x86/float_dsp: port vector_fmul_window to yasm

Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

2 files changed, 63 insertions, 73 deletions

diff --git a/libavutil/x86/float_dsp.asm b/libavutil/x86/float_dsp.asm
index a5fcffb30b..aec0273235 100644
--- a/libavutil/x86/float_dsp.asm
+++ b/libavutil/x86/float_dsp.asm
@@ -192,6 +192,61 @@ VECTOR_DMUL_SCALAR
 %endif
 
 ;-----------------------------------------------------------------------------
+; vector_fmul_window(float *dst, const float *src0,
+;                    const float *src1, const float *win, int len);
+;-----------------------------------------------------------------------------
+%macro VECTOR_FMUL_WINDOW 0
+cglobal vector_fmul_window, 5, 6, 6, dst, src0, src1, win, len, len1
+    shl     lend, 2
+    lea    len1q, [lenq - mmsize]
+    add    src0q, lenq
+    add     dstq, lenq
+    add     winq, lenq
+    neg     lenq
+.loop
+    mova      m0, [winq  + lenq]
+    mova      m4, [src0q + lenq]
+%if cpuflag(sse)
+    mova      m1, [winq  + len1q]
+    mova      m5, [src1q + len1q]
+    shufps    m1, m1, 0x1b
+    shufps    m5, m5, 0x1b
+    mova      m2, m0
+    mova      m3, m1
+    mulps     m2, m4
+    mulps     m3, m5
+    mulps     m1, m4
+    mulps     m0, m5
+    addps     m2, m3
+    subps     m1, m0
+    shufps    m2, m2, 0x1b
+%else
+    pswapd    m1, [winq  + len1q]
+    pswapd    m5, [src1q + len1q]
+    mova      m2, m0
+    mova      m3, m1
+    pfmul     m2, m4
+    pfmul     m3, m5
+    pfmul     m1, m4
+    pfmul     m0, m5
+    pfadd     m2, m3
+    pfsub     m1, m0
+    pswapd    m2, m2
+%endif
+    mova      [dstq + lenq], m1
+    mova      [dstq + len1q], m2
+    sub       len1q, mmsize
+    add       lenq,  mmsize
+    jl .loop
+    REP_RET
+%endmacro
+
+INIT_MMX 3dnowext
+VECTOR_FMUL_WINDOW
+INIT_XMM sse
+VECTOR_FMUL_WINDOW
+
+;-----------------------------------------------------------------------------
 ; vector_fmul_add(float *dst, const float *src0, const float *src1,
 ;                 const float *src2, int len)
 ;-----------------------------------------------------------------------------
diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c
index 88ffbc11b5..64b3a4d6bd 100644
--- a/libavutil/x86/float_dsp_init.c
+++ b/libavutil/x86/float_dsp_init.c
@@ -44,6 +44,11 @@ void ff_vector_dmul_scalar_sse2(double *dst, const double *src,
 void ff_vector_dmul_scalar_avx(double *dst, const double *src,
                                double mul, int len);
 
+void ff_vector_fmul_window_3dnowext(float *dst, const float *src0,
+                                    const float *src1, const float *win, int len);
+void ff_vector_fmul_window_sse(float *dst, const float *src0,
+                               const float *src1, const float *win, int len);
+
 void ff_vector_fmul_add_sse(float *dst, const float *src0, const float *src1,
                             const float *src2, int len);
 void ff_vector_fmul_add_avx(float *dst, const float *src0, const float *src1,
@@ -60,88 +65,18 @@ float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
 
 void ff_butterflies_float_sse(float *src0, float *src1, int len);
 
