AVX2 code for Low bit-depth 12-tap filter
Functions:
av1_convolve_y_sr_avx2(): SSE2 = 11.36x AVX2 = 20.53x
av1_convolve_x_sr_avx2(): SSE2 = 8.42x AVX2 = 17.88x
Change-Id: Icf5fa818c469ffe45b240e0bc603ba2b495fb0e9
diff --git a/aom_dsp/x86/convolve_avx2.h b/aom_dsp/x86/convolve_avx2.h
index 09a7e94..8924033 100644
--- a/aom_dsp/x86/convolve_avx2.h
+++ b/aom_dsp/x86/convolve_avx2.h
@@ -517,6 +517,30 @@
coeffs[3] = _mm256_shuffle_epi32(coeff, 0xff);
}
+static INLINE void prepare_coeffs_12taps(
+ const InterpFilterParams *const filter_params, const int subpel_q4,
+ __m256i *const coeffs /* [4] */) {
+ const int16_t *filter = av1_get_interp_filter_subpel_kernel(
+ filter_params, subpel_q4 & SUBPEL_MASK);
+
+ __m128i coeff_8 = _mm_loadu_si128((__m128i *)filter);
+ __m256i coeff = _mm256_broadcastsi128_si256(coeff_8);
+
+ // coeffs 0 1 0 1 0 1 0 1
+ coeffs[0] = _mm256_shuffle_epi32(coeff, 0x00);
+ // coeffs 2 3 2 3 2 3 2 3
+ coeffs[1] = _mm256_shuffle_epi32(coeff, 0x55);
+ // coeffs 4 5 4 5 4 5 4 5
+ coeffs[2] = _mm256_shuffle_epi32(coeff, 0xaa);
+ // coeffs 6 7 6 7 6 7 6 7
+ coeffs[3] = _mm256_shuffle_epi32(coeff, 0xff);
+ // coeffs 8 9 10 11 0 0 0 0
+ coeff_8 = _mm_loadl_epi64((__m128i *)(filter + 8));
+ coeff = _mm256_broadcastq_epi64(coeff_8);
+ coeffs[4] = _mm256_shuffle_epi32(coeff, 0x00); // coeffs 8 9 8 9 8 9 8 9
+ coeffs[5] = _mm256_shuffle_epi32(coeff, 0x55); // coeffs 10 11 10 11.. 10 11
+}
+
static INLINE __m256i convolve_lowbd(const __m256i *const s,
const __m256i *const coeffs) {
const __m256i res_01 = _mm256_maddubs_epi16(s[0], coeffs[0]);
@@ -566,6 +590,22 @@
return res;
}
+static INLINE __m256i convolve_12taps(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_4 = _mm256_madd_epi16(s[4], coeffs[4]);
+ const __m256i res_5 = _mm256_madd_epi16(s[5], coeffs[5]);
+
+ const __m256i res1 = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1),
+ _mm256_add_epi32(res_2, res_3));
+ const __m256i res = _mm256_add_epi32(_mm256_add_epi32(res_4, res_5), res1);
+
+ 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]);
diff --git a/av1/common/x86/convolve_avx2.c b/av1/common/x86/convolve_avx2.c
index 3eeae67..c7d1141 100644
--- a/av1/common/x86/convolve_avx2.c
+++ b/av1/common/x86/convolve_avx2.c
@@ -22,351 +22,497 @@
int dst_stride, int w, int h,
const InterpFilterParams *filter_params_y,
const int subpel_y_qn) {
- if (filter_params_y->taps > 8) {
- if (w < 4 || h < 4) {
- av1_convolve_y_sr_c(src, src_stride, dst, dst_stride, w, h,
- filter_params_y, subpel_y_qn);
- } else {
- av1_convolve_y_sr_12tap_sse2(src, src_stride, dst, dst_stride, w, h,
- filter_params_y, subpel_y_qn);
- }
+ int i, j, vert_tap = SUBPEL_TAPS;
+ // right shift is F-1 because we are already dividing
+ // filter co-efficients by 2
+ const int right_shift_bits = (FILTER_BITS - 1);
+ __m128i right_shift = _mm_cvtsi32_si128(right_shift_bits);
+ __m256i right_shift_const = _mm256_set1_epi16((1 << right_shift_bits) >> 1);
+
+ __m256i coeffs[6], s[12];
+ __m128i d[10];
+
+ // Condition for checking valid vert_filt taps
+ const int16_t *const filter = av1_get_interp_filter_subpel_kernel(
+ filter_params_y, subpel_y_qn & SUBPEL_MASK);
+ if (filter_params_y->taps == 12) {
+ vert_tap = 12;
+ } else if (!(filter[0] | filter[1] | filter[6] | filter[7])) {
+ vert_tap = 4;
+ } else if (!(filter[0] | filter[7])) {
+ vert_tap = 6;
+ }
+
+ if (vert_tap == 6)
+ prepare_coeffs_6t_lowbd(filter_params_y, subpel_y_qn, coeffs);
+ else if (vert_tap == 12) {
