PR Netflix#1475

Author: StormBytePP

libvmaf/src/feature/x86/adm_avx2.c

@@ -819,12 +819,12 @@ void adm_decouple_avx2(AdmBuffer *buf, int w, int h, int stride,
             __m256 od_inv_64 = _mm256_mul_ps(inv_64, _mm256_cvtepi32_ps(od));
             __m256 rst_d_f = _mm256_mul_ps(kd_inv_32768, od_inv_64);
 
-            __m256i gt0_rst_h_f = (__m256i)(_mm256_cmp_ps(rst_h_f, _mm256_setzero_ps(), 14));
-            __m256i lt0_rst_h_f = (__m256i)(_mm256_cmp_ps(rst_h_f, _mm256_setzero_ps(), 1));
-            __m256i gt0_rst_v_f = (__m256i)(_mm256_cmp_ps(rst_v_f, _mm256_setzero_ps(), 14));
-            __m256i lt0_rst_v_f = (__m256i)(_mm256_cmp_ps(rst_v_f, _mm256_setzero_ps(), 1));
-            __m256i gt0_rst_d_f = (__m256i)(_mm256_cmp_ps(rst_d_f, _mm256_setzero_ps(), 14));
-            __m256i lt0_rst_d_f = (__m256i)(_mm256_cmp_ps(rst_d_f, _mm256_setzero_ps(), 1));
+            __m256i gt0_rst_h_f = _mm256_castps_si256(_mm256_cmp_ps(rst_h_f, _mm256_setzero_ps(), 14));
+            __m256i lt0_rst_h_f = _mm256_castps_si256(_mm256_cmp_ps(rst_h_f, _mm256_setzero_ps(), 1));
+            __m256i gt0_rst_v_f = _mm256_castps_si256(_mm256_cmp_ps(rst_v_f, _mm256_setzero_ps(), 14));
+            __m256i lt0_rst_v_f = _mm256_castps_si256(_mm256_cmp_ps(rst_v_f, _mm256_setzero_ps(), 1));
+            __m256i gt0_rst_d_f = _mm256_castps_si256(_mm256_cmp_ps(rst_d_f, _mm256_setzero_ps(), 14));
+            __m256i lt0_rst_d_f = _mm256_castps_si256(_mm256_cmp_ps(rst_d_f, _mm256_setzero_ps(), 1));
 
             __m256i mask_min_max_h = _mm256_or_si256(gt0_rst_h_f, lt0_rst_h_f);
             __m256i mask_min_max_v = _mm256_or_si256(gt0_rst_v_f, lt0_rst_v_f);
@@ -834,7 +834,7 @@ void adm_decouple_avx2(AdmBuffer *buf, int w, int h, int stride,
             __m256i mask_rst_v = _mm256_and_si256(mask_min_max_v, angle_flag);
             __m256i mask_rst_d = _mm256_and_si256(mask_min_max_d, angle_flag);
 
-	    __m256d adm_gain_d = _mm256_set1_pd(adm_enhn_gain_limit);
+	        __m256d adm_gain_d = _mm256_set1_pd(adm_enhn_gain_limit);
             __m256d rst_h_gainlo_d = _mm256_mul_pd(_mm256_cvtepi32_pd(_mm256_extractf128_si256(rst_h, 0)), adm_gain_d);
             __m256d rst_h_gainhi_d = _mm256_mul_pd(_mm256_cvtepi32_pd(_mm256_extractf128_si256(rst_h, 1)), adm_gain_d);
             __m256i rst_h_gain = _mm256_insertf128_si256(_mm256_castsi128_si256(_mm256_cvtpd_epi32(rst_h_gainlo_d)), _mm256_cvtpd_epi32(rst_h_gainhi_d),1);
@@ -2088,15 +2088,15 @@ float adm_cm_avx2(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_stride
             }
             accum_inner_h_lo_256 = _mm256_add_epi64(accum_inner_h_lo_256, accum_inner_h_hi_256);
             __m128i r2_h = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_h_lo_256), _mm256_extracti128_si256(accum_inner_h_lo_256, 1));
-            int64_t res_h = r2_h[0] + r2_h[1];
+            int64_t res_h = _mm_extract_epi64(r2_h, 0) + _mm_extract_epi64(r2_h, 1);
 
             accum_inner_v_lo_256 = _mm256_add_epi64(accum_inner_v_lo_256, accum_inner_v_hi_256);
             __m128i r2_v = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_v_lo_256), _mm256_extracti128_si256(accum_inner_v_lo_256, 1));
-            int64_t res_v = r2_v[0] + r2_v[1];
+            int64_t res_v = _mm_extract_epi64(r2_v, 0) + _mm_extract_epi64(r2_v, 1);
 
