Optimize av1_dist_wtd_convolve functions Introduced 4 tap filter usage for av1_dist_wtd_convolve functions. When tested for multiple test cases observed 0.3% average reduction in encoder time for speed = 1 preset. Approximately on average 20% improvement is seen at unit test level. Change-Id: Iaefe21c13660985840c6ba90cf30e8205ded098e
diff --git a/aom_dsp/x86/convolve_avx2.h b/aom_dsp/x86/convolve_avx2.h index e4873fc..4a1068e 100644 --- a/aom_dsp/x86/convolve_avx2.h +++ b/aom_dsp/x86/convolve_avx2.h
@@ -126,6 +126,122 @@ s[6] = s[7]; \ } +#define DIST_WTD_CONVOLVE_HORIZONTAL_FILTER_8TAP \ + for (i = 0; i < im_h; i += 2) { \ + __m256i data = _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)src_h)); \ + if (i + 1 < im_h) \ + data = _mm256_inserti128_si256( \ + data, _mm_loadu_si128((__m128i *)(src_h + src_stride)), 1); \ + src_h += (src_stride << 1); \ + __m256i res = convolve_lowbd_x(data, coeffs_x, 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 DIST_WTD_CONVOLVE_VERTICAL_FILTER_8TAP \ + __m256i s[8]; \ + __m256i s0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); \ + __m256i s1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); \ + __m256i s2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); \ + __m256i s3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); \ + __m256i s4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride)); \ + __m256i s5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride)); \ + \ + s[0] = _mm256_unpacklo_epi16(s0, s1); \ + s[1] = _mm256_unpacklo_epi16(s2, s3); \ + s[2] = _mm256_unpacklo_epi16(s4, s5); \ + \ + s[4] = _mm256_unpackhi_epi16(s0, s1); \ + s[5] = _mm256_unpackhi_epi16(s2, s3); \ + s[6] = _mm256_unpackhi_epi16(s4, s5); \ + \ + 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); \ + \ + const __m256i res_a = convolve(s, coeffs_y); \ + const __m256i res_a_round = _mm256_sra_epi32( \ + _mm256_add_epi32(res_a, round_const_v), round_shift_v); \ + \ + if (w - j > 4) { \ + const __m256i res_b = convolve(s + 4, coeffs_y); \ + const __m256i res_b_round = _mm256_sra_epi32( \ + _mm256_add_epi32(res_b, round_const_v), round_shift_v); \ + const __m256i res_16b = _mm256_packs_epi32(res_a_round, res_b_round); \ + const __m256i res_unsigned = _mm256_add_epi16(res_16b, offset_const); \ + \ + if (do_average) { \ + const __m256i data_ref_0 = load_line2_avx2( \ + &dst[i * dst_stride + j], &dst[i * dst_stride + j + dst_stride]); \ + const __m256i comp_avg_res = \ + comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); \ + \ + const __m256i round_result = convolve_rounding( \ + &comp_avg_res, &offset_const, &rounding_const, rounding_shift); \ + \ + const __m256i res_8 = _mm256_packus_epi16(round_result, round_result); \ + const __m128i res_0 = _mm256_castsi256_si128(res_8); \ + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); \ + \ + _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); \ + _mm_storel_epi64( \ + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); \ + } else { \ + const __m128i res_0 = _mm256_castsi256_si128(res_unsigned); \ + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); \ + \ + const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1); \ + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), \ + res_1); \ + } \ + } else { \ + const __m256i res_16b = _mm256_packs_epi32(res_a_round, res_a_round); \ + const __m256i res_unsigned = _mm256_add_epi16(res_16b, offset_const); \ + \ + if (do_average) { \ + const __m256i data_ref_0 = load_line2_avx2( \ + &dst[i * dst_stride + j], &dst[i * dst_stride + j + dst_stride]); \ + \ + const __m256i comp_avg_res = \ + comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); \ + \ + const __m256i round_result = convolve_rounding( \ + &comp_avg_res, &offset_const, &rounding_const, rounding_shift); \ + \ + const __m256i res_8 = _mm256_packus_epi16(round_result, round_result); \ + const __m128i res_0 = _mm256_castsi256_si128(res_8); \ + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); \ + \ + *(uint32_t *)(&dst0[i * dst_stride0 + j]) = _mm_cvtsi128_si32(res_0); \ + *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = \ + _mm_cvtsi128_si32(res_1); \ + \ + } else { \ + const __m128i res_0 = _mm256_castsi256_si128(res_unsigned); \ + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); \ + \ + const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1); \ + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), \ + 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] */) {
