altivec patches by Romain Dolbeau

Originally committed as revision 1423 to svn://svn.ffmpeg.org/ffmpeg/trunk

5 files changed, 191 insertions, 81 deletions

diff --git a/libavcodec/ppc/dsputil_altivec.c b/libavcodec/ppc/dsputil_altivec.c
index 5f14ed0eb9..cf0100d0a4 100644
--- a/libavcodec/ppc/dsputil_altivec.c
+++ b/libavcodec/ppc/dsputil_altivec.c
@@ -26,15 +26,16 @@
 
 int pix_abs16x16_x2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
-    int s, i;
-    vector unsigned char *tv, zero;
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned char zero = (const vector unsigned char)(0);
+    vector unsigned char *tv;
     vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
     vector unsigned int sad;
     vector signed int sumdiffs;
 
     s = 0;
-    zero = vec_splat_u8(0);
-    sad = vec_splat_u32(0);
+    sad = (vector unsigned int)(0);
     for(i=0;i<16;i++) {
         /*
            Read unaligned pixels into our vectors. The vectors are as follows:
@@ -72,16 +73,17 @@ int pix_abs16x16_x2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 
 int pix_abs16x16_y2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
-    int s, i;
-    vector unsigned char *tv, zero;
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned char zero = (const vector unsigned char)(0);
+    vector unsigned char *tv;
     vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
     vector unsigned int sad;
     vector signed int sumdiffs;
     uint8_t *pix3 = pix2 + line_size;
 
     s = 0;
-    zero = vec_splat_u8(0);
-    sad = vec_splat_u32(0);
+    sad = (vector unsigned int)(0);
 
     /*
        Due to the fact that pix3 = pix2 + line_size, the pix3 of one
@@ -131,20 +133,21 @@ int pix_abs16x16_y2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 
 int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
-    int s, i;
+    int i;
+    int s __attribute__((aligned(16)));
     uint8_t *pix3 = pix2 + line_size;
-    vector unsigned char *tv, avgv, t5, zero;
+    const vector unsigned char zero = (const vector unsigned char)(0);
+    const vector unsigned short two = (const vector unsigned short)(2);
+    vector unsigned char *tv, avgv, t5;
     vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
     vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
     vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
-    vector unsigned short avghv, avglv, two;
+    vector unsigned short avghv, avglv;
     vector unsigned short t1, t2, t3, t4;
     vector unsigned int sad;
     vector signed int sumdiffs;
 
-    zero = vec_splat_u8(0);
-    two = vec_splat_u16(2);
-    sad = vec_splat_u32(0);
+    sad = (vector unsigned int)(0);
     
     s = 0;
 
@@ -231,13 +234,14 @@ int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 
 int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
-    int i, s;
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned int zero = (const vector unsigned int)(0);
     vector unsigned char perm1, perm2, *pix1v, *pix2v;
     vector unsigned char t1, t2, t3,t4, t5;
-    vector unsigned int sad, zero;
+    vector unsigned int sad;
     vector signed int sumdiffs;
     
-    zero = (vector unsigned int) (0);
     sad = (vector unsigned int) (0);
 
 
@@ -272,14 +276,15 @@ int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 
 int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
-    int i, s;
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned int zero = (const vector unsigned int)(0);
     vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
     vector unsigned char t1, t2, t3,t4, t5;
-    vector unsigned int sad, zero;
+    vector unsigned int sad;
     vector signed int sumdiffs;
 
-    zero = (vector unsigned int) (0);
-    sad = (vector unsigned int) (0);
+    sad = (vector unsigned int)(0);
     permclear = (vector unsigned char) (255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
 
     for(i=0;i<8;i++) {
@@ -315,14 +320,15 @@ int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 
 int pix_norm1_altivec(uint8_t *pix, int line_size)
 {
-    int s, i;
-    vector unsigned char *tv, zero;
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned int zero = (const vector unsigned int)(0);
+    vector unsigned char *tv;
     vector unsigned char pixv;
     vector unsigned int sv;
     vector signed int sum;
-        
-    zero = vec_splat_u8(0);
-    sv = vec_splat_u32(0);
+    
+    sv = (vector unsigned int)(0);
     
     s = 0;
     for (i = 0; i < 16; i++) {
@@ -343,17 +349,122 @@ int pix_norm1_altivec(uint8_t *pix, int line_size)
     return s;
 }
 