             accum_inner_d_lo_256 = _mm256_add_epi64(accum_inner_d_lo_256, accum_inner_d_hi_256);
             __m128i r2_d = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_d_lo_256), _mm256_extracti128_si256(accum_inner_d_lo_256, 1));
-            int64_t res_d = r2_d[0] + r2_d[1];
+            int64_t res_d = _mm_extract_epi64(r2_d, 0) + _mm_extract_epi64(r2_d, 1);
 
             for (j = end_col_mod6; j < end_col; ++j) {
                 xh = src->band_h[i * src_stride + j] * i_rfactor[0];
@@ -2521,13 +2521,13 @@ float i4_adm_cm_avx2(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_str
             }
 
             __m128i r2_h = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_h_256), _mm256_extracti128_si256(accum_inner_h_256, 1));
-            int64_t res_h = r2_h[0] + r2_h[1];
+            int64_t res_h = _mm_extract_epi64(r2_h, 0) + _mm_extract_epi64(r2_h, 1);
 
             __m128i r2_v = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_v_256), _mm256_extracti128_si256(accum_inner_v_256, 1));
-            int64_t res_v = r2_v[0] + r2_v[1];
+            int64_t res_v = _mm_extract_epi64(r2_v, 0) + _mm_extract_epi64(r2_v, 1);
 
             __m128i r2_d = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_d_256), _mm256_extracti128_si256(accum_inner_d_256, 1));
-            int64_t res_d = r2_d[0] + r2_d[1];
+            int64_t res_d = _mm_extract_epi64(r2_d, 0) + _mm_extract_epi64(r2_d, 1);
 
             for (j = end_col_mod2; j < end_col; ++j)
             {
@@ -3586,15 +3586,15 @@ float adm_csf_den_scale_avx2(const adm_dwt_band_t *src, int w, int h,
 
         accum_inner_h_lo = _mm256_add_epi64(accum_inner_h_lo, accum_inner_h_hi);
         __m128i h_r2 = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_h_lo), _mm256_extracti128_si256(accum_inner_h_lo, 1));
-        uint64_t h_r1 = h_r2[0] + h_r2[1];
+        uint64_t h_r1 = _mm_extract_epi64(h_r2, 0) + _mm_extract_epi64(h_r2, 1);
 
         accum_inner_v_lo = _mm256_add_epi64(accum_inner_v_lo, accum_inner_v_hi);
         __m128i v_r2 = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_v_lo), _mm256_extracti128_si256(accum_inner_v_lo, 1));
-        uint64_t v_r1 = v_r2[0] + v_r2[1];
+        uint64_t v_r1 = _mm_extract_epi64(v_r2, 0) + _mm_extract_epi64(v_r2, 1);
 
         accum_inner_d_lo = _mm256_add_epi64(accum_inner_d_lo, accum_inner_d_hi);
         __m128i d_r2 = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_d_lo), _mm256_extracti128_si256(accum_inner_d_lo, 1));
-        uint64_t d_r1 = d_r2[0] + d_r2[1];
+        uint64_t d_r1 = _mm_extract_epi64(d_r2, 0) + _mm_extract_epi64(d_r2, 1);
 
         for (int j = right_mod_8; j < right; ++j) {
             uint16_t h = (uint16_t)abs(src_h[j]);
@@ -3992,13 +3992,13 @@ float adm_csf_den_s123_avx2(const i4_adm_dwt_band_t *src, int scale, int w, int
             accum_inner_d_256 = _mm256_add_epi64(accum_inner_d_256, d_cu);
         }
         __m128i h_r2 = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_h_256), _mm256_extracti128_si256(accum_inner_h_256, 1));
-        uint64_t h_r1 = h_r2[0] + h_r2[1];
+        uint64_t h_r1 = _mm_extract_epi64(h_r2, 0) + _mm_extract_epi64(h_r2, 1);
 
         __m128i d_r2 = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_d_256), _mm256_extracti128_si256(accum_inner_d_256, 1));
-        uint64_t d_r1 = d_r2[0] + d_r2[1];
+        uint64_t d_r1 = _mm_extract_epi64(d_r2, 0) + _mm_extract_epi64(d_r2, 1);
 
