MMX in intra 8x8 prediction replaced with SSE2

8x8 Intra predictor implemented with MMX is replaced with SSE2.

Change-Id: I0c90e7c1e1e6942489ac2bfe58903b728aac7a52

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 93f821e66d9619cc1c7a6b5f546b9c351210c784..b1daa3bb93da1a4a6402ea18d7422b94d3761c3f 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -236,16 +236,11 @@ INTRA_PRED_TEST(C, TestIntraPred8, vpx_dc_predictor_8x8_c,
                 vpx_d153_predictor_8x8_c, vpx_d207_predictor_8x8_c,
                 vpx_d63_predictor_8x8_c, vpx_tm_predictor_8x8_c)
 
-#if HAVE_SSE && CONFIG_USE_X86INC
-INTRA_PRED_TEST(SSE, TestIntraPred8, vpx_dc_predictor_8x8_sse,
-                vpx_dc_left_predictor_8x8_sse, vpx_dc_top_predictor_8x8_sse,
-                vpx_dc_128_predictor_8x8_sse, vpx_v_predictor_8x8_sse, NULL,
-                NULL, NULL, NULL, NULL, NULL, NULL, NULL)
-#endif  // HAVE_SSE && CONFIG_USE_X86INC
-
 #if HAVE_SSE2 && CONFIG_USE_X86INC
-INTRA_PRED_TEST(SSE2, TestIntraPred8, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
-                NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_8x8_sse2)
+INTRA_PRED_TEST(SSE2, TestIntraPred8, vpx_dc_predictor_8x8_sse2,
+                vpx_dc_left_predictor_8x8_sse2, vpx_dc_top_predictor_8x8_sse2,
+                vpx_dc_128_predictor_8x8_sse2, vpx_v_predictor_8x8_sse2, NULL,
+                NULL, NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_8x8_sse2)
 #endif  // HAVE_SSE2 && CONFIG_USE_X86INC
 
 #if HAVE_SSSE3 && CONFIG_USE_X86INC

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index d2ea2b3eeb748084bba3c079e6093950df157bb4..a05cf8fa79722ab1c628b8a7d9d9f2a82e7af19f 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -142,22 +142,22 @@ add_proto qw/void vpx_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vpx_d153_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vpx_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_v_predictor_8x8 neon msa/, "$sse_x86inc";
+specialize qw/vpx_v_predictor_8x8 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa/, "$sse_x86inc";
+specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_dc_top_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_top_predictor_8x8 neon msa/, "$sse_x86inc";
+specialize qw/vpx_dc_top_predictor_8x8 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_dc_left_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_left_predictor_8x8 neon msa/, "$sse_x86inc";
+specialize qw/vpx_dc_left_predictor_8x8 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_dc_128_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_128_predictor_8x8 neon msa/, "$sse_x86inc";
+specialize qw/vpx_dc_128_predictor_8x8 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_d207_predictor_16x16/, "$ssse3_x86inc";

diff --git a/vpx_dsp/x86/intrapred_sse2.asm b/vpx_dsp/x86/intrapred_sse2.asm
index 08490b81287f27b53a0848e4a7fd9e30936abcf2..c3b7d2b7ef57fe1d43e001951ace79f0ddeb3eae 100644 (file)
--- a/vpx_dsp/x86/intrapred_sse2.asm
+++ b/vpx_dsp/x86/intrapred_sse2.asm
@@ -85,7 +85,7 @@ cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset
   RESTORE_GOT
   RET
 
-INIT_MMX sse
+INIT_XMM sse2
 cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
@@ -99,8 +99,8 @@ cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
   paddw                 m0, m2
   paddw                 m0, [GLOBAL(pw_8)]
   psraw                 m0, 4
-  pshufw                m0, m0, 0x0
-  packuswb              m0, m0
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
   movq    [dstq          ], m0
   movq    [dstq+strideq  ], m0
   movq    [dstq+strideq*2], m0
@@ -114,8 +114,8 @@ cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
   RESTORE_GOT
   RET
 
-INIT_MMX sse
-cglobal dc_top_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+INIT_XMM sse2
+cglobal dc_top_predictor_8x8, 3, 5, 2, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
   pxor                  m1, m1
@@ -125,8 +125,8 @@ cglobal dc_top_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
   psadbw                m0, m1
   paddw                 m0, [GLOBAL(pw2_8)]
   psraw                 m0, 3
-  pshufw                m0, m0, 0x0
-  packuswb              m0, m0
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
   movq    [dstq          ], m0
   movq    [dstq+strideq  ], m0
   movq    [dstq+strideq*2], m0
@@ -140,10 +140,11 @@ cglobal dc_top_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
   RESTORE_GOT
   RET
 
-INIT_MMX sse
-cglobal dc_left_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+INIT_XMM sse2
+cglobal dc_left_predictor_8x8, 2, 5, 2, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
+  movifnidn          leftq, leftmp
   pxor                  m1, m1
   movq                  m0, [leftq]
   DEFINE_ARGS dst, stride, stride3
@@ -151,8 +152,8 @@ cglobal dc_left_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
   psadbw                m0, m1
   paddw                 m0, [GLOBAL(pw2_8)]
   psraw                 m0, 3
-  pshufw                m0, m0, 0x0
-  packuswb              m0, m0
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
   movq    [dstq          ], m0
   movq    [dstq+strideq  ], m0
   movq    [dstq+strideq*2], m0
@@ -180,8 +181,8 @@ cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset
   RESTORE_GOT
   RET
 
-INIT_MMX sse
-cglobal dc_128_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+INIT_XMM sse2
+cglobal dc_128_predictor_8x8, 2, 5, 1, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
   DEFINE_ARGS dst, stride, stride3
@@ -463,7 +464,7 @@ cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
   movd      [dstq+strideq], m0
   RET
 
-INIT_MMX sse
+INIT_XMM sse2
 cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above
   movq                  m0, [aboveq]
   DEFINE_ARGS dst, stride, stride3