6ce9700 100644
--- a/av1/common/convolve.c
+++ b/av1/common/convolve.c
@@ -401,7 +401,6 @@
for (int x = 0; x < w; ++x) {
int32_t sum = (1 << (bd + FILTER_BITS - 1));
for (int k = 0; k < filter_params_x->taps; ++k) {
- assert((x_filter[k] % 2) == 0);
sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
}
assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
@@ -419,7 +418,6 @@
for (int x = 0; x < w; ++x) {
CONV_BUF_TYPE sum = 1 << offset_bits;
for (int k = 0; k < filter_params_y->taps; ++k) {
- assert((y_filter[k] % 2) == 0);
sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
}
assert(0 <= sum && sum < (1 << (offset_bits + 2)));
@@ -456,7 +454,6 @@
for (int x = 0; x < w; ++x) {
CONV_BUF_TYPE res = 0;
for (int k = 0; k < filter_params_y->taps; ++k) {
- assert((y_filter[k] % 2) == 0);
res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
}
res *= (1 << bits);
@@ -490,7 +487,6 @@
for (int x = 0; x < w; ++x) {
CONV_BUF_TYPE res = 0;
for (int k = 0; k < filter_params_x->taps; ++k) {
- assert((x_filter[k] % 2) == 0);
res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
}
res = (1 << bits) * ROUND_POWER_OF_TWO(res, conv_params->round_0);
@@ -554,7 +550,6 @@
for (int x = 0; x < w; ++x) {
int32_t sum = (1 << (bd + FILTER_BITS - 1));
for (int k = 0; k < filter_params_x->taps; ++k) {
- assert((x_filter[k] % 2) == 0);
sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
}
assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
@@ -572,7 +567,6 @@
for (int x = 0; x < w; ++x) {
CONV_BUF_TYPE sum = 1 << offset_bits;
for (int k = 0; k < filter_params_y->taps; ++k) {
- assert((y_filter[k] % 2) == 0);
sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
}
assert(0 <= sum && sum < (1 << (offset_bits + 2)));
@@ -602,7 +596,6 @@
for (int x = 0; x < w; ++x) {
CONV_BUF_TYPE res = 0;
for (int k = 0; k < filter_params_y->taps; ++k) {
- assert((y_filter[k] % 2) == 0);
res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
}
dst[y * dst_stride + x] =
@@ -630,7 +623,6 @@
for (int x = 0; x < w; ++x) {
CONV_BUF_TYPE res = 0;
for (int k = 0; k < filter_params_x->taps; ++k) {
- assert((x_filter[k] % 2) == 0);
res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
}
res = ROUND_POWER_OF_TWO(res, conv_params->round_0);
diff --git a/av1/common/reconinter.c b/av1/common/reconinter.c
index 0a033f1..46d2129 100644
--- a/av1/common/reconinter.c
+++ b/av1/common/reconinter.c
@@ -740,7 +740,7 @@
// CONFIG_HIGHBITDEPTH or just 8 otherwise.
#define INTER_PRED_BYTES_PER_PIXEL 4
- DECLARE_ALIGNED(16, uint8_t,
+ DECLARE_ALIGNED(32, uint8_t,
tmp_buf[INTER_PRED_BYTES_PER_PIXEL * MAX_SB_SQUARE]);
#undef INTER_PRED_BYTES_PER_PIXEL
@@ -988,7 +988,7 @@
for (idx = 0; idx < b8_w; idx += b4_w) {
MB_MODE_INFO *this_mbmi = &xd->mi[row * xd->mi_stride + col]->mbmi;
is_compound = has_second_ref(this_mbmi);
- DECLARE_ALIGNED(16, int32_t, tmp_dst[8 * 8]);
+ DECLARE_ALIGNED(32, int32_t, tmp_dst[8 * 8]);
int tmp_dst_stride = 8;
assert(w <= 8 && h <= 8);
ConvolveParams conv_params = get_conv_params_no_round(
@@ -1124,7 +1124,7 @@
uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
uint8_t *pre[2];
SubpelParams subpel_params[2];
- DECLARE_ALIGNED(16, int32_t, tmp_dst[MAX_SB_SIZE * MAX_SB_SIZE]);
+ DECLARE_ALIGNED(32, int32_t, tmp_dst[MAX_SB_SIZE * MAX_SB_SIZE]);
for (ref = 0; ref < 1 + is_compound; ++ref) {
#if CONFIG_INTRABC
diff --git a/av1/common/x86/convolve_2d_avx2.c b/av1/common/x86/convolve_2d_avx2.c
index 9ee6e4b..a5df334 100644
--- a/av1/common/x86/convolve_2d_avx2.c
+++ b/av1/common/x86/convolve_2d_avx2.c
@@ -39,10 +39,10 @@
int i, j;
const int fo_vert = filter_params_y->taps / 2 - 1;
const int fo_horiz = filter_params_x->taps / 2 - 1;
- const int do_average = conv_params->do_average;
const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+ const __m256i avg_mask = _mm256_set1_epi32(conv_params->do_average ? -1 : 0);
- __m256i filt[4], s[4];
+ __m256i filt[4], s[4], coeffs[4];
filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
@@ -51,33 +51,7 @@
/* Horizontal filter */
{
- const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
-
- const __m128i coeffs_x8 = _mm_loadu_si128((__m128i *)x_filter);
- // since not all compilers yet support _mm256_set_m128i()
- const __m256i coeffs_x = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_x8), coeffs_x8, 1);
-
- // right shift all filter co-efficients by 1 to reduce the bits required.
- // This extra right shift will be taken care of at the end while rounding
- // the result.
