Merge "WORKAROUND FIX FOR GCC4.9.1"

2 files changed, 84 insertions, 35 deletions

diff --git a/test/lpf_8_test.cc b/test/lpf_8_test.cc
index abc410798..b9f3d3d5a 100644
--- a/test/lpf_8_test.cc
+++ b/test/lpf_8_test.cc
@@ -522,6 +522,13 @@ INSTANTIATE_TEST_CASE_P(
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 #endif
 
+#if HAVE_AVX2 && (!CONFIG_VP9_HIGHBITDEPTH)
+INSTANTIATE_TEST_CASE_P(
+    AVX2_C_COMPARE_SINGLE, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&vp9_lpf_horizontal_16_avx2, &vp9_lpf_horizontal_16_c, 8)));
+#endif
+
 #if HAVE_SSE2
 #if CONFIG_VP9_HIGHBITDEPTH
 INSTANTIATE_TEST_CASE_P(
@@ -584,4 +591,5 @@ INSTANTIATE_TEST_CASE_P(
                    &vp9_lpf_vertical_8_dual_c, 8)));
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 #endif
+
 }  // namespace
diff --git a/vp9/common/x86/vp9_loopfilter_intrin_avx2.c b/vp9/common/x86/vp9_loopfilter_intrin_avx2.c
index 439c028f2..0cb0912ad 100644
--- a/vp9/common/x86/vp9_loopfilter_intrin_avx2.c
+++ b/vp9/common/x86/vp9_loopfilter_intrin_avx2.c
@@ -9,6 +9,7 @@
  */
 
 #include <immintrin.h>  /* AVX2 */
+#include "vpx_ports/mem.h"
 
 static void mb_lpf_horizontal_edge_w_avx2_8(unsigned char *s, int p,
         const unsigned char *_blimit, const unsigned char *_limit,
@@ -392,6 +393,11 @@ static void mb_lpf_horizontal_edge_w_avx2_8(unsigned char *s, int p,
     }
 }
 
+DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {
+  0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128,
+  8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
+};
+
 static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
         const unsigned char *_blimit, const unsigned char *_limit,
         const unsigned char *_thresh) {
@@ -401,6 +407,9 @@ static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
     __m128i p7, p6, p5;
     __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
     __m128i q5, q6, q7;
+    __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4,
+            q256_4, p256_3, q256_3, p256_2, q256_2, p256_1, q256_1,
+            p256_0, q256_0;
 
     const __m128i thresh = _mm_broadcastb_epi8(
             _mm_cvtsi32_si128((int) _thresh[0]));
@@ -409,16 +418,37 @@ static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
     const __m128i blimit = _mm_broadcastb_epi8(
             _mm_cvtsi32_si128((int) _blimit[0]));
 
-    p4 = _mm_loadu_si128((__m128i *) (s - 5 * p));
-    p3 = _mm_loadu_si128((__m128i *) (s - 4 * p));
-    p2 = _mm_loadu_si128((__m128i *) (s - 3 * p));
-    p1 = _mm_loadu_si128((__m128i *) (s - 2 * p));
-    p0 = _mm_loadu_si128((__m128i *) (s - 1 * p));
-    q0 = _mm_loadu_si128((__m128i *) (s - 0 * p));
-    q1 = _mm_loadu_si128((__m128i *) (s + 1 * p));
-    q2 = _mm_loadu_si128((__m128i *) (s + 2 * p));
-    q3 = _mm_loadu_si128((__m128i *) (s + 3 * p));
-    q4 = _mm_loadu_si128((__m128i *) (s + 4 * p));
+    p256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 5 * p)));
+    p256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 4 * p)));
+    p256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 3 * p)));
+    p256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 2 * p)));
+    p256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 1 * p)));
+    q256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 0 * p)));
+    q256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 1 * p)));
+    q256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 2 * p)));
+    q256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 3 * p)));
+    q256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 4 * p)));
+
+    p4 = _mm256_castsi256_si128(p256_4);
+    p3 = _mm256_castsi256_si128(p256_3);
+    p2 = _mm256_castsi256_si128(p256_2);
+    p1 = _mm256_castsi256_si128(p256_1);
+    p0 = _mm256_castsi256_si128(p256_0);
+    q0 = _mm256_castsi256_si128(q256_0);
+    q1 = _mm256_castsi256_si128(q256_1);
+    q2 = _mm256_castsi256_si128(q256_2);
+    q3 = _mm256_castsi256_si128(q256_3);
+    q4 = _mm256_castsi256_si128(q256_4);
 