diff --git a/av1/common/x86/jnt_convolve_avx2.c b/av1/common/x86/jnt_convolve_avx2.c index cc46ef3..23cd6ab 100644 --- a/av1/common/x86/jnt_convolve_avx2.c +++ b/av1/common/x86/jnt_convolve_avx2.c
@@ -44,9 +44,7 @@ CONV_BUF_TYPE *dst = conv_params->dst; int dst_stride = conv_params->dst_stride; const int bd = 8; - 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, is_horiz_4tap = 0; const int bits = FILTER_BITS - conv_params->round_1; const __m256i wt = unpack_weights_avx2(conv_params); const int do_average = conv_params->do_average; @@ -58,18 +56,10 @@ const int rounding_shift = 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; const __m256i rounding_const = _mm256_set1_epi16((1 << rounding_shift) >> 1); - __m256i filt[4], coeffs[4]; assert(bits >= 0); 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); - const __m256i round_const = _mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1); const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0 - 1); @@ -77,51 +67,121 @@ (void)filter_params_y; (void)subpel_y_q4; - for (i = 0; i < h; i += 2) { - const uint8_t *src_data = src_ptr + i * src_stride; - CONV_BUF_TYPE *dst_data = dst + i * dst_stride; - for (j = 0; j < w; j += 8) { - const __m256i data = - load_line2_avx2(&src_data[j], &src_data[j + src_stride]); + __m256i filt[4], coeffs[4]; - __m256i res = convolve_lowbd_x(data, coeffs, filt); + filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2); + filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32)); - res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift); + prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs); - res = _mm256_slli_epi16(res, bits); + // Condition for checking valid horz_filt taps + if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs[0], coeffs[3]), 0))) + is_horiz_4tap = 1; - const __m256i res_unsigned = _mm256_add_epi16(res, offset_const); + // horz_filt as 4 tap + if (is_horiz_4tap) { + const int fo_horiz = 1; + const uint8_t *const src_ptr = src - fo_horiz; + for (i = 0; i < h; i += 2) { + const uint8_t *src_data = src_ptr + i * src_stride; + CONV_BUF_TYPE *dst_data = dst + i * dst_stride; + for (j = 0; j < w; j += 8) { + const __m256i data = + load_line2_avx2(&src_data[j], &src_data[j + src_stride]); - // Accumulate values into the destination buffer - if (do_average) { - const __m256i data_ref_0 = - load_line2_avx2(&dst_data[j], &dst_data[j + dst_stride]); - const __m256i comp_avg_res = - comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); + __m256i res = convolve_lowbd_x_4tap(data, coeffs + 1, filt); + res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift); + res = _mm256_slli_epi16(res, bits); - const __m256i round_result = convolve_rounding( - &comp_avg_res, &offset_const, &rounding_const, rounding_shift); + const __m256i res_unsigned = _mm256_add_epi16(res, offset_const); - const __m256i res_8 = _mm256_packus_epi16(round_result, round_result); - const __m128i res_0 = _mm256_castsi256_si128(res_8); - const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + // Accumulate values into the destination buffer + if (do_average) { + const __m256i data_ref_0 = + load_line2_avx2(&dst_data[j], &dst_data[j + dst_stride]); + const __m256i comp_avg_res = + comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); - if (w > 4) { - _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); - _mm_storel_epi64( - (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); + const __m256i round_result = convolve_rounding( + &comp_avg_res, &offset_const, &rounding_const, rounding_shift); + + const __m256i