Mercurial > libavcodec.hg

changeset 12399:020540442072 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Convert ff_imdct_half_sse() to yasm. This is to avoid split asm sections that attempt to preserve some registers between sections.
author alexc
date Sun, 22 Aug 2010 14:39:58 +0000
parents 31736ceab3aa
children 4f13b2ded34d
files x86/fft_mmx.asm x86/fft_sse.c
diffstat 2 files changed, 195 insertions(+), 108 deletions(-) [+]
line wrap: on
line diff
--- a/x86/fft_mmx.asm	Sat Aug 21 19:50:36 2010 +0000
+++ b/x86/fft_mmx.asm	Sun Aug 22 14:39:58 2010 +0000
@@ -29,6 +29,23 @@
 
 %include "x86inc.asm"
 
+%ifdef ARCH_X86_64
+%define pointer resq
+%else
+%define pointer resd
+%endif
+
+struc FFTContext
+    .nbits:    resd 1
+    .reverse:  resd 1
+    .revtab:   pointer 1
+    .tmpbuf:   pointer 1
+    .mdctsize: resd 1
+    .mdctbits: resd 1
+    .tcos:     pointer 1
+    .tsin:     pointer 1
+endstruc
+
 SECTION_RODATA
 
 %define M_SQRT1_2 0.70710678118654752440
@@ -428,6 +445,16 @@
 %define SECTION_REL
 %endif
 
+%macro FFT_DISPATCH 2; clobbers 5 GPRs, 8 XMMs
+    lea r2, [dispatch_tab%1]
+    mov r2, [r2 + (%2q-2)*gprsize]
+%ifdef PIC
+    lea r3, [$$]
+    add r2, r3
+%endif
+    call r2
+%endmacro ; FFT_DISPATCH
+
 %macro DECL_FFT 2-3 ; nbits, cpu, suffix
 %xdefine list_of_fft fft4%2 SECTION_REL, fft8%2 SECTION_REL
 %if %1==5
@@ -464,13 +491,7 @@
 ; On x86_32, this function does the register saving and restoring for all of fft.
 ; The others pass args in registers and don't spill anything.
 cglobal fft_dispatch%3%2, 2,5,8, z, nbits
-    lea r2, [dispatch_tab%3%2]
-    mov r2, [r2 + (nbitsq-2)*gprsize]
-%ifdef PIC
-    lea r3, [$$]
-    add r2, r3
-%endif
-    call r2
+    FFT_DISPATCH %3%2, nbits
     RET
 %endmacro ; DECL_FFT
 