+ prepare_coeffs_12taps(filter_params_y, subpel_y_qn, coeffs);
} else {
- int i, j, vert_tap = SUBPEL_TAPS;
- // 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);
+ prepare_coeffs_lowbd(filter_params_y, subpel_y_qn, coeffs);
+ }
- __m256i coeffs[4], s[8];
- __m128i d[6];
+ // vert_filt as 4 tap
+ if (vert_tap == 4) {
+ 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));
- // Condition for checking valid vert_filt taps
- const int16_t *const filter = av1_get_interp_filter_subpel_kernel(
- filter_params_y, subpel_y_qn & SUBPEL_MASK);
- if (!(filter[0] | filter[1] | filter[6] | filter[7])) {
- vert_tap = 4;
- } else if (!(filter[0] | filter[7])) {
- vert_tap = 6;
- }
+ // 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);
- if (vert_tap == 6)
- prepare_coeffs_6t_lowbd(filter_params_y, subpel_y_qn, coeffs);
- else
- prepare_coeffs_lowbd(filter_params_y, subpel_y_qn, coeffs);
+ const __m256i src_12a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[1]), _mm256_castsi128_si256(d[2]), 0x20);
- // vert_filt as 4 tap
- if (vert_tap == 4) {
- 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(d[2]), _mm256_castsi128_si256(d[3]), 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(d[3]), _mm256_castsi128_si256(d[4]), 0x20);
- const __m256i src_12a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(d[1]), _mm256_castsi128_si256(d[2]), 0x20);
+ s[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
+ s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
- const __m256i src_23a = _mm256_permute2x128_si256(
- _mm256_castsi128_si256(d[2]), _mm256_castsi128_si256(d[3]), 0x20);
+ s[3] = _mm256_unpackhi_epi8(src_01a, src_12a);
+ s[4] = _mm256_unpackhi_epi8(src_23a, src_34a);
- 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[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];
- 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
- *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);
- }
- }
- s[0] = s[1];
- s[1] = s[2];
-
- s[3] = s[4];
- s[4] = s[5];
- }
- }
- } else if (vert_tap == 6) {
- const int fo_vert = vert_tap / 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));
- // 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);
-
- src6 = _mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(data + 4 * src_stride)));
- const __m256i src_34a =
- _mm256_permute2x128_si256(_mm256_castsi128_si256(d[3]), src6, 0x20);
-
- s[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
- s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
-
- 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_45a = _mm256_permute2x128_si256(
- src6,
- _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[2] = _mm256_unpacklo_epi8(src_45a, src_56a);
- s[5] = _mm256_unpackhi_epi8(src_45a, src_56a);
-
- const __m256i res_lo = convolve_lowbd_6tap(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_lowbd_6tap(s + 3, 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);
- // 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
- *(uint16_t *)p_1 = (uint16_t)_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));
+ for (i = 0; i < h; i += 2) {
+ data = &src_ptr[i * src_stride + j];
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 + 6 * src_stride)));
- const __m256i src_56a =
- _mm256_permute2x128_si256(_mm256_castsi128_si256(d[5]), src6, 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[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[5] = _mm256_unpackhi_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);
