Optimize av1_convolve_sr functions
Introduced 4 tap filter usage for convolve_sr functions.
When tested for multiple test cases observed 0.3%
average reduction in encoder time averaged for presets 3 and 4.
Approximately on average 20% improvement is seen at
unit test level.
Change-Id: If507760fce2407150583f12d3325bfde4f133d84
diff --git a/aom_dsp/x86/convolve_avx2.h b/aom_dsp/x86/convolve_avx2.h
index 644f3dd..e4873fc 100644
--- a/aom_dsp/x86/convolve_avx2.h
+++ b/aom_dsp/x86/convolve_avx2.h
@@ -34,6 +34,98 @@
2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8,
};
+#define CONVOLVE_SR_HORIZONTAL_FILTER_8TAP \
+ for (i = 0; i < (im_h - 2); i += 2) { \
+ __m256i data = _mm256_castsi128_si256( \
+ _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j])); \
+ data = _mm256_inserti128_si256( \
+ data, \
+ _mm_loadu_si128( \
+ (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]), \
+ 1); \
+ \
+ __m256i res = convolve_lowbd_x(data, coeffs_h, filt); \
+ res = \
+ _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+ _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); \
+ } \
+ \
+ __m256i data_1 = _mm256_castsi128_si256( \
+ _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j])); \
+ \
+ __m256i res = convolve_lowbd_x(data_1, coeffs_h, filt); \
+ \
+ res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+ \
+ _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+
+#define CONVOLVE_SR_VERTICAL_FILTER_8TAP \
+ __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)); \
+ \
+ __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); \
+ \
+ 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); \
+ \
+ for (i = 0; i < h; i += 2) { \
+ const int16_t *data = &im_block[i * im_stride]; \
+ \
+ const __m256i s6 = _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); \
+ const __m256i s7 = _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); \
+ \
+ s[3] = _mm256_unpacklo_epi16(s6, s7); \
+ s[7] = _mm256_unpackhi_epi16(s6, s7); \
+ \
+ __m256i res_a = convolve(s, coeffs_v); \
+ __m256i res_b = convolve(s + 4, coeffs_v); \
+ \
+ res_a = \
+ _mm256_sra_epi32(_mm256_add_epi32(res_a, sum_round_v), sum_shift_v); \
+ res_b = \
+ _mm256_sra_epi32(_mm256_add_epi32(res_b, sum_round_v), sum_shift_v); \
+ \
+ 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); \
+ \
+ const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); \
+ 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); \
+ \
+ __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(p_0, res_0); \
+ xx_storel_32(p_1, res_1); \
+ } else { \
+ *(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]; \
+ }
+
static INLINE void prepare_coeffs_lowbd(
const InterpFilterParams *const filter_params, const int subpel_q4,
__m256i *const coeffs /* [4] */) {
@@ -95,6 +187,17 @@
return res;
}
+static INLINE __m256i convolve_lowbd_4tap(const __m256i *const s,
+ const __m256i *const coeffs) {
+ const __m256i res_23 = _mm256_maddubs_epi16(s[0], coeffs[0]);
+ const __m256i res_45 = _mm256_maddubs_epi16(s[1], coeffs[1]);
+
+ // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+ const __m256i res = _mm256_add_epi16(res_45, res_23);
+
+ return res;
+}
+
static INLINE __m256i convolve(const __m256i *const s,
const __m256i *const coeffs) {
const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]);
@@ -130,6 +233,17 @@
return convolve_lowbd(s, coeffs);
}
+static INLINE __m256i convolve_lowbd_x_4tap(const __m256i data,
+ const __m256i *const coeffs,
+ const __m256i *const filt) {
+ __m256i s[2];
+
+ s[0] = _mm256_shuffle_epi8(data, filt[0]);
+ s[1] = _mm256_shuffle_epi8(data, filt[1]);
+
+ return convolve_lowbd_4tap(s, coeffs);
+}
+
static INLINE void add_store_aligned_256(CONV_BUF_TYPE *const dst,
const __m256i *const res,
const int do_average) {
diff --git a/av1/common/x86/convolve_2d_avx2.c b/av1/common/x86/convolve_2d_avx2.c
index 5a69ef6..ae12a60 100644
--- a/av1/common/x86/convolve_2d_avx2.c
+++ b/av1/common/x86/convolve_2d_avx2.c
@@ -27,31 +27,15 @@
const int subpel_x_q4, const int subpel_y_q4,
ConvolveParams *conv_params) {
const int bd = 8;
-
- 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;
- 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;
-
+ int i, is_horiz_4tap = 0, is_vert_4tap = 0;
+ DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
const int bits =
FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