- // Since all filter co-efficients are even, this change will not affect the
- // end result
- const __m256i coeffs_x_1 = _mm256_srai_epi16(coeffs_x, 1);
-
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0200u));
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0604u));
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0a08u));
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0e0cu));
+ prepare_coeffs(filter_params_x, subpel_x_q4, coeffs);
const __m256i round_const =
_mm256_set1_epi16(((1 << (conv_params->round_0 - 1)) >> 1) +
@@ -93,52 +67,20 @@
_mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + (j + 8)]),
1);
- // filter the source buffer
- s[0] = _mm256_shuffle_epi8(data, filt[0]);
- s[1] = _mm256_shuffle_epi8(data, filt[1]);
- s[2] = _mm256_shuffle_epi8(data, filt[2]);
- s[3] = _mm256_shuffle_epi8(data, filt[3]);
+ __m256i res = convolve_x(data, coeffs, filt);
- const __m256i res_0 = _mm256_maddubs_epi16(s[0], coeff_01);
- const __m256i res_1 = _mm256_maddubs_epi16(s[1], coeff_23);
- const __m256i res_2 = _mm256_maddubs_epi16(s[2], coeff_45);
- const __m256i res_3 = _mm256_maddubs_epi16(s[3], coeff_67);
-
- const __m256i res_a = _mm256_add_epi16(res_0, res_2);
- const __m256i res_b = _mm256_add_epi16(res_1, res_3);
-
- __m256i res = _mm256_add_epi16(res_a, res_b);
res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift);
res = _mm256_permute4x64_epi64(res, 216);
// 0 1 2 3 8 9 10 11 4 5 6 7 12 13 14 15
- _mm256_storeu_si256((__m256i *)&im_block[i * im_stride + j], res);
+ _mm256_store_si256((__m256i *)&im_block[i * im_stride + j], res);
}
}
}
/* Vertical filter */
{
- const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
-
- const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter);
- const __m256i coeffs_y = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_y8), coeffs_y8, 1);
-
- // coeffs 0 1 0 1 2 3 2 3
- const __m256i tmp_0 = _mm256_unpacklo_epi32(coeffs_y, coeffs_y);
- // coeffs 4 5 4 5 6 7 6 7
- const __m256i tmp_1 = _mm256_unpackhi_epi32(coeffs_y, coeffs_y);
-
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 = _mm256_unpacklo_epi64(tmp_0, tmp_0);
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 = _mm256_unpackhi_epi64(tmp_0, tmp_0);
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 = _mm256_unpacklo_epi64(tmp_1, tmp_1);
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 = _mm256_unpackhi_epi64(tmp_1, tmp_1);
+ prepare_coeffs_y_2d(filter_params_y, subpel_y_q4, coeffs);
const __m256i round_const = _mm256_set1_epi32(
((1 << conv_params->round_1) >> 1) -
@@ -149,69 +91,38 @@
for (j = 0; j < w; j += 16) {
// Filter 0 1 2 3 4 5 6 7
const int16_t *data = &im_block[i * im_stride + j];
- const __m256i src_0 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 0 * im_stride),
- *(__m256i *)(data + 1 * im_stride));
- const __m256i src_1 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 2 * im_stride),
- *(__m256i *)(data + 3 * im_stride));
- const __m256i src_2 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 4 * im_stride),
- *(__m256i *)(data + 5 * im_stride));
- const __m256i src_3 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 6 * im_stride),
- *(__m256i *)(data + 7 * im_stride));
+ s[0] = _mm256_unpacklo_epi16(*(__m256i *)(data + 0 * im_stride),
+ *(__m256i *)(data + 1 * im_stride));
+ s[1] = _mm256_unpacklo_epi16(*(__m256i *)(data + 2 * im_stride),
+ *(__m256i *)(data + 3 * im_stride));
+ s[2] = _mm256_unpacklo_epi16(*(__m256i *)(data + 4 * im_stride),
+ *(__m256i *)(data + 5 * im_stride));
+ s[3] = _mm256_unpacklo_epi16(*(__m256i *)(data + 6 * im_stride),
+ *(__m256i *)(data + 7 * im_stride));
- const __m256i res_0 = _mm256_madd_epi16(src_0, coeff_01);
- const __m256i res_1 = _mm256_madd_epi16(src_1, coeff_23);
- const __m256i res_2 = _mm256_madd_epi16(src_2, coeff_45);
- const __m256i res_3 = _mm256_madd_epi16(src_3, coeff_67);
-
- const __m256i res_a = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1),
- _mm256_add_epi32(res_2, res_3));
+ const __m256i res_a = convolve_y_2d(s, coeffs);
// Filter 8 9 10 11 12 13 14 15
- const __m256i src_4 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 0 * im_stride),
- *(__m256i *)(data + 1 * im_stride));
- const __m256i src_5 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 2 * im_stride),
- *(__m256i *)(data + 3 * im_stride));
- const __m256i src_6 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 4 * im_stride),
- *(__m256i *)(data + 5 * im_stride));
- const __m256i src_7 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 6 * im_stride),
- *(__m256i *)(data + 7 * im_stride));
+ s[0] = _mm256_unpackhi_epi16(*(__m256i *)(data + 0 * im_stride),
+ *(__m256i *)(data + 1 * im_stride));
+ s[1] = _mm256_unpackhi_epi16(*(__m256i *)(data + 2 * im_stride),
+ *(__m256i *)(data + 3 * im_stride));
+ s[2] = _mm256_unpackhi_epi16(*(__m256i *)(data + 4 * im_stride),
+ *(__m256i *)(data + 5 * im_stride));
+ s[3] = _mm256_unpackhi_epi16(*(__m256i *)(data + 6 * im_stride),
+ *(__m256i *)(data + 7 * im_stride));
- const __m256i res_4 = _mm256_madd_epi16(src_4, coeff_01);
- const __m256i res_5 = _mm256_madd_epi16(src_5, coeff_23);
- const __m256i res_6 = _mm256_madd_epi16(src_6, coeff_45);
- const __m256i res_7 = _mm256_madd_epi16(src_7, coeff_67);
-
- const __m256i res_b = _mm256_add_epi32(_mm256_add_epi32(res_4, res_5),
- _mm256_add_epi32(res_6, res_7));
+ const __m256i res_b = convolve_y_2d(s, coeffs);
const __m256i res_a_round =
_mm256_sra_epi32(_mm256_add_epi32(res_a, round_const), round_shift);
const __m256i res_b_round =
_mm256_sra_epi32(_mm256_add_epi32(res_b, round_const), round_shift);
- // Accumulate values into the destination buffer
- __m256i *const p = (__m256i *)&dst[i * dst_stride + j];
- if (do_average) {
- _mm256_storeu_si256(
- p + 0, _mm256_add_epi32(_mm256_loadu_si256(p + 0), res_a_round));
- if (w - j > 8) {
- _mm256_storeu_si256(
- p + 1,
- _mm256_add_epi32(_mm256_loadu_si256(p + 1), res_b_round));
- }
- } else {
- _mm256_storeu_si256(p + 0, res_a_round);
- if (w - j > 8) {
- _mm256_storeu_si256(p + 1, res_b_round);
- }
+ add_store_aligned(&dst[i * dst_stride + j], &res_a_round, &avg_mask);
+ if (w - j > 8) {
+ add_store_aligned(&dst[i * dst_stride + j + 8], &res_b_round,
+ &avg_mask);
}
}
}
@@ -226,192 +137,123 @@
ConvolveParams *conv_params) {
const int bd = 8;
- DECLARE_ALIGNED(32, int16_t,
- im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
+ 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 = MAX_SB_SIZE;
+ int im_stride = 8;
int i, j;
const int fo_vert = filter_params_y->taps / 2 - 1;
const int fo_horiz = filter_params_x->taps / 2 - 1;
const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
- __m256i filt[4], s[4];
+ __m256i filt[4], coeffs_h[4], coeffs_v[4];
filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
- /* Horizontal filter */
- {
- const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+ prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_h);
+ prepare_coeffs_y_2d(filter_params_y, subpel_y_q4, coeffs_v);
- const __m128i coeffs_x8 = _mm_loadu_si128((__m128i *)x_filter);
- // since not all compilers yet support _mm256_set_m128i()
- const __m256i coeffs_x = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_x8), coeffs_x8, 1);
+ const __m256i round_const_h = _mm256_set1_epi16(
+ ((1 << (conv_params->round_0 - 1)) >> 1) + (1 << (bd + FILTER_BITS - 2)));
+ const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0 - 1);
- // right shift all filter co-efficients by 1 to reduce the bits required.
