diff mbox series

[FFmpeg-devel] avutil/x86/float_dsp: add fma3 for scalarproduct

Message ID CAPYw7P4ivOABgsb6YzPZnqkEf0FFpma7WvWx0jhAcTeP6=1Ujg@mail.gmail.com
State New
Headers show
Series [FFmpeg-devel] avutil/x86/float_dsp: add fma3 for scalarproduct | expand

Checks

Context Check Description
andriy/make_x86 success Make finished
andriy/make_fate_x86 success Make fate finished

Commit Message

Paul B Mahol Sept. 12, 2022, 3:22 p.m. UTC 
Patch attached.
diff mbox series

Patch

From f7c47b8eefa1c06a74d17f13b4e9010785dc6430 Mon Sep 17 00:00:00 2001
From: Paul B Mahol <onemda@gmail.com>
Date: Wed, 20 Jan 2021 16:58:31 +0100
Subject: [PATCH] avutil/x86/float_dsp: add fma3 for scalarproduct

Signed-off-by: Paul B Mahol <onemda@gmail.com>
---
 libavutil/x86/float_dsp.asm    | 127 +++++++++++++++++++++++++++++++++
 libavutil/x86/float_dsp_init.c |   2 +
 2 files changed, 129 insertions(+)

diff --git a/libavutil/x86/float_dsp.asm b/libavutil/x86/float_dsp.asm
index cca4d019c7..8f8e6dddf5 100644
--- a/libavutil/x86/float_dsp.asm
+++ b/libavutil/x86/float_dsp.asm
@@ -440,6 +440,133 @@  cglobal scalarproduct_float, 3,3,2, v1, v2, offset
 %endif
     RET
 
+INIT_YMM fma3
+cglobal scalarproduct_float, 3,5,8, v1, v2, size, len, offset
+    xor   offsetq, offsetq
+    xorps      m0, m0
+    shl     sized, 2
+    mov      lenq, sizeq
+    cmp      lenq, 32
+    jl   .l16
+    cmp      lenq, 64
+    jl   .l32
+    xorps    m1, m1
+    cmp      lenq, 128
+    jl   .l64
+    and    lenq, ~127
+    xorps    m2, m2
+    xorps    m3, m3
+.loop128:
+    movups   m4, [v1q+offsetq]
+    movups   m5, [v1q+offsetq + 32]
+    movups   m6, [v1q+offsetq + 64]
+    movups   m7, [v1q+offsetq + 96]
+    fmaddps  m0, m4, [v2q+offsetq     ], m0
+    fmaddps  m1, m5, [v2q+offsetq + 32], m1
+    fmaddps  m2, m6, [v2q+offsetq + 64], m2
+    fmaddps  m3, m7, [v2q+offsetq + 96], m3
+    add   offsetq, 128
+    cmp   offsetq, lenq
+    jl .loop128
+    addps    m0, m2
+    addps    m1, m3
+    mov      lenq, sizeq
+    and      lenq, 127
+    cmp      lenq, 64
+    jge .l64
+    addps    m0, m1
+    cmp      lenq, 32
+    jge .l32
+    vextractf128 xmm2, m0, 1
+    addps    xmm0, xmm2
+    cmp      lenq, 16
+    jge .l16
+    movhlps  xmm1, xmm0
+    addps    xmm0, xmm1
+    movss    xmm1, xmm0
+    shufps   xmm0, xmm0, 1
+    addss    xmm0, xmm1
+%if ARCH_X86_64 == 0
+    movss r0m, xm0
+    fld dword r0m
+%endif
+    RET
+.l64:
+    and    lenq, ~63
+    add    lenq, offsetq
+.loop64:
+    movups   m4, [v1q+offsetq]
+    movups   m5, [v1q+offsetq + 32]
+    fmaddps  m0, m4, [v2q+offsetq], m0
+    fmaddps  m1, m5, [v2q+offsetq + 32], m1
+    add   offsetq, 64
+    cmp   offsetq, lenq
+    jl .loop64
+    addps    m0, m1
+    mov      lenq, sizeq
+    and      lenq, 63
+    cmp      lenq, 32
+    jge .l32
+    vextractf128 xmm2, m0, 1
+    addps    xmm0, xmm2
+    cmp      lenq, 16
+    jge .l16
+    movhlps  xmm1, xmm0
+    addps    xmm0, xmm1
+    movss    xmm1, xmm0
+    shufps   xmm0, xmm0, 1
+    addss    xmm0, xmm1
+%if ARCH_X86_64 == 0
+    movss r0m, xm0
+    fld dword r0m
+%endif
+    RET
+.l32:
+    and    lenq, ~31
+    add    lenq, offsetq
+.loop32:
+    movups   m4, [v1q+offsetq]
+    fmaddps  m0, m4, [v2q+offsetq], m0
+    add   offsetq, 32
+    cmp   offsetq, lenq
+    jl .loop32
+    vextractf128 xmm2, m0, 1
+    addps    xmm0, xmm2
+    mov      lenq, sizeq
+    and      lenq, 31
+    cmp      lenq, 16
+    jge .l16
+    movhlps  xmm1, xmm0
+    addps    xmm0, xmm1
+    movss    xmm1, xmm0
+    shufps   xmm0, xmm0, 1
+    addss    xmm0, xmm1
+%if ARCH_X86_64 == 0
+    movss r0m, xm0
+    fld dword r0m
+%endif
+    RET
+.l16:
+    and    lenq, ~15
+    add    lenq, offsetq
+.loop16:
+    movaps   xmm1, [v1q+offsetq]
+    mulps    xmm1, [v2q+offsetq]
+    addps    xmm0, xmm1
+    add   offsetq, 16
+    cmp   offsetq, lenq
+    jl .loop16
+    movhlps  xmm1, xmm0
+    addps    xmm0, xmm1
+    movss    xmm1, xmm0
+    shufps   xmm0, xmm0, 1
+    addss    xmm0, xmm1
+%if ARCH_X86_64 == 0
+    movss r0m, xm0
+    fld dword r0m
+%endif
+    RET
+
 ;-----------------------------------------------------------------------------
 ; void ff_butterflies_float(float *src0, float *src1, int len);
 ;-----------------------------------------------------------------------------
diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c
index ad17bc2044..ad6b506259 100644
--- a/libavutil/x86/float_dsp_init.c
+++ b/libavutil/x86/float_dsp_init.c
@@ -74,6 +74,7 @@  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);
+float ff_scalarproduct_float_fma3(const float *v1, const float *v2, int order);
 
 void ff_butterflies_float_sse(float *av_restrict src0, float *av_restrict src1, int len);
 
@@ -112,5 +113,6 @@  av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
         fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_fma3;
         fdsp->vector_fmul_add    = ff_vector_fmul_add_fma3;
         fdsp->vector_dmac_scalar = ff_vector_dmac_scalar_fma3;
+        fdsp->scalarproduct_float = ff_scalarproduct_float_fma3;
     }
 }
-- 
2.37.2