         __m128i v_r2 = _mm_add_epi64(_mm256_castsi256_si128(accum_inner_v_256), _mm256_extracti128_si256(accum_inner_v_256, 1));
-        uint64_t v_r1 = v_r2[0] + v_r2[1];
+        uint64_t v_r1 = _mm_extract_epi64(v_r2, 0) + _mm_extract_epi64(v_r2, 1);
 
         for (int j = right_mod_4; j < right; ++j)
         {

libvmaf/src/feature/x86/adm_avx512.c

@@ -1746,17 +1746,17 @@ float adm_cm_avx512(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_stri
             accum_inner_h_lo_512 = _mm512_add_epi64(accum_inner_h_lo_512, accum_inner_h_hi_512);
             __m256i r4_h = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_h_lo_512), _mm512_extracti64x4_epi64(accum_inner_h_lo_512, 1));
             __m128i r2_h = _mm_add_epi64(_mm256_castsi256_si128(r4_h), _mm256_extracti128_si256(r4_h, 1));
-            int64_t res_h = r2_h[0] + r2_h[1];
+            int64_t res_h = _mm_extract_epi64(r2_h, 0) + _mm_extract_epi64(r2_h, 1);
 
             accum_inner_v_lo_512 = _mm512_add_epi64(accum_inner_v_lo_512, accum_inner_v_hi_512);
             __m256i r4_v = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_v_lo_512), _mm512_extracti64x4_epi64(accum_inner_v_lo_512, 1));
             __m128i r2_v = _mm_add_epi64(_mm256_castsi256_si128(r4_v), _mm256_extracti128_si256(r4_v, 1));
-            int64_t res_v = r2_v[0] + r2_v[1];
+            int64_t res_v = _mm_extract_epi64(r2_v, 0) + _mm_extract_epi64(r2_v, 1);
 
             accum_inner_d_lo_512 = _mm512_add_epi64(accum_inner_d_lo_512, accum_inner_d_hi_512);
             __m256i r4_d = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_d_lo_512), _mm512_extracti64x4_epi64(accum_inner_d_lo_512, 1));
             __m128i r2_d = _mm_add_epi64(_mm256_castsi256_si128(r4_d), _mm256_extracti128_si256(r4_d, 1));
-            int64_t res_d = r2_d[0] + r2_d[1];
+            int64_t res_d = _mm_extract_epi64(r2_d, 0) + _mm_extract_epi64(r2_d, 1);
 
             for (j = end_col_mod14; j < end_col; ++j) {
                 xh = src->band_h[i * src_stride + j] * i_rfactor[0];
@@ -2168,15 +2168,15 @@ float i4_adm_cm_avx512(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_s
 
             __m256i r4_h = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_h_512), _mm512_extracti64x4_epi64(accum_inner_h_512, 1));
             __m128i r2_h = _mm_add_epi64(_mm256_castsi256_si128(r4_h), _mm256_extracti128_si256(r4_h, 1));
-            int64_t res_h = r2_h[0] + r2_h[1];
+            int64_t res_h = _mm_extract_epi64(r2_h, 0) + _mm_extract_epi64(r2_h, 1);
 
             __m256i r4_v = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_v_512), _mm512_extracti64x4_epi64(accum_inner_v_512, 1));
             __m128i r2_v = _mm_add_epi64(_mm256_castsi256_si128(r4_v), _mm256_extracti128_si256(r4_v, 1));
-            int64_t res_v = r2_v[0] + r2_v[1];
+            int64_t res_v = _mm_extract_epi64(r2_v, 0) + _mm_extract_epi64(r2_v, 1);
 
             __m256i r4_d = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_d_512), _mm512_extracti64x4_epi64(accum_inner_d_512, 1));
             __m128i r2_d = _mm_add_epi64(_mm256_castsi256_si128(r4_d), _mm256_extracti128_si256(r4_d, 1));
-            int64_t res_d = r2_d[0] + r2_d[1];
+            int64_t res_d = _mm_extract_epi64(r2_d, 0) + _mm_extract_epi64(r2_d, 1);
 
             for (j = end_col_mod6; j < end_col; ++j)
             {
@@ -3858,17 +3858,17 @@ float adm_csf_den_scale_avx512(const adm_dwt_band_t *src, int w, int h,
         accum_inner_h_lo = _mm512_add_epi64(accum_inner_h_lo, accum_inner_h_hi);
         __m256i h_r4 = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_h_lo), _mm512_extracti64x4_epi64(accum_inner_h_lo, 1));
         __m128i h_r2 = _mm_add_epi64(_mm256_castsi256_si128(h_r4), _mm256_extracti64x2_epi64(h_r4, 1));
-        uint64_t h_r1 = h_r2[0] + h_r2[1];
+        uint64_t h_r1 = _mm_extract_epi64(h_r2, 0) + _mm_extract_epi64(h_r2, 1);
 