- // This extra right shift will be taken care of at the end while rounding
- // the result.
- // Since all filter co-efficients are even, this change will not affect
- // the end result
- const __m256i coeffs_x_1 = _mm256_srai_epi16(coeffs_x, 1);
+ const __m256i round_const_v = _mm256_set1_epi32(
+ ((1 << conv_params->round_1) >> 1) -
+ (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)) +
+ ((1 << (2 * FILTER_BITS - conv_params->round_0)) >> 1));
+ const __m128i round_shift_v =
+ _mm_cvtsi32_si128(2 * FILTER_BITS - conv_params->round_0);
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0200u));
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0604u));
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0a08u));
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0e0cu));
+ for (j = 0; j < w; j += 8) {
+ for (i = 0; i < im_h; i += 2) {
+ __m256i data = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
- const __m256i round_const =
- _mm256_set1_epi16(((1 << (conv_params->round_0 - 1)) >> 1) +
- (1 << (bd + FILTER_BITS - 2)));
- const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0 - 1);
-
- for (i = 0; i < im_h; ++i) {
- for (j = 0; j < w; j += 16) {
- // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 8 9 10 11 12 13 14 15 16 17
- // 18 19 20 21 22 23
- const __m256i data = _mm256_inserti128_si256(
- _mm256_loadu_si256((__m256i *)&src_ptr[(i * src_stride) + j]),
- _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + (j + 8)]),
+ // Load the next line
+ if (i + 1 < im_h)
+ data = _mm256_inserti128_si256(
+ data, _mm_loadu_si128(
+ (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),
1);
- // filter the source buffer
- s[0] = _mm256_shuffle_epi8(data, filt[0]);
- s[1] = _mm256_shuffle_epi8(data, filt[1]);
- s[2] = _mm256_shuffle_epi8(data, filt[2]);
- s[3] = _mm256_shuffle_epi8(data, filt[3]);
+ __m256i res = convolve_x(data, coeffs_h, filt);
- const __m256i res_0 = _mm256_maddubs_epi16(s[0], coeff_01);
- const __m256i res_1 = _mm256_maddubs_epi16(s[1], coeff_23);
- const __m256i res_2 = _mm256_maddubs_epi16(s[2], coeff_45);
- const __m256i res_3 = _mm256_maddubs_epi16(s[3], coeff_67);
+ res =
+ _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h);
- const __m256i res_a = _mm256_add_epi16(res_0, res_2);
- const __m256i res_b = _mm256_add_epi16(res_1, res_3);
-
- __m256i res = _mm256_add_epi16(res_a, res_b);
- res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift);
-
- // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
- _mm256_storeu_si256((__m256i *)&im_block[i * im_stride + j], res);
- }
+ _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
}
- }
- /* Vertical filter */
- {
- const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+ /* Vertical filter */
+ {
+ __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));
+ __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));
+ __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));
+ __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));
+ __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));
+ __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));
- const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter);
- const __m256i coeffs_y = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_y8), coeffs_y8, 1);
+ __m256i s[8];
+ s[0] = _mm256_unpacklo_epi16(src_0, src_1);
+ s[1] = _mm256_unpacklo_epi16(src_2, src_3);
+ s[2] = _mm256_unpacklo_epi16(src_4, src_5);
- // coeffs 0 1 0 1 2 3 2 3
- const __m256i tmp_0 = _mm256_unpacklo_epi32(coeffs_y, coeffs_y);
- // coeffs 4 5 4 5 6 7 6 7
- const __m256i tmp_1 = _mm256_unpackhi_epi32(coeffs_y, coeffs_y);
+ s[4] = _mm256_unpackhi_epi16(src_0, src_1);
+ s[5] = _mm256_unpackhi_epi16(src_2, src_3);
+ s[6] = _mm256_unpackhi_epi16(src_4, src_5);
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 = _mm256_unpacklo_epi64(tmp_0, tmp_0);
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 = _mm256_unpackhi_epi64(tmp_0, tmp_0);
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 = _mm256_unpacklo_epi64(tmp_1, tmp_1);
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 = _mm256_unpackhi_epi64(tmp_1, tmp_1);
+ for (i = 0; i < h; i += 2) {
+ const int16_t *data = &im_block[i * im_stride];
- const __m256i round_const = _mm256_set1_epi32(
- ((1 << conv_params->round_1) >> 1) -
- (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)) +
- ((1 << (2 * FILTER_BITS - conv_params->round_0)) >> 1));
- const __m128i round_shift =
- _mm_cvtsi32_si128(2 * FILTER_BITS - conv_params->round_0);
+ const __m256i s6 =
+ _mm256_loadu_si256((__m256i *)(data + 6 * im_stride));
+ const __m256i s7 =
+ _mm256_loadu_si256((__m256i *)(data + 7 * im_stride));
- for (i = 0; i < h; ++i) {
- for (j = 0; j < w; j += 16) {
- // Filter 0 1 2 3 8 9 10 11
- const int16_t *data = &im_block[i * im_stride + j];
- const __m256i src_0 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 0 * im_stride),
- *(__m256i *)(data + 1 * im_stride));
- const __m256i src_1 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 2 * im_stride),
- *(__m256i *)(data + 3 * im_stride));
- const __m256i src_2 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 4 * im_stride),
- *(__m256i *)(data + 5 * im_stride));