-int pix_sum_altivec(UINT8 * pix, int line_size)
+/**
+ * Sum of Squared Errors for a 8x8 block.
+ * AltiVec-enhanced.
+ * It's the pix_abs8x8_altivec code above w/ squaring added.
+ */
+int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size)
+{
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned int zero = (const vector unsigned int)(0);
+    vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
+    vector unsigned char t1, t2, t3,t4, t5;
+    vector unsigned int sum;
+    vector signed int sumsqr;
+    
+    sum = (vector unsigned int)(0);
+    permclear = (vector unsigned char)(0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
+    
+    for(i=0;i<8;i++) {
+	/* Read potentially unaligned pixels into t1 and t2
+	   Since we're reading 16 pixels, and actually only want 8,
+	   mask out the last 8 pixels. The 0s don't change the sum. */
+        perm1 = vec_lvsl(0, pix1);
+        pix1v = (vector unsigned char *) pix1;
+        perm2 = vec_lvsl(0, pix2);
+        pix2v = (vector unsigned char *) pix2;
+        t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
+        t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
+
+        /*
+          Since we want to use unsigned chars, we can take advantage
+          of the fact that abs(a-b)^2 = (a-b)^2.
+        */
+        
+	/* Calculate abs differences vector */ 
+        t3 = vec_max(t1, t2);
+        t4 = vec_min(t1, t2);
+        t5 = vec_sub(t3, t4);
+        
+        /* Square the values and add them to our sum */
+        sum = vec_msum(t5, t5, sum);
+        
+        pix1 += line_size;
+        pix2 += line_size;
+    }
+    
+    /* Sum up the four partial sums, and put the result into s */
+    sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
+    sumsqr = vec_splat(sumsqr, 3);
+    vec_ste(sumsqr, 0, &s);
+    
+    return s;
+}
+
+/**
+ * Sum of Squared Errors for a 16x16 block.
+ * AltiVec-enhanced.
+ * It's the pix_abs16x16_altivec code above w/ squaring added.
+ */
+int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size)
 {
+    int i;
+    int s __attribute__((aligned(16)));
+    const vector unsigned int zero = (const vector unsigned int)(0);
+    vector unsigned char perm1, perm2, *pix1v, *pix2v;
+    vector unsigned char t1, t2, t3,t4, t5;
+    vector unsigned int sum;
+    vector signed int sumsqr;
+    
+    sum = (vector unsigned int)(0);
+    
+    for(i=0;i<16;i++) {
+	/* Read potentially unaligned pixels into t1 and t2 */
+        perm1 = vec_lvsl(0, pix1);
+        pix1v = (vector unsigned char *) pix1;
+        perm2 = vec_lvsl(0, pix2);
+        pix2v = (vector unsigned char *) pix2;
+        t1 = vec_perm(pix1v[0], pix1v[1], perm1);
+        t2 = vec_perm(pix2v[0], pix2v[1], perm2);
+
+        /*
+          Since we want to use unsigned chars, we can take advantage
+          of the fact that abs(a-b)^2 = (a-b)^2.
+        */
+        
+	/* Calculate abs differences vector */ 
+        t3 = vec_max(t1, t2);
+        t4 = vec_min(t1, t2);
+        t5 = vec_sub(t3, t4);
+        
+        /* Square the values and add them to our sum */
+        sum = vec_msum(t5, t5, sum);
+        
+        pix1 += line_size;
+        pix2 += line_size;
+    }
+    
+    /* Sum up the four partial sums, and put the result into s */
+    sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
+    sumsqr = vec_splat(sumsqr, 3);
+    vec_ste(sumsqr, 0, &s);
+    
+    return s;
+}
 
+int pix_sum_altivec(UINT8 * pix, int line_size)
+{
+    const vector unsigned int zero = (const vector unsigned int)(0);
     vector unsigned char perm, *pixv;
     vector unsigned char t1;
-    vector unsigned int sad, zero;
+    vector unsigned int sad;
     vector signed int sumdiffs;
 
-    int s, i;
-
-    zero = (vector unsigned int) (0);
+    int i;
+    int s __attribute__((aligned(16)));
+    
     sad = (vector unsigned int) (0);
     
     for (i = 0; i < 16; i++) {
@@ -380,7 +491,7 @@ void get_pixels_altivec(DCTELEM *restrict block, const UINT8 *pixels, int line_s
 {
     int i;
     vector unsigned char perm, bytes, *pixv;
-    vector unsigned char zero = (vector unsigned char) (0);
+    const vector unsigned char zero = (const vector unsigned char) (0);
     vector signed short shorts;
 
     for(i=0;i<8;i++)
@@ -407,7 +518,7 @@ void diff_pixels_altivec(DCTELEM *restrict block, const UINT8 *s1,
 {
     int i;
     vector unsigned char perm, bytes, *pixv;
-    vector unsigned char zero = (vector unsigned char) (0);
+    const vector unsigned char zero = (const vector unsigned char) (0);
     vector signed short shorts1, shorts2;
 