- __m256i filt[4], coeffs_h[4], coeffs_v[4];
-
assert(conv_params->round_0 > 0);
- filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2);
- filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
- filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
- filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
-
- prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_h);
- prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_v);
-
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);
@@ -65,58 +49,96 @@
((1 << (offset_bits - conv_params->round_1)) >> 1));
const __m128i round_shift_v = _mm_cvtsi32_si128(bits);
- 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]));
+ __m256i filt[4], coeffs_h[4], coeffs_v[4];
- // Load the next line
- if (i + 1 < im_h)
+ filt[0] = _mm256_load_si256((__m256i const *)(filt_global_avx2));
+ filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+
+ prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_h);
+ prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_v);
+
+ // Condition for checking valid horz_filt taps
+ if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs_h[0], coeffs_h[3]), 0)))
+ is_horiz_4tap = 1;
+
+ // Condition for checking valid vert_filt taps
+ if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs_v[0], coeffs_v[3]), 0)))
+ is_vert_4tap = 1;
+
+ // horz_filt as 4 tap and vert_filt as 8 tap
+ if (is_horiz_4tap) {
+ int im_h = h + filter_params_y->taps - 1;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const int fo_horiz = 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+ // horz-filter
+ for (int j = 0; j < w; j += 8) {
+ for (i = 0; i < (im_h - 2); i += 2) {
+ __m256i data = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
+
+ // Load the next line
data = _mm256_inserti128_si256(
data,
_mm_loadu_si128(
(__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),
1);
+ __m256i res = convolve_lowbd_x_4tap(data, coeffs_h + 1, filt);
- __m256i res = convolve_lowbd_x(data, coeffs_h, filt);
+ res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h),
+ round_shift_h);
+ _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+ }
+ __m256i data_1 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
+
+ __m256i res = convolve_lowbd_x_4tap(data_1, coeffs_h + 1, filt);
res =
_mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h);
-
_mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
- }
- /* Vertical filter */
- {
+ // vert filter
+ CONVOLVE_SR_VERTICAL_FILTER_8TAP;
+ }
+ } else if (is_vert_4tap) {
+ int im_h = h + 3;
+ const int fo_vert = 1;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+ filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+ filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+ for (int j = 0; j < w; j += 8) {
+ // horz_filter
+ CONVOLVE_SR_HORIZONTAL_FILTER_8TAP;
+ // vert_filter
+ __m256i s[6];
__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));
- __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);
-
- 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);
+ s[3] = _mm256_unpackhi_epi16(src_0, src_1);
+ s[4] = _mm256_unpackhi_epi16(src_2, src_3);
for (i = 0; i < h; i += 2) {
const int16_t *data = &im_block[i * im_stride];
- const __m256i s6 =
- _mm256_loadu_si256((__m256i *)(data + 6 * im_stride));
- const __m256i s7 =
- _mm256_loadu_si256((__m256i *)(data + 7 * im_stride));
+ const __m256i s4 =
+ _mm256_loadu_si256((__m256i *)(data + 4 * im_stride));
+ const __m256i s5 =
+ _mm256_loadu_si256((__m256i *)(data + 5 * im_stride));
- s[3] = _mm256_unpacklo_epi16(s6, s7);
- s[7] = _mm256_unpackhi_epi16(s6, s7);
+ s[2] = _mm256_unpacklo_epi16(s4, s5);
+ s[5] = _mm256_unpackhi_epi16(s4, s5);
- __m256i res_a = convolve(s, coeffs_v);
- __m256i res_b = convolve(s + 4, coeffs_v);
+ __m256i res_a = convolve_4tap(s, coeffs_v + 1);
+ __m256i res_b = convolve_4tap(s + 3, coeffs_v + 1);
// Combine V round and 2F-H-V round into a single rounding
res_a =
@@ -154,13 +176,25 @@
s[0] = s[1];
s[1] = s[2];
- s[2] = s[3];
-
+ s[3] = s[4];
s[4] = s[5];
- s[5] = s[6];
- s[6] = s[7];
}
}
+ } else {
+ int j;
+ int im_h = h + filter_params_y->taps - 1;
+ 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;
+
+ filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+ filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+ for (j = 0; j < w; j += 8) {