- const __m256i src_3 =
- _mm256_unpacklo_epi16(*(__m256i *)(data + 6 * im_stride),
- *(__m256i *)(data + 7 * im_stride));
+ s[3] = _mm256_unpacklo_epi16(s6, s7);
+ s[7] = _mm256_unpackhi_epi16(s6, s7);
- const __m256i res_0 = _mm256_madd_epi16(src_0, coeff_01);
- const __m256i res_1 = _mm256_madd_epi16(src_1, coeff_23);
- const __m256i res_2 = _mm256_madd_epi16(src_2, coeff_45);
- const __m256i res_3 = _mm256_madd_epi16(src_3, coeff_67);
-
- const __m256i res_a = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1),
- _mm256_add_epi32(res_2, res_3));
-
- // Filter 4 5 6 7 12 13 14 15
- const __m256i src_4 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 0 * im_stride),
- *(__m256i *)(data + 1 * im_stride));
- const __m256i src_5 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 2 * im_stride),
- *(__m256i *)(data + 3 * im_stride));
- const __m256i src_6 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 4 * im_stride),
- *(__m256i *)(data + 5 * im_stride));
- const __m256i src_7 =
- _mm256_unpackhi_epi16(*(__m256i *)(data + 6 * im_stride),
- *(__m256i *)(data + 7 * im_stride));
-
- const __m256i res_4 = _mm256_madd_epi16(src_4, coeff_01);
- const __m256i res_5 = _mm256_madd_epi16(src_5, coeff_23);
- const __m256i res_6 = _mm256_madd_epi16(src_6, coeff_45);
- const __m256i res_7 = _mm256_madd_epi16(src_7, coeff_67);
-
- const __m256i res_b = _mm256_add_epi32(_mm256_add_epi32(res_4, res_5),
- _mm256_add_epi32(res_6, res_7));
+ const __m256i res_a = convolve_y_2d(s, coeffs_v);
+ const __m256i res_b = convolve_y_2d(s + 4, coeffs_v);
// Combine V round and 2F-H-V round into a single rounding
- const __m256i res_a_round =
- _mm256_sra_epi32(_mm256_add_epi32(res_a, round_const), round_shift);
- const __m256i res_b_round =
- _mm256_sra_epi32(_mm256_add_epi32(res_b, round_const), round_shift);
+ const __m256i res_a_round = _mm256_sra_epi32(
+ _mm256_add_epi32(res_a, round_const_v), round_shift_v);
+ const __m256i res_b_round = _mm256_sra_epi32(
+ _mm256_add_epi32(res_b, round_const_v), round_shift_v);
/* rounding code */
// 16 bit conversion
const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
// 8 bit conversion and saturation to uint8
- __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);
- res_8b = _mm256_permute4x64_epi64(res_8b, 216);
- // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
- const __m128i res = _mm256_castsi256_si128(res_8b);
+ const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);
+
+ const __m128i res_0 = _mm256_castsi256_si128(res_8b);
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
// Store values into the destination buffer
- __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
- if (w - j > 8) {
- _mm_storeu_si128(p, res);
- } else if (w - j > 4) {
- _mm_storel_epi64(p, res);
+ __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
+ __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];
+ if (w - j > 4) {
+ _mm_storel_epi64(p_0, res_0);
+ _mm_storel_epi64(p_1, res_1);
} else if (w == 4) {
- xx_storel_32(&dst[i * dst_stride + j], res);
+ xx_storel_32(p_0, res_0);
+ xx_storel_32(p_1, res_1);
} else {
- *(uint16_t *)p = _mm_cvtsi128_si32(res);
+ *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
}
+
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+
+ s[4] = s[5];
+ s[5] = s[6];
+ s[6] = s[7];
}
}
}
diff --git a/av1/common/x86/convolve_avx2.c b/av1/common/x86/convolve_avx2.c
index dd1025d..7fd76f8 100644
--- a/av1/common/x86/convolve_avx2.c
+++ b/av1/common/x86/convolve_avx2.c
@@ -339,142 +339,142 @@
}
}
-DECLARE_ALIGNED(32, static const uint8_t, g_shuf1[32]) = {
- 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15,
- 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15
-};
-
void av1_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst0,
int dst_stride0, int w, int h,
InterpFilterParams *filter_params_x,
InterpFilterParams *filter_params_y,
const int subpel_x_q4, const int subpel_y_q4,
ConvolveParams *conv_params) {
- if (w < 16) {
- av1_convolve_y_sse2(src, src_stride, dst0, dst_stride0, w, h,
- filter_params_x, filter_params_y, subpel_x_q4,
- subpel_y_q4, conv_params);
- return;
- }
- {
- CONV_BUF_TYPE *dst = conv_params->dst;
- int dst_stride = conv_params->dst_stride;
- int i, j;
- const int fo_vert = filter_params_y->taps / 2 - 1;
- const int do_average = conv_params->do_average;
- const uint8_t *const src_ptr = src - fo_vert * src_stride;
- const int bits =
- FILTER_BITS - conv_params->round_0 - (conv_params->round_1 - 1);
- const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
- const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter);
- const __m256i coeffs_y = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_y8), coeffs_y8, 1);
+ CONV_BUF_TYPE *dst = conv_params->dst;
+ int dst_stride = conv_params->dst_stride;
+ int i, j;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride;
+ const int bits =
+ FILTER_BITS - conv_params->round_0 - (conv_params->round_1 - 1);
+ const __m256i avg_mask = _mm256_set1_epi32(conv_params->do_average ? -1 : 0);
+ __m256i coeffs[4], s[8];
- (void)conv_params;
+ prepare_coeffs(filter_params_y, subpel_y_q4, coeffs);
- // right shift all filter co-efficients by 1 to reduce the bits required.