-#if HAVE_6REGS && HAVE_INLINE_ASM
-static void vector_fmul_window_3dnowext(float *dst, const float *src0,
-                                        const float *src1, const float *win,
-                                        int len)
-{
-    x86_reg i = -len * 4;
-    x86_reg j =  len * 4 - 8;
-    __asm__ volatile (
-        "1:                             \n"
-        "pswapd (%5, %1), %%mm1         \n"
-        "movq   (%5, %0), %%mm0         \n"
-        "pswapd (%4, %1), %%mm5         \n"
-        "movq   (%3, %0), %%mm4         \n"
-        "movq      %%mm0, %%mm2         \n"
-        "movq      %%mm1, %%mm3         \n"
-        "pfmul     %%mm4, %%mm2         \n" // src0[len + i] * win[len + i]
-        "pfmul     %%mm5, %%mm3         \n" // src1[j]       * win[len + j]
-        "pfmul     %%mm4, %%mm1         \n" // src0[len + i] * win[len + j]
-        "pfmul     %%mm5, %%mm0         \n" // src1[j]       * win[len + i]
-        "pfadd     %%mm3, %%mm2         \n"
-        "pfsub     %%mm0, %%mm1         \n"
-        "pswapd    %%mm2, %%mm2         \n"
-        "movq      %%mm1, (%2, %0)      \n"
-        "movq      %%mm2, (%2, %1)      \n"
-        "sub          $8, %1            \n"
-        "add          $8, %0            \n"
-        "jl           1b                \n"
-        "femms                          \n"
-        : "+r"(i), "+r"(j)
-        : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
-    );
-}
-
-static void vector_fmul_window_sse(float *dst, const float *src0,
-                                   const float *src1, const float *win, int len)
-{
-    x86_reg i = -len * 4;
-    x86_reg j =  len * 4 - 16;
-    __asm__ volatile (
-        "1:                             \n"
-        "movaps      (%5, %1), %%xmm1   \n"
-        "movaps      (%5, %0), %%xmm0   \n"
-        "movaps      (%4, %1), %%xmm5   \n"
-        "movaps      (%3, %0), %%xmm4   \n"
-        "shufps $0x1b, %%xmm1, %%xmm1   \n"
-        "shufps $0x1b, %%xmm5, %%xmm5   \n"
-        "movaps        %%xmm0, %%xmm2   \n"
-        "movaps        %%xmm1, %%xmm3   \n"
-        "mulps         %%xmm4, %%xmm2   \n" // src0[len + i] * win[len + i]
-        "mulps         %%xmm5, %%xmm3   \n" // src1[j]       * win[len + j]
-        "mulps         %%xmm4, %%xmm1   \n" // src0[len + i] * win[len + j]
-        "mulps         %%xmm5, %%xmm0   \n" // src1[j]       * win[len + i]
-        "addps         %%xmm3, %%xmm2   \n"
-        "subps         %%xmm0, %%xmm1   \n"
-        "shufps $0x1b, %%xmm2, %%xmm2   \n"
-        "movaps        %%xmm1, (%2, %0) \n"
-        "movaps        %%xmm2, (%2, %1) \n"
-        "sub              $16, %1       \n"
-        "add              $16, %0       \n"
-        "jl                1b           \n"
-        : "+r"(i), "+r"(j)
-        : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
-    );
-}
-#endif /* HAVE_6REGS && HAVE_INLINE_ASM */
-
 av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
 {
     int cpu_flags = av_get_cpu_flags();
 
-#if HAVE_6REGS && HAVE_INLINE_ASM
-    if (INLINE_AMD3DNOWEXT(cpu_flags)) {
-        fdsp->vector_fmul_window  = vector_fmul_window_3dnowext;
-    }
-    if (INLINE_SSE(cpu_flags)) {
-        fdsp->vector_fmul_window = vector_fmul_window_sse;
+    if (EXTERNAL_AMD3DNOWEXT(cpu_flags)) {
+        fdsp->vector_fmul_window = ff_vector_fmul_window_3dnowext;
     }
-#endif
     if (EXTERNAL_SSE(cpu_flags)) {
         fdsp->vector_fmul = ff_vector_fmul_sse;
         fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_sse;
         fdsp->vector_fmul_scalar = ff_vector_fmul_scalar_sse;
+        fdsp->vector_fmul_window = ff_vector_fmul_window_sse;
         fdsp->vector_fmul_add    = ff_vector_fmul_add_sse;
         fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_sse;
         fdsp->scalarproduct_float = ff_scalarproduct_float_sse;