x86/float_dsp: add ff_vector_fmul_reverse_avx2

~20% faster than AVX.
Signed-off-by: James Almer <jamrial@gmail.com>

libavutil/x86/float_dsp.asm

@@ -22,6 +22,9 @@
 
 %include "x86util.asm"
 
+SECTION_RODATA 32
+pd_reverse: dd 7, 6, 5, 4, 3, 2, 1, 0
+
 SECTION .text
 
 ;-----------------------------------------------------------------------------
@@ -359,10 +362,16 @@ VECTOR_FMUL_ADD
 ;-----------------------------------------------------------------------------
 %macro VECTOR_FMUL_REVERSE 0
 cglobal vector_fmul_reverse, 4,4,2, dst, src0, src1, len
+%if cpuflag(avx2)
+    mova    m2, [pd_reverse]
+%endif
     lea       lenq, [lend*4 - 2*mmsize]
 ALIGN 16
 .loop:
-%if cpuflag(avx)
+%if cpuflag(avx2)
+    vpermd  m0, m2, [src1q]
+    vpermd  m1, m2, [src1q+mmsize]
+%elif cpuflag(avx)
     vmovaps     xmm0, [src1q + 16]
     vinsertf128 m0, m0, [src1q], 1
     vshufps     m0, m0, m0, q0123
@@ -391,6 +400,10 @@ VECTOR_FMUL_REVERSE
 INIT_YMM avx
 VECTOR_FMUL_REVERSE
 %endif
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+VECTOR_FMUL_REVERSE
+%endif
 
 ; float scalarproduct_float_sse(const float *v1, const float *v2, int len)
 INIT_XMM sse

libavutil/x86/float_dsp_init.c

@@ -67,6 +67,8 @@ void ff_vector_fmul_reverse_sse(float *dst, const float *src0,
                                 const float *src1, int len);
 void ff_vector_fmul_reverse_avx(float *dst, const float *src0,
                                 const float *src1, int len);
+void ff_vector_fmul_reverse_avx2(float *dst, const float *src0,
+                                 const float *src1, int len);
 
 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
 
@@ -101,6 +103,9 @@ av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
         fdsp->vector_fmul_add    = ff_vector_fmul_add_avx;
         fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_avx;
     }
+    if (EXTERNAL_AVX2_FAST(cpu_flags)) {
+        fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_avx2;
+    }
     if (EXTERNAL_FMA3_FAST(cpu_flags)) {
         fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_fma3;
         fdsp->vector_fmul_add    = ff_vector_fmul_add_fma3;