altivec float optimizations

Originally committed as revision 5999 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 196 insertions, 2 deletions

diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index f1c5c032f0..80f2201c14 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -351,7 +351,8 @@ OBJS-$(TARGET_ALTIVEC)                 += ppc/dsputil_altivec.o      \
                                           ppc/fdct_altivec.o         \
                                           ppc/h264_altivec.o 	     \
                                           ppc/snow_altivec.o 	     \
-                                          ppc/vc1dsp_altivec.o
+                                          ppc/vc1dsp_altivec.o       \
+                                          ppc/float_altivec.o
 
 CFLAGS += $(CFLAGS-yes)
 OBJS += $(OBJS-yes)
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index 156d9c7ab4..f022153866 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -36,6 +36,7 @@ void dsputil_h264_init_ppc(DSPContext* c, AVCodecContext *avctx);
 void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx);
 void vc1dsp_init_altivec(DSPContext* c, AVCodecContext *avctx);
 void snow_init_altivec(DSPContext* c, AVCodecContext *avctx);
+void float_init_altivec(DSPContext* c, AVCodecContext *avctx);
 
 #endif
 
@@ -270,7 +271,7 @@ void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
         dsputil_init_altivec(c, avctx);
         snow_init_altivec(c, avctx);
         vc1dsp_init_altivec(c, avctx);
-
+        float_init_altivec(c, avctx);
         c->gmc1 = gmc1_altivec;
 