- // This extra right shift will be taken care of at the end while rounding
- // the result. Since all filter co-efficients are even, this change will not
- // affect the end result
- const __m256i coeffs_y_1 = _mm256_srai_epi16(coeffs_y, 1);
+ (void)conv_params;
+ (void)filter_params_x;
+ (void)subpel_x_q4;
+ (void)dst0;
+ (void)dst_stride0;
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0200u));
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0604u));
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0a08u));
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0e0cu));
+ for (j = 0; j < w; j += 16) {
+ const uint8_t *data = &src_ptr[j];
+ __m256i src6;
- const __m256i shuf = _mm256_load_si256((__m256i const *)g_shuf1);
+ // Load lines a and b. Line a to lower 128, line b to upper 128
+ const __m256i src_01a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+ 0x20);
- (void)filter_params_x;
- (void)subpel_x_q4;
- (void)dst0;
- (void)dst_stride0;
+ const __m256i src_12a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+ 0x20);
- for (i = 0; i < h; ++i) {
- for (j = 0; j < w; j += 16) {
- const uint8_t *data = &src_ptr[i * src_stride + j];
- // Load lines a and b. Line a to lower 128, line b to upper 128
- const __m256i src_01a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
- 0x20);
- const __m256i src_23a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
- 0x20);
- const __m256i src_45a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
- 0x20);
- const __m256i src_67a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 6 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
- 0x20);
+ const __m256i src_23a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+ 0x20);
- // Permute across lanes. (a_lo a_hi b_lo b_hi -> a_lo b_lo a_hi b_hi)
- const __m256i src_01b = _mm256_permute4x64_epi64(src_01a, 0xd8);
- const __m256i src_23b = _mm256_permute4x64_epi64(src_23a, 0xd8);
- const __m256i src_45b = _mm256_permute4x64_epi64(src_45a, 0xd8);
- const __m256i src_67b = _mm256_permute4x64_epi64(src_67a, 0xd8);
- // Interleave a and b within lanes.
- const __m256i src_01 = _mm256_shuffle_epi8(src_01b, shuf);
- const __m256i src_23 = _mm256_shuffle_epi8(src_23b, shuf);
- const __m256i src_45 = _mm256_shuffle_epi8(src_45b, shuf);
- const __m256i src_67 = _mm256_shuffle_epi8(src_67b, shuf);
+ const __m256i src_34a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+ 0x20);
- // Filter source pixels
- const __m256i res_01 = _mm256_maddubs_epi16(src_01, coeff_01);
- const __m256i res_23 = _mm256_maddubs_epi16(src_23, coeff_23);
- const __m256i res_45 = _mm256_maddubs_epi16(src_45, coeff_45);
- const __m256i res_67 = _mm256_maddubs_epi16(src_67, coeff_67);
+ const __m256i src_45a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+ 0x20);
- // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
- __m256i res = _mm256_add_epi16(_mm256_add_epi16(res_01, res_45),
- _mm256_add_epi16(res_23, res_67));
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 6 * src_stride)));
+ const __m256i src_56a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+ src6, 0x20);
- const __m256i res_01_32b =
- _mm256_cvtepi16_epi32(_mm256_castsi256_si128(res));
- const __m256i res_23_32b =
- _mm256_cvtepi16_epi32(_mm256_extracti128_si256(res, 1));
+ s[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
+ s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
+ s[2] = _mm256_unpacklo_epi8(src_45a, src_56a);
- const __m256i res_01_shift = _mm256_slli_epi32(res_01_32b, bits);
- const __m256i res_23_shift = _mm256_slli_epi32(res_23_32b, bits);
+ s[4] = _mm256_unpackhi_epi8(src_01a, src_12a);
+ s[5] = _mm256_unpackhi_epi8(src_23a, src_34a);
+ s[6] = _mm256_unpackhi_epi8(src_45a, src_56a);
- // Accumulate values into the destination buffer
- __m256i *const p = (__m256i *)&dst[i * dst_stride + j];
- if (do_average) {
- const __m256i dst_lo = _mm256_loadu_si256(p + 0);
- const __m256i dst_hi = _mm256_loadu_si256(p + 1);
- const __m256i res_lo = _mm256_add_epi32(dst_lo, res_01_shift);
- const __m256i res_hi = _mm256_add_epi32(dst_hi, res_23_shift);
- _mm256_storeu_si256(p + 0, res_lo);
- if (w - j > 8) {
- _mm256_storeu_si256(p + 1, res_hi);
- }
- } else {
- _mm256_storeu_si256(p + 0, res_01_shift);
- if (w - j > 8) {
- _mm256_storeu_si256(p + 1, res_23_shift);
- }
- }
+ for (i = 0; i < h; i += 2) {
+ data = &src_ptr[i * src_stride + j];
+ const __m256i src_67a = _mm256_permute2x128_si256(
+ src6, _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+ 0x20);
+
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 8 * src_stride)));
+ const __m256i src_78a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+ src6, 0x20);
+
+ s[3] = _mm256_unpacklo_epi8(src_67a, src_78a);
