Did not port the 3dnowext version. Added an AVX version. Used SSE2 instead
of SSE to save 2 instructions when loading/reversing by using pshufd instead
of shufps.
---
Sandy Bridge: C-2641 SSE2-961 AVX-820
Athlon64: C-5162 SSE2-2509
libavutil/float_dsp.h | 8 ++--
libavutil/x86/float_dsp.asm | 50 ++++++++++++++++++++++++
libavutil/x86/float_dsp_init.c | 81 ++++------------------------------------
3 files changed, 62 insertions(+), 77 deletions(-)
diff --git a/libavutil/float_dsp.h b/libavutil/float_dsp.h
index d0ceaaf..1cc70cc 100644
--- a/libavutil/float_dsp.h
+++ b/libavutil/float_dsp.h
@@ -90,13 +90,13 @@ typedef struct AVFloatDSPContext {
* @param dst result vector
* constraints: 16-byte aligned
* @param src0 first source vector
- * constraints: 16-byte aligned
+ * constraints: 32-byte aligned
* @param src1 second source vector
- * constraints: 16-byte aligned
+ * constraints: 32-byte aligned
* @param win half-window vector
- * constraints: 16-byte aligned
+ * constraints: 32-byte aligned
* @param len length of vector
- * constraints: multiple of 4
+ * constraints: multiple of 8
*/
void (*vector_fmul_window)(float *dst, const float *src0,
const float *src1, const float *win, int len);
diff --git a/libavutil/x86/float_dsp.asm b/libavutil/x86/float_dsp.asm
index 4113fd9..ae9b0e4 100644
--- a/libavutil/x86/float_dsp.asm
+++ b/libavutil/x86/float_dsp.asm
@@ -162,3 +162,53 @@ VECTOR_DMUL_SCALAR
INIT_YMM avx
VECTOR_DMUL_SCALAR
%endif
+
+;------------------------------------------------------------------------------
+; void ff_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, len1, len0
+ lea len0q, [len1d*4]
+ lea len1q, [len1d*4-mmsize]
+ add dstq, len0q
+ add src0q, len0q
+ add winq, len0q
+ neg len0q
+.loop:
+ mova m0, [winq +len0q]
+ mova m4, [src0q+len0q]
+%if cpuflag(avx)
+ vperm2f128 m1, m1, [winq +len1q], q0203
+ vperm2f128 m5, m5, [src1q+len1q], q0203
+ shufps m1, m1, m1, q0123
+ shufps m1, m5, m5, q0123
+%else
+ pshufd m1, [winq +len1q], q0123
+ pshufd m5, [src1q+len1q], q0123
+%endif
+ mulps m2, m0, m4
+ mulps m3, m1, m5
+ mulps m1, m1, m4
+ mulps m0, m0, m5
+ addps m2, m2, m3
+ subps m1, m1, m0
+ shufps m2, m2, m2, q0123
+%if cpuflag(avx)
+ vextractf128 [dstq+len1q ], m2, 1
+ vextractf128 [dstq+len1q+16], m2, 0
+%else
+ mova [dstq+len1q], m2
+%endif
+ mova [dstq+len0q], m1
+ sub len1q, mmsize
+ add len0q, mmsize
+ jl .loop
+ REP_RET
+%endmacro
+
+INIT_XMM sse2
+VECTOR_FMUL_WINDOW
+INIT_YMM avx
+VECTOR_FMUL_WINDOW
diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c
index 56fb32b..5078af3 100644
--- a/libavutil/x86/float_dsp_init.c
+++ b/libavutil/x86/float_dsp_init.c
@@ -41,84 +41,17 @@ extern void ff_vector_dmul_scalar_sse2(double *dst, const
double *src,
extern void ff_vector_dmul_scalar_avx(double *dst, const double *src,
double mul, int len);
-#if HAVE_6REGS
-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 */
+extern void ff_vector_fmul_window_sse2(float *dst, const float *src0,
+ const float *src1, const float *win,
+ int len);
+extern void ff_vector_fmul_window_avx(float *dst, const float *src0,
+ const float *src1, const float *win,
+ int len);
void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
{
int mm_flags = av_get_cpu_flags();
-#if HAVE_6REGS
- if (INLINE_AMD3DNOWEXT(mm_flags)) {
- fdsp->vector_fmul_window = vector_fmul_window_3dnowext;
- }
- if (INLINE_SSE(mm_flags)) {
- fdsp->vector_fmul_window = vector_fmul_window_sse;
- }
-#endif
if (EXTERNAL_SSE(mm_flags)) {
fdsp->vector_fmul = ff_vector_fmul_sse;
fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_sse;
@@ -126,10 +59,12 @@ void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
}
if (EXTERNAL_SSE2(mm_flags)) {
fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_sse2;
+ fdsp->vector_fmul_window = ff_vector_fmul_window_sse2;
}
if (EXTERNAL_AVX(mm_flags)) {
fdsp->vector_fmul = ff_vector_fmul_avx;
fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_avx;
fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_avx;
+ fdsp->vector_fmul_window = ff_vector_fmul_window_avx;
}
}
--
1.7.1
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
