faster dequant

diff --git a/common/quant.c b/common/quant.c
index ed1148cb4148ac9ba53d2f5918cdcad2f1939f05..024dc6e1d93bf3d88aab2f65115125a964dad262 100644 (file)
--- a/common/quant.c
+++ b/common/quant.c
@@ -209,9 +209,9 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
 #ifdef ARCH_X86
         pf->quant_4x4 = x264_quant_4x4_mmx;
         pf->quant_8x8 = x264_quant_8x8_mmx;
-#endif
         pf->dequant_4x4 = x264_dequant_4x4_mmx;
         pf->dequant_8x8 = x264_dequant_8x8_mmx;
+#endif
     }
 
     if( cpu&X264_CPU_MMXEXT )
@@ -227,6 +227,8 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
         pf->quant_4x4_dc = x264_quant_4x4_dc_sse2;
         pf->quant_4x4 = x264_quant_4x4_sse2;
         pf->quant_8x8 = x264_quant_8x8_sse2;
+        pf->dequant_4x4 = x264_dequant_4x4_sse2;
+        pf->dequant_8x8 = x264_dequant_8x8_sse2;
     }
 #endif
 

diff --git a/common/x86/deblock-a.asm b/common/x86/deblock-a.asm
index 8071678b4fd028dfeda41aa45132bfcc12bfff8f..fd87234dc1824d4b30327d67b6d05691480f75a6 100644 (file)
--- a/common/x86/deblock-a.asm
+++ b/common/x86/deblock-a.asm
@@ -29,34 +29,6 @@ pb_a1: times 16 db 0xa1
 
 SECTION .text
 
-%macro INIT_MMX 0
-    %undef movq
-    %define m0 mm0
-    %define m1 mm1
-    %define m2 mm2
-    %define m3 mm3
-    %define m4 mm4
-    %define m5 mm5
-    %define m6 mm6
-    %define m7 mm7
-    %undef  m8
-    %undef  m9
-%endmacro
-
-%macro INIT_XMM 0
-    %define movq movdqa
-    %define m0 xmm0
-    %define m1 xmm1
-    %define m2 xmm2
-    %define m3 xmm3
-    %define m4 xmm4
-    %define m5 xmm5
-    %define m6 xmm6
-    %define m7 xmm7
-    %define m8 xmm8
-    %define m9 xmm9
-%endmacro
-
 ; expands to [base],...,[base+7*stride]
 %define PASS8ROWS(base, base3, stride, stride3) \
     [base], [base+stride], [base+stride*2], [base3], \

diff --git a/common/x86/quant-a.asm b/common/x86/quant-a.asm
index 11e78ea6221b1256831e481e4e3b338f37fa85fe..886b34000b5a37c7d1016b03597c8b553ab54a7f 100644 (file)
--- a/common/x86/quant-a.asm
+++ b/common/x86/quant-a.asm
@@ -24,7 +24,8 @@
 %include "x86inc.asm"
 
 SECTION_RODATA
-pd_1:  times 2 dd 1
+pw_1:     times 8 dw 1
+pd_10000: times 4 dd 1<<16
 
 SECTION .text
 
@@ -133,76 +134,74 @@ QUANT_AC x264_quant_8x8_ssse3, SSSE3_QUANT_1x8, 8, 16
 ; dequant
 ;=============================================================================
 
-%macro DEQUANT16_L_1x4 3
+%macro DEQUANT16_L 3
 ;;; %1      dct[y][x]
 ;;; %2,%3   dequant_mf[i_mf][y][x]
-;;; mm5     i_qbits
-
-    movq     mm1, %2
-    movq     mm2, %3
-    movq     mm0, %1
-    packssdw mm1, mm2
-    pmullw   mm0, mm1
-    psllw    mm0, mm5
-    movq     %1,  mm0
+;;; m5      i_qbits
+
+    movq     m1, %2
+    movq     m2, %3
+    movq     m0, %1
+    packssdw m1, m2
+    pmullw   m0, m1
+    psllw    m0, m5
+    movq     %1, m0
 %endmacro
 
-%macro DEQUANT32_R_1x4 3
+%macro DEQUANT32_R 3
 ;;; %1      dct[y][x]
 ;;; %2,%3   dequant_mf[i_mf][y][x]
-;;; mm5     -i_qbits
-;;; mm6     f as dwords
-;;; mm7     0
-
-    movq      mm0, %1
-    movq      mm1, mm0
-    punpcklwd mm0, mm0
-    punpckhwd mm1, mm1
-
-    movq      mm2, mm0
-    movq      mm3, mm1
-    pmulhw    mm0, %2
-    pmulhw    mm1, %3
-    pmullw    mm2, %2
-    pmullw    mm3, %3
-    pslld     mm0, 16
-    pslld     mm1, 16
-    paddd     mm0, mm2
-    paddd     mm1, mm3
-
-    paddd     mm0, mm6
-    paddd     mm1, mm6
-    psrad     mm0, mm5
-    psrad     mm1, mm5
-
-    packssdw  mm0, mm1
-    movq      %1,  mm0
+;;; m4      f
+;;; m5      -i_qbits
+;;; m6      1
+;;; m7      0
+
+    movq      m0, %1
+    movq      m1, m0
+    movq      m2, %2
+    movq      m3, %3
+    punpcklwd m0, m4
+    punpckhwd m1, m4
+    por       m2, m6 ; FIXME munge precomputed arrays?
+    por       m3, m6
+    pmaddwd   m0, m2
+    pmaddwd   m1, m3
+    psrad     m0, m5
+    psrad     m1, m5
+    packssdw  m0, m1
+    movq      %1, m0
 %endmacro
 