res_8 = _mm256_packus_epi16(round_result, round_result); + const __m128i res_0 = _mm256_castsi256_si128(res_8); + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + + if (w > 4) { + _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); + _mm_storel_epi64( + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); + } else { + *(uint32_t *)(&dst0[i * dst_stride0 + j]) = + _mm_cvtsi128_si32(res_0); + *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = + _mm_cvtsi128_si32(res_1); + } } else { - *(uint32_t *)(&dst0[i * dst_stride0 + j]) = _mm_cvtsi128_si32(res_0); - *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = - _mm_cvtsi128_si32(res_1); - } - } else { - const __m128i res_0 = _mm256_castsi256_si128(res_unsigned); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); + const __m128i res_0 = _mm256_castsi256_si128(res_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); - const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), - res_1); + const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), + 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)); + for (i = 0; i < h; i += 2) { + const uint8_t *src_data = src_ptr + i * src_stride; + CONV_BUF_TYPE *dst_data = dst + i * dst_stride; + for (j = 0; j < w; j += 8) { + const __m256i data = + load_line2_avx2(&src_data[j], &src_data[j + src_stride]); + + __m256i res = convolve_lowbd_x(data, coeffs, filt); + + res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift); + + res = _mm256_slli_epi16(res, bits); + + const __m256i res_unsigned = _mm256_add_epi16(res, offset_const); + + // Accumulate values into the destination buffer + if (do_average) { + const __m256i data_ref_0 = + load_line2_avx2(&dst_data[j], &dst_data[j + dst_stride]); + const __m256i comp_avg_res = + comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); + + const __m256i round_result = convolve_rounding( + &comp_avg_res, &offset_const, &rounding_const, rounding_shift); + + const __m256i res_8 = _mm256_packus_epi16(round_result, round_result); + const __m128i res_0 = _mm256_castsi256_si128(res_8); + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + + if (w > 4) { + _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); + _mm_storel_epi64( + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); + } else { + *(uint32_t *)(&dst0[i * dst_stride0 + j]) = + _mm_cvtsi128_si32(res_0); + *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = + _mm_cvtsi128_si32(res_1); + } + } else { + const __m128i res_0 = _mm256_castsi256_si128(res_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); + + const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), + res_1); + } } } } @@ -136,9 +196,7 @@ CONV_BUF_TYPE *dst = conv_params->dst; int dst_stride = conv_params->dst_stride; const int bd = 8; - 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; // +1 to compensate for dividing the filter coeffs by 2 const int left_shift = FILTER_BITS - conv_params->round_0 + 1; const __m256i round_const = @@ -168,171 +226,368 @@ (void)filter_params_x; (void)subpel_x_q4; - for (j = 0; j < w; j += 16) { - const uint8_t *data = &src_ptr[j]; - __m256i src6; - // Load lines a and b. Line a to lower 128, line b to upper 128 - { - __m256i src_ab[7]; - __m256i src_a[7]; - src_a[0] = _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); - for (int kk = 0; kk < 6; ++kk) { - data += src_stride; - src_a[kk + 1] = - _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); - src_ab[kk] = _mm256_permute2x128_si256(src_a[kk], src_a[kk + 1], 0x20); + // Condition for checking valid vert_filt taps + if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs[0], coeffs[3]), 0))) + is_vert_4tap = 1; + + 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]; + __m256i src4; + // Load lines a and b. Line a to lower 128, line b to upper 128 + { + __m256i src_ab[4]; + __m256i src_a[5]; + src_a[0] = _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); + for (int kk = 0; kk < 4; ++kk) { + data += src_stride; + src_a[kk + 