x86: synth filter float: implement SSE2 version

Timings for Arrandale:
          C    SSE
win32:  2108   334
win64:  1152   322

Factorizing the inner loop with a call/jmp is a >15 cycles cost, even with
the jmp destination being aligned.

Unrolling for ARCH_X86_64 is a 20 cycles gain.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

1 files changed, 154 insertions, 0 deletions

diff --git a/libavcodec/x86/dcadsp.asm b/libavcodec/x86/dcadsp.asm
index b9e8b8bf0f..0c73bbb1b6 100644
--- a/libavcodec/x86/dcadsp.asm
+++ b/libavcodec/x86/dcadsp.asm
@@ -175,3 +175,157 @@ cglobal dca_lfe_fir%1, 3,3,6-%1, out, in, cf0
 INIT_XMM sse
 DCA_LFE_FIR 0
 DCA_LFE_FIR 1
+
+INIT_XMM sse2
+%macro INNER_LOOP   1
+    ; reading backwards:  ptr1=synth_buf+j+i   ptr2=synth_big+j-i
+    ;~ a += window[i + j     ]*(-synth_buf[15 - i + j      ])
+    ;~ b += window[i + j + 16]*( synth_buf[     i + j      ])
+    pshufd        m5, [ptr2 + j + (15-3)*4], q0123
+    mova          m6, [ptr1 + j]
+%if ARCH_X86_64
+    pshufd       m11, [ptr2 + j + (15-3)*4 - mmsize], q0123
+    mova         m12, [ptr1 + j + mmsize]
+%endif
+    mulps         m6, [win  + %1 + j + 16*4]
+    mulps         m5, [win  + %1 + j]
+%if ARCH_X86_64
+    mulps        m12, [win  + %1 + j + mmsize + 16*4]
+    mulps        m11, [win  + %1 + j + mmsize]
+%endif
+    addps         m2, m6
+    subps         m1, m5
+%if ARCH_X86_64
+    addps         m8, m12
+    subps         m7, m11
+%endif
+    ;~ c += window[i + j + 32]*( synth_buf[16 + i + j      ])
+    ;~ d += window[i + j + 48]*( synth_buf[31 - i + j      ])
+    pshufd        m6, [ptr2 + j + (31-3)*4], q0123
+    mova          m5, [ptr1 + j + 16*4]
+%if ARCH_X86_64
+    pshufd       m12, [ptr2 + j + (31-3)*4 - mmsize], q0123
+    mova         m11, [ptr1 + j + mmsize + 16*4]
+%endif
+    mulps         m5, [win  + %1 + j + 32*4]
+    mulps         m6, [win  + %1 + j + 48*4]
+%if ARCH_X86_64
+    mulps        m11, [win  + %1 + j + mmsize + 32*4]
+    mulps        m12, [win  + %1 + j + mmsize + 48*4]
+%endif
+    addps         m3, m5
+    addps         m4, m6
+%if ARCH_X86_64
+    addps         m9, m11
+    addps        m10, m12
+%endif
+    sub            j, 64*4
+%endmacro
+
+; void ff_synth_filter_inner_sse2(float *synth_buf, float synth_buf2[32],
+;                                 const float window[512], float out[32],
+;                                 intptr_t offset, float scale)
+cglobal synth_filter_inner, 0,6+4*ARCH_X86_64,7+6*ARCH_X86_64, \
+                              synth_buf, synth_buf2, window, out, off, scale
+%define scale m0
+%if ARCH_X86_32 || WIN64
+    movd       scale, scalem
+; Make sure offset is in a register and not on the stack
+%define OFFQ  r4q
+%else
+%define OFFQ  offq
+%endif
+    pshufd        m0, m0, 0
+    ; prepare inner counter limit 1
+    mov          r5q, 480
+    sub          r5q, offmp
+    and          r5q, -64
+    shl          r5q, 2
+    mov         OFFQ, r5q
+%define i        r5q
+    mov            i, 16*4-(ARCH_X86_64+1)*mmsize  ; main loop counter
+
+%define buf2     synth_buf2q
+%if ARCH_X86_32
+    mov         buf2, synth_buf2mp
+%endif
+.mainloop
+    ; m1=a  m2=b  m3=c  m4=d
+    pxor          m3, m3
+    pxor          m4, m4
+    mova          m1, [buf2 + i]
+    mova          m2, [buf2 + i + 16*4]
+%if ARCH_X86_32
+%define ptr1     r0q
+%define ptr2     r1q
+%define win      r2q
+%define j        r3q
+    mov          win, windowm
+    mov         ptr1, synth_bufm
+    add          win, i
+    add         ptr1, i
+%else
+%define ptr1     r6q
+%define ptr2     r7q ; must be loaded
+%define win      r8q
+%define j        r9q
+%if ARCH_X86_64
+    pxor          m9, m9
+    pxor         m10, m10
+    mova          m7, [buf2 + i + mmsize]
+    mova          m8, [buf2 + i + mmsize + 16*4]
+%endif
+    lea          win, [windowq + i]
+    lea         ptr1, [synth_bufq + i]
+%endif
+    mov         ptr2, synth_bufmp
+    ; prepare the inner loop counter
+    mov            j, OFFQ
+    sub         ptr2, i
+.loop1:
+    INNER_LOOP  0
+    jge       .loop1
+
+    mov            j, 448*4
+    sub            j, OFFQ
+    jz          .end
+    sub         ptr1, j
+    sub         ptr2, j
+    add          win, OFFQ ; now at j-64, so define OFFSET
+    sub            j, 64*4
+.loop2:
+    INNER_LOOP  64*4
+    jge       .loop2
+
+.end:
+%if ARCH_X86_32
+    mov         buf2, synth_buf2m ; needed for next iteration anyway
+    mov         outq, outmp       ; j, which will be set again during it
+%endif
+    ;~ out[i     ] = a*scale;
+    ;~ out[i + 16] = b*scale;
+    mulps         m1, scale
+    mulps         m2, scale
+%if ARCH_X86_64
+    mulps         m7, scale
+    mulps         m8, scale
+%endif
+    ;~ synth_buf2[i     ] = c;
+    ;~ synth_buf2[i + 16] = d;
+    mova   [buf2 + i +  0*4], m3
+    mova   [buf2 + i + 16*4], m4
+%if ARCH_X86_64
+    mova   [buf2 + i +  0*4 + mmsize], m9
+    mova   [buf2 + i + 16*4 + mmsize], m10
+%endif
+    ;~ out[i     ] = a;
+    ;~ out[i + 16] = a;
+    mova   [outq + i +  0*4], m1
+    mova   [outq + i + 16*4], m2
+%if ARCH_X86_64
+    mova   [outq + i +  0*4 + mmsize], m7
+    mova   [outq + i + 16*4 + mmsize], m8
+%endif
+    sub            i, (ARCH_X86_64+1)*mmsize
+    jge    .mainloop
+    RET