 #ifdef CONFIG_ENCODERS
diff --git a/libavcodec/ppc/float_altivec.c b/libavcodec/ppc/float_altivec.c
new file mode 100644
index 0000000000..89d3ccb5f3
--- /dev/null
+++ b/libavcodec/ppc/float_altivec.c
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "../dsputil.h"
+
+#include "gcc_fixes.h"
+
+#include "dsputil_altivec.h"
+
+static void vector_fmul_altivec(float *dst, const float *src, int len)
+{
+    int i;
+    vector float d0, d1, s, zero = (vector float)vec_splat_u32(0);
+    for(i=0; i<len-7; i+=8) {
+        d0 = vec_ld(0, dst+i);
+        s = vec_ld(0, src+i);
+        d1 = vec_ld(16, dst+i);
+        d0 = vec_madd(d0, s, zero);
+        d1 = vec_madd(d1, vec_ld(16,src+i), zero);
+        vec_st(d0, 0, dst+i);
+        vec_st(d1, 16, dst+i);
+    }
+}
+
+static void vector_fmul_reverse_altivec(float *dst, const float *src0,
+                                        const float *src1, int len)
+{
+    int i;
+    vector float d, s0, s1, h0, l0,
+                 s2, s3, zero = (vector float)vec_splat_u32(0);
+    src1 += len-4;
+    for(i=0; i<len-7; i+=8) {
+        s1 = vec_ld(0, src1-i);              // [a,b,c,d]
+        s0 = vec_ld(0, src0+i);
+        l0 = vec_mergel(s1, s1);             // [c,c,d,d]
+        s3 = vec_ld(-16, src1-i);
+        h0 = vec_mergeh(s1, s1);             // [a,a,b,b]
+        s2 = vec_ld(16, src0+i);
+        s1 = vec_mergeh(vec_mergel(l0,h0),   // [d,b,d,b]
+                        vec_mergeh(l0,h0));  // [c,a,c,a]
+                                             // [d,c,b,a]
+        l0 = vec_mergel(s3, s3);
+        d = vec_madd(s0, s1, zero);
+        h0 = vec_mergeh(s3, s3);
+        vec_st(d, 0, dst+i);
+        s3 = vec_mergeh(vec_mergel(l0,h0),
+                        vec_mergeh(l0,h0));
+        d = vec_madd(s2, s3, zero);
+        vec_st(d, 16, dst+i);
+    }
+}
+
+static void vector_fmul_add_add_altivec(float *dst, const float *src0,
+                                        const float *src1, const float *src2,
+                                        int src3, int len, int step)
+{
+    int i;
+    vector float d, s0, s1, s2, t0, t1, edges;
+    vector unsigned char align = vec_lvsr(0,dst),
+                         mask = vec_lvsl(0, dst);
+
+    t0 = vec_ld(0, dst);
+#if 0 //FIXME: there is still something wrong
+    if (step == 2) {
+        int y;
+        vector float d0, d1, s3, t2;
+        vector unsigned int sel =
+                vec_mergeh(vec_splat_u32(-1), vec_splat_u32(0));
+        t1 = vec_ld(16, dst);
+        for (i=0,y=0; i<len-3; i+=4,y+=8) {
+
+            s0 = vec_ld(0,src0+i);
+            s1 = vec_ld(0,src1+i);
+            s2 = vec_ld(0,src2+i);
+
+//          t0 = vec_ld(0, dst+y);  //[x x x|a]
+//          t1 = vec_ld(16, dst+y); //[b c d|e]
+            t2 = vec_ld(31, dst+y); //[f g h|x]
+
+            d = vec_madd(s0,s1,s2); // [A B C D]
+
+                                                 // [A A B B]
+
+                                                 // [C C D D]
+
+            d0 = vec_perm(t0, t1, mask); // [a b c d]
+
+            d0 = vec_sel(vec_mergeh(d, d), d0, sel);   // [A b B d]
+
+            edges = vec_perm(t1, t0, mask);
+
+            t0 = vec_perm(edges, d0, align); // [x x x|A]
+
+            t1 = vec_perm(d0, edges, align); // [b B d|e]
+
+            vec_stl(t0, 0, dst+y);
+
+            d1 = vec_perm(t1, t2, mask); // [e f g h]
+
+            d1 = vec_sel(vec_mergel(d, d), d1, sel); // [C f D h]
+
+            edges = vec_perm(t2, t1, mask);
+
+            t1 = vec_perm(edges, d1, align); // [b B d|C]
+
+            t2 = vec_perm(d1, edges, align); // [f D h|x]
+
+            vec_stl(t1, 16, dst+y);
+
+            t0 = t1;
+
+            vec_stl(t2, 31, dst+y);
+
+            t1 = t2;
+        }
+    } else
+    #endif
+    if (step == 1 && src3 == 0)
+        for (i=0; i<len-3; i+=4) {
+            t1 = vec_ld(15, dst+i);
+            s0 = vec_ld(0, src0+i);
+            s1 = vec_ld(0, src1+i);
+            s2 = vec_ld(0, src2+i);
+            edges = vec_perm(t1 ,t0, mask);
+            d = vec_madd(s0,s1,s2);
+            t0 = vec_perm(edges, d, align);
+            t1 = vec_perm(d, edges, align);
+            vec_st(t0, 0, dst+i);
+            vec_st(t1, 15, dst+i);
+            t0 = t1;
+        }
+    else
+        ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step);
+}
+
+void float_to_int16_altivec(int16_t *dst, const float *src, int len)
+{
+    int i;
+    vector float s0, s1;
+    vector signed int t0, t1;
+    vector signed short d0, d1, d;
+    vector unsigned char align;
+    if(((long)dst)&15) //FIXME
+    for(i=0; i<len-7; i+=8) {
+        s0 = vec_ld(0, src+i);
+        s1 = vec_ld(16, src+i);
+        t0 = vec_cts(s0, 0);
+        d0 = vec_ld(0, dst+i);
+        t1 = vec_cts(s1, 0);
+        d1 = vec_ld(15, dst+i);
+        d = vec_packs(t0,t1);
+        d1 = vec_perm(d1, d0, vec_lvsl(0,dst+i));
+        align = vec_lvsr(0, dst+i);
+        d0 = vec_perm(d1, d, align);
+        d1 = vec_perm(d, d1, align);
+        vec_st(d0, 0, dst+i);
+        vec_st(d1,15, dst+i);
+    }
+    else
+    for(i=0; i<len-7; i+=8) {
+        s0 = vec_ld(0, src+i);
+        s1 = vec_ld(16, src+i);
+        t0 = vec_cts(s0, 0);
+        t1 = vec_cts(s1, 0);
+        d = vec_packs(t0,t1);
+        vec_st(d, 0, dst+i);
+    }
+}
+
+void float_init_altivec(DSPContext* c, AVCodecContext *avctx)
+{
+    c->vector_fmul = vector_fmul_altivec;
+    c->vector_fmul_reverse = vector_fmul_reverse_altivec;
+    c->vector_fmul_add_add = vector_fmul_add_add_altivec;
+    if(!(avctx->flags & CODEC_FLAG_BITEXACT))
+        c->float_to_int16 = float_to_int16_altivec;
+}