+ 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);
- 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_lowbd(s, coeffs);
+ 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_lo = _mm256_sra_epi16(
- _mm256_add_epi16(res_lo, 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_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
+ __m256i res_8b_hi = _mm256_packus_epi16(res_16b_hi, res_16b_hi);
- if (w - j > 8) {
- const __m256i res_hi = convolve_lowbd(s + 4, coeffs);
+ __m256i res_a = _mm256_unpacklo_epi64(res_8b_lo, res_8b_hi);
- /* 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);
+ const __m128i res_0 = _mm256_castsi256_si128(res_a);
+ const __m128i res_1 = _mm256_extracti128_si256(res_a, 1);
- __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],
+ _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 {
- 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
- *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);
- }
+ __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 = (uint16_t)_mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = (uint16_t)_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[3] = s[4];
+ s[4] = s[5];
+ }
+ }
+ } else if (vert_tap == 6) {
+ const int fo_vert = vert_tap / 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));
+ // 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);
+
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride)));
+ const __m256i src_34a =
+ _mm256_permute2x128_si256(_mm256_castsi128_si256(d[3]), src6, 0x20);
+
+ s[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
+ s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
+
+ 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_45a = _mm256_permute2x128_si256(
+ src6,
+ _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[2] = _mm256_unpacklo_epi8(src_45a, src_56a);
+ s[5] = _mm256_unpackhi_epi8(src_45a, src_56a);
+
+ const __m256i res_lo = convolve_lowbd_6tap(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_lowbd_6tap(s + 3, 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);
+ // 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);
+ }
+ }
+ s[0] = s[1];
+ s[1] = s[2];
+ s[3] = s[4];
+ s[4] = s[5];
+ }
+ }
+ } else if (vert_tap == 12) { // vert_tap == 12
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_vert * src_stride;
+ const __m256i v_zero = _mm256_setzero_si256();
+ right_shift = _mm_cvtsi32_si128(FILTER_BITS);
+ right_shift_const = _mm256_set1_epi32((1 << FILTER_BITS) >> 1);
+
+ for (j = 0; j < w; j += 8) {
+ const uint8_t *data = &src_ptr[j];
+ __m256i src10;
+
+ d[0] = _mm_loadl_epi64((__m128i *)(data + 0 * src_stride));
+ d[1] = _mm_loadl_epi64((__m128i *)(data + 1 * src_stride));
+ d[2] = _mm_loadl_epi64((__m128i *)(data + 2 * src_stride));
+ d[3] = _mm_loadl_epi64((__m128i *)(data + 3 * src_stride));
+ d[4] = _mm_loadl_epi64((__m128i *)(data + 4 * src_stride));
+ d[5] = _mm_loadl_epi64((__m128i *)(data + 5 * src_stride));
+ d[6] = _mm_loadl_epi64((__m128i *)(data + 6 * src_stride));
+ d[7] = _mm_loadl_epi64((__m128i *)(data + 7 * src_stride));
+ d[8] = _mm_loadl_epi64((__m128i *)(data + 8 * src_stride));
+ d[9] = _mm_loadl_epi64((__m128i *)(data + 9 * 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);
+
+ const __m256i src_56a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[5]), _mm256_castsi128_si256(d[6]), 0x20);
+
+ const __m256i src_67a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[6]), _mm256_castsi128_si256(d[7]), 0x20);
+