+ CONVOLVE_SR_HORIZONTAL_FILTER_8TAP;
+
+ CONVOLVE_SR_VERTICAL_FILTER_8TAP;
+ }
}
}
diff --git a/av1/common/x86/convolve_avx2.c b/av1/common/x86/convolve_avx2.c
index 0e91ea9..21b9fe4 100644
--- a/av1/common/x86/convolve_avx2.c
+++ b/av1/common/x86/convolve_avx2.c
@@ -23,153 +23,239 @@
const InterpFilterParams *filter_params_y,
const int subpel_x_q4, const int subpel_y_q4,
ConvolveParams *conv_params) {
- int i, j;
- const int fo_vert = filter_params_y->taps / 2 - 1;
- const uint8_t *const src_ptr = src - fo_vert * src_stride;
-
+ int i, j, is_vert_4tap = 0;
// 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];
assert(conv_params->round_0 <= FILTER_BITS);
assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
- prepare_coeffs_lowbd(filter_params_y, subpel_y_q4, coeffs);
-
(void)filter_params_x;
(void)subpel_x_q4;
(void)conv_params;
+ __m256i coeffs[4], s[8];
+ __m128i d[6];
- for (j = 0; j < w; j += 16) {
- const uint8_t *data = &src_ptr[j];
- __m256i src6;
+ prepare_coeffs_lowbd(filter_params_y, subpel_y_q4, coeffs);
- // 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);
+ // Condition for checking valid vert_filt taps
+ if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs[0], coeffs[3]), 0)))
+ is_vert_4tap = 1;
- 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);
+ // vert_filt as 4 tap
+ if (is_vert_4tap) {
+ const int fo_vert = 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride;
+ for (j = 0; j < w; j += 16) {
+ const uint8_t *data = &src_ptr[j];
+ d[0] = _mm_loadu_si128((__m128i *)(data + 0 * src_stride));
+ d[1] = _mm_loadu_si128((__m128i *)(data + 1 * src_stride));
+ d[2] = _mm_loadu_si128((__m128i *)(data + 2 * src_stride));
+ d[3] = _mm_loadu_si128((__m128i *)(data + 3 * src_stride));
+ d[4] = _mm_loadu_si128((__m128i *)(data + 4 * src_stride));
- 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);
+ // 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(d[0]), _mm256_castsi128_si256(d[1]), 0x20);
- 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_12a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[1]), _mm256_castsi128_si256(d[2]), 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_23a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[2]), _mm256_castsi128_si256(d[3]), 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);
+ const __m256i src_34a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[3]), _mm256_castsi128_si256(d[4]), 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[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
+ s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
- 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);
+ s[3] = _mm256_unpackhi_epi8(src_01a, src_12a);
+ s[4] = _mm256_unpackhi_epi8(src_23a, src_34a);
- 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);
+ for (i = 0; i < h; i += 2) {
+ data = &src_ptr[i * src_stride + j];
+ d[5] = _mm_loadu_si128((__m128i *)(data + 5 * src_stride));
+ const __m256i src_45a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[4]), _mm256_castsi128_si256(d[5]), 0x20);
+
+ d[4] = _mm_loadu_si128((__m128i *)(data + 6 * src_stride));
+ const __m256i src_56a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[5]), _mm256_castsi128_si256(d[4]), 0x20);
+
+ s[2] = _mm256_unpacklo_epi8(src_45a, src_56a);
+ s[5] = _mm256_unpackhi_epi8(src_45a, src_56a);
+
+ const __m256i res_lo = convolve_lowbd_4tap(s, coeffs + 1);
+ /* 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_lowbd_4tap(s + 3, coeffs + 1);
+
+ /* rounding code */
+ // shift by F - 1
+ 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_hi = _mm256_packus_epi16(res_16b_hi, res_16b_hi);
+
+ __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 {
+ __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[3] = s[4];
+ s[4] = s[5];
+ }
+ }
+ } else {
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride;
+
+ for (j = 0; j < w; j += 16) {
+ const uint8_t *data = &src_ptr[j];
+ __m256i src6;
+
+ d[0] = _mm_loadu_si128((__m128i *)(data + 0 * src_stride));
+ d[1] = _mm_loadu_si128((__m128i *)(data + 1 * src_stride));
+ d[2] = _mm_loadu_si128((__m128i *)(data + 2 * src_stride));
+ d[3] = _mm_loadu_si128((__m128i *)(data + 3 * src_stride));
+ d[4] = _mm_loadu_si128((__m128i *)(data + 4 * src_stride));
+ d[5] = _mm_loadu_si128((__m128i *)(data + 5 * src_stride));
