Speed up tm_predictor_4x4

tm_predictor_4x4 is implemented with SSE2 using XMM registers.
Speed up by ~25% in ./test_intra_pred_speed.

Change-Id: I25074b78d476a2cb17f81cf654bdfd80df2070e0

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 5d59e83f751cd97618a124a79546c6d837f1bb01..d44a64a0b0bbe24eff351f98a3aa405c0dff8dfb 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -191,9 +191,14 @@ INTRA_PRED_TEST(C, TestIntraPred4, vpx_dc_predictor_4x4_c,
 INTRA_PRED_TEST(SSE, TestIntraPred4, vpx_dc_predictor_4x4_sse,
                 vpx_dc_left_predictor_4x4_sse, vpx_dc_top_predictor_4x4_sse,
                 vpx_dc_128_predictor_4x4_sse, vpx_v_predictor_4x4_sse, NULL,
-                NULL, NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_4x4_sse)
+                NULL, NULL, NULL, NULL, NULL, NULL, NULL)
 #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)
+#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,

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index b369b054819f2b489a72ebef8f59f5f039c4915e..03dcd4eeb08b48fab23e8c3ebc58b32f1e24f4c5 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -97,7 +97,7 @@ add_proto qw/void vpx_ve_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, cons
 specialize qw/vpx_ve_predictor_4x4/;
 
 add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse_x86inc";
+specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse_x86inc";

diff --git a/vpx_dsp/x86/intrapred_sse2.asm b/vpx_dsp/x86/intrapred_sse2.asm
index 22b5731886c530979e03716d354ea01b8308ca37..04b39a5839652a15f673462ae3273c53388f1122 100644 (file)
--- a/vpx_dsp/x86/intrapred_sse2.asm
+++ b/vpx_dsp/x86/intrapred_sse2.asm
@@ -515,35 +515,34 @@ cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
   jnz .loop
   REP_RET
 
-INIT_MMX sse
-cglobal tm_predictor_4x4, 4, 4, 4, dst, stride, above, left
+INIT_XMM sse2
+cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left
   pxor                  m1, m1
-  movd                  m2, [aboveq-1]
-  movd                  m0, [aboveq]
-  punpcklbw             m2, m1
+  movq                  m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x
   punpcklbw             m0, m1
-  pshufw                m2, m2, 0x0
-  DEFINE_ARGS dst, stride, line, left
-  mov                lineq, -2
-  add                leftq, 4
-  psubw                 m0, m2
-.loop:
-  movd                  m2, [leftq+lineq*2]
-  movd                  m3, [leftq+lineq*2+1]
+  pshuflw               m2, m0, 0x0   ; [63:0] tl tl tl tl [word]
+  psrldq                m0, 2
+  psubw                 m0, m2        ; [63:0] t1-tl t2-tl t3-tl t4-tl [word]
+  movd                  m2, [leftq]
   punpcklbw             m2, m1
-  punpcklbw             m3, m1
-  pshufw                m2, m2, 0x0
-  pshufw                m3, m3, 0x0
-  paddw                 m2, m0
+  pshuflw               m4, m2, 0x0   ; [63:0] l1 l1 l1 l1 [word]
+  pshuflw               m3, m2, 0x55  ; [63:0] l2 l2 l2 l2 [word]
+  paddw                 m4, m0
   paddw                 m3, m0
-  packuswb              m2, m2
+  packuswb              m4, m4
   packuswb              m3, m3
-  movd      [dstq        ], m2
+  movd      [dstq        ], m4
   movd      [dstq+strideq], m3
   lea                 dstq, [dstq+strideq*2]
-  inc                lineq
-  jnz .loop
-  REP_RET
+  pshuflw               m4, m2, 0xaa
+  pshuflw               m3, m2, 0xff
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m4
+  packuswb              m3, m3
+  movd      [dstq        ], m4
+  movd      [dstq+strideq], m3
+  RET
 
 INIT_XMM sse2
 cglobal tm_predictor_8x8, 4, 4, 4, dst, stride, above, left