Line data    Source code

       1             : /*
       2             : * Copyright(c) 2019 Intel Corporation
       3             : * SPDX - License - Identifier: BSD - 2 - Clause - Patent
       4             : */
       5             : 
       6             : #ifndef EbHighbdIntraPrediction_SSE2_h
       7             : #define EbHighbdIntraPrediction_SSE2_h
       8             : 
       9             : #include <emmintrin.h>
      10             : #include "EbDefinitions.h"
      11             : #include "aom_dsp_rtcd.h"
      12             : 
      13           0 : static INLINE __m128i dc_sum_4x32bit(const __m128i src) {
      14             :     __m128i sum, sum_hi;
      15           0 :     sum_hi = _mm_srli_si128(src, 8);
      16           0 :     sum = _mm_add_epi32(src, sum_hi);
      17           0 :     sum_hi = _mm_srli_si128(sum, 4);
      18           0 :     return _mm_add_epi32(sum, sum_hi);
      19             : }
      20             : 
      21           0 : static INLINE __m128i dc_sum_4x16bit(const __m128i src) {
      22             :     __m128i sum, sum_hi;
      23           0 :     const __m128i src_hi = _mm_srli_si128(src, 4);
      24           0 :     sum = _mm_add_epi16(src, src_hi);
      25           0 :     sum_hi = _mm_srli_si128(sum, 2);
      26           0 :     sum = _mm_add_epi16(sum, sum_hi);
      27             : 
      28           0 :     return sum;
      29             : }
      30             : 
      31           0 : static INLINE __m128i dc_sum_4x16bit_large(const __m128i src) {
      32             :     // Unpack to avoid 12-bit overflow.
      33           0 :     const __m128i src_32 = _mm_unpacklo_epi16(src, _mm_setzero_si128());
      34           0 :     return dc_sum_4x32bit(src_32);
      35             : }
      36             : 
      37           0 : static INLINE __m128i dc_sum_8x16bit(const __m128i src) {
      38           0 :     const __m128i src_hi = _mm_srli_si128(src, 8);
      39           0 :     const __m128i sum = _mm_add_epi16(src, src_hi);
      40           0 :     return dc_sum_4x16bit(sum);
      41             : }
      42             : 
      43           0 : static INLINE __m128i dc_sum_8x16bit_large(const __m128i src) {
      44           0 :     const __m128i src_hi = _mm_srli_si128(src, 8);
      45           0 :     const __m128i sum = _mm_add_epi16(src, src_hi);
      46           0 :     return dc_sum_4x16bit_large(sum);
      47             : }
      48             : 
      49           0 : static INLINE __m128i dc_sum_4(const uint16_t *const src) {
      50           0 :     const __m128i s = _mm_loadl_epi64((const __m128i *)src);
      51           0 :     return dc_sum_4x16bit(s);
      52             : }
      53             : 
      54           0 : static INLINE __m128i dc_sum_8(const uint16_t *const src) {
      55           0 :     const __m128i s = _mm_loadu_si128((const __m128i *)src);
      56           0 :     return dc_sum_8x16bit(s);
      57             : }
      58             : 
      59             : #endif // EbHighbdIntraPrediction_SSE2_h