+ // 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(d[0]), _mm256_castsi128_si256(d[1]), 0x20);
+
+ const __m256i src_12a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[1]), _mm256_castsi128_si256(d[2]), 0x20);
+
+ const __m256i src_23a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[2]), _mm256_castsi128_si256(d[3]), 0x20);
+
+ const __m256i src_34a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[3]), _mm256_castsi128_si256(d[4]), 0x20);
+
+ const __m256i src_45a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[4]), _mm256_castsi128_si256(d[5]), 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);
+ _mm_loadu_si128((__m128i *)(data + 6 * src_stride)));
+ const __m256i src_56a =
+ _mm256_permute2x128_si256(_mm256_castsi128_si256(d[5]), src6, 0x20);
- s[3] = _mm256_unpacklo_epi8(src_67a, src_78a);
- s[7] = _mm256_unpackhi_epi8(src_67a, src_78a);
+ 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_lo = convolve_lowbd(s, coeffs);
+ 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);
- /* 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);
+ 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);
- if (w - j > 8) {
- const __m256i res_hi = convolve_lowbd(s + 4, coeffs);
+ 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_lowbd(s, coeffs);
/* rounding code */
// shift by F - 1
- const __m256i res_16b_hi = _mm256_sra_epi16(
- _mm256_add_epi16(res_hi, right_shift_const), right_shift);
+ 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_hi = _mm256_packus_epi16(res_16b_hi, res_16b_hi);
+ __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
- __m256i res_a = _mm256_unpacklo_epi64(res_8b_lo, res_8b_hi);
+ if (w - j > 8) {
+ const __m256i res_hi = convolve_lowbd(s + 4, coeffs);
- const __m128i res_0 = _mm256_castsi256_si128(res_a);
- const __m128i res_1 = _mm256_extracti128_si256(res_a, 1);
+ /* rounding code */
+ // shift by F - 1
+ 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_hi = _mm256_packus_epi16(res_16b_hi, res_16b_hi);
- _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],
+ __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 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 {
- __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);
+ 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 {
+ __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];
}
-
- s[0] = s[1];
- s[1] = s[2];
- s[2] = s[3];
-
- s[4] = s[5];
- s[5] = s[6];
- s[6] = s[7];
}
}
}
@@ -180,26 +266,14 @@
const InterpFilterParams *filter_params_y,
const int subpel_x_q4, const int subpel_y_q4,
ConvolveParams *conv_params) {
- int i, j;
- const int fo_horiz = filter_params_x->taps / 2 - 1;
- const uint8_t *const src_ptr = src - fo_horiz;
const int bits = FILTER_BITS - conv_params->round_0;
- __m256i filt[4], coeffs[4];
-
- filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2);
- filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
- filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
- filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
-
- prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs);
-
const __m256i round_0_const =
_mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1);
const __m128i round_0_shift = _mm_cvtsi32_si128(conv_params->round_0 - 1);
const __m256i round_const = _mm256_set1_epi16((1 << bits) >> 1);
const __m128i round_shift = _mm_cvtsi32_si128(bits);
-
+ int i, is_horiz_4tap = 0;
(void)filter_params_y;
(void)subpel_y_q4;
@@ -208,51 +282,101 @@
((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
assert(conv_params->round_0 > 0);
- if (w <= 8) {
- for (i = 0; i < h; i += 2) {
- const __m256i data = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride]))),
- _mm256_castsi128_si256(_mm_loadu_si128(
- (__m128i *)(&src_ptr[i * src_stride + src_stride]))),
- 0x20);
+ __m256i coeffs[4], filt[4];
+ filt[0] = _mm256_load_si256((__m256i const *)(filt_global_avx2));
+ filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
- __m256i res_16b = convolve_lowbd_x(data, coeffs, filt);
+ prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs);
- res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
- round_0_shift);
+ // Condition for checking valid horz_filt taps
+ if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs[0], coeffs[3]), 0)))
+ is_horiz_4tap = 1;
- res_16b =
- _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const), round_shift);
