* README: add some stuff to read

* configure.ac: check for ieee754.h header
* liboil/colorspace/argb_paint.c: (argb_paint_u8_ref),
(argb_paint_u8_fast): change algorithm to be more accurate
* liboil/colorspace/argb_paint_i386.c: (argb_paint_u8_mmx): same
* liboil/conv/conv_3dnow.c: (conv_f32_s16_3dnow),
(conv_s32_f32_3dnow): make these implementations actually work
* liboil/conv/conv_bitstuff.c: compile fixes
* liboil/conv/conv_misc.c: (conv_f64_s16_table),
(conv_f32_s16_table): add a f32 impl
* liboil/simdpack/diffsquaresum_f64.c: (diffsquaresum_f64_ref),
(diffsquaresum_f64_i10_simple), (diffsquaresum_f64_i10_fast):
fix implementation

9 files changed, 182 insertions, 54 deletions

diff --git a/ChangeLog b/ChangeLog
index 3b5555b..af63a24 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,19 @@
+2005-01-07  David Schleef  <ds@schleef.org>
+
+	* README: add some stuff to read
+	* configure.ac: check for ieee754.h header
+	* liboil/colorspace/argb_paint.c: (argb_paint_u8_ref),
+	(argb_paint_u8_fast): change algorithm to be more accurate
+	* liboil/colorspace/argb_paint_i386.c: (argb_paint_u8_mmx): same
+	* liboil/conv/conv_3dnow.c: (conv_f32_s16_3dnow), 
+	(conv_s32_f32_3dnow): make these implementations actually work
+	* liboil/conv/conv_bitstuff.c: compile fixes
+	* liboil/conv/conv_misc.c: (conv_f64_s16_table),
+	(conv_f32_s16_table): add a f32 impl
+	* liboil/simdpack/diffsquaresum_f64.c: (diffsquaresum_f64_ref),
+	(diffsquaresum_f64_i10_simple), (diffsquaresum_f64_i10_fast):
+	fix implementation
+
 2005-01-05  David Schleef  <ds@schleef.org>
 
 	* testsuite/stride.c: (main), (hist): new test
diff --git a/README b/README
index e3c5f80..12cfe35 100644
--- a/README
+++ b/README
@@ -1,15 +1,53 @@
 
-Ideas:
+ABI warning
+===========
 
-implementations should be flagged as to whether or not they contain
-assembly, and also whether or not they are the product of the
-alternate optimization.  This way, testing code can automatically
-flag functions where the alternate optimization is faster than the
-C or asm code, indicating that the compiler output should be turned
-into assembly.
+In general, liboil is not ABI stable.  However, portions of liboil
+are guaranteed to be stable through the 0.3.x series, and using
+a compatibility library, through the 0.4.x series as well.  This
+policy is designed to provide almost all applications with ABI
+stability for the symbols most used in liboil.
 
+Any symbols declared by including <liboil/liboil.h> follow this
+ABI policy.  This includes all liboil function classes as well
+as a limited number of core functionality, such as oil_init().
 
 
+
+ABI Implementation
+==================
+
+The liboil-0.3.x series create the shared library liboil-0.3.so.
+Applications that use liboil are linked against this library, and
+will load liboil-0.3.so at runtime.
+
+The liboil-0.4.x series will create two shared libraries,
+liboil-0.4.so and a compaitiblity library liboil-0.3.so that
+implements functionality removed in the 0.3->0.4 transition.
+The 0.3 shared library will be binary compatible with the
+liboil-0.3.x releases.  Thus, applictions compiled with a
+liboil-0.3.x release will continue to function after the shared
+library is upgraded to a 0.4.x release.  Applications compiled
+with 0.4.x will use liboil-0.4.so directly.
+
+Distributions can use this compatibility library in order to
+smooth transitions between liboil major releases.  For example,
+consider the case where AppA depends on libB and libC, and both
+libraries use liboil.  When liboil-0.4 is released, the distro
+can start using it immediately, and AppA will continue to work
+correctly.  Then, at a convenient time, libB and libC can
+independently be recompiled using liboil-0.4.x, and the packages
+will no longer depend on liboil-0.3.so.
+
+The traditional way of dealing with these changes is to
+migrate all packages to the new library as quickly as possible.
+This is unduly cumbersome.
+
+
+
+Random Ideas
+============
+
 Why is trans8x8_f64 so slow on powerpc (compared to trans8x8_u16)?
 