+ s[7] = _mm256_unpackhi_epi8(src_67a, src_78a);
+
+ const __m256i res_lo = convolve(s, coeffs);
+
+ const __m256i res_lo_0_32b =
+ _mm256_cvtepi16_epi32(_mm256_castsi256_si128(res_lo));
+ const __m256i res_lo_0_shift = _mm256_slli_epi32(res_lo_0_32b, bits);
+
+ // Accumulate values into the destination buffer
+ add_store_aligned(&dst[i * dst_stride + j], &res_lo_0_shift, &avg_mask);
+
+ const __m256i res_lo_1_32b =
+ _mm256_cvtepi16_epi32(_mm256_extracti128_si256(res_lo, 1));
+ const __m256i res_lo_1_shift = _mm256_slli_epi32(res_lo_1_32b, bits);
+
+ add_store_aligned(&dst[i * dst_stride + j + dst_stride], &res_lo_1_shift,
+ &avg_mask);
+
+ if (w - j > 8) {
+ const __m256i res_hi = convolve(s + 4, coeffs);
+
+ const __m256i res_hi_0_32b =
+ _mm256_cvtepi16_epi32(_mm256_castsi256_si128(res_hi));
+ const __m256i res_hi_0_shift = _mm256_slli_epi32(res_hi_0_32b, bits);
+
+ add_store_aligned(&dst[i * dst_stride + j + 8], &res_hi_0_shift,
+ &avg_mask);
+
+ const __m256i res_hi_1_32b =
+ _mm256_cvtepi16_epi32(_mm256_extracti128_si256(res_hi, 1));
+ const __m256i res_hi_1_shift = _mm256_slli_epi32(res_hi_1_32b, bits);
+
+ add_store_aligned(&dst[i * dst_stride + j + 8 + dst_stride],
+ &res_hi_1_shift, &avg_mask);
}
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+
+ s[4] = s[5];
+ s[5] = s[6];
+ s[6] = s[7];
}
}
}
@@ -485,122 +485,148 @@
InterpFilterParams *filter_params_y,
const int subpel_x_q4, const int subpel_y_q4,
ConvolveParams *conv_params) {
- if (w < 16) {
- av1_convolve_y_sr_sse2(src, src_stride, dst, dst_stride, w, h,
- filter_params_x, filter_params_y, subpel_x_q4,
- subpel_y_q4, conv_params);
- return;
- }
- {
- int i, j;
- const int fo_vert = filter_params_y->taps / 2 - 1;
- const uint8_t *const src_ptr = src - fo_vert * src_stride;
- const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
- const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter);
- const __m256i coeffs_y = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_y8), coeffs_y8, 1);
- // right shift is F-1 because we are already dividing
- // filter co-efficients by 2
- const int right_shift_bits = (FILTER_BITS - 1);
- const __m128i right_shift = _mm_cvtsi32_si128(right_shift_bits);
- const __m256i right_shift_const =
- _mm256_set1_epi16((1 << right_shift_bits) >> 1);
+ int i, j;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride;
- // right shift all filter co-efficients by 1 to reduce the bits required.
- // This extra right shift will be taken care of at the end while rounding
- // the result.
- // Since all filter co-efficients are even, this change will not affect the
- // end result
- const __m256i coeffs_y_1 = _mm256_srai_epi16(coeffs_y, 1);
+ // right shift is F-1 because we are already dividing
+ // filter co-efficients by 2
+ const int right_shift_bits = (FILTER_BITS - 1);
+ const __m128i right_shift = _mm_cvtsi32_si128(right_shift_bits);
+ const __m256i right_shift_const =
+ _mm256_set1_epi16((1 << right_shift_bits) >> 1);
+ __m256i coeffs[4], s[8];
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0200u));
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0604u));
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0a08u));
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 =
- _mm256_shuffle_epi8(coeffs_y_1, _mm256_set1_epi16(0x0e0cu));
+ prepare_coeffs(filter_params_y, subpel_y_q4, coeffs);
- const __m256i shuf = _mm256_load_si256((__m256i const *)g_shuf1);
+ (void)filter_params_x;
+ (void)subpel_x_q4;
+ (void)conv_params;
- (void)filter_params_x;
- (void)subpel_x_q4;
+ for (j = 0; j < w; j += 16) {
+ const uint8_t *data = &src_ptr[j];
+ __m256i src6;
- for (i = 0; i < h; ++i) {
- for (j = 0; j < w; j += 16) {
- const uint8_t *data = &src_ptr[i * src_stride + j];
- // Load lines a and b. Line a to lower 128, line b to upper 128
- const __m256i src_01a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
- 0x20);
- const __m256i src_23a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
- 0x20);
- const __m256i src_45a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
- 0x20);
- const __m256i src_67a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 6 * src_stride))),
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
- 0x20);
+ // Load lines a and b. Line a to lower 128, line b to upper 128
+ const __m256i src_01a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+ 0x20);
- // Permute across lanes. (a_lo a_hi b_lo b_hi -> a_lo b_lo a_hi b_hi)
- const __m256i src_01b = _mm256_permute4x64_epi64(src_01a, 0xd8);
- const __m256i src_23b = _mm256_permute4x64_epi64(src_23a, 0xd8);
- const __m256i src_45b = _mm256_permute4x64_epi64(src_45a, 0xd8);
- const __m256i src_67b = _mm256_permute4x64_epi64(src_67a, 0xd8);
- // Interleave a and b within lanes.