+ const __m256i src_78a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[7]), _mm256_castsi128_si256(d[8]), 0x20);
+
+ const __m256i src_89a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(d[8]), _mm256_castsi128_si256(d[9]), 0x20);
+
+ src10 = _mm256_castsi128_si256(
+ _mm_loadl_epi64((__m128i *)(data + 10 * src_stride)));
+ const __m256i src_910a =
+ _mm256_permute2x128_si256(_mm256_castsi128_si256(d[9]), src10, 0x20);
+
+ const __m256i src_01 = _mm256_unpacklo_epi8(src_01a, v_zero);
+ const __m256i src_12 = _mm256_unpacklo_epi8(src_12a, v_zero);
+ const __m256i src_23 = _mm256_unpacklo_epi8(src_23a, v_zero);
+ const __m256i src_34 = _mm256_unpacklo_epi8(src_34a, v_zero);
+ const __m256i src_45 = _mm256_unpacklo_epi8(src_45a, v_zero);
+ const __m256i src_56 = _mm256_unpacklo_epi8(src_56a, v_zero);
+ const __m256i src_67 = _mm256_unpacklo_epi8(src_67a, v_zero);
+ const __m256i src_78 = _mm256_unpacklo_epi8(src_78a, v_zero);
+ const __m256i src_89 = _mm256_unpacklo_epi8(src_89a, v_zero);
+ const __m256i src_910 = _mm256_unpacklo_epi8(src_910a, v_zero);
+
+ s[0] = _mm256_unpacklo_epi16(src_01, src_12);
+ s[1] = _mm256_unpacklo_epi16(src_23, src_34);
+ s[2] = _mm256_unpacklo_epi16(src_45, src_56);
+ s[3] = _mm256_unpacklo_epi16(src_67, src_78);
+ s[4] = _mm256_unpacklo_epi16(src_89, src_910);
+
+ s[6] = _mm256_unpackhi_epi16(src_01, src_12);
+ s[7] = _mm256_unpackhi_epi16(src_23, src_34);
+ s[8] = _mm256_unpackhi_epi16(src_45, src_56);
+ s[9] = _mm256_unpackhi_epi16(src_67, src_78);
+ s[10] = _mm256_unpackhi_epi16(src_89, src_910);
+
+ for (i = 0; i < h; i += 2) {
+ data = &src_ptr[i * src_stride + j];
+ const __m256i src_1011a = _mm256_permute2x128_si256(
+ src10,
+ _mm256_castsi128_si256(
+ _mm_loadl_epi64((__m128i *)(data + 11 * src_stride))),
+ 0x20);
+
+ src10 = _mm256_castsi128_si256(
+ _mm_loadl_epi64((__m128i *)(data + 12 * src_stride)));
+
+ const __m256i src_1112a = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadl_epi64((__m128i *)(data + 11 * src_stride))),
+ src10, 0x20);
+
+ const __m256i src_1011 = _mm256_unpacklo_epi8(src_1011a, v_zero);
+ const __m256i src_1112 = _mm256_unpacklo_epi8(src_1112a, v_zero);
+
+ s[5] = _mm256_unpacklo_epi16(src_1011, src_1112);
+ s[11] = _mm256_unpackhi_epi16(src_1011, src_1112);
+
+ const __m256i res_lo = convolve_12taps(s, coeffs);
+
+ const __m256i res_32b_lo = _mm256_sra_epi32(
+ _mm256_add_epi32(res_lo, right_shift_const), right_shift);
+ // 8 bit conversion and saturation to uint8
+ __m256i res_16b_lo = _mm256_packs_epi32(res_32b_lo, res_32b_lo);
+ __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
+
+ if (w - j > 4) {
+ const __m256i res_hi = convolve_12taps(s + 6, coeffs);
+
+ const __m256i res_32b_hi = _mm256_sra_epi32(
+ _mm256_add_epi32(res_hi, right_shift_const), right_shift);
+ __m256i res_16b_hi = _mm256_packs_epi32(res_32b_hi, res_32b_hi);
+ // 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_epi32(res_8b_lo, res_8b_hi);
+
+ const __m128i res_0 = _mm256_extracti128_si256(res_a, 0);
+ const __m128i res_1 = _mm256_extracti128_si256(res_a, 1);
+
+ _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 {
+ const __m128i res_0 = _mm256_extracti128_si256(res_8b_lo, 0);
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b_lo, 1);
+ if (w - j > 2) {
+ *(uint32_t *)&dst[i * dst_stride + j] =
+ (uint32_t)_mm_cvtsi128_si32(res_0);
+ *(uint32_t *)&dst[i * dst_stride + j + dst_stride] =
+ (uint32_t)_mm_cvtsi128_si32(res_1);
+ } else {
+ *(uint16_t *)&dst[i * dst_stride + j] =
+ (uint16_t)_mm_cvtsi128_si32(res_0);
