x86/aacpsdsp: add ff_ps_hybrid_synthesis_deint_{sse,sse4}

About 2x faster than the c version.

libavcodec/x86/aacpsdsp.asm

@@ -166,6 +166,129 @@ align 16
     jl .loop
     REP_RET
 
+;***********************************************************
+;void ps_hybrid_synthesis_deint_sse4(float out[2][38][64],
+;                                    float (*in)[32][2],
+;                                    int i, int len)
+;***********************************************************
+%macro HYBRID_SYNTHESIS_DEINT 0
+cglobal ps_hybrid_synthesis_deint, 3, 7, 5, out, in, i, len, out0, out1, tmp
+%if cpuflag(sse4)
+%define MOVH movsd
+%else
+%define MOVH movlps
+%endif
+    movsxdifnidn        iq, id
+    mov               lend, 32 << 3
+    lea               outq, [outq+iq*4]
+    mov               tmpd, id
+    shl               tmpd, 8
+    add                inq, tmpq
+    mov               tmpd, 64
+    sub               tmpd, id
+    mov                 id, tmpd
+
+    test                id, 1
+    jne .loop4
+    test                id, 2
+    jne .loop8
+
+align 16
+.loop16:
+    mov              out0q, outq
+    mov              out1q, 38*64*4
+    add              out1q, out0q
+    mov               tmpd, lend
+
+.inner_loop16:
+    movaps              m0, [inq]
+    movaps              m1, [inq+lenq]
+    movaps              m2, [inq+lenq*2]
+    movaps              m3, [inq+3*32*2*4]
+    TRANSPOSE4x4PS 0, 1, 2, 3, 4
+    movaps         [out0q], m0
+    movaps         [out1q], m1
+    movaps    [out0q+lenq], m2
+    movaps    [out1q+lenq], m3
+    lea              out0q, [out0q+lenq*2]
+    lea              out1q, [out1q+lenq*2]
+    add                inq, mmsize
+    sub               tmpd, mmsize
+    jg .inner_loop16
+    add               outq, 16
+    add                inq, 3*32*2*4
+    sub                 id, 4
+    jg .loop16
+    RET
+
+align 16
+.loop8:
+    mov              out0q, outq
+    mov              out1q, 38*64*4
+    add              out1q, out0q
+    mov               tmpd, lend
+
+.inner_loop8:
+    movaps              m0, [inq]
+    movaps              m1, [inq+lenq]
+    SBUTTERFLYPS 0, 1, 2
+    SBUTTERFLYPD 0, 1, 2
+    MOVH           [out0q], m0
+    MOVH           [out1q], m1
+    movhps    [out0q+lenq], m0
+    movhps    [out1q+lenq], m1
+    lea              out0q, [out0q+lenq*2]
+    lea              out1q, [out1q+lenq*2]
+    add                inq, mmsize
+    sub               tmpd, mmsize
+    jg .inner_loop8
+    add               outq, 8
+    add                inq, lenq
+    sub                 id, 2
+    jg .loop16
+    RET
+
+align 16
+.loop4:
+    mov              out0q, outq
+    mov              out1q, 38*64*4
+    add              out1q, out0q
+    mov               tmpd, lend
+
+.inner_loop4:
+    movaps              m0, [inq]
+    movss          [out0q], m0
+%if cpuflag(sse4)
+    extractps      [out1q], m0, 1
+    extractps [out0q+lenq], m0, 2
+    extractps [out1q+lenq], m0, 3
+%else
+    movhlps             m1, m0
+    movss     [out0q+lenq], m1
+    shufps              m0, m0, 0xb1
+    movss          [out1q], m0
+    movhlps             m1, m0
+    movss     [out1q+lenq], m1
+%endif
+    lea              out0q, [out0q+lenq*2]
+    lea              out1q, [out1q+lenq*2]
+    add                inq, mmsize
+    sub               tmpd, mmsize
+    jg .inner_loop4
+    add               outq, 4
+    sub                 id, 1
+    test                id, 2
+    jne .loop8
+    cmp                 id, 4
+    jge .loop16
+    RET
+%endmacro
+
+INIT_XMM sse
+HYBRID_SYNTHESIS_DEINT
+INIT_XMM sse4
+HYBRID_SYNTHESIS_DEINT
+
 ;*******************************************************************
 ;void ff_ps_hybrid_analysis_<opt>(float (*out)[2], float (*in)[2],
 ;                                 const float (*filter)[8][2],

libavcodec/x86/aacpsdsp_init.c

@@ -40,6 +40,10 @@ void ff_ps_stereo_interpolate_sse3(float (*l)[2], float (*r)[2],
 void ff_ps_stereo_interpolate_ipdopd_sse3(float (*l)[2], float (*r)[2],
                                           float h[2][4], float h_step[2][4],
                                           int len);
+void ff_ps_hybrid_synthesis_deint_sse(float out[2][38][64], float (*in)[32][2],
+                                      int i, int len);
+void ff_ps_hybrid_synthesis_deint_sse4(float out[2][38][64], float (*in)[32][2],
+                                       int i, int len);
 
 av_cold void ff_psdsp_init_x86(PSDSPContext *s)
 {
@@ -48,6 +52,7 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s)
     if (EXTERNAL_SSE(cpu_flags)) {
         s->add_squares            = ff_ps_add_squares_sse;
         s->mul_pair_single        = ff_ps_mul_pair_single_sse;
+        s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse;
         s->hybrid_analysis        = ff_ps_hybrid_analysis_sse;
     }
     if (EXTERNAL_SSE3(cpu_flags)) {
@@ -56,4 +61,7 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s)
         s->stereo_interpolate[1]  = ff_ps_stereo_interpolate_ipdopd_sse3;
         s->hybrid_analysis        = ff_ps_hybrid_analysis_sse3;
     }
+    if (EXTERNAL_SSE4(cpu_flags)) {
+        s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse4;
+    }
 }

libavutil/x86/x86util.asm

@@ -71,6 +71,12 @@
     SWAP %1, %3, %2
 %endmacro
 
+%macro SBUTTERFLYPD 3
+    movlhps m%3, m%1, m%2
+    movhlps m%2, m%2, m%1
+    SWAP %1, %3
+%endmacro
+
 %macro TRANSPOSE4x4B 5
     SBUTTERFLY bw, %1, %2, %5
     SBUTTERFLY bw, %3, %4, %5
@@ -117,12 +123,9 @@
 %macro TRANSPOSE4x4PS 5
     SBUTTERFLYPS %1, %2, %5
     SBUTTERFLYPS %3, %4, %5
-    movlhps m%5, m%1, m%3
+    SBUTTERFLYPD %1, %3, %5
-    movhlps m%3, m%1
+    SBUTTERFLYPD %2, %4, %5
-    SWAP %5, %1
+    SWAP %2, %3
-    movlhps m%5, m%2, m%4
-    movhlps m%4, m%2
-    SWAP %5, %2, %3
 %endmacro
 
 %macro TRANSPOSE8x4D 9-11