     {
         const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
@@ -534,23 +564,35 @@ static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
             flat = _mm_cmpeq_epi8(flat, zero);
             flat = _mm_and_si128(flat, mask);
 
-            p5 = _mm_loadu_si128((__m128i *) (s - 6 * p));
-            q5 = _mm_loadu_si128((__m128i *) (s + 5 * p));
+            p256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 6 * p)));
+            q256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 5 * p)));
+            p5 = _mm256_castsi256_si128(p256_5);
+            q5 = _mm256_castsi256_si128(q256_5);
             flat2 = _mm_max_epu8(
                     _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
                     _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
 
             flat2 = _mm_max_epu8(work, flat2);
-            p6 = _mm_loadu_si128((__m128i *) (s - 7 * p));
-            q6 = _mm_loadu_si128((__m128i *) (s + 6 * p));
+            p256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 7 * p)));
+            q256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 6 * p)));
+            p6 = _mm256_castsi256_si128(p256_6);
+            q6 = _mm256_castsi256_si128(q256_6);
             work = _mm_max_epu8(
                     _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
                     _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
 
             flat2 = _mm_max_epu8(work, flat2);
 
-            p7 = _mm_loadu_si128((__m128i *) (s - 8 * p));
-            q7 = _mm_loadu_si128((__m128i *) (s + 7 * p));
+            p256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 8 * p)));
+            q256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 7 * p)));
+            p7 = _mm256_castsi256_si128(p256_7);
+            q7 = _mm256_castsi256_si128(q256_7);
             work = _mm_max_epu8(
                     _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
                     _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
@@ -566,29 +608,28 @@ static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
         {
             const __m256i eight = _mm256_set1_epi16(8);
             const __m256i four = _mm256_set1_epi16(4);
-            __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4,
-                    q256_4, p256_3, q256_3, p256_2, q256_2, p256_1, q256_1,
-                    p256_0, q256_0;
             __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
                     pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p,
                     res_q;
 
-            p256_7 = _mm256_cvtepu8_epi16(p7);
-            p256_6 = _mm256_cvtepu8_epi16(p6);
-            p256_5 = _mm256_cvtepu8_epi16(p5);
-            p256_4 = _mm256_cvtepu8_epi16(p4);
-            p256_3 = _mm256_cvtepu8_epi16(p3);
-            p256_2 = _mm256_cvtepu8_epi16(p2);
-            p256_1 = _mm256_cvtepu8_epi16(p1);
-            p256_0 = _mm256_cvtepu8_epi16(p0);
-            q256_0 = _mm256_cvtepu8_epi16(q0);
-            q256_1 = _mm256_cvtepu8_epi16(q1);
-            q256_2 = _mm256_cvtepu8_epi16(q2);
-            q256_3 = _mm256_cvtepu8_epi16(q3);
-            q256_4 = _mm256_cvtepu8_epi16(q4);
-            q256_5 = _mm256_cvtepu8_epi16(q5);
-            q256_6 = _mm256_cvtepu8_epi16(q6);
-            q256_7 = _mm256_cvtepu8_epi16(q7);
+            const __m256i filter = _mm256_load_si256(
+                                  (__m256i const *)filt_loopfilter_avx2);
+            p256_7 = _mm256_shuffle_epi8(p256_7, filter);
+            p256_6 = _mm256_shuffle_epi8(p256_6, filter);
+            p256_5 = _mm256_shuffle_epi8(p256_5, filter);
+            p256_4 = _mm256_shuffle_epi8(p256_4, filter);
+            p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+            p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+            p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+            p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+            q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+            q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+            q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+            q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+            q256_4 = _mm256_shuffle_epi8(q256_4, filter);
+            q256_5 = _mm256_shuffle_epi8(q256_5, filter);
+            q256_6 = _mm256_shuffle_epi8(q256_6, filter);
+            q256_7 = _mm256_shuffle_epi8(q256_7, filter);
 
             pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
                     _mm256_add_epi16(p256_4, p256_3));