SSE2 speed up of h_predictor_4x4

Relocate h_predictor_4x4 from SSSE3 to SSE2 with XMM registers.
Speed up by ~25% in ./test_intra_pred_speed.

Change-Id: I64e14c13b482a471449be3559bfb0da45cf88d9d

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index d44a64a..58d1097 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -195,15 +195,16 @@ INTRA_PRED_TEST(SSE, TestIntraPred4, vpx_dc_predictor_4x4_sse,
 #endif  // HAVE_SSE && CONFIG_USE_X86INC
 
 #if HAVE_SSE2 && CONFIG_USE_X86INC
-INTRA_PRED_TEST(SSE2, TestIntraPred4, NULL, NULL, NULL, NULL, NULL, NULL,
-                NULL, NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_4x4_sse2)
+INTRA_PRED_TEST(SSE2, TestIntraPred4, NULL, NULL, NULL, NULL, NULL,
+                vpx_h_predictor_4x4_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
+                vpx_tm_predictor_4x4_sse2)
 #endif  // HAVE_SSE2 && CONFIG_USE_X86INC
 
 #if HAVE_SSSE3 && CONFIG_USE_X86INC
 INTRA_PRED_TEST(SSSE3, TestIntraPred4, NULL, NULL, NULL, NULL, NULL,
-                vpx_h_predictor_4x4_ssse3, vpx_d45_predictor_4x4_ssse3, NULL,
-                NULL, vpx_d153_predictor_4x4_ssse3,
-                vpx_d207_predictor_4x4_ssse3, vpx_d63_predictor_4x4_ssse3, NULL)
+                NULL, vpx_d45_predictor_4x4_ssse3, NULL, NULL,
+                vpx_d153_predictor_4x4_ssse3, vpx_d207_predictor_4x4_ssse3,
+                vpx_d63_predictor_4x4_ssse3, NULL)
 #endif  // HAVE_SSSE3 && CONFIG_USE_X86INC
 
 #if HAVE_DSPR2

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 4de85a4..d11b327 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -76,7 +76,7 @@ add_proto qw/void vpx_d63f_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vpx_d63f_predictor_4x4/;
 
 add_proto qw/void vpx_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_h_predictor_4x4 neon dspr2 msa/, "$ssse3_x86inc";
+specialize qw/vpx_h_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_he_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_he_predictor_4x4/;

diff --git a/vpx_dsp/x86/intrapred_sse2.asm b/vpx_dsp/x86/intrapred_sse2.asm
index 04b39a5..5ef7ae3 100644 (file)
--- a/vpx_dsp/x86/intrapred_sse2.asm
+++ b/vpx_dsp/x86/intrapred_sse2.asm
@@ -516,6 +516,22 @@ cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
   REP_RET
 
 INIT_XMM sse2
+cglobal h_predictor_4x4, 2, 4, 4, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0
+  pshufd                m1, m0, 0x1
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m1
+  pshufd                m2, m0, 0x2
+  lea                 dstq, [dstq+strideq*2]
+  pshufd                m3, m0, 0x3
+  movd      [dstq        ], m2
+  movd      [dstq+strideq], m3
+  RET
+
+INIT_XMM sse2
 cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left
   pxor                  m1, m1
   movq                  m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x

diff --git a/vpx_dsp/x86/intrapred_ssse3.asm b/vpx_dsp/x86/intrapred_ssse3.asm
index 88df9b2..744c40c 100644 (file)
--- a/vpx_dsp/x86/intrapred_ssse3.asm
+++ b/vpx_dsp/x86/intrapred_ssse3.asm
@@ -34,24 +34,6 @@ sh_b2333: db 2, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
 SECTION .text
 
 INIT_MMX ssse3
-cglobal h_predictor_4x4, 2, 4, 3, dst, stride, line, left
-  movifnidn          leftq, leftmp
-  add                leftq, 4
-  mov                lineq, -2
-  pxor                  m0, m0
-.loop:
-  movd                  m1, [leftq+lineq*2  ]
-  movd                  m2, [leftq+lineq*2+1]
-  pshufb                m1, m0
-  pshufb                m2, m0
-  movd      [dstq        ], m1
-  movd      [dstq+strideq], m2
-  lea                 dstq, [dstq+strideq*2]
-  inc                lineq
-  jnz .loop
-  REP_RET
-
-INIT_MMX ssse3
 cglobal h_predictor_8x8, 2, 4, 3, dst, stride, line, left
   movifnidn          leftq, leftmp
   add                leftq, 8