@@ -481,3 +502,170 @@
 DECL_FFT 4, _3dn2
 DECL_FFT 4, _3dn2, _interleave
 
+INIT_XMM
+%undef mulps
+%undef addps
+%undef subps
+%undef unpcklps
+%undef unpckhps
+
+%macro PREROTATER 5 ;-2*k, 2*k, input+n4, tcos+n8, tsin+n8
+    movaps   xmm0, [%3+%2*4]
+    movaps   xmm1, [%3+%1*4-0x10]
+    movaps   xmm2, xmm0
+    shufps   xmm0, xmm1, 0x88
+    shufps   xmm1, xmm2, 0x77
+    movlps   xmm4, [%4+%2*2]
+    movlps   xmm5, [%5+%2*2+0x0]
+    movhps   xmm4, [%4+%1*2-0x8]
+    movhps   xmm5, [%5+%1*2-0x8]
+    movaps   xmm2, xmm0
+    movaps   xmm3, xmm1
+    mulps    xmm0, xmm5
+    mulps    xmm1, xmm4
+    mulps    xmm2, xmm4
+    mulps    xmm3, xmm5
+    subps    xmm1, xmm0
+    addps    xmm2, xmm3
+    movaps   xmm0, xmm1
+    unpcklps xmm1, xmm2
+    unpckhps xmm0, xmm2
+%endmacro
+
+%macro PREROTATEW 3 ;addr1, addr2, xmm
+    movlps   %1,   %3
+    movhps   %2,   %3
+%endmacro
+
+%macro CMUL 6 ;j, xmm0, xmm1, 3, 4, 5
+    movaps   xmm6, [%4+%1*2]
+    movaps   %2,   [%4+%1*2+0x10]
+    movaps   %3,   xmm6
+    movaps   xmm7, %2
+    mulps    xmm6, [%5+%1*1]
+    mulps    %2,   [%6+%1*1]
+    mulps    %3,   [%6+%1*1]
+    mulps    xmm7, [%5+%1*1]
+    subps    %2,   xmm6
+    addps    %3,   xmm7
+%endmacro
+
+%macro POSROTATESHUF 5 ;j, k, z+n8, tcos+n8, tsin+n8
+.post:
+    CMUL     %1,   xmm0, xmm1, %3, %4, %5
+    CMUL     %2,   xmm4, xmm5, %3, %4, %5
+    shufps   xmm1, xmm1, 0x1b
+    shufps   xmm5, xmm5, 0x1b
+    movaps   xmm6, xmm4
+    unpckhps xmm4, xmm1
+    unpcklps xmm6, xmm1
+    movaps   xmm2, xmm0
+    unpcklps xmm0, xmm5
+    unpckhps xmm2, xmm5
+    movaps   [%3+%2*2],      xmm6
+    movaps   [%3+%2*2+0x10], xmm4
+    movaps   [%3+%1*2],      xmm0
+    movaps   [%3+%1*2+0x10], xmm2
+    sub      %2,   0x10
+    add      %1,   0x10
+    jl       .post
+%endmacro
+
+cglobal imdct_half_sse, 3,7,8; FFTContext *s, FFTSample *output, const FFTSample *input
+%ifdef ARCH_X86_64
+%define rrevtab r10
+%define rtcos   r11
+%define rtsin   r12
+    push  r10
+    push  r11
+    push  r12
+    push  r13
+    push  r14
+%else
+%define rrevtab r6
+%define rtsin   r6
+%define rtcos   r5
+%endif
+    mov   r3d, [r0+FFTContext.mdctsize]
+    add   r2, r3
+    shr   r3, 1
+    mov   rtcos, [r0+FFTContext.tcos]
+    mov   rtsin, [r0+FFTContext.tsin]
+    add   rtcos, r3
+    add   rtsin, r3
+%ifndef ARCH_X86_64
+    push  rtcos
+    push  rtsin
+%endif
+    shr   r3, 1
+    mov   rrevtab, [r0+FFTContext.revtab]
+    add   rrevtab, r3
+%ifndef ARCH_X86_64
+    push  rrevtab
+%endif
+
+    sub   r3, 4
+%ifdef ARCH_X86_64
+    xor   r4, r4
+    sub   r4, r3
+%endif
+.pre:
+%ifndef ARCH_X86_64
+;unspill
+    xor   r4, r4
+    sub   r4, r3
+    mov   rtsin, [esp+4]
+    mov   rtcos, [esp+8]
+%endif
+
+    PREROTATER r4, r3, r2, rtcos, rtsin
+%ifdef ARCH_X86_64
+    movzx  r5,  word [rrevtab+r4*1-4]
+    movzx  r6,  word [rrevtab+r4*1-2]
+    movzx  r13, word [rrevtab+r3*1]
+    movzx  r14, word [rrevtab+r3*1+2]
+    PREROTATEW [r1+r5 *8], [r1+r6 *8], xmm0
+    PREROTATEW [r1+r13*8], [r1+r14*8], xmm1
+    add    r4, 4
+%else
+    mov    r6, [esp]
+    movzx  r5, word [r6+r4*1-4]
+    movzx  r4, word [r6+r4*1-2]
+    PREROTATEW [r1+r5*8], [r1+r4*8], xmm0
+    movzx  r5, word [r6+r3*1]
+    movzx  r4, word [r6+r3*1+2]
+    PREROTATEW [r1+r5*8], [r1+r4*8], xmm1
+%endif
+    sub    r3, 4
+    jns    .pre
+
+    mov  r5, r0
+    mov  r6, r1
+    mov  r0, r1
+    mov  r1d, [r5+FFTContext.nbits]
+
+    FFT_DISPATCH _sse, r1
+
+    mov  r0d, [r5+FFTContext.mdctsize]
+    add  r6, r0
+    shr  r0, 1
+%ifndef ARCH_X86_64
+%define rtcos r2
+%define rtsin r3
+    mov  rtcos, [esp+8]
+    mov  rtsin, [esp+4]
+%endif
+    neg  r0
+    mov  r1, -16
+    sub  r1, r0
+    POSROTATESHUF r0, r1, r6, rtcos, rtsin
+%ifdef ARCH_X86_64
+    pop  r14
+    pop  r13
+    pop  r12
+    pop  r11
+    pop  r10
+%else
+    add esp, 12
+%endif
+    RET
--- a/x86/fft_sse.c	Sat Aug 21 19:50:36 2010 +0000
+++ b/x86/fft_sse.c	Sun Aug 22 14:39:58 2010 +0000
@@ -71,107 +71,6 @@
     memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
 }
 
