Enable sse2 version of sad8x4/4x8

The encoding time for bus at CIF goes from 661s to 625s. This commit
also enabled unit test of sad8x4/4x8 in sad_test.cc.

Change-Id: If3d10ebb56bda584bdb69bcf056599d580b12cb1

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 9555a9a..7698c1e 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -308,6 +308,8 @@ const sad_m_by_n_fn_t sad_16x16_c_vp9 = vp9_sad16x16_c;
 const sad_m_by_n_fn_t sad_8x16_c_vp9 = vp9_sad8x16_c;
 const sad_m_by_n_fn_t sad_16x8_c_vp9 = vp9_sad16x8_c;
 const sad_m_by_n_fn_t sad_8x8_c_vp9 = vp9_sad8x8_c;
+const sad_m_by_n_fn_t sad_8x4_c_vp9 = vp9_sad8x4_c;
+const sad_m_by_n_fn_t sad_4x8_c_vp9 = vp9_sad4x8_c;
 const sad_m_by_n_fn_t sad_4x4_c_vp9 = vp9_sad4x4_c;
 #endif
 const sad_m_by_n_test_param_t c_tests[] = {
@@ -325,6 +327,8 @@ const sad_m_by_n_test_param_t c_tests[] = {
   make_tuple(8, 16, sad_8x16_c_vp9),
   make_tuple(16, 8, sad_16x8_c_vp9),
   make_tuple(8, 8, sad_8x8_c_vp9),
+  make_tuple(8, 4, sad_8x4_c_vp9),
+  make_tuple(4, 8, sad_4x8_c_vp9),
   make_tuple(4, 4, sad_4x4_c_vp9),
 #endif
 };
@@ -404,8 +408,10 @@ INSTANTIATE_TEST_CASE_P(MMX, SADTest, ::testing::ValuesIn(mmx_tests));
 #if HAVE_SSE
 #if CONFIG_VP9_ENCODER
 const sad_m_by_n_fn_t sad_4x4_sse_vp9 = vp9_sad4x4_sse;
+const sad_m_by_n_fn_t sad_4x8_sse_vp9 = vp9_sad4x8_sse;
 INSTANTIATE_TEST_CASE_P(SSE, SADTest, ::testing::Values(
-                        make_tuple(4, 4, sad_4x4_sse_vp9)));
+                        make_tuple(4, 4, sad_4x4_sse_vp9),
+                        make_tuple(4, 8, sad_4x8_sse_vp9)));
 
 const sad_n_by_n_by_4_fn_t sad_4x4x4d_sse = vp9_sad4x4x4d_sse;
 INSTANTIATE_TEST_CASE_P(SSE, SADx4Test, ::testing::Values(
@@ -428,6 +434,7 @@ const sad_m_by_n_fn_t sad_16x16_sse2_vp9 = vp9_sad16x16_sse2;
 const sad_m_by_n_fn_t sad_8x16_sse2_vp9 = vp9_sad8x16_sse2;
 const sad_m_by_n_fn_t sad_16x8_sse2_vp9 = vp9_sad16x8_sse2;
 const sad_m_by_n_fn_t sad_8x8_sse2_vp9 = vp9_sad8x8_sse2;
+const sad_m_by_n_fn_t sad_8x4_sse2_vp9 = vp9_sad8x4_sse2;
 #endif
 const sad_m_by_n_test_param_t sse2_tests[] = {
 #if CONFIG_VP8_ENCODER
@@ -444,6 +451,7 @@ const sad_m_by_n_test_param_t sse2_tests[] = {
   make_tuple(8, 16, sad_8x16_sse2_vp9),
   make_tuple(16, 8, sad_16x8_sse2_vp9),
   make_tuple(8, 8, sad_8x8_sse2_vp9),
+  make_tuple(8, 4, sad_8x4_sse2_vp9),
 #endif
 };
 INSTANTIATE_TEST_CASE_P(SSE2, SADTest, ::testing::ValuesIn(sse2_tests));

diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index f281e08..8c78b7b 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -387,10 +387,10 @@ specialize vp9_sad8x8 mmx sse2
 
 # TODO(jingning): need to covert these functions into mmx/sse2 form
 prototype unsigned int vp9_sad8x4 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad"
-specialize vp9_sad8x4
+specialize vp9_sad8x4 sse2
 
 prototype unsigned int vp9_sad4x8 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad"
-specialize vp9_sad4x8
+specialize vp9_sad4x8 sse
 
 prototype unsigned int vp9_sad4x4 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad"
 specialize vp9_sad4x4 mmx sse

diff --git a/vp9/encoder/x86/vp9_sad_sse2.asm b/vp9/encoder/x86/vp9_sad_sse2.asm
index ea92377..8fb7d41 100644 (file)
--- a/vp9/encoder/x86/vp9_sad_sse2.asm
+++ b/vp9/encoder/x86/vp9_sad_sse2.asm
@@ -166,29 +166,46 @@ cglobal sad8x%1, 4, 7, 5, src, src_stride, ref, ref_stride, \
 INIT_XMM sse2
 SAD8XN 16 ; sad8x16_sse2
 SAD8XN  8 ; sad8x8_sse2
+SAD8XN  4 ; sad8x4_sse2
 
-; unsigned int vp9_sad4x4_sse(uint8_t *src, int src_stride,
-;                             uint8_t *ref, int ref_stride);
-INIT_MMX sse
-cglobal sad4x4, 4, 4, 8, src, src_stride, ref, ref_stride
+; unsigned int vp9_sad4x{4, 8}_sse(uint8_t *src, int src_stride,
+;                                  uint8_t *ref, int ref_stride);
+%macro SAD4XN 1
+cglobal sad4x%1, 4, 7, 7, src, src_stride, ref, ref_stride, \
+                          src_stride3, ref_stride3, n_rows
   movsxdifnidn src_strideq, src_strided
   movsxdifnidn ref_strideq, ref_strided
-  movd                  m0, [refq]
-  movd                  m1, [refq+ref_strideq]
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movd                  m1, [refq]
+  movd                  m2, [refq+ref_strideq]
+  movd                  m3, [refq+ref_strideq*2]
+  movd                  m4, [refq+ref_stride3q]
+  punpckldq             m1, m2
+  punpckldq             m3, m4
   movd                  m2, [srcq]
-  movd                  m3, [srcq+src_strideq]
-  lea                 refq, [refq+ref_strideq*2]
-  lea                 srcq, [srcq+src_strideq*2]
-  movd                  m4, [refq]
-  movd                  m5, [refq+ref_strideq]
-  movd                  m6, [srcq]
-  movd                  m7, [srcq+src_strideq]
-  punpckldq             m0, m1
-  punpckldq             m2, m3
-  punpckldq             m4, m5
-  punpckldq             m6, m7
-  psadbw                m0, m2
-  psadbw                m4, m6
-  paddd                 m0, m4
+  movd                  m5, [srcq+src_strideq]
+  movd                  m4, [srcq+src_strideq*2]
+  movd                  m6, [srcq+src_stride3q]
+  punpckldq             m2, m5
+  punpckldq             m4, m6
+  psadbw                m1, m2
+  psadbw                m3, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
   movd                 eax, m0
   RET
+%endmacro
+
+INIT_MMX sse
+SAD4XN  8 ; sad4x8_sse
+SAD4XN  4 ; sad4x4_sse