 
diff --git a/configure.ac b/configure.ac
index 7429ca8..e63a86e 100644
--- a/configure.ac
+++ b/configure.ac
@@ -86,6 +86,9 @@ AC_CHECK_LIB(m, lrint,
 AC_CHECK_LIB(m, lrintf,
   AC_DEFINE(HAVE_LRINTF, 1, [Define if lrintf() is available]))
 
+AC_CHECK_HEADER(ieee754.h,
+  AC_DEFINE(HAVE_IEEE754_H, 1, [Define if ieee754.h exists]))
+
 AS_COMPILER_FLAG(-Wall, LIBOIL_CFLAGS="$LIBOIL_CFLAGS -Wall")
 if test "x$LIBOIL_CVS" = "xyes"
 then
diff --git a/liboil/colorspace/argb_paint.c b/liboil/colorspace/argb_paint.c
index 6c7598a..7bbb168 100644
--- a/liboil/colorspace/argb_paint.c
+++ b/liboil/colorspace/argb_paint.c
@@ -35,8 +35,8 @@
 OIL_DEFINE_CLASS (argb_paint_u8, "uint8_t *i_4xn, uint8_t *s1_4, uint8_t *s2_n, int n");
 
 
-#define imult(a,b) (((a)*(b) + (((a)*(b)) >> 8))>>8)
-#define apply(a,b,c) (imult(a,255-c) + imult(b,c))
+#define div255(x) (((x + 128) + ((x + 128)>>8))>>8)
+#define blend(x,y,a) div255((x)*(a) + (y)*(255-(a)))
 
 static void
 argb_paint_u8_ref (uint8_t *dest, uint8_t *color, uint8_t *alpha, int n)
@@ -44,10 +44,10 @@ argb_paint_u8_ref (uint8_t *dest, uint8_t *color, uint8_t *alpha, int n)
   int i;
 
   for(i=0;i<n;i++){
-    dest[0] = apply(dest[0],color[0],alpha[0]);
-    dest[1] = apply(dest[1],color[1],alpha[0]);
-    dest[2] = apply(dest[2],color[2],alpha[0]);
-    dest[3] = apply(dest[3],color[3],alpha[0]);
+    dest[0] = blend(color[0],dest[0],alpha[0]);
+    dest[1] = blend(color[1],dest[1],alpha[0]);
+    dest[2] = blend(color[2],dest[2],alpha[0]);
+    dest[3] = blend(color[3],dest[3],alpha[0]);
     dest+=4;
     alpha++;
   }
@@ -68,10 +68,10 @@ argb_paint_u8_fast (uint8_t *dest, uint8_t *color, uint8_t *alpha, int n)
       dest[2] = color[2];
       dest[3] = color[3];
     } else {
-      dest[0] = apply(dest[0],color[0],alpha[0]);
-      dest[1] = apply(dest[1],color[1],alpha[0]);
-      dest[2] = apply(dest[2],color[2],alpha[0]);
-      dest[3] = apply(dest[3],color[3],alpha[0]);
+      dest[0] = blend(color[0],dest[0],alpha[0]);
+      dest[1] = blend(color[1],dest[1],alpha[0]);
+      dest[2] = blend(color[2],dest[2],alpha[0]);
+      dest[3] = blend(color[3],dest[3],alpha[0]);
     }
     dest+=4;
     alpha++;
diff --git a/liboil/colorspace/argb_paint_i386.c b/liboil/colorspace/argb_paint_i386.c
index f6785a4..3753589 100644
--- a/liboil/colorspace/argb_paint_i386.c
+++ b/liboil/colorspace/argb_paint_i386.c
@@ -35,11 +35,12 @@
 OIL_DECLARE_CLASS (argb_paint_u8);
 
 
-#define imult(a,b) (((a)*(b) + (((a)*(b)) >> 8))>>8)
-#define apply(a,b,c) (imult(a,255-c) + imult(b,c))
+#define div255(x) (((x + 128) + ((x + 128)>>8))>>8)
+#define blend(x,y,a) div255((x)*(a) + (y)*(255-(a)))
 
 static short constants[][4] = {
-  { 255, 255, 255, 255 }
+  { 255, 255, 255, 255 },
+  { 128, 128, 128, 128 }
 };
 
