x86/float_dsp: add SSE2 and AVX versions of scalarproduct_double

Signed-off-by: James Almer <jamrial@gmail.com>

libavutil/x86/float_dsp.asm

@@ -567,6 +567,58 @@ cglobal scalarproduct_float, 3,5,8, v1, v2, size, len, offset
 %endif
     RET
 
+;---------------------------------------------------------------------------------
+; double scalarproduct_double(const double *v1, const double *v2, size_t len)
+;---------------------------------------------------------------------------------
+%macro SCALARPRODUCT_DOUBLE 0
+cglobal scalarproduct_double, 3,3,8, v1, v2, offset
+    shl offsetq, 3
+    add     v1q, offsetq
+    add     v2q, offsetq
+    neg offsetq
+    xorpd    m0, m0
+    xorpd    m1, m1
+    movapd   m2, m0
+    movapd   m3, m1
+align 16
+.loop:
+    movapd   m4, [v1q+offsetq+mmsize*0]
+    movapd   m5, [v1q+offsetq+mmsize*1]
+    movapd   m6, [v1q+offsetq+mmsize*2]
+    movapd   m7, [v1q+offsetq+mmsize*3]
+    mulpd    m4, [v2q+offsetq+mmsize*0]
+    mulpd    m5, [v2q+offsetq+mmsize*1]
+    mulpd    m6, [v2q+offsetq+mmsize*2]
+    mulpd    m7, [v2q+offsetq+mmsize*3]
+    addpd    m0, m4
+    addpd    m1, m5
+    addpd    m2, m6
+    addpd    m3, m7
+    add offsetq, mmsize*4
+    jl .loop
+    addpd    m0, m1
+    addpd    m2, m3
+    addpd    m0, m2
+%if mmsize == 32
+    vextractf128 xm1, m0, 1
+    addpd   xm0, xm1
+%endif
+    movhlps xm1, xm0
+    addsd   xm0, xm1
+%if ARCH_X86_64 == 0
+    movsd   r0m, xm0
+    fld qword r0m
+%endif
+    RET
+%endmacro
+
+INIT_XMM sse2
+SCALARPRODUCT_DOUBLE
+%if HAVE_AVX_EXTERNAL
+INIT_YMM avx
+SCALARPRODUCT_DOUBLE
+%endif
+
 ;-----------------------------------------------------------------------------
 ; void ff_butterflies_float(float *src0, float *src1, int len);
 ;-----------------------------------------------------------------------------

libavutil/x86/float_dsp_init.c

@@ -73,6 +73,9 @@ void ff_vector_fmul_reverse_avx2(float *dst, const float *src0,
 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
 float ff_scalarproduct_float_fma3(const float *v1, const float *v2, int order);
 
+double ff_scalarproduct_double_sse2(const double *v1, const double *v2, size_t order);
+double ff_scalarproduct_double_avx(const double *v1, const double *v2, size_t order);
+
 void ff_butterflies_float_sse(float *restrict src0, float *restrict src1, int len);
 
 av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
@@ -93,6 +96,7 @@ av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
         fdsp->vector_dmul = ff_vector_dmul_sse2;
         fdsp->vector_dmac_scalar = ff_vector_dmac_scalar_sse2;
         fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_sse2;
+        fdsp->scalarproduct_double = ff_scalarproduct_double_sse2;
     }
     if (EXTERNAL_AVX_FAST(cpu_flags)) {
         fdsp->vector_fmul = ff_vector_fmul_avx;
@@ -102,6 +106,7 @@ av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
         fdsp->vector_dmac_scalar = ff_vector_dmac_scalar_avx;
         fdsp->vector_fmul_add    = ff_vector_fmul_add_avx;
         fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_avx;
+        fdsp->scalarproduct_double = ff_scalarproduct_double_avx;
     }
     if (EXTERNAL_AVX2_FAST(cpu_flags)) {
         fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_avx2;