- const __m256i src_01 = _mm256_shuffle_epi8(src_01b, shuf);
- const __m256i src_23 = _mm256_shuffle_epi8(src_23b, shuf);
- const __m256i src_45 = _mm256_shuffle_epi8(src_45b, shuf);
- const __m256i src_67 = _mm256_shuffle_epi8(src_67b, shuf);
+ const __m256i src_12a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+ 0x20);
- // Filter source pixels
- const __m256i res_01 = _mm256_maddubs_epi16(src_01, coeff_01);
- const __m256i res_23 = _mm256_maddubs_epi16(src_23, coeff_23);
- const __m256i res_45 = _mm256_maddubs_epi16(src_45, coeff_45);
- const __m256i res_67 = _mm256_maddubs_epi16(src_67, coeff_67);
+ const __m256i src_23a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+ 0x20);
- // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
- __m256i res_16b = _mm256_add_epi16(_mm256_add_epi16(res_01, res_45),
- _mm256_add_epi16(res_23, res_67));
+ const __m256i src_34a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+ 0x20);
+
+ const __m256i src_45a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+ 0x20);
+
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 6 * src_stride)));
+ const __m256i src_56a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+ src6, 0x20);
+
+ s[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
+ s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
+ s[2] = _mm256_unpacklo_epi8(src_45a, src_56a);
+
+ s[4] = _mm256_unpackhi_epi8(src_01a, src_12a);
+ s[5] = _mm256_unpackhi_epi8(src_23a, src_34a);
+ s[6] = _mm256_unpackhi_epi8(src_45a, src_56a);
+
+ for (i = 0; i < h; i += 2) {
+ data = &src_ptr[i * src_stride + j];
+ const __m256i src_67a = _mm256_permute2x128_si256(
+ src6, _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+ 0x20);
+
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 8 * src_stride)));
+ const __m256i src_78a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+ src6, 0x20);
+
+ s[3] = _mm256_unpacklo_epi8(src_67a, src_78a);
+ s[7] = _mm256_unpackhi_epi8(src_67a, src_78a);
+
+ const __m256i res_lo = convolve(s, coeffs);
+
+ /* rounding code */
+ // shift by F - 1
+ const __m256i res_16b_lo = _mm256_sra_epi16(
+ _mm256_add_epi16(res_lo, right_shift_const), right_shift);
+ // 8 bit conversion and saturation to uint8
+ __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
+
+ if (w - j > 8) {
+ const __m256i res_hi = convolve(s + 4, coeffs);
/* rounding code */
// shift by F - 1
- __m256i res_16b_shift = _mm256_sra_epi16(
- _mm256_add_epi16(res_16b, right_shift_const), right_shift);
+ const __m256i res_16b_hi = _mm256_sra_epi16(
+ _mm256_add_epi16(res_hi, right_shift_const), right_shift);
// 8 bit conversion and saturation to uint8
- __m256i res_8b = _mm256_packus_epi16(res_16b_shift, res_16b_shift);
- res_8b = _mm256_permute4x64_epi64(res_8b, 216);
- // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
- __m128i res = _mm256_castsi256_si128(res_8b);
+ __m256i res_8b_hi = _mm256_packus_epi16(res_16b_hi, res_16b_hi);
- // Store values into the destination buffer
- if (w - j > 8) {
- _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
- } else if (w - j > 4) {
- _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res);
+ __m256i res_a = _mm256_unpacklo_epi64(res_8b_lo, res_8b_hi);
+
+ const __m128i res_0 = _mm256_castsi256_si128(res_a);
+ const __m128i res_1 = _mm256_extracti128_si256(res_a, 1);
+
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_0);
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ res_1);
+ } else {
+ const __m128i res_0 = _mm256_castsi256_si128(res_8b_lo);
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b_lo, 1);
+ if (w - j > 4) {
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_0);
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ res_1);
+ } else if (w - j > 2) {
+ xx_storel_32(&dst[i * dst_stride + j], res_0);
+ xx_storel_32(&dst[i * dst_stride + j + dst_stride], res_1);
} else {
- xx_storel_32(&dst[i * dst_stride + j], res);
+ __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
+ __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];
+ *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
}
}
+
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+
+ s[4] = s[5];
+ s[5] = s[6];
+ s[6] = s[7];
}
}
}
@@ -615,44 +641,18 @@
int dst_stride = conv_params->dst_stride;
int i, j;
const int fo_horiz = filter_params_x->taps / 2 - 1;
- const int do_average = conv_params->do_average;
const uint8_t *const src_ptr = src - fo_horiz;
const int bits = FILTER_BITS - conv_params->round_1;
+ const __m256i avg_mask = _mm256_set1_epi32(conv_params->do_average ? -1 : 0);
- __m256i filt[4], s[4];
+ __m256i filt[4], coeffs[4];
- filt[0] = _mm256_loadu_si256((__m256i const *)filt1_global_avx2);
- filt[1] = _mm256_loadu_si256((__m256i const *)filt2_global_avx2);
- filt[2] = _mm256_loadu_si256((__m256i const *)filt3_global_avx2);
- filt[3] = _mm256_loadu_si256((__m256i const *)filt4_global_avx2);
+ filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+ filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+ filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
- const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
-
- const __m128i coeffs_x8 = _mm_loadu_si128((__m128i *)x_filter);
- // since not all compilers yet support _mm256_set_m128i()
- const __m256i coeffs_x =
- _mm256_insertf128_si256(_mm256_castsi128_si256(coeffs_x8), coeffs_x8, 1);
-
- // right shift all filter co-efficients by 1 to reduce the bits required.
- // This extra right shift will be taken care of at the end while rounding the
- // result.
- // Since all filter co-efficients are even, this change will not affect the
- // end result
- const __m256i coeffs_x_1 = _mm256_srai_epi16(coeffs_x, 1);
-
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0200u));
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0604u));
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0a08u));
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0e0cu));
+ prepare_coeffs(filter_params_x, subpel_x_q4, coeffs);
const __m256i round_const =
_mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1);
@@ -671,21 +671,8 @@
_mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]),
_MM_SHUFFLE(2, 1, 1, 0));
- // filter the source buffer
- s[0] = _mm256_shuffle_epi8(data, filt[0]);
- s[1] = _mm256_shuffle_epi8(data, filt[1]);
- s[2] = _mm256_shuffle_epi8(data, filt[2]);
- s[3] = _mm256_shuffle_epi8(data, filt[3]);
+ __m256i res = convolve_x(data, coeffs, filt);
- const __m256i res_0 = _mm256_maddubs_epi16(s[0], coeff_01);
- const __m256i res_1 = _mm256_maddubs_epi16(s[1], coeff_23);