1] = + _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); + src_ab[kk] = + _mm256_permute2x128_si256(src_a[kk], src_a[kk + 1], 0x20); + } + src4 = src_a[4]; + s[0] = _mm256_unpacklo_epi8(src_ab[0], src_ab[1]); + s[1] = _mm256_unpacklo_epi8(src_ab[2], src_ab[3]); + + s[3] = _mm256_unpackhi_epi8(src_ab[0], src_ab[1]); + s[4] = _mm256_unpackhi_epi8(src_ab[2], src_ab[3]); } - src6 = src_a[6]; - s[0] = _mm256_unpacklo_epi8(src_ab[0], src_ab[1]); - s[1] = _mm256_unpacklo_epi8(src_ab[2], src_ab[3]); - s[2] = _mm256_unpacklo_epi8(src_ab[4], src_ab[5]); - s[4] = _mm256_unpackhi_epi8(src_ab[0], src_ab[1]); - s[5] = _mm256_unpackhi_epi8(src_ab[2], src_ab[3]); - s[6] = _mm256_unpackhi_epi8(src_ab[4], src_ab[5]); - } - for (i = 0; i < h; i += 2) { - data = &src_ptr[(i + 7) * src_stride + j]; - const __m256i src7 = - _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); - const __m256i src_67a = _mm256_permute2x128_si256(src6, src7, 0x20); + for (i = 0; i < h; i += 2) { + data = &src_ptr[(i + 5) * src_stride + j]; + const __m256i src5 = + _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); + const __m256i src_45a = _mm256_permute2x128_si256(src4, src5, 0x20); - src6 = _mm256_castsi128_si256( - _mm_loadu_si128((__m128i *)(data + src_stride))); - const __m256i src_78a = _mm256_permute2x128_si256(src7, src6, 0x20); + src4 = _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + src_stride))); + const __m256i src_56a = _mm256_permute2x128_si256(src5, src4, 0x20); - s[3] = _mm256_unpacklo_epi8(src_67a, src_78a); - s[7] = _mm256_unpackhi_epi8(src_67a, src_78a); + s[2] = _mm256_unpacklo_epi8(src_45a, src_56a); + s[5] = _mm256_unpackhi_epi8(src_45a, src_56a); - __m256i res_lo = convolve_lowbd(s, coeffs); + __m256i res_lo = convolve_lowbd_4tap(s, coeffs + 1); - res_lo = _mm256_add_epi16(res_lo, offset_const_1); + res_lo = _mm256_add_epi16(res_lo, offset_const_1); - const __m256i res_lo_0_32b = _mm256_unpacklo_epi16(res_lo, zero); - const __m256i res_lo_0_shift = - _mm256_slli_epi32(res_lo_0_32b, left_shift); - const __m256i res_lo_0_round = _mm256_sra_epi32( - _mm256_add_epi32(res_lo_0_shift, round_const), round_shift); + const __m256i res_lo_0_32b = _mm256_unpacklo_epi16(res_lo, zero); + const __m256i res_lo_0_shift = + _mm256_slli_epi32(res_lo_0_32b, left_shift); + const __m256i res_lo_0_round = _mm256_sra_epi32( + _mm256_add_epi32(res_lo_0_shift, round_const), round_shift); - const __m256i res_lo_1_32b = _mm256_unpackhi_epi16(res_lo, zero); - const __m256i res_lo_1_shift = - _mm256_slli_epi32(res_lo_1_32b, left_shift); - const __m256i res_lo_1_round = _mm256_sra_epi32( - _mm256_add_epi32(res_lo_1_shift, round_const), round_shift); + const __m256i res_lo_1_32b = _mm256_unpackhi_epi16(res_lo, zero); + const __m256i res_lo_1_shift = + _mm256_slli_epi32(res_lo_1_32b, left_shift); + const __m256i res_lo_1_round = _mm256_sra_epi32( + _mm256_add_epi32(res_lo_1_shift, round_const), round_shift); - const __m256i res_lo_round = - _mm256_packs_epi32(res_lo_0_round, res_lo_1_round); + const __m256i res_lo_round = + _mm256_packs_epi32(res_lo_0_round, res_lo_1_round); - const __m256i res_lo_unsigned = - _mm256_add_epi16(res_lo_round, offset_const_2); + const __m256i res_lo_unsigned = + _mm256_add_epi16(res_lo_round, offset_const_2); - if (w - j < 16) { - if (do_average) { - const __m256i data_ref_0 = load_line2_avx2( - &dst[i * dst_stride + j], &dst[i * dst_stride + j + dst_stride]); - const __m256i comp_avg_res = comp_avg(&data_ref_0, &res_lo_unsigned, - &wt, use_dist_wtd_comp_avg); + if (w - j < 16) { + if (do_average) { + const __m256i data_ref_0 = + load_line2_avx2(&dst[i * dst_stride + j], + &dst[i * dst_stride + j + dst_stride]); + const __m256i comp_avg_res = comp_avg(&data_ref_0, &res_lo_unsigned, + &wt, use_dist_wtd_comp_avg); - const __m256i round_result = convolve_rounding( - &comp_avg_res, &offset_const, &rounding_const, rounding_shift); + const __m256i round_result = convolve_rounding( + &comp_avg_res, &offset_const, &rounding_const, rounding_shift); - const __m256i res_8 = _mm256_packus_epi16(round_result, round_result); - const __m128i res_0 = _mm256_castsi256_si128(res_8); - const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + const __m256i res_8 = + _mm256_packus_epi16(round_result, round_result); + const __m128i res_0 = _mm256_castsi256_si128(res_8); + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); - if (w - j > 4) { - _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); - _mm_storel_epi64( - (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); + if (w - j > 4) { + _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); + _mm_storel_epi64( + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), + res_1); + } else { + *(uint32_t *)(&dst0[i * dst_stride0 + j]) = + _mm_cvtsi128_si32(res_0); + *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = + _mm_cvtsi128_si32(res_1); + } } else { - *(uint32_t *)(&dst0[i * dst_stride0 + j]) = - _mm_cvtsi128_si32(res_0); - *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = - _mm_cvtsi128_si32(res_1); + const __m128i res_0 = _mm256_castsi256_si128(res_lo_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); + + const __m128i res_1 = _mm256_extracti128_si256(res_lo_unsigned, 1); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), + res_1); } } else { - const __m128i res_0 = _mm256_castsi256_si128(res_lo_unsigned); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); + __m256i res_hi = convolve_lowbd_4tap(s + 3, coeffs + 1); - const __m128i res_1 = _mm256_extracti128_si256(res_lo_unsigned, 1); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), - res_1); + res_hi = _mm256_add_epi16(res_hi, offset_const_1); + + const __m256i res_hi_0_32b = _mm256_unpacklo_epi16(res_hi, zero); + const __m256i res_hi_0_shift = + _mm256_slli_epi32(res_hi_0_32b, left_shift); + const __m256i res_hi_0_round = _mm256_sra_epi32( + _mm256_add_epi32(res_hi_0_shift, round_const), round_shift); + + const __m256i res_hi_1_32b = _mm256_unpackhi_epi16(res_hi, zero); + const __m256i res_hi_1_shift = + _mm256_slli_epi32(res_hi_1_32b, left_shift); + const __m256i res_hi_1_round = _mm256_sra_epi32( + _mm256_add_epi32(res_hi_1_shift, round_const), round_shift); + + const __m256i res_hi_round = + _mm256_packs_epi32(res_hi_0_round, res_hi_1_round); + + const __m256i res_hi_unsigned = + _mm256_add_epi16(res_hi_round, offset_const_2); + + if (do_average) { + const __m256i data_ref_0_lo = + load_line2_avx2(&dst[i * dst_stride + j], + &dst[i * dst_stride + j + dst_stride]); + + const __m256i data_ref_0_hi = + load_line2_avx2(&dst[i * dst_stride + j + 8], + &dst[i * dst_stride + j + 8 + dst_stride]); + + const __m256i comp_avg_res_lo = comp_avg( + &data_ref_0_lo, &res_lo_unsigned, &wt, use_dist_wtd_comp_avg); + + const __m256i comp_avg_res_hi = comp_avg( + &data_ref_0_hi, &res_hi_unsigned, &wt, use_dist_wtd_comp_avg); + + const __m256i round_result_lo = + convolve_rounding(&comp_avg_res_lo, &offset_const, + &rounding_const, rounding_shift); + + const __m256i round_result_hi = + convolve_rounding(&comp_avg_res_hi, &offset_const, + &rounding_const, rounding_shift); + + const __m256i res_8 = + _mm256_packus_epi16(round_result_lo, round_result_hi); + const __m128i res_0 = _mm256_castsi256_si128(res_8); + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + + _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); + _mm_store_si128( + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); + + } else { + const __m128i res_lo_0 = _mm256_castsi256_si128(res_lo_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_lo_0); + + const __m128i res_lo_1 = + _mm256_extracti128_si256(res_lo_unsigned, 1); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), + res_lo_1); + + const __m128i res_hi_0 = _mm256_castsi256_si128(res_hi_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + 8]), + res_hi_0); + + const __m128i res_hi_1 = + _mm256_extracti128_si256(res_hi_unsigned, 1); + _mm_store_si128( + (__m128i *)(&dst[i * dst_stride + j + 8 + dst_stride]), + res_hi_1); + } } - } else { - __m256i res_hi = convolve_lowbd(s + 4, coeffs); + s[0] = s[1]; + s[1] = s[2]; - res_hi = _mm256_add_epi16(res_hi, offset_const_1); - - const __m256i res_hi_0_32b = _mm256_unpacklo_epi16(res_hi, zero); - const __m256i res_hi_0_shift = - _mm256_slli_epi32(res_hi_0_32b, left_shift); - const __m256i res_hi_0_round = _mm256_sra_epi32( - _mm256_add_epi32(res_hi_0_shift, round_const), round_shift); - - const __m256i res_hi_1_32b = _mm256_unpackhi_epi16(res_hi, zero); - const __m256i res_hi_1_shift = - _mm256_slli_epi32(res_hi_1_32b, left_shift); - const __m256i res_hi_1_round = _mm256_sra_epi32( - _mm256_add_epi32(res_hi_1_shift, round_const), round_shift); - - const __m256i res_hi_round = - _mm256_packs_epi32(res_hi_0_round, res_hi_1_round); - - const __m256i res_hi_unsigned = - _mm256_add_epi16(res_hi_round, offset_const_2); - - if (do_average) { - const __m256i data_ref_0_lo = load_line2_avx2( - &dst[i * dst_stride + j], &dst[i * dst_stride + j + dst_stride]); - - const __m256i data_ref_0_hi = - load_line2_avx2(&dst[i * dst_stride + j + 8], - &dst[i * dst_stride + j + 8 + dst_stride]); - - const __m256i comp_avg_res_lo = comp_avg( - &data_ref_0_lo, &res_lo_unsigned, &wt, use_dist_wtd_comp_avg); - - const __m256i comp_avg_res_hi = comp_avg( - &data_ref_0_hi, &res_hi_unsigned, &wt, use_dist_wtd_comp_avg); - - const __m256i round_result_lo = convolve_rounding( - &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift); - - const __m256i round_result_hi = convolve_rounding( - &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift); - - const __m256i res_8 = - _mm256_packus_epi16(round_result_lo, round_result_hi); - const __m128i res_0 = _mm256_castsi256_si128(res_8); - const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); - - _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); - _mm_store_si128( - (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); - - } else { - const __m128i res_lo_0 = _mm256_castsi256_si128(res_lo_unsigned); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_lo_0); - - const __m128i res_lo_1 = _mm256_extracti128_si256(res_lo_unsigned, 1); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), - res_lo_1); - - const __m128i res_hi_0 = _mm256_castsi256_si128(res_hi_unsigned); - _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + 8]), res_hi_0); - - const __m128i res_hi_1 = _mm256_extracti128_si256(res_hi_unsigned, 1); - _mm_store_si128( - (__m128i *)(&dst[i * dst_stride + j + 8 + dst_stride]), res_hi_1); - } + s[3] = s[4]; + s[4] = s[5]; } - s[0] = s[1]; - s[1] = s[2]; - s[2] = s[3]; + } + } 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; + // Load lines a and b. Line a to lower 128, line b to upper 128 + { + __m256i src_ab[7]; + __m256i src_a[7]; + src_a[0] = _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); + for (int kk = 0; kk < 6; ++kk) { + data += src_stride; + src_a[kk + 1] = + _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); + src_ab[kk] = + _mm256_permute2x128_si256(src_a[kk], src_a[kk + 1], 0x20); + } + src6 = src_a[6]; + s[0] = _mm256_unpacklo_epi8(src_ab[0], src_ab[1]); + s[1] = _mm256_unpacklo_epi8(src_ab[2], src_ab[3]); + s[2] = _mm256_unpacklo_epi8(src_ab[4], src_ab[5]); + s[4] = _mm256_unpackhi_epi8(src_ab[0], src_ab[1]); + s[5] = _mm256_unpackhi_epi8(src_ab[2], src_ab[3]); + s[6] = _mm256_unpackhi_epi8(src_ab[4], src_ab[5]); + } - s[4] = s[5]; - s[5] = s[6]; - s[6] = s[7]; + for (i = 0; i < h; i += 2) { + data = &src_ptr[(i + 7) * src_stride + j]; + const __m256i src7 = + _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data)); + const __m256i src_67a = _mm256_permute2x128_si256(src6, src7, 0x20); + + src6 = _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + src_stride))); + const __m256i src_78a = _mm256_permute2x128_si256(src7, src6, 0x20); + + s[3] = _mm256_unpacklo_epi8(src_67a, src_78a); + s[7] = _mm256_unpackhi_epi8(src_67a, src_78a); + + __m256i res_lo = convolve_lowbd(s, coeffs); + + res_lo = _mm256_add_epi16(res_lo, offset_const_1); + + const __m256i res_lo_0_32b = _mm256_unpacklo_epi16(res_lo, zero); + const __m256i res_lo_0_shift = + _mm256_slli_epi32(res_lo_0_32b, left_shift); + const __m256i res_lo_0_round = _mm256_sra_epi32( + _mm256_add_epi32(res_lo_0_shift, round_const), round_shift); + + const __m256i res_lo_1_32b = _mm256_unpackhi_epi16(res_lo, zero); + const __m256i res_lo_1_shift = + _mm256_slli_epi32(res_lo_1_32b, left_shift); + const __m256i res_lo_1_round = _mm256_sra_epi32( + _mm256_add_epi32(res_lo_1_shift, round_const), round_shift); + + const __m256i res_lo_round = + _mm256_packs_epi32(res_lo_0_round, res_lo_1_round); + + const __m256i res_lo_unsigned = + _mm256_add_epi16(res_lo_round, offset_const_2); + + if (w - j < 16) { + if (do_average) { + const __m256i data_ref_0 = + load_line2_avx2(&dst[i * dst_stride + j], + &dst[i * dst_stride + j + dst_stride]); + const __m256i comp_avg_res = comp_avg(&data_ref_0, &res_lo_unsigned, + &wt, use_dist_wtd_comp_avg); + + const __m256i round_result = convolve_rounding( + &comp_avg_res, &offset_const, &rounding_const, rounding_shift); + + const __m256i res_8 = + _mm256_packus_epi16(round_result, round_result); + const __m128i res_0 = _mm256_castsi256_si128(res_8); + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + + if (w - j > 4) { + _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); + _mm_storel_epi64( + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), + res_1); + } else { + *(uint32_t *)(&dst0[i * dst_stride0 + j]) = + _mm_cvtsi128_si32(res_0); + *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = + _mm_cvtsi128_si32(res_1); + } + } else { + const __m128i res_0 = _mm256_castsi256_si128(res_lo_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0); + + const __m128i res_1 = _mm256_extracti128_si256(res_lo_unsigned, 1); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), + res_1); + } + } else { + __m256i res_hi = convolve_lowbd(s + 4, coeffs); + + res_hi = _mm256_add_epi16(res_hi, offset_const_1); + + const __m256i res_hi_0_32b = _mm256_unpacklo_epi16(res_hi, zero); + const __m256i res_hi_0_shift = + _mm256_slli_epi32(res_hi_0_32b, left_shift); + const __m256i res_hi_0_round = _mm256_sra_epi32( + _mm256_add_epi32(res_hi_0_shift, round_const), round_shift); + + const __m256i res_hi_1_32b = _mm256_unpackhi_epi16(res_hi, zero); + const __m256i res_hi_1_shift = + _mm256_slli_epi32(res_hi_1_32b, left_shift); + const __m256i res_hi_1_round = _mm256_sra_epi32( + _mm256_add_epi32(res_hi_1_shift, round_const), round_shift); + + const __m256i res_hi_round = + _mm256_packs_epi32(res_hi_0_round, res_hi_1_round); + + const __m256i res_hi_unsigned = + _mm256_add_epi16(res_hi_round, offset_const_2); + + if (do_average) { + const __m256i data_ref_0_lo = + load_line2_avx2(&dst[i * dst_stride + j], + &dst[i * dst_stride + j + dst_stride]); + + const __m256i data_ref_0_hi = + load_line2_avx2(&dst[i * dst_stride + j + 8], + &dst[i * dst_stride + j + 8 + dst_stride]); + + const __m256i comp_avg_res_lo = comp_avg( + &data_ref_0_lo, &res_lo_unsigned, &wt, use_dist_wtd_comp_avg); + + const __m256i comp_avg_res_hi = comp_avg( + &data_ref_0_hi, &res_hi_unsigned, &wt, use_dist_wtd_comp_avg); + + const __m256i round_result_lo = + convolve_rounding(&comp_avg_res_lo, &offset_const, + &rounding_const, rounding_shift); + + const __m256i round_result_hi = + convolve_rounding(&comp_avg_res_hi, &offset_const, + &rounding_const, rounding_shift); + + const __m256i res_8 = + _mm256_packus_epi16(round_result_lo, round_result_hi); + const __m128i res_0 = _mm256_castsi256_si128(res_8); + const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); + + _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); + _mm_store_si128( + (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); + + } else { + const __m128i res_lo_0 = _mm256_castsi256_si128(res_lo_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_lo_0); + + const __m128i res_lo_1 = + _mm256_extracti128_si256(res_lo_unsigned, 1); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), + res_lo_1); + + const __m128i res_hi_0 = _mm256_castsi256_si128(res_hi_unsigned); + _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + 8]), + res_hi_0); + + const __m128i res_hi_1 = + _mm256_extracti128_si256(res_hi_unsigned, 1); + _mm_store_si128( + (__m128i *)(&dst[i * dst_stride + j + 8 + dst_stride]), + res_hi_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]; + } } } } @@ -348,12 +603,9 @@ 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; const __m256i wt = unpack_weights_avx2(conv_params); const int do_average = conv_params->do_average; const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; @@ -364,18 +616,9 @@ const int rounding_shift = 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; const __m256i rounding_const = _mm256_set1_epi16((1 << rounding_shift) >> 1); - __m256i filt[4], s[8], coeffs_x[4], coeffs_y[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_x); - prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y); - 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); @@ -385,9 +628,29 @@ (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1))); const __m128i round_shift_v = _mm_cvtsi32_si128(conv_params->round_1); - for (j = 0; j < w; j += 8) { - /* Horizontal filter */ - { + __m256i filt[4], coeffs_x[4], coeffs_y[4]; + + 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_x); + prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y); + + // Condition for checking valid horz_filt taps + if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs_x[0], coeffs_x[3]), 0))) + is_horiz_4tap = 1; + + // Condition for checking valid vert_filt taps + if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs_y[0], coeffs_y[3]), 0))) + is_vert_4tap = 1; + + 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; + for (int j = 0; j < w; j += 8) { + /* Horizontal filter */ const uint8_t *src_h = src_ptr + j; for (i = 0; i < im_h; i += 2) { __m256i data = @@ -396,49 +659,59 @@ data = _mm256_inserti128_si256( data, _mm_loadu_si128((__m128i *)(src_h + src_stride)), 1); src_h += (src_stride << 1); - __m256i res = convolve_lowbd_x(data, coeffs_x, filt); + __m256i res = convolve_lowbd_x_4tap(data, coeffs_x + 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); } + DIST_WTD_CONVOLVE_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; - /* Vertical filter */ - { + 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) { + /* Horizontal filter */ + const uint8_t *src_h = src_ptr + j; + DIST_WTD_CONVOLVE_HORIZONTAL_FILTER_8TAP; + + /* Vertical filter */ + __m256i s[6]; __m256i s0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); __m256i s1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); __m256i s2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); __m256i s3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); - __m256i s4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride)); - __m256i s5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride)); s[0] = _mm256_unpacklo_epi16(s0, s1); s[1] = _mm256_unpacklo_epi16(s2, s3); - s[2] = _mm256_unpacklo_epi16(s4, s5); - s[4] = _mm256_unpackhi_epi16(s0, s1); - s[5] = _mm256_unpackhi_epi16(s2, s3); - s[6] = _mm256_unpackhi_epi16(s4, s5); + s[3] = _mm256_unpackhi_epi16(s0, s1); + s[4] = _mm256_unpackhi_epi16(s2, s3); 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); - const __m256i res_a = convolve(s, coeffs_y); + const __m256i res_a = convolve_4tap(s, coeffs_y + 1); const __m256i res_a_round = _mm256_sra_epi32( _mm256_add_epi32(res_a, round_const_v), round_shift_v); if (w - j > 4) { - const __m256i res_b = convolve(s + 4, coeffs_y); + const __m256i res_b = convolve_4tap(s + 3, coeffs_y + 1); const __m256i res_b_round = _mm256_sra_epi32( _mm256_add_epi32(res_b, round_const_v), round_shift_v); const __m256i res_16b = _mm256_packs_epi32(res_a_round, res_b_round); @@ -504,16 +777,28 @@ res_1); } } - 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 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 (int j = 0; j < w; j += 8) { + /* Horizontal filter */ + const uint8_t *src_h = src_ptr + j; + DIST_WTD_CONVOLVE_HORIZONTAL_FILTER_8TAP; + + DIST_WTD_CONVOLVE_VERTICAL_FILTER_8TAP; + } } }