+ *(uint16_t *)&dst[i * dst_stride + j + dst_stride] =
+ (uint16_t)_mm_cvtsi128_si32(res_1);
+ }
+ }
+ 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];
+ }
+ }
+ } 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 + 6 * src_stride)));
+ const __m256i src_56a =
+ _mm256_permute2x128_si256(_mm256_castsi128_si256(d[5]), 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_lowbd(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_lowbd(s + 4, 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);
+ // 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = (uint16_t)_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];
}
}
}
@@ -377,58 +523,91 @@
const InterpFilterParams *filter_params_x,
const int subpel_x_qn,
ConvolveParams *conv_params) {
- if (filter_params_x->taps > 8) {
- if (w < 4) {
- av1_convolve_x_sr_c(src, src_stride, dst, dst_stride, w, h,
- filter_params_x, subpel_x_qn, conv_params);
- } else {
- av1_convolve_x_sr_12tap_sse2(src, src_stride, dst, dst_stride, w, h,
- filter_params_x, subpel_x_qn, conv_params);
- }
+ const int bits = FILTER_BITS - conv_params->round_0;
+ const __m128i round_shift = _mm_cvtsi32_si128(bits);
+ __m256i round_0_const =
+ _mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1);
+ __m128i round_0_shift = _mm_cvtsi32_si128(conv_params->round_0 - 1);
+ __m256i round_const = _mm256_set1_epi16((1 << bits) >> 1);
+ int i, horiz_tap = SUBPEL_TAPS;
+
+ assert(bits >= 0);
+ assert((FILTER_BITS - conv_params->round_1) >= 0 ||
+ ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
+ assert(conv_params->round_0 > 0);
+
+ __m256i coeffs[6], filt[4];
+ filt[0] = _mm256_load_si256((__m256i const *)(filt_global_avx2));
+ filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+
+ const int16_t *const filter = av1_get_interp_filter_subpel_kernel(
+ filter_params_x, subpel_x_qn & SUBPEL_MASK);
+ if (filter_params_x->taps == 12) {
+ horiz_tap = 12;
+ } else if (!(filter[0] | filter[1] | filter[6] | filter[7])) {
+ horiz_tap = 4;
+ } else if (!(filter[0] | filter[7])) {
+ horiz_tap = 6;
+ }
+
+ if (horiz_tap == 6)
+ prepare_coeffs_6t_lowbd(filter_params_x, subpel_x_qn, coeffs);
+ else if (horiz_tap == 12) {
+ prepare_coeffs_12taps(filter_params_x, subpel_x_qn, coeffs);
} else {
- const int bits = FILTER_BITS - conv_params->round_0;
+ prepare_coeffs_lowbd(filter_params_x, subpel_x_qn, 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, horiz_tap = SUBPEL_TAPS;
+ // horz_filt as 4 tap
+ if (horiz_tap == 4) {
+ 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);
- assert(bits >= 0);
- assert((FILTER_BITS - conv_params->round_1) >= 0 ||
- ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
- assert(conv_params->round_0 > 0);
+ __m256i res_16b = convolve_lowbd_x_4tap(data, coeffs + 1, filt);
- __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));
+ res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
+ round_0_shift);
- const int16_t *const filter = av1_get_interp_filter_subpel_kernel(
- filter_params_x, subpel_x_qn & SUBPEL_MASK);
- if (!(filter[0] | filter[1] | filter[6] | filter[7])) {
- horiz_tap = 4;
- } else if (!(filter[0] | filter[7])) {
- horiz_tap = 6;
- }
+ res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const),
+ round_shift);
- if (horiz_tap == 6)
- prepare_coeffs_6t_lowbd(filter_params_x, subpel_x_qn, coeffs);
- else
- prepare_coeffs_lowbd(filter_params_x, subpel_x_qn, coeffs);
+ /* rounding code */
+ // 8 bit conversion and saturation to uint8
