jnt_convolve_ssse3.c (9983B)
1 /* 2 * Copyright (c) 2018, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include <tmmintrin.h> 13 14 #include "config/av1_rtcd.h" 15 16 #include "aom_dsp/aom_filter.h" 17 #include "aom_dsp/x86/convolve_sse2.h" 18 19 void av1_dist_wtd_convolve_2d_ssse3( 20 const uint8_t *src, int src_stride, uint8_t *dst0, int dst_stride0, int w, 21 int h, const InterpFilterParams *filter_params_x, 22 const InterpFilterParams *filter_params_y, const int subpel_x_qn, 23 const int subpel_y_qn, ConvolveParams *conv_params) { 24 CONV_BUF_TYPE *dst = conv_params->dst; 25 int dst_stride = conv_params->dst_stride; 26 const int bd = 8; 27 28 DECLARE_ALIGNED(16, int16_t, 29 im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]); 30 int im_h = h + filter_params_y->taps - 1; 31 int im_stride = MAX_SB_SIZE; 32 int i, j; 33 const int fo_vert = filter_params_y->taps / 2 - 1; 34 const int fo_horiz = filter_params_x->taps / 2 - 1; 35 const int do_average = conv_params->do_average; 36 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 37 const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz; 38 39 const __m128i zero = _mm_setzero_si128(); 40 41 const int w0 = conv_params->fwd_offset; 42 const int w1 = conv_params->bck_offset; 43 const __m128i wt0 = _mm_set1_epi16(w0); 44 const __m128i wt1 = _mm_set1_epi16(w1); 45 const __m128i wt = _mm_unpacklo_epi16(wt0, wt1); 46 47 const int offset_0 = 48 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 49 const int offset = (1 << offset_0) + (1 << (offset_0 - 1)); 50 const __m128i offset_const = _mm_set1_epi16(offset); 51 const int rounding_shift = 52 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 53 const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1); 54 55 /* Horizontal filter */ 56 { 57 const int16_t *x_filter = av1_get_interp_filter_subpel_kernel( 58 filter_params_x, subpel_x_qn & SUBPEL_MASK); 59 const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter); 60 61 // coeffs 0 1 0 1 2 3 2 3 62 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x); 63 // coeffs 4 5 4 5 6 7 6 7 64 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x); 65 66 // coeffs 0 1 0 1 0 1 0 1 67 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); 68 // coeffs 2 3 2 3 2 3 2 3 69 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); 70 // coeffs 4 5 4 5 4 5 4 5 71 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); 72 // coeffs 6 7 6 7 6 7 6 7 73 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); 74 75 const __m128i round_const = _mm_set1_epi32( 76 ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1))); 77 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0); 78 79 for (i = 0; i < im_h; ++i) { 80 for (j = 0; j < w; j += 8) { 81 const __m128i data = 82 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); 83 84 const __m128i src_lo = _mm_unpacklo_epi8(data, zero); 85 const __m128i src_hi = _mm_unpackhi_epi8(data, zero); 86 87 // Filter even-index pixels 88 const __m128i res_0 = _mm_madd_epi16(src_lo, coeff_01); 89 const __m128i src_2 = _mm_alignr_epi8(src_hi, src_lo, 4); 90 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); 91 const __m128i src_4 = _mm_alignr_epi8(src_hi, src_lo, 8); 92 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); 93 const __m128i src_6 = _mm_alignr_epi8(src_hi, src_lo, 12); 94 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); 95 96 __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4), 97 _mm_add_epi32(res_2, res_6)); 98 res_even = 99 _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift); 100 101 // Filter odd-index pixels 102 const __m128i src_1 = _mm_alignr_epi8(src_hi, src_lo, 2); 103 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); 104 const __m128i src_3 = _mm_alignr_epi8(src_hi, src_lo, 6); 105 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); 106 const __m128i src_5 = _mm_alignr_epi8(src_hi, src_lo, 10); 107 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); 108 const __m128i src_7 = _mm_alignr_epi8(src_hi, src_lo, 14); 109 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); 110 111 __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5), 112 _mm_add_epi32(res_3, res_7)); 113 res_odd = 114 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift); 115 116 // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7 117 __m128i res = _mm_packs_epi32(res_even, res_odd); 118 _mm_store_si128((__m128i *)&im_block[i * im_stride + j], res); 119 } 120 } 121 } 122 123 /* Vertical filter */ 124 { 125 const int16_t *y_filter = av1_get_interp_filter_subpel_kernel( 126 filter_params_y, subpel_y_qn & SUBPEL_MASK); 127 const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter); 128 129 // coeffs 0 1 0 1 2 3 2 3 130 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y); 131 // coeffs 4 5 4 5 6 7 6 7 132 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y); 133 134 // coeffs 0 1 0 1 0 1 0 1 135 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); 136 // coeffs 2 3 2 3 2 3 2 3 137 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); 138 // coeffs 4 5 4 5 4 5 4 5 139 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); 140 // coeffs 6 7 6 7 6 7 6 7 141 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); 142 143 const __m128i round_const = _mm_set1_epi32( 144 ((1 << conv_params->round_1) >> 1) - 145 (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1))); 146 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1); 147 148 for (i = 0; i < h; ++i) { 149 for (j = 0; j < w; j += 8) { 150 // Filter even-index pixels 151 const int16_t *data = &im_block[i * im_stride + j]; 152 const __m128i src_0 = 153 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), 154 *(__m128i *)(data + 1 * im_stride)); 155 const __m128i src_2 = 156 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), 157 *(__m128i *)(data + 3 * im_stride)); 158 const __m128i src_4 = 159 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), 160 *(__m128i *)(data + 5 * im_stride)); 161 const __m128i src_6 = 162 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), 163 *(__m128i *)(data + 7 * im_stride)); 164 165 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); 166 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); 167 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); 168 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); 169 170 const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2), 171 _mm_add_epi32(res_4, res_6)); 172 173 // Filter odd-index pixels 174 const __m128i src_1 = 175 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), 176 *(__m128i *)(data + 1 * im_stride)); 177 const __m128i src_3 = 178 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), 179 *(__m128i *)(data + 3 * im_stride)); 180 const __m128i src_5 = 181 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), 182 *(__m128i *)(data + 5 * im_stride)); 183 const __m128i src_7 = 184 _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride), 185 *(__m128i *)(data + 7 * im_stride)); 186 187 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); 188 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); 189 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); 190 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); 191 192 const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3), 193 _mm_add_epi32(res_5, res_7)); 194 195 // Rearrange pixels back into the order 0 ... 7 196 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); 197 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd); 198 199 const __m128i res_lo_round = 200 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); 201 const __m128i res_hi_round = 202 _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift); 203 204 const __m128i res_16b = _mm_packs_epi32(res_lo_round, res_hi_round); 205 const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const); 206 207 // Accumulate values into the destination buffer 208 if (do_average) { 209 const __m128i data_ref_0 = 210 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 211 212 const __m128i comp_avg_res = 213 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 214 215 const __m128i round_result = convolve_rounding( 216 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 217 218 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 219 220 if (w > 4) 221 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); 222 else 223 *(int *)(&dst0[i * dst_stride0 + j]) = _mm_cvtsi128_si32(res_8); 224 } else { 225 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned); 226 } 227 } 228 } 229 } 230 }