 static void
@@ -55,10 +56,10 @@ argb_paint_u8_mmx (uint8_t *dest, uint8_t *color, uint8_t *alpha, int n)
       "  movq (%0), %%mm1\n"
       "  punpcklbw %%mm0, %%mm1\n"
       "  movb (%2), %%al\n"
-      "  je 3f\n"
+      "  je 4f\n"
       "  cmpl $255, %1\n"
       "  jne 2f\n"
-      "  movd %%mm3, (%0)\n"
+      "  movq %%mm3, %%mm2\n"
       "  jmp 3f\n"
       "2:\n"
       "  movd %1, %%mm2\n"
@@ -68,13 +69,15 @@ argb_paint_u8_mmx (uint8_t *dest, uint8_t *color, uint8_t *alpha, int n)
       "  pmullw %%mm1, %%mm4\n"
       "  pmullw %%mm3, %%mm2\n"
       "  paddw %%mm4, %%mm2\n"
+      "  paddw 8(%4), %%mm2\n"
       "  movq %%mm2, %%mm1\n"
       "  psrlw $8, %%mm1\n"
       "  paddw %%mm1, %%mm2\n"
       "  psrlw $8, %%mm2\n"
+      "3: \n"
       "  packuswb %%mm0, %%mm2\n"
       "  movd %%mm2, (%0)\n"
-      "3:\n"
+      "4:\n"
       "  add $4, %0\n"
       "  add $1, %2\n"
       "  decl %3\n"
@@ -85,3 +88,4 @@ argb_paint_u8_mmx (uint8_t *dest, uint8_t *color, uint8_t *alpha, int n)
 }
 OIL_DEFINE_IMPL_FULL (argb_paint_u8_mmx, argb_paint_u8, OIL_IMPL_FLAG_MMX|OIL_IMPL_FLAG_SSE);
 
+
diff --git a/liboil/conv/conv_3dnow.c b/liboil/conv/conv_3dnow.c
index 54da52f..af7df49 100644
--- a/liboil/conv/conv_3dnow.c
+++ b/liboil/conv/conv_3dnow.c
@@ -34,45 +34,52 @@
 
 #ifdef __GNUC__
 /* suboptimal */
-static void conv_f32_s32_3dnow(float *dst, int dst_stride, int32_t *src, int src_stride, int n)
+static void
+conv_f32_s16_3dnow(float *dst, int dst_stride, int16_t *src, int src_stride,
+    int n)
 {
 	int i;
 
-	if (n & 1) 
-	  *dst++ = (float) *src++;
-	n /= 2;
-
 	for(i=0;i<n;i++){
 		asm volatile(
-			"	pi2fd 0(%0), %%mm0		\n"
-			"	movq %%mm0, 0(%1)		\n"
+                        "  xor %%eax, %%eax \n"
+                        "  movw 0(%0), %%eax \n"
+                        "  movd %%eax, %%mm0 \n"
+			"  pi2fd 0(%0), %%mm0 \n"
+			"  movd %%mm0, 0(%1) \n"
 			:
-			: "a" (src), "c" (dst)
-			: "mm0" 
+			: "r" (src), "r" (dst)
+			: "eax", "mm0" 
 		);
 		dst = OIL_OFFSET(dst, dst_stride);
 		src = OIL_OFFSET(src, src_stride);
 	}
         asm volatile ("emms");
 }
-OIL_DEFINE_IMPL_FULL(conv_f32_s32_3dnow, conv_f32_s32,
+OIL_DEFINE_IMPL_FULL(conv_f32_s16_3dnow, conv_f32_s16,
 	OIL_IMPL_FLAG_3DNOW);
 
 /* suboptimal */
-static void conv_s32_f32_3dnow (int32_t *dst, int dst_stride, float *src, int src_stride, int n)
+static void
+conv_s32_f32_3dnow (int32_t *dst, int dst_stride, float *src, int src_stride,
+    int n)
 {
 	int i;
-
-	if (n & 1) 
-	  *src++ = (int32_t) *dst++;
-	n /= 2;
+        const float constants[][2] = {
+          { -0.5, -0.5 },
+          { -1.0, -1.0 }
+        };
 
 	for(i=0;i<n;i++){
 		asm volatile(
-			"	pf2id 0(%0), %%mm0		\n"
-			"	movq %%mm0, 0(%1)		\n"
+			"  movq 0(%0), %%mm0 \n"
+			"  pfadd 0(%2), %%mm0 \n"
+			"  pf2id %%mm0, %%mm1 \n"
+			"  pfcmpgt 0(%2), %%mm0 \n"
+			"  paddd %%mm0, %%mm1 \n"
+			"  movd %%mm1, 0(%1) \n"
 			:
-			: "a" (src), "c" (dst)
+			: "r" (src), "r" (dst), "r" (constants)
 			: "mm0" 
 		);
 		dst = OIL_OFFSET(dst, dst_stride);
diff --git a/liboil/conv/conv_bitstuff.c b/liboil/conv/conv_bitstuff.c
index 82c7b9d..a535bfe 100644
--- a/liboil/conv/conv_bitstuff.c
+++ b/liboil/conv/conv_bitstuff.c
@@ -29,6 +29,7 @@
 #include "config.h"
 #endif
 #include <liboil/liboilfunction.h>
+#include <conv.h>
 