+ __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
- // horz_filt as 4 tap
- if (horiz_tap == 4) {
- 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);
+ 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
+ *(uint16_t *)p_1 = (uint16_t)_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);
@@ -442,68 +621,65 @@
// 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 = (uint16_t)_mm_cvtsi128_si32(res_0);
- *(uint16_t *)p_1 = (uint16_t)_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);
- }
+ // 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 if (horiz_tap == 6) {
- const int fo_horiz = horiz_tap / 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));
+ }
+ } else if (horiz_tap == 6) {
+ const int fo_horiz = horiz_tap / 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);
+ 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_6tap(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);
+
+ 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_6tap(data, coeffs, filt);
@@ -517,67 +693,195 @@
// 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);
- }
+ // 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 (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);
+ }
+ }
+ } else if (horiz_tap == 12) { // horiz_tap == 12
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+ const uint8_t *const src_ptr = src - fo_horiz;
+ const __m256i v_zero = _mm256_setzero_si256();
+ round_0_const = _mm256_set1_epi32((1 << (conv_params->round_0)) >> 1);
+ round_const = _mm256_set1_epi32((1 << bits) >> 1);
+ round_0_shift = _mm_cvtsi32_si128(conv_params->round_0);
+ __m256i s[6];
- __m256i res_16b = convolve_lowbd_x_6tap(data, coeffs, filt);
+ if (w <= 4) {
+ 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);
+ // row0 0..7 row1 0..7
+ const __m256i s_16l = _mm256_unpacklo_epi8(data, v_zero);
+ // row0 8..F row1 8..F
+ const __m256i s_16h = _mm256_unpackhi_epi8(data, v_zero);
- res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
- round_0_shift);
+ // row0 00 00 01 01 .. 03 03 row1 00 00 01 01 .. 03 03
+ const __m256i s_ll = _mm256_unpacklo_epi16(s_16l, s_16l);
+ // row0 04 04 .. 07 07 row1 04 04 .. 07 07
+ const __m256i s_lh = _mm256_unpackhi_epi16(s_16l, s_16l);
- res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const),
- round_shift);
+ // row0 08 08 09 09 .. 0B 0B row1 08 08 09 09 .. 0B 0B
+ const __m256i s_hl = _mm256_unpacklo_epi16(s_16h, s_16h);
+ // row0 0C 0C .. 0F 0F row1 0C 0C .. 0F 0F
+ const __m256i s_hh = _mm256_unpackhi_epi16(s_16h, s_16h);
- /* rounding code */
- // 8 bit conversion and saturation to uint8
- __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
+ // 00 01 01 02 02 03 03 04 10 11 11 12 12 13 13 14
+ s[0] = _mm256_alignr_epi8(s_lh, s_ll, 2);
+ // 02 03 03 04 04 05 05 06 12 13 13 14 14 15 15 16
+ s[1] = _mm256_alignr_epi8(s_lh, s_ll, 10);
+ // 04 05 05 06 06 07 07 08 14 15 15 16 16 17 17 18
+ s[2] = _mm256_alignr_epi8(s_hl, s_lh, 2);
+ // 06 07 07 08 08 09 09 0A 16 17 17 18 18 19 19 1A
+ s[3] = _mm256_alignr_epi8(s_hl, s_lh, 10);
+ // 08 09 09 0A 0A 0B 0B 0C 18 19 19 1A 1A 1B 1B 1C
+ s[4] = _mm256_alignr_epi8(s_hh, s_hl, 2);
+ // 0A 0B 0B 0C 0C 0D 0D 0E 1A 1B 1B 1C 1C 1D 1D 1E
+ s[5] = _mm256_alignr_epi8(s_hh, s_hl, 10);
- // 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 __m256i res_lo = convolve_12taps(s, coeffs);
+
+ __m256i res_32b_lo = _mm256_sra_epi32(
+ _mm256_add_epi32(res_lo, round_0_const), round_0_shift);
+
+ // 00 01 02 03 10 12 13 14