         accum_inner_v_lo = _mm512_add_epi64(accum_inner_v_lo, accum_inner_v_hi);
         __m256i v_r4 = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_v_lo), _mm512_extracti64x4_epi64(accum_inner_v_lo, 1));
         __m128i v_r2 = _mm_add_epi64(_mm256_castsi256_si128(v_r4), _mm256_extracti64x2_epi64(v_r4, 1));
-        uint64_t v_r1 = v_r2[0] + v_r2[1];
+        uint64_t v_r1 = _mm_extract_epi64(v_r2, 0) + _mm_extract_epi64(v_r2, 1);
 
         accum_inner_d_lo = _mm512_add_epi64(accum_inner_d_lo, accum_inner_d_hi);
         __m256i d_r4 = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_d_lo), _mm512_extracti64x4_epi64(accum_inner_d_lo, 1));
         __m128i d_r2 = _mm_add_epi64(_mm256_castsi256_si128(d_r4), _mm256_extracti64x2_epi64(d_r4, 1));
-        uint64_t d_r1 = d_r2[0] + d_r2[1];
+        uint64_t d_r1 = _mm_extract_epi64(d_r2, 0) + _mm_extract_epi64(d_r2, 1);
 
         for (int j = right_mod_16; j < right; ++j) {
             uint16_t h = (uint16_t)abs(src_h[j]);
@@ -3985,15 +3985,15 @@ float adm_csf_den_s123_avx512(const i4_adm_dwt_band_t *src, int scale, int w, in
         }
         __m256i h_r4 = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_h_512), _mm512_extracti64x4_epi64(accum_inner_h_512, 1));
         __m128i h_r2 = _mm_add_epi64(_mm256_castsi256_si128(h_r4), _mm256_extracti64x2_epi64(h_r4, 1));
-        uint64_t h_r1 = h_r2[0] + h_r2[1];
+        uint64_t h_r1 = _mm_extract_epi64(h_r2, 0) + _mm_extract_epi64(h_r2, 1);
 
         __m256i d_r4 = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_d_512), _mm512_extracti64x4_epi64(accum_inner_d_512, 1));
         __m128i d_r2 = _mm_add_epi64(_mm256_castsi256_si128(d_r4), _mm256_extracti64x2_epi64(d_r4, 1));
-        uint64_t d_r1 = d_r2[0] + d_r2[1];
+        uint64_t d_r1 = _mm_extract_epi64(d_r2, 0) + _mm_extract_epi64(d_r2, 1);
 
         __m256i v_r4 = _mm256_add_epi64(_mm512_castsi512_si256(accum_inner_v_512), _mm512_extracti64x4_epi64(accum_inner_v_512, 1));
         __m128i v_r2 = _mm_add_epi64(_mm256_castsi256_si128(v_r4), _mm256_extracti64x2_epi64(v_r4, 1));
-        uint64_t v_r1 = v_r2[0] + v_r2[1];
+        uint64_t v_r1 = _mm_extract_epi64(v_r2, 0) + _mm_extract_epi64(v_r2, 1);
 
         for (int j = right_mod_8; j < right; ++j)
         {

libvmaf/src/feature/x86/motion_avx2.c

@@ -534,7 +534,8 @@ void sad_avx2(VmafPicture *pic_a, VmafPicture *pic_b, uint64_t *sad)
         a += (pic_a->stride[0] / 2);
         b += (pic_b->stride[0] / 2);
     }
-    uint64_t r1 = final_accum[0] + final_accum[1] + final_accum[2] + final_accum[3];
+    uint64_t r1 = _mm256_extract_epi64(final_accum, 0) + _mm256_extract_epi64(final_accum, 1) +
+                  _mm256_extract_epi64(final_accum, 2) + _mm256_extract_epi64(final_accum, 3);
 
     *sad += r1;
 }

libvmaf/src/feature/x86/motion_avx512.c

@@ -450,7 +450,7 @@ void sad_avx512(VmafPicture *pic_a, VmafPicture *pic_b, uint64_t *sad)
     }
     __m256i r4 = _mm256_add_epi64(_mm512_castsi512_si256(final_accum), _mm512_extracti64x4_epi64(final_accum, 1));
     __m128i r2 = _mm_add_epi64(_mm256_castsi256_si128(r4), _mm256_extracti64x2_epi64(r4, 1));
-    uint64_t r1 = r2[0] + r2[1];
+    uint64_t r1 = _mm_extract_epi64(r2, 0) + _mm_extract_epi64(r2, 1);
 
     *sad += r1;