- const __m256i res_2 = _mm256_maddubs_epi16(s[2], coeff_45);
- const __m256i res_3 = _mm256_maddubs_epi16(s[3], coeff_67);
-
- const __m256i res_a = _mm256_add_epi16(res_0, res_2);
- const __m256i res_b = _mm256_add_epi16(res_1, res_3);
-
- __m256i res = _mm256_add_epi16(res_a, res_b);
res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift);
const __m256i res_lo_round =
@@ -697,21 +684,10 @@
const __m256i res_hi_shift = _mm256_slli_epi32(res_hi_round, bits);
// Accumulate values into the destination buffer
- __m256i *const p = (__m256i *)&dst[i * dst_stride + j];
- if (do_average) {
- const __m256i dst_lo = _mm256_loadu_si256(p + 0);
- const __m256i dst_hi = _mm256_loadu_si256(p + 1);
- const __m256i res_lo = _mm256_add_epi32(dst_lo, res_lo_shift);
- const __m256i res_hi = _mm256_add_epi32(dst_hi, res_hi_shift);
- _mm256_storeu_si256(p + 0, res_lo);
- if (w - j > 8) {
- _mm256_storeu_si256(p + 1, res_hi);
- }
- } else {
- _mm256_storeu_si256(p + 0, res_lo_shift);
- if (w - j > 8) {
- _mm256_storeu_si256(p + 1, res_hi_shift);
- }
+ add_store_aligned(&dst[i * dst_stride + j], &res_lo_shift, &avg_mask);
+ if (w - j > 8) {
+ add_store_aligned(&dst[i * dst_stride + j + 8], &res_hi_shift,
+ &avg_mask);
}
}
}
@@ -723,102 +699,60 @@
InterpFilterParams *filter_params_y,
const int subpel_x_q4, const int subpel_y_q4,
ConvolveParams *conv_params) {
- if (w < 4) {
- av1_convolve_x_sr_sse2(src, src_stride, dst, dst_stride, w, h,
- filter_params_x, filter_params_y, subpel_x_q4,
- subpel_y_q4, conv_params);
- return;
- }
- {
- int i, j;
- const int fo_horiz = filter_params_x->taps / 2 - 1;
- const uint8_t *const src_ptr = src - fo_horiz;
+ int i, j;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_horiz;
- __m256i filt[4], s[4];
+ __m256i filt[4], coeffs[4];
- filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
- filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
- filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
- filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+ filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+ filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+ filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
- const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
- *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+ prepare_coeffs(filter_params_x, subpel_x_q4, coeffs);
- const __m128i coeffs_x8 = _mm_loadu_si128((__m128i *)x_filter);
- // since not all compilers yet support _mm256_set_m128i()
- const __m256i coeffs_x = _mm256_insertf128_si256(
- _mm256_castsi128_si256(coeffs_x8), coeffs_x8, 1);
+ const __m256i round_const =
+ _mm256_set1_epi16(((1 << (conv_params->round_0 - 1)) >> 1) +
+ ((1 << (FILTER_BITS - 1)) >> 1));
+ const __m128i round_shift = _mm_cvtsi32_si128(FILTER_BITS - 1);
- // right shift all filter co-efficients by 1 to reduce the bits required.
- // This extra right shift will be taken care of at the end while rounding
- // the result.
- // Since all filter co-efficients are even, this change will not affect the
- // end result
- const __m256i coeffs_x_1 = _mm256_srai_epi16(coeffs_x, 1);
+ (void)filter_params_y;
+ (void)subpel_y_q4;
- // coeffs 0 1 0 1 0 1 0 1
- const __m256i coeff_01 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0200u));
- // coeffs 2 3 2 3 2 3 2 3
- const __m256i coeff_23 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0604u));
- // coeffs 4 5 4 5 4 5 4 5
- const __m256i coeff_45 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0a08u));
- // coeffs 6 7 6 7 6 7 6 7
- const __m256i coeff_67 =
- _mm256_shuffle_epi8(coeffs_x_1, _mm256_set1_epi16(0x0e0cu));
+ for (i = 0; i < h; ++i) {
+ for (j = 0; j < w; j += 16) {
+ // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 8 9 10 11 12 13 14 15 16 17 18
+ // 19 20 21 22 23
+ const __m256i data = _mm256_inserti128_si256(
+ _mm256_loadu_si256((__m256i *)&src_ptr[(i * src_stride) + j]),
+ _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + (j + 8)]), 1);
- const __m256i round_const =
- _mm256_set1_epi16(((1 << (conv_params->round_0 - 1)) >> 1) +
- ((1 << (FILTER_BITS - 1)) >> 1));
- const __m128i round_shift = _mm_cvtsi32_si128(FILTER_BITS - 1);
+ __m256i res_16b = convolve_x(data, coeffs, filt);
- (void)filter_params_y;
- (void)subpel_y_q4;
+ // Combine V round and 2F-H-V round into a single rounding
+ res_16b =
+ _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const), round_shift);
- for (i = 0; i < h; ++i) {
- for (j = 0; j < w; j += 16) {
- // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 8 9 10 11 12 13 14 15 16 17 18
- // 19 20 21 22 23
- const __m256i data = _mm256_inserti128_si256(
- _mm256_loadu_si256((__m256i *)&src_ptr[(i * src_stride) + j]),
- _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + (j + 8)]),
- 1);
+ /* rounding code */
+ // 8 bit conversion and saturation to uint8
+ __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
- // filter the source buffer
- s[0] = _mm256_shuffle_epi8(data, filt[0]);
- s[1] = _mm256_shuffle_epi8(data, filt[1]);
- s[2] = _mm256_shuffle_epi8(data, filt[2]);
- s[3] = _mm256_shuffle_epi8(data, filt[3]);
-
- const __m256i res_0 = _mm256_maddubs_epi16(s[0], coeff_01);
- const __m256i res_1 = _mm256_maddubs_epi16(s[1], coeff_23);
- const __m256i res_2 = _mm256_maddubs_epi16(s[2], coeff_45);
- const __m256i res_3 = _mm256_maddubs_epi16(s[3], coeff_67);
-
- const __m256i res_a = _mm256_add_epi16(res_0, res_2);
- const __m256i res_b = _mm256_add_epi16(res_1, res_3);
-
- __m256i res_16b = _mm256_add_epi16(res_a, res_b);
- // Combine V round and 2F-H-V round into a single rounding
- res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const),
- round_shift);
-
- /* rounding code */
- // 8 bit conversion and saturation to uint8
- __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
- res_8b = _mm256_permute4x64_epi64(res_8b, 216);
+ // Store values into the destination buffer
+ if (w - j > 8) {
// 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+ res_8b = _mm256_permute4x64_epi64(res_8b, 216);
__m128i res = _mm256_castsi256_si128(res_8b);
-
- // Store values into the destination buffer
- if (w - j > 8) {
- _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
- } else if (w - j > 4) {
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+ } else {
+ __m128i res = _mm256_castsi256_si128(res_8b);
+ if (w - j > 4) {
_mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res);
- } else {
+ } else if (w - j > 2) {
xx_storel_32(&dst[i * dst_stride + j], res);
+ } else {
+ __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
+ *(uint16_t *)p = _mm_cvtsi128_si32(res);
}
}
}