     for(i=0;i<4;i++)
diff --git a/libavcodec/ppc/dsputil_altivec.h b/libavcodec/ppc/dsputil_altivec.h
index d4d259d9ed..cef29b8806 100644
--- a/libavcodec/ppc/dsputil_altivec.h
+++ b/libavcodec/ppc/dsputil_altivec.h
@@ -23,8 +23,37 @@ extern int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 extern int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_norm1_altivec(uint8_t *pix, int line_size);
+extern int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size);
+extern int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_sum_altivec(UINT8 * pix, int line_size);
 extern void diff_pixels_altivec(DCTELEM* block, const UINT8* s1, const UINT8* s2, int stride);
 extern void get_pixels_altivec(DCTELEM* block, const UINT8 * pixels, int line_size);
 
 extern int has_altivec(void);
+
+
+
+// used to build registers permutation vectors (vcprm)
+// the 's' are for words in the _s_econd vector
+#define WORD_0 0x00,0x01,0x02,0x03
+#define WORD_1 0x04,0x05,0x06,0x07
+#define WORD_2 0x08,0x09,0x0a,0x0b
+#define WORD_3 0x0c,0x0d,0x0e,0x0f
+#define WORD_s0 0x10,0x11,0x12,0x13
+#define WORD_s1 0x14,0x15,0x16,0x17
+#define WORD_s2 0x18,0x19,0x1a,0x1b
+#define WORD_s3 0x1c,0x1d,0x1e,0x1f
+
+#define vcprm(a,b,c,d) (const vector unsigned char)(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d)
+
+// vcprmle is used to keep the same index as in the SSE version.
+// it's the same as vcprm, with the index inversed
+// ('le' is Little Endian)
+#define vcprmle(a,b,c,d) vcprm(d,c,b,a)
+
+// used to build inverse/identity vectors (vcii)
+// n is _n_egative, p is _p_ositive
+#define FLOAT_n -1.
+#define FLOAT_p 1.
+
+#define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d)
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index 733d0c156e..8713e891be 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -42,9 +42,12 @@ void dsputil_init_ppc(DSPContext* c, unsigned mask)
         c->pix_abs16x16 = pix_abs16x16_altivec;
         c->pix_abs8x8 = pix_abs8x8_altivec;
         c->pix_norm1 = pix_norm1_altivec;
+        c->sse[1]= sse8_altivec;
+        c->sse[0]= sse16_altivec;
         c->pix_sum = pix_sum_altivec;
         c->diff_pixels = diff_pixels_altivec;
         c->get_pixels = get_pixels_altivec;
+
     } else
 #endif
     {
diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c
index 1a926b77ca..eb31484148 100644
--- a/libavcodec/ppc/fft_altivec.c
+++ b/libavcodec/ppc/fft_altivec.c
@@ -1,7 +1,7 @@
 /*
  * FFT/IFFT transforms
  * AltiVec-enabled
- * Copyright (c) 2002 Romain Dolbeau <romain@dolbeau.org>
+ * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
  * Based on code Copyright (c) 2002 Fabrice Bellard.
  *
  * This library is free software; you can redistribute it and/or
@@ -22,31 +22,6 @@
 
 #include "dsputil_altivec.h"
 
