Line data    Source code

       1             : /*
       2             : * Copyright(c) 2019 Intel Corporation
       3             : * SPDX - License - Identifier: BSD - 2 - Clause - Patent
       4             : */
       5             : 
       6             : #ifndef EbPictureOperators_SSE2_h
       7             : #define EbPictureOperators_SSE2_h
       8             : 
       9             : #include <emmintrin.h>
      10             : #include "EbDefinitions.h"
      11             : 
      12             : #ifdef __cplusplus
      13             : extern "C" {
      14             : #endif
      15             : 
      16             :     //-----
      17             :     extern void zero_out_coeff4x4_sse(
      18             :         int16_t *coeff_buffer,
      19             :         uint32_t coeff_stride,
      20             :         uint32_t coeff_origin_index,
      21             :         uint32_t area_width,
      22             :         uint32_t area_height);
      23             : 
      24             :     extern void zero_out_coeff8x8_sse2(
      25             :         int16_t *coeff_buffer,
      26             :         uint32_t coeff_stride,
      27             :         uint32_t coeff_origin_index,
      28             :         uint32_t area_width,
      29             :         uint32_t area_height);
      30             : 
      31             :     extern void zero_out_coeff16x16_sse2(
      32             :         int16_t *coeff_buffer,
      33             :         uint32_t coeff_stride,
      34             :         uint32_t coeff_origin_index,
      35             :         uint32_t area_width,
      36             :         uint32_t area_height);
      37             : 
      38             :     extern void zero_out_coeff32x32_sse2(
      39             :         int16_t *coeff_buffer,
      40             :         uint32_t coeff_stride,
      41             :         uint32_t coeff_origin_index,
      42             :         uint32_t area_width,
      43             :         uint32_t area_height);
      44             : 
      45             :     extern void residual_kernel16bit_sse2_intrin(
      46             :         uint16_t *input,
      47             :         uint32_t  input_stride,
      48             :         uint16_t *pred,
      49             :         uint32_t  pred_stride,
      50             :         int16_t  *residual,
      51             :         uint32_t  residual_stride,
      52             :         uint32_t  area_width,
      53             :         uint32_t  area_height);
      54             : 
      55             :     void picture_addition_kernel4x4_sse_intrin(
      56             :         uint8_t  *pred_ptr,
      57             :         uint32_t  pred_stride,
      58             :         int16_t  *residual_ptr,
      59             :         uint32_t  residual_stride,
      60             :         uint8_t  *recon_ptr,
      61             :         uint32_t  recon_stride,
      62             :         uint32_t  width,
      63             :         uint32_t  height);
      64             : 
      65             :     void picture_addition_kernel8x8_sse2_intrin(
      66             :         uint8_t  *pred_ptr,
      67             :         uint32_t  pred_stride,
      68             :         int16_t  *residual_ptr,
      69             :         uint32_t  residual_stride,
      70             :         uint8_t  *recon_ptr,
      71             :         uint32_t  recon_stride,
      72             :         uint32_t  width,
      73             :         uint32_t  height);
      74             : 
      75             :     void picture_addition_kernel16x16_sse2_intrin(
      76             :         uint8_t  *pred_ptr,
      77             :         uint32_t  pred_stride,
      78             :         int16_t  *residual_ptr,
      79             :         uint32_t  residual_stride,
      80             :         uint8_t  *recon_ptr,
      81             :         uint32_t  recon_stride,
      82             :         uint32_t  width,
      83             :         uint32_t  height);
      84             : 
      85             :     void picture_addition_kernel32x32_sse2_intrin(
      86             :         uint8_t  *pred_ptr,
      87             :         uint32_t  pred_stride,
      88             :         int16_t  *residual_ptr,
      89             :         uint32_t  residual_stride,
      90             :         uint8_t  *recon_ptr,
      91             :         uint32_t  recon_stride,
      92             :         uint32_t  width,
      93             :         uint32_t  height);
      94             : 
      95             :     void picture_addition_kernel64x64_sse2_intrin(
      96             :         uint8_t  *pred_ptr,
      97             :         uint32_t  pred_stride,
      98             :         int16_t  *residual_ptr,
      99             :         uint32_t  residual_stride,
     100             :         uint8_t  *recon_ptr,
     101             :         uint32_t  recon_stride,
     102             :         uint32_t  width,
     103             :         uint32_t  height);
     104             : 
     105             :     void residual_kernel_sub_sampled4x4_sse_intrin(
     106             :         uint8_t  *input,
     107             :         uint32_t  input_stride,
     108             :         uint8_t  *pred,
     109             :         uint32_t  pred_stride,
     110             :         int16_t  *residual,
     111             :         uint32_t  residual_stride,
     112             :         uint32_t  area_width,
     113             :         uint32_t  area_height,
     114             :         uint8_t   last_line);
     115             : 
     116             :     void residual_kernel_sub_sampled8x8_sse2_intrin(
     117             :         uint8_t  *input,
     118             :         uint32_t  input_stride,
     119             :         uint8_t  *pred,
     120             :         uint32_t  pred_stride,
     121             :         int16_t  *residual,
     122             :         uint32_t  residual_stride,
     123             :         uint32_t  area_width,
     124             :         uint32_t  area_height,
     125             :         uint8_t   last_line);
     126             : 
     127             :     void residual_kernel_sub_sampled16x16_sse2_intrin(