+ res_32b_lo = _mm256_sra_epi32(_mm256_add_epi32(res_32b_lo, round_const),
+ round_shift);
+ // 8 bit conversion and saturation to uint8
+ // 00 01 02 03 00 01 02 03 10 11 12 13 10 11 12 13
+ __m256i res_16b_lo = _mm256_packs_epi32(res_32b_lo, res_32b_lo);
+ // 00 01 02 03 00 01 02 03 00 01 02 03 00 01 02 03
+ // 10 11 12 13 10 11 12 13 10 11 12 13 10 11 12 13
+ __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
+
+ // 00 01 02 03 00 01 02 03 00 01 02 03 00 01 02 03
+ const __m128i res_0 = _mm256_extracti128_si256(res_8b_lo, 0);
+ // 10 11 12 13 10 11 12 13 10 11 12 13 10 11 12 13
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b_lo, 1);
+ if (w > 2) {
+ // 00 01 02 03
+ *(uint32_t *)&dst[i * dst_stride] =
+ (uint32_t)_mm_cvtsi128_si32(res_0);
+ // 10 11 12 13
+ *(uint32_t *)&dst[i * dst_stride + dst_stride] =
+ (uint32_t)_mm_cvtsi128_si32(res_1);
+ } else {
+ // 00 01
+ *(uint16_t *)&dst[i * dst_stride] =
+ (uint16_t)_mm_cvtsi128_si32(res_0);
+ // 10 11
+ *(uint16_t *)&dst[i * dst_stride + dst_stride] =
+ (uint16_t)_mm_cvtsi128_si32(res_1);
}
}
} else {
- 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) {
+ for (i = 0; i < h; i++) {
+ for (int j = 0; j < w; j += 8) {
const __m256i data = _mm256_permute2x128_si256(
_mm256_castsi128_si256(
- _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride]))),
+ _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride + j]))),
_mm256_castsi128_si256(_mm_loadu_si128(
- (__m128i *)(&src_ptr[i * src_stride + src_stride]))),
+ (__m128i *)(&src_ptr[i * src_stride + j + 4]))),
0x20);
+ // row0 0..7 4..B
+ const __m256i s_16l = _mm256_unpacklo_epi8(data, v_zero);
+ // row0 8..F C..13
+ const __m256i s_16h = _mm256_unpackhi_epi8(data, v_zero);
+
+ // row0 00 00 01 01 .. 03 03 04 04 05 05 .. 07 07
+ const __m256i s_ll = _mm256_unpacklo_epi16(s_16l, s_16l);
+ // row0 04 04 .. 07 07 08 08 .. 0B 0B
+ const __m256i s_lh = _mm256_unpackhi_epi16(s_16l, s_16l);
+
+ // row0 08 08 09 09 .. 0B 0B 0C 0C 0D 0D .. 0F 0F
+ const __m256i s_hl = _mm256_unpacklo_epi16(s_16h, s_16h);
+ // row0 0C 0C 0D 0D .. 0F 0F 10 10 11 11 .. 13 13
+ const __m256i s_hh = _mm256_unpackhi_epi16(s_16h, s_16h);
+
+ s[0] = _mm256_alignr_epi8(s_lh, s_ll, 2);
+ s[1] = _mm256_alignr_epi8(s_lh, s_ll, 10);
+ s[2] = _mm256_alignr_epi8(s_hl, s_lh, 2);
+ s[3] = _mm256_alignr_epi8(s_hl, s_lh, 10);
+ s[4] = _mm256_alignr_epi8(s_hh, s_hl, 2);
+ s[5] = _mm256_alignr_epi8(s_hh, s_hl, 10);
+
+ const __m256i res_lo = convolve_12taps(s, coeffs);
+
+ __m256i res_32b_lo = _mm256_sra_epi32(
+ _mm256_add_epi32(res_lo, round_0_const), round_0_shift);
+
+ res_32b_lo = _mm256_sra_epi32(
+ _mm256_add_epi32(res_32b_lo, round_const), round_shift);
+ // 8 bit conversion and saturation to uint8
+ __m256i res_16b_lo = _mm256_packs_epi32(res_32b_lo, res_32b_lo);
+ __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
+ const __m128i res_0 = _mm256_extracti128_si256(res_8b_lo, 0);
+ const __m128i res_1 = _mm256_extracti128_si256(res_8b_lo, 1);
+ *(uint32_t *)&dst[i * dst_stride + j] =
+ (uint32_t)_mm_cvtsi128_si32(res_0);
+ *(uint32_t *)&dst[i * dst_stride + j + 4] =
+ (uint32_t)_mm_cvtsi128_si32(res_1);
+ }
+ }
+ }
+ } else {
+ 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);
+
+ 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(data, coeffs, filt);
@@ -591,51 +895,11 @@
// 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(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);
- }
+ // 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);
}
}
}
