ref: 589f3c7bc8cd8b7e2dea136f6e0a9847668ad7d6
parent: 64e46a033fa4bc86f3cde4fa5424e8fb0970cafa
parent: e86c7c863e465a2f532cc281ea73392b72ba5c78
author: Jian Zhou <zhoujian@google.com>
date: Fri Dec 4 16:43:10 EST 2015
Merge changes Ie48229c2,Ib9f18468,I0c90e7c1 * changes: Speed up h_predictor_16x16 Speed up h_predictor_8x8 MMX in intra 8x8 prediction replaced with SSE2
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -236,23 +236,19 @@
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,
+ vpx_h_predictor_8x8_sse2, NULL, NULL, NULL, NULL, NULL,
+ NULL, vpx_tm_predictor_8x8_sse2)
#endif // HAVE_SSE2 && CONFIG_USE_X86INC
#if HAVE_SSSE3 && CONFIG_USE_X86INC
INTRA_PRED_TEST(SSSE3, TestIntraPred8, NULL, NULL, NULL, NULL, NULL,
- vpx_h_predictor_8x8_ssse3, vpx_d45_predictor_8x8_ssse3, NULL,
- NULL, vpx_d153_predictor_8x8_ssse3,
- vpx_d207_predictor_8x8_ssse3, vpx_d63_predictor_8x8_ssse3, NULL)
+ NULL, vpx_d45_predictor_8x8_ssse3, NULL, NULL,
+ vpx_d153_predictor_8x8_ssse3, vpx_d207_predictor_8x8_ssse3,
+ vpx_d63_predictor_8x8_ssse3, NULL)
#endif // HAVE_SSSE3 && CONFIG_USE_X86INC
#if HAVE_DSPR2
@@ -294,13 +290,13 @@
vpx_dc_left_predictor_16x16_sse2,
vpx_dc_top_predictor_16x16_sse2,
vpx_dc_128_predictor_16x16_sse2, vpx_v_predictor_16x16_sse2,
- NULL, NULL, NULL, NULL, NULL, NULL, NULL,
+ vpx_h_predictor_16x16_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
vpx_tm_predictor_16x16_sse2)
#endif // HAVE_SSE2 && CONFIG_USE_X86INC
#if HAVE_SSSE3 && CONFIG_USE_X86INC
INTRA_PRED_TEST(SSSE3, TestIntraPred16, NULL, NULL, NULL, NULL, NULL,
- vpx_h_predictor_16x16_ssse3, vpx_d45_predictor_16x16_ssse3,
+ NULL, vpx_d45_predictor_16x16_ssse3,
NULL, NULL, vpx_d153_predictor_16x16_ssse3,
vpx_d207_predictor_16x16_ssse3, vpx_d63_predictor_16x16_ssse3,
NULL)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -130,7 +130,7 @@
specialize qw/vpx_d63e_predictor_8x8/;
add_proto qw/void vpx_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_h_predictor_8x8 neon dspr2 msa/, "$ssse3_x86inc";
+specialize qw/vpx_h_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
add_proto qw/void vpx_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
specialize qw/vpx_d117_predictor_8x8/;
@@ -142,22 +142,22 @@
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/, "$ssse3_x86inc";
@@ -178,7 +178,7 @@
specialize qw/vpx_d63e_predictor_16x16/;
add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_h_predictor_16x16 neon dspr2 msa/, "$ssse3_x86inc";
+specialize qw/vpx_h_predictor_16x16 neon dspr2 msa/, "$sse2_x86inc";
add_proto qw/void vpx_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
specialize qw/vpx_d117_predictor_16x16/;
--- a/vpx_dsp/x86/intrapred_sse2.asm
+++ b/vpx_dsp/x86/intrapred_sse2.asm
@@ -85,7 +85,7 @@
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 @@
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 @@
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 @@
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 @@
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 @@
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 @@
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 @@
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
@@ -531,6 +532,53 @@
movd [dstq ], m2
movd [dstq+strideq], m3
RET
+
+INIT_XMM sse2
+cglobal h_predictor_8x8, 2, 5, 3, dst, stride, line, left
+ movifnidn leftq, leftmp
+ mov lineq, -2
+ DEFINE_ARGS dst, stride, line, left, stride3
+ lea stride3q, [strideq*3]
+ movq m0, [leftq ]
+ punpcklbw m0, m0 ; l1 l1 l2 l2 ... l8 l8
+.loop:
+ pshuflw m1, m0, 0x0 ; l1 l1 l1 l1 l1 l1 l1 l1
+ pshuflw m2, m0, 0x55 ; l2 l2 l2 l2 l2 l2 l2 l2
+ movq [dstq ], m1
+ movq [dstq+strideq], m2
+ pshuflw m1, m0, 0xaa
+ pshuflw m2, m0, 0xff
+ movq [dstq+strideq*2], m1
+ movq [dstq+stride3q ], m2
+ pshufd m0, m0, 0xe ; [63:0] l5 l5 l6 l6 l7 l7 l8 l8
+ inc lineq
+ lea dstq, [dstq+strideq*4]
+ jnz .loop
+ REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_16x16, 2, 5, 3, dst, stride, line, left
+ movifnidn leftq, leftmp
+ mov lineq, -4
+ DEFINE_ARGS dst, stride, line, left, stride3
+ lea stride3q, [strideq*3]
+.loop:
+ movd m0, [leftq]
+ punpcklbw m0, m0
+ punpcklbw m0, m0 ; l1 to l4 each repeated 4 times
+ pshufd m1, m0, 0x0 ; l1 repeated 16 times
+ pshufd m2, m0, 0x55 ; l2 repeated 16 times
+ mova [dstq ], m1
+ mova [dstq+strideq ], m2
+ pshufd m1, m0, 0xaa
+ pshufd m2, m0, 0xff
+ mova [dstq+strideq*2], m1
+ mova [dstq+stride3q ], m2
+ inc lineq
+ lea leftq, [leftq+4 ]
+ lea dstq, [dstq+strideq*4]
+ jnz .loop
+ REP_RET
INIT_XMM sse2
cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left
--- a/vpx_dsp/x86/intrapred_ssse3.asm
+++ b/vpx_dsp/x86/intrapred_ssse3.asm
@@ -33,42 +33,6 @@
SECTION .text
-INIT_MMX ssse3
-cglobal h_predictor_8x8, 2, 4, 3, dst, stride, line, left
- movifnidn leftq, leftmp
- add leftq, 8
- mov lineq, -4
- pxor m0, m0
-.loop:
- movd m1, [leftq+lineq*2 ]
- movd m2, [leftq+lineq*2+1]
- pshufb m1, m0
- pshufb m2, m0
- movq [dstq ], m1
- movq [dstq+strideq], m2
- lea dstq, [dstq+strideq*2]
- inc lineq
- jnz .loop
- REP_RET
-
-INIT_XMM ssse3
-cglobal h_predictor_16x16, 2, 4, 3, dst, stride, line, left
- movifnidn leftq, leftmp
- add leftq, 16
- mov lineq, -8
- pxor m0, m0
-.loop:
- movd m1, [leftq+lineq*2 ]
- movd m2, [leftq+lineq*2+1]
- pshufb m1, m0
- pshufb m2, m0
- mova [dstq ], m1
- mova [dstq+strideq], m2
- lea dstq, [dstq+strideq*2]
- inc lineq
- jnz .loop
- REP_RET
-
INIT_XMM ssse3
cglobal h_predictor_32x32, 2, 4, 3, dst, stride, line, left
movifnidn leftq, leftmp