     128             :         uint8_t  *input,
     129             :         uint32_t  input_stride,
     130             :         uint8_t  *pred,
     131             :         uint32_t  pred_stride,
     132             :         int16_t  *residual,
     133             :         uint32_t  residual_stride,
     134             :         uint32_t  area_width,
     135             :         uint32_t  area_height,
     136             :         uint8_t   last_line);
     137             : 
     138             :     void residual_kernel_sub_sampled32x32_sse2_intrin(
     139             :         uint8_t  *input,
     140             :         uint32_t  input_stride,
     141             :         uint8_t  *pred,
     142             :         uint32_t  pred_stride,
     143             :         int16_t  *residual,
     144             :         uint32_t  residual_stride,
     145             :         uint32_t  area_width,
     146             :         uint32_t  area_height,
     147             :         uint8_t   last_line);
     148             : 
     149             :     void residual_kernel_sub_sampled64x64_sse2_intrin(
     150             :         uint8_t  *input,
     151             :         uint32_t  input_stride,
     152             :         uint8_t  *pred,
     153             :         uint32_t  pred_stride,
     154             :         int16_t  *residual,
     155             :         uint32_t  residual_stride,
     156             :         uint32_t  area_width,
     157             :         uint32_t  area_height,
     158             :         uint8_t   last_line);
     159             : 
     160             :     void picture_addition_kernel16bit_sse2_intrin(
     161             :         uint16_t *pred_ptr,
     162             :         uint32_t  pred_stride,
     163             :         int16_t  *residual_ptr,
     164             :         uint32_t  residual_stride,
     165             :         uint16_t  *recon_ptr,
     166             :         uint32_t  recon_stride,
     167             :         uint32_t  width,
     168             :         uint32_t  height);
     169             : 
     170  1195020300 :     static INLINE int32_t Hadd32_SSE2_INTRIN(const __m128i src) {
     171  1195020300 :         const __m128i dst0 = _mm_add_epi32(src, _mm_srli_si128(src, 8));
     172  2390050600 :         const __m128i dst1 = _mm_add_epi32(dst0, _mm_srli_si128(dst0, 4));
     173             : 
     174  1195020300 :         return _mm_cvtsi128_si32(dst1);
     175             :     }
     176             : 
     177             :     uint64_t spatial_full_distortion_kernel4x_n_sse2_intrin(
     178             :         uint8_t   *input,
     179             :         uint32_t   input_offset,
     180             :         uint32_t   input_stride,
     181             :         uint8_t   *recon,
     182             :         uint32_t   recon_offset,
     183             :         uint32_t   recon_stride,
     184             :         uint32_t   area_width,
     185             :         uint32_t   area_height);
     186             : 
     187             :     uint64_t spatial_full_distortion_kernel8x_n_sse2_intrin(
     188             :         uint8_t   *input,
     189             :         uint32_t   input_offset,
     190             :         uint32_t   input_stride,
     191             :         uint8_t   *recon,
     192             :         uint32_t   recon_offset,
     193             :         uint32_t   recon_stride,
     194             :         uint32_t   area_width,
     195             :         uint32_t   area_height);
     196             : 
     197             :     uint64_t spatial_full_distortion_kernel16x_n_sse2_intrin(
     198             :         uint8_t   *input,
     199             :         uint32_t   input_offset,
     200             :         uint32_t   input_stride,
     201             :         uint8_t   *recon,
     202             :         uint32_t   recon_offset,
     203             :         uint32_t   recon_stride,
     204             :         uint32_t   area_width,
     205             :         uint32_t   area_height);
     206             : 
     207             :     uint64_t spatial_full_distortion_kernel32x_n_sse2_intrin(
     208             :         uint8_t   *input,
     209             :         uint32_t   input_offset,
     210             :         uint32_t   input_stride,
     211             :         uint8_t   *recon,
     212             :         uint32_t   recon_offset,
     213             :         uint32_t   recon_stride,
     214             :         uint32_t   area_width,
     215             :         uint32_t   area_height);
     216             : 
     217             :     uint64_t spatial_full_distortion_kernel64x_n_sse2_intrin(
     218             :         uint8_t   *input,
     219             :         uint32_t   input_offset,
     220             :         uint32_t   input_stride,
     221             :         uint8_t   *recon,
     222             :         uint32_t   recon_offset,
     223             :         uint32_t   recon_stride,
     224             :         uint32_t   area_width,
     225             :         uint32_t   area_height);
     226             : 
     227             :     uint64_t spatial_full_distortion_kernel128x_n_sse2_intrin(
     228             :         uint8_t   *input,
     229             :         uint32_t   input_offset,
     230             :         uint32_t   input_stride,
     231             :         uint8_t   *recon,
     232             :         uint32_t   recon_offset,
     233             :         uint32_t   recon_stride,
     234             :         uint32_t   area_width,
     235             :         uint32_t   area_height);
     236             : 
     237             :     void picture_addition_sse2(
     238             :         uint8_t  *pred_ptr,
     239             :         uint32_t  pred_stride,
     240             :         int16_t *residual_ptr,
     241             :         uint32_t  residual_stride,
     242             :         uint8_t  *recon_ptr,
     243             :         uint32_t  recon_stride,
     244             :         uint32_t  width,
     245             :         uint32_t  height);
     246             : 
     247             : #ifdef __cplusplus
     248             : }
     249             : #endif
     250             : #endif // EbPictureOperators_SSE2_h