-void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
-{
-    av_unused x86_reg i, j, k, l;
-    long n = 1 << s->mdct_bits;
-    long n2 = n >> 1;
-    long n4 = n >> 2;
-    long n8 = n >> 3;
-    const uint16_t *revtab = s->revtab + n8;
-    const FFTSample *tcos = s->tcos;
-    const FFTSample *tsin = s->tsin;
-    FFTComplex *z = (FFTComplex *)output;
-
-    /* pre rotation */
-    for(k=n8-2; k>=0; k-=2) {
-        __asm__ volatile(
-            "movaps     (%2,%1,2), %%xmm0 \n" // { z[k].re,    z[k].im,    z[k+1].re,  z[k+1].im  }
-            "movaps  -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
-            "movaps        %%xmm0, %%xmm2 \n"
-            "shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re,    z[k+1].re,  z[-k-2].re, z[-k-1].re }
-            "shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im,  z[k].im    }
-            "movlps       (%3,%1), %%xmm4 \n"
-            "movlps       (%4,%1), %%xmm5 \n"
-            "movhps     -8(%3,%0), %%xmm4 \n" // { cos[k],     cos[k+1],   cos[-k-2],  cos[-k-1]  }
-            "movhps     -8(%4,%0), %%xmm5 \n" // { sin[k],     sin[k+1],   sin[-k-2],  sin[-k-1]  }
-            "movaps        %%xmm0, %%xmm2 \n"
-            "movaps        %%xmm1, %%xmm3 \n"
-            "mulps         %%xmm5, %%xmm0 \n" // re*sin
-            "mulps         %%xmm4, %%xmm1 \n" // im*cos
-            "mulps         %%xmm4, %%xmm2 \n" // re*cos
-            "mulps         %%xmm5, %%xmm3 \n" // im*sin
-            "subps         %%xmm0, %%xmm1 \n" // -> re
-            "addps         %%xmm3, %%xmm2 \n" // -> im
-            "movaps        %%xmm1, %%xmm0 \n"
-            "unpcklps      %%xmm2, %%xmm1 \n" // { z[k],    z[k+1]  }
-            "unpckhps      %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
-            ::"r"(-4*k), "r"(4*k),
-              "r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
-        );
-#if ARCH_X86_64
-        // if we have enough regs, don't let gcc make the luts latency-bound
-        // but if not, latency is faster than spilling
-        __asm__("movlps %%xmm0, %0 \n"
-            "movhps %%xmm0, %1 \n"
-            "movlps %%xmm1, %2 \n"
-            "movhps %%xmm1, %3 \n"
-            :"=m"(z[revtab[-k-2]]),
-             "=m"(z[revtab[-k-1]]),
-             "=m"(z[revtab[ k  ]]),
-             "=m"(z[revtab[ k+1]])
-        );
-#else
-        __asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
-        __asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
-        __asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k  ]]));
-        __asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
-#endif
-    }
-
-    ff_fft_dispatch_sse(z, s->nbits);
-
-    /* post rotation + reinterleave + reorder */
-
-#define CMUL(j,xmm0,xmm1)\
-        "movaps   (%2,"#j",2), %%xmm6 \n"\
-        "movaps 16(%2,"#j",2), "#xmm0"\n"\
-        "movaps        %%xmm6, "#xmm1"\n"\
-        "movaps        "#xmm0",%%xmm7 \n"\
-        "mulps      (%3,"#j"), %%xmm6 \n"\
-        "mulps      (%4,"#j"), "#xmm0"\n"\
-        "mulps      (%4,"#j"), "#xmm1"\n"\
-        "mulps      (%3,"#j"), %%xmm7 \n"\
-        "subps         %%xmm6, "#xmm0"\n"\
-        "addps         %%xmm7, "#xmm1"\n"
-
-    j = -n2;
-    k = n2-16;
-    __asm__ volatile(
-        "1: \n"
-        CMUL(%0, %%xmm0, %%xmm1)
-        CMUL(%1, %%xmm4, %%xmm5)
-        "shufps    $0x1b, %%xmm1, %%xmm1 \n"
-        "shufps    $0x1b, %%xmm5, %%xmm5 \n"
-        "movaps   %%xmm4, %%xmm6 \n"
-        "unpckhps %%xmm1, %%xmm4 \n"
-        "unpcklps %%xmm1, %%xmm6 \n"
-        "movaps   %%xmm0, %%xmm2 \n"
-        "unpcklps %%xmm5, %%xmm0 \n"
-        "unpckhps %%xmm5, %%xmm2 \n"
-        "movaps   %%xmm6,   (%2,%1,2) \n"
-        "movaps   %%xmm4, 16(%2,%1,2) \n"
-        "movaps   %%xmm0,   (%2,%0,2) \n"
-        "movaps   %%xmm2, 16(%2,%0,2) \n"
-        "sub $16, %1 \n"
-        "add $16, %0 \n"
-        "jl 1b \n"
-        :"+&r"(j), "+&r"(k)
-        :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
-        :"memory"
-    );
-}
-
 void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
 {
     x86_reg j, k;