 #ifdef HAVE_IEEE754_H
 
@@ -49,7 +50,7 @@ static void conv_f32_u8_bitstuff(float *dst, int dest_stride, uint8_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f32_u8_bitstuff, conv_f32_u8_class);
+OIL_DEFINE_IMPL(conv_f32_u8_bitstuff, conv_f32_u8);
 
 static void conv_f32_s8_bitstuff(float *dst, int dest_stride, int8_t *src,
 	int src_stride, int n)
@@ -66,7 +67,7 @@ static void conv_f32_s8_bitstuff(float *dst, int dest_stride, int8_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f32_s8_bitstuff, conv_f32_s8_class);
+OIL_DEFINE_IMPL(conv_f32_s8_bitstuff, conv_f32_s8);
 
 static void conv_f32_u16_bitstuff(float *dst, int dest_stride, uint16_t *src,
 	int src_stride, int n)
@@ -83,7 +84,7 @@ static void conv_f32_u16_bitstuff(float *dst, int dest_stride, uint16_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f32_u16_bitstuff, conv_f32_u16_class);
+OIL_DEFINE_IMPL(conv_f32_u16_bitstuff, conv_f32_u16);
 
 static void conv_f32_s16_bitstuff(float *dst, int dest_stride, int16_t *src,
 	int src_stride, int n)
@@ -100,7 +101,7 @@ static void conv_f32_s16_bitstuff(float *dst, int dest_stride, int16_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f32_s16_bitstuff, conv_f32_s16_class);
+OIL_DEFINE_IMPL(conv_f32_s16_bitstuff, conv_f32_s16);
 
 
 #define signbit_S32(x) (((uint32_t)(x))>>31)
@@ -124,7 +125,7 @@ static void conv_s16_f32_bitstuff(int16_t *dst, int dest_stride, float *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_s16_f32_bitstuff, conv_s16_f32_class);
+OIL_DEFINE_IMPL(conv_s16_f32_bitstuff, conv_s16_f32);
 
 
 #if 0
@@ -143,7 +144,7 @@ static void conv_f64_u8_bitstuff(float *dst, int dest_stride, uint8_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f64_u8_bitstuff, conv_f64_u8_class);
+OIL_DEFINE_IMPL(conv_f64_u8_bitstuff, conv_f64_u8);
 
 static void conv_f64_s8_bitstuff(float *dst, int dest_stride, int8_t *src,
 	int src_stride, int n)
@@ -160,7 +161,7 @@ static void conv_f64_s8_bitstuff(float *dst, int dest_stride, int8_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f64_s8_bitstuff, conv_f64_s8_class);
+OIL_DEFINE_IMPL(conv_f64_s8_bitstuff, conv_f64_s8);
 
 static void conv_f64_u16_bitstuff(float *dst, int dest_stride, uint16_t *src,
 	int src_stride, int n)
@@ -177,7 +178,7 @@ static void conv_f64_u16_bitstuff(float *dst, int dest_stride, uint16_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f64_u16_bitstuff, conv_f64_u16_class);
+OIL_DEFINE_IMPL(conv_f64_u16_bitstuff, conv_f64_u16);
 
 static void conv_f64_s16_bitstuff(float *dst, int dest_stride, int16_t *src,
 	int src_stride, int n)
@@ -194,7 +195,7 @@ static void conv_f64_s16_bitstuff(float *dst, int dest_stride, int16_t *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f64_s16_bitstuff, conv_f64_s16_class);
+OIL_DEFINE_IMPL(conv_f64_s16_bitstuff, conv_f64_s16);
 #endif
 
 /* This implementation is slightly inaccurate */
@@ -225,7 +226,7 @@ static void conv_s16_f64_bitstuff(int16_t *dst, int dest_stride, float *src,
 		OIL_INCREMENT (src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_s16_f64_bitstuff, conv_s16_f64_class);
+OIL_DEFINE_IMPL(conv_s16_f64_bitstuff, conv_s16_f64);
 
