From f7813a5324be39d13ab536c245d15dfc602a7849 Mon Sep 17 00:00:00 2001
From: Tim Redfern <tim@eclectronics.org>
Date: Sun, 29 Dec 2013 12:19:38 +0000
Subject: basic type mechanism working

---
 ffmpeg/libavutil/x86/float_dsp.asm | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

(limited to 'ffmpeg/libavutil/x86/float_dsp.asm')

diff --git a/ffmpeg/libavutil/x86/float_dsp.asm b/ffmpeg/libavutil/x86/float_dsp.asm
index 004e6cf..49d4876 100644
--- a/ffmpeg/libavutil/x86/float_dsp.asm
+++ b/ffmpeg/libavutil/x86/float_dsp.asm
@@ -263,3 +263,28 @@ cglobal scalarproduct_float, 3,3,2, v1, v2, offset
 %endif
     RET
 
+;-----------------------------------------------------------------------------
+; void ff_butterflies_float(float *src0, float *src1, int len);
+;-----------------------------------------------------------------------------
+INIT_XMM sse
+cglobal butterflies_float, 3,3,3, src0, src1, len
+%if ARCH_X86_64
+    movsxd    lenq, lend
+%endif
+    test      lenq, lenq
+    jz .end
+    shl       lenq, 2
+    add      src0q, lenq
+    add      src1q, lenq
+    neg       lenq
+.loop:
+    mova        m0, [src0q + lenq]
+    mova        m1, [src1q + lenq]
+    subps       m2, m0, m1
+    addps       m0, m0, m1
+    mova        [src1q + lenq], m2
+    mova        [src0q + lenq], m0
+    add       lenq, mmsize
+    jl .loop
+.end:
+    REP_RET
-- 
cgit v1.2.3