-%macro DEQUANT_LOOP 2
-    mov  t0d, 8*(%2-2)
+%macro DEQUANT_LOOP 3
+%if 8*(%2-2*%3)
+    mov t0d, 8*(%2-2*%3)
 %%loop:
-    %1 [r0+t0+8], [r1+t0*2+16], [r1+t0*2+24]
-    %1 [r0+t0  ], [r1+t0*2   ], [r1+t0*2+ 8]
-    sub  t0d, 16
-    jge  %%loop
+    %1 [r0+t0+8*%3], [r1+t0*2+16*%3], [r1+t0*2+24*%3]
+    %1 [r0+t0     ], [r1+t0*2      ], [r1+t0*2+ 8*%3]
+    sub t0d, 16*%3
+    jge %%loop
     rep ret
+%else
+    %1 [r0+8*%3], [r1+16*%3], [r1+24*%3]
+    %1 [r0     ], [r1      ], [r1+ 8*%3]
+    ret
+%endif
 %endmacro
 
 ;-----------------------------------------------------------------------------
 ; void x264_dequant_4x4_mmx( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
 ;-----------------------------------------------------------------------------
-%macro DEQUANT_WxH 3
+%macro DEQUANT_WxH 4
 cglobal %1, 0,3
 %ifdef ARCH_X86_64
     %define t0  r4
     %define t0d r4d
     imul r4d, r2d, 0x2b
     shr  r4d, 8     ; i_qbits = i_qp / 6
-    lea  r3d, [r4d*3]
-    sub  r2, r3
-    sub  r2, r3     ; i_mf = i_qp % 6
-    shl  r2, %3+2
+    lea  r3, [r4*3]
+    sub  r2d, r3d
+    sub  r2d, r3d   ; i_mf = i_qp % 6
+    shl  r2d, %3+2
     add  r1, r2     ; dequant_mf[i_mf]
 %else
     %define t0  r2
@@ -222,19 +221,27 @@ cglobal %1, 0,3
     jl   .rshift32  ; negative qbits => rightshift
 
 .lshift:
-    movd mm5, t0d
-    DEQUANT_LOOP DEQUANT16_L_1x4, %2
+    movd m5, t0d
+    DEQUANT_LOOP DEQUANT16_L, %2, %4
 
 .rshift32:
     neg   t0d
-    movd  mm5, t0d
+    movd  m5, t0d
     picgetgot t0d
-    movq  mm6, [pd_1 GLOBAL]
-    pxor  mm7, mm7
-    pslld mm6, mm5
-    psrld mm6, 1
-    DEQUANT_LOOP DEQUANT32_R_1x4, %2
+    movq  m4, [pw_1 GLOBAL]
+    movq  m6, [pd_10000 GLOBAL]
+    psllw m4, m5
+    pxor  m7, m7
+    psrlw m4, 1
+    DEQUANT_LOOP DEQUANT32_R, %2, %4
 %endmacro
 
-DEQUANT_WxH x264_dequant_4x4_mmx, 4, 4
-DEQUANT_WxH x264_dequant_8x8_mmx, 16, 6
+%ifndef ARCH_X86_64
+INIT_MMX
+DEQUANT_WxH x264_dequant_4x4_mmx, 4,  4, 1
+DEQUANT_WxH x264_dequant_8x8_mmx, 16, 6, 1
+%endif
+INIT_XMM
+DEQUANT_WxH x264_dequant_4x4_sse2, 4,  4, 2
+DEQUANT_WxH x264_dequant_8x8_sse2, 16, 6, 2
+

diff --git a/common/x86/quant.h b/common/x86/quant.h
index 8532fde9a3fbb4fa9ed9d1c7a08effa2e0b96c65..587286c3139254ab37fa17664d55ac92ada310e4 100644 (file)
--- a/common/x86/quant.h
+++ b/common/x86/quant.h
@@ -35,5 +35,7 @@ void x264_quant_4x4_ssse3( int16_t dct[4][4], uint16_t mf[16], uint16_t bias[16]
 void x264_quant_8x8_ssse3( int16_t dct[8][8], uint16_t mf[64], uint16_t bias[64] );
 void x264_dequant_4x4_mmx( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp );
 void x264_dequant_8x8_mmx( int16_t dct[8][8], int dequant_mf[6][8][8], int i_qp );
+void x264_dequant_4x4_sse2( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp );
+void x264_dequant_8x8_sse2( int16_t dct[8][8], int dequant_mf[6][8][8], int i_qp );
 
 #endif

diff --git a/common/x86/x86inc.asm b/common/x86/x86inc.asm
index b8fbe13a6c64483dce1d3771e4b2fbe0acc54906..256aa2e4a29a8335a914a6161a05ede972aeb708 100644 (file)
--- a/common/x86/x86inc.asm
+++ b/common/x86/x86inc.asm
@@ -328,3 +328,31 @@ SECTION ".note.GNU-stack" noalloc noexec nowrite progbits
 %assign FENC_STRIDE 16
 %assign FDEC_STRIDE 32
 
+%macro INIT_MMX 0
+    %undef  movq
+    %define m0 mm0
+    %define m1 mm1
+    %define m2 mm2
+    %define m3 mm3
+    %define m4 mm4
+    %define m5 mm5
+    %define m6 mm6
+    %define m7 mm7
+    %undef  m8
+    %undef  m9
+%endmacro
+
+%macro INIT_XMM 0
+    %define movq movdqa
+    %define m0 xmm0
+    %define m1 xmm1
+    %define m2 xmm2
+    %define m3 xmm3
+    %define m4 xmm4
+    %define m5 xmm5
+    %define m6 xmm6
+    %define m7 xmm7
+    %define m8 xmm8
+    %define m9 xmm9
+%endmacro
+