+ // horz_filt as 4 tap
+ if (is_horiz_4tap) {
+ const int fo_horiz = 1;
+ const uint8_t *const src_ptr = src - fo_horiz;
+ if (w <= 8) {
+ for (i = 0; i < h; i += 2) {
+ const __m256i data = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride]))),
+ _mm256_castsi128_si256(_mm_loadu_si128(
+ (__m128i *)(&src_ptr[i * src_stride + src_stride]))),
+ 0x20);
- /* rounding code */
- // 8 bit conversion and saturation to uint8
- __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
+ __m256i res_16b = convolve_lowbd_x_4tap(data, coeffs + 1, filt);
- const __m128i res_0 = _mm256_castsi256_si128(res_8b);
- const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
- if (w > 4) {
- _mm_storel_epi64((__m128i *)&dst[i * dst_stride], res_0);
- _mm_storel_epi64((__m128i *)&dst[i * dst_stride + dst_stride], res_1);
- } else if (w > 2) {
- xx_storel_32(&dst[i * dst_stride], res_0);
- xx_storel_32(&dst[i * dst_stride + dst_stride], res_1);
- } else {
- __m128i *const p_0 = (__m128i *)&dst[i * dst_stride];
- __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + dst_stride];
- *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
- *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
+ res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
+ round_0_shift);
+
+ 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);
+
+ const __m128i res_0 = _mm256_castsi256_si128(res_8b);
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
+
+ if (w > 4) {
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride], res_0);
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + dst_stride], res_1);
+ } else if (w > 2) {
+ xx_storel_32(&dst[i * dst_stride], res_0);
+ xx_storel_32(&dst[i * dst_stride + dst_stride], res_1);
+ } else {
+ __m128i *const p_0 = (__m128i *)&dst[i * dst_stride];
+ __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + dst_stride];
+ *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
+ }
+ }
+ } else {
+ for (i = 0; i < h; ++i) {
+ for (int 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);
+
+ __m256i res_16b = convolve_lowbd_x_4tap(data, coeffs + 1, filt);
+
+ res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
+ round_0_shift);
+
+ 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);
+
+ // Store values into the destination buffer
+ // 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);
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+ }
}
}
} else {
- 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 int fo_horiz = filter_params_x->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_horiz;
+ filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+ filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+ if (w <= 8) {
+ for (i = 0; i < h; i += 2) {
+ const __m256i data = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride]))),
+ _mm256_castsi128_si256(_mm_loadu_si128(
+ (__m128i *)(&src_ptr[i * src_stride + src_stride]))),
+ 0x20);
__m256i res_16b = convolve_lowbd_x(data, coeffs, filt);
@@ -266,11 +390,49 @@
// 8 bit conversion and saturation to uint8
__m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
- // Store values into the destination buffer
- // 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);
- _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+ const __m128i res_0 = _mm256_castsi256_si128(res_8b);
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
+ if (w > 4) {
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride], res_0);
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + dst_stride], res_1);
+ } else if (w > 2) {
+ xx_storel_32(&dst[i * dst_stride], res_0);
+ xx_storel_32(&dst[i * dst_stride + dst_stride], res_1);
+ } else {
+ __m128i *const p_0 = (__m128i *)&dst[i * dst_stride];
+ __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + dst_stride];
+ *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
+ }
+ }
+ } else {
+ for (i = 0; i < h; ++i) {
+ for (int 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);
+
+ __m256i res_16b = convolve_lowbd_x(data, coeffs, filt);
+
+ res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
+ round_0_shift);
+
+ 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);
+
+ // Store values into the destination buffer
+ // 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);
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+ }
}
}
}