-// used to build registers permutation vectors (vcprm)
-// the 's' are for words in the _s_econd vector
-#define WORD_0 0x00,0x01,0x02,0x03
-#define WORD_1 0x04,0x05,0x06,0x07
-#define WORD_2 0x08,0x09,0x0a,0x0b
-#define WORD_3 0x0c,0x0d,0x0e,0x0f
-#define WORD_s0 0x10,0x11,0x12,0x13
-#define WORD_s1 0x14,0x15,0x16,0x17
-#define WORD_s2 0x18,0x19,0x1a,0x1b
-#define WORD_s3 0x1c,0x1d,0x1e,0x1f
-
-#define vcprm(a,b,c,d) (const vector unsigned char)(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d)
-
-// vcprmle is used to keep the same index as in the SSE version.
-// it's the same as vcprm, with the index inversed
-// ('le' is Little Endian)
-#define vcprmle(a,b,c,d) vcprm(d,c,b,a)
-
-// used to build inverse/identity vectors (vcii)
-// n is _n_egative, p is _p_ositive
-#define FLOAT_n -1.
-#define FLOAT_p 1.
-
-#define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d)
-
 /**
  * Do a complex FFT with the parameters defined in fft_init(). The
  * input data must be permuted before with s->revtab table. No
@@ -55,16 +30,8 @@
  * This code assumes that the 'z' pointer is 16 bytes-aligned
  * It also assumes all FFTComplex are 8 bytes-aligned pair of float
  * The code is exactly the same as the SSE version, except
- * that successive MUL + ADD/SUB have been fusionned into
+ * that successive MUL + ADD/SUB have been merged into
  * fused multiply-add ('vec_madd' in altivec)
- *
- * To test this code you can use fft-test in libavcodec ; use
- * the following line in libavcodec to compile (MacOS X):
- * #####
- * gcc -I. -Ippc -no-cpp-precomp -pipe -O3 -fomit-frame-pointer -mdynamic-no-pic -Wall
- *     -faltivec -DARCH_POWERPC -DHAVE_ALTIVEC -DCONFIG_DARWIN fft-test.c fft.c
- *     ppc/fft_altivec.c ppc/dsputil_altivec.c mdct.c -DHAVE_LRINTF -o fft-test
- * #####
  */
 void fft_calc_altivec(FFTContext *s, FFTComplex *z)
 {
diff --git a/libavcodec/ppc/mpegvideo_altivec.c b/libavcodec/ppc/mpegvideo_altivec.c
index bcbc1e6baf..ccd32c4e4d 100644
--- a/libavcodec/ppc/mpegvideo_altivec.c
+++ b/libavcodec/ppc/mpegvideo_altivec.c
@@ -100,7 +100,7 @@ int dct_quantize_altivec(MpegEncContext* s,
     int lastNonZero;
     vector float row0, row1, row2, row3, row4, row5, row6, row7;
     vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
-    const vector float zero = {FOUR_INSTANCES(0.0f)};
+    const vector float zero = (vector float)(FOUR_INSTANCES(0.0f));
 
     // Load the data into the row/alt vectors
     {
@@ -144,18 +144,18 @@ int dct_quantize_altivec(MpegEncContext* s,
 		// in the vector local variables, as floats, which we'll use during the
 		// quantize step...
     {
-        const vector float vec_0_298631336 = {FOUR_INSTANCES(0.298631336f)};
-        const vector float vec_0_390180644 = {FOUR_INSTANCES(-0.390180644f)};
-        const vector float vec_0_541196100 = {FOUR_INSTANCES(0.541196100f)};
-        const vector float vec_0_765366865 = {FOUR_INSTANCES(0.765366865f)};
-        const vector float vec_0_899976223 = {FOUR_INSTANCES(-0.899976223f)};
-        const vector float vec_1_175875602 = {FOUR_INSTANCES(1.175875602f)};
-        const vector float vec_1_501321110 = {FOUR_INSTANCES(1.501321110f)};
-        const vector float vec_1_847759065 = {FOUR_INSTANCES(-1.847759065f)};
-        const vector float vec_1_961570560 = {FOUR_INSTANCES(-1.961570560f)};
-        const vector float vec_2_053119869 = {FOUR_INSTANCES(2.053119869f)};
-        const vector float vec_2_562915447 = {FOUR_INSTANCES(-2.562915447f)};
-        const vector float vec_3_072711026 = {FOUR_INSTANCES(3.072711026f)};
+        const vector float vec_0_298631336 = (vector float)(FOUR_INSTANCES(0.298631336f));
+        const vector float vec_0_390180644 = (vector float)(FOUR_INSTANCES(-0.390180644f));
+        const vector float vec_0_541196100 = (vector float)(FOUR_INSTANCES(0.541196100f));
+        const vector float vec_0_765366865 = (vector float)(FOUR_INSTANCES(0.765366865f));
+        const vector float vec_0_899976223 = (vector float)(FOUR_INSTANCES(-0.899976223f));
+        const vector float vec_1_175875602 = (vector float)(FOUR_INSTANCES(1.175875602f));
+        const vector float vec_1_501321110 = (vector float)(FOUR_INSTANCES(1.501321110f));
+        const vector float vec_1_847759065 = (vector float)(FOUR_INSTANCES(-1.847759065f));
+        const vector float vec_1_961570560 = (vector float)(FOUR_INSTANCES(-1.961570560f));
+        const vector float vec_2_053119869 = (vector float)(FOUR_INSTANCES(2.053119869f));
+        const vector float vec_2_562915447 = (vector float)(FOUR_INSTANCES(-2.562915447f));
+        const vector float vec_3_072711026 = (vector float)(FOUR_INSTANCES(3.072711026f));
 
 
         int whichPass, whichHalf;