 #endif
 
diff --git a/liboil/conv/conv_misc.c b/liboil/conv/conv_misc.c
index 73898be..b32250e 100644
--- a/liboil/conv/conv_misc.c
+++ b/liboil/conv/conv_misc.c
@@ -38,6 +38,43 @@ static void
 conv_f64_s16_table(double *dest, int dest_stride, short *src,
 	int src_stride, int n)
 {
+	static double ints_high[256];
+	static double ints_low[256];
+	static int init = 0;
+	int i;
+	unsigned int idx;
+	if(!init){
+		for(i=0;i<256;i++){
+			ints_high[i]=256.0*((i<128)?i:i-256);
+			ints_low[i]=i;
+		}
+		init = 1;
+	}
+
+	if(n&1){
+		idx = (unsigned short)*src;
+		*dest = ints_high[(idx>>8)] + ints_low[(idx&0xff)];
+		OIL_INCREMENT(dest, dest_stride);
+		OIL_INCREMENT(src, src_stride);
+		n-=1;
+	}
+	for(i=0;i<n;i+=2){
+		idx = (unsigned short)*src;
+		*dest = ints_high[(idx>>8)] + ints_low[(idx&0xff)];
+		OIL_INCREMENT(dest, dest_stride);
+		OIL_INCREMENT(src, src_stride);
+		idx = (unsigned short)*src;
+		*dest = ints_high[(idx>>8)] + ints_low[(idx&0xff)];
+		OIL_INCREMENT(dest, dest_stride);
+		OIL_INCREMENT(src, src_stride);
+	}
+}
+OIL_DEFINE_IMPL(conv_f64_s16_table, conv_f64_s16);
+
+static void
+conv_f32_s16_table(float *dest, int dest_stride, short *src,
+	int src_stride, int n)
+{
 	static float ints_high[256];
 	static float ints_low[256];
 	static int init = 0;
@@ -69,6 +106,7 @@ conv_f64_s16_table(double *dest, int dest_stride, short *src,
 		OIL_INCREMENT(src, src_stride);
 	}
 }
-OIL_DEFINE_IMPL(conv_f64_s16_table, conv_f64_s16);
+OIL_DEFINE_IMPL(conv_f32_s16_table, conv_f32_s16);
+
 
 
diff --git a/liboil/simdpack/diffsquaresum_f64.c b/liboil/simdpack/diffsquaresum_f64.c
index 633f90a..f503ccf 100644
--- a/liboil/simdpack/diffsquaresum_f64.c
+++ b/liboil/simdpack/diffsquaresum_f64.c
@@ -47,7 +47,8 @@ diffsquaresum_f64_ref(double *dest, double *src1, int sstr1, double *src2,
 	int i;
 
 	for(i=0;i<n;i++){
-		x = OIL_OFFSET(src1, i*sstr1) - OIL_OFFSET(src2, i*sstr2);
+		x = OIL_GET(src1, i*sstr1, double) -
+                  OIL_GET(src2, i*sstr2, double);
 		x = x*x;
 		tmp = sum;
 		sum += x;
@@ -68,7 +69,8 @@ diffsquaresum_f64_i10_simple(double *dest, double *src1, int sstr1, double *src2
 	int i;
 
 	for(i=0;i<n;i++){
-		x = OIL_OFFSET(src1, i*sstr1) - OIL_OFFSET(src2, i*sstr2);
+		x = OIL_GET(src1, i*sstr1, double) -
+                  OIL_GET(src2, i*sstr2, double);
 		x = x*x;
 		sum += x;
 	}
@@ -78,6 +80,25 @@ diffsquaresum_f64_i10_simple(double *dest, double *src1, int sstr1, double *src2
 OIL_DEFINE_IMPL (diffsquaresum_f64_i10_simple, diffsquaresum_f64);
 
 static void
+diffsquaresum_f64_i10_fast(double *dest, double *src1, int sstr1, double *src2,
+    int sstr2, int n)
+{
+	double sum0 = 0;
+	double x;
+
+	while(n>0){
+		x = *src1 - *src2;
+		sum0 += x * x;
+		OIL_INCREMENT (src1, sstr1);
+		OIL_INCREMENT (src2, sstr2);
+		n--;
+	}
+
+	*dest = sum0;
+}
+OIL_DEFINE_IMPL (diffsquaresum_f64_i10_fast, diffsquaresum_f64);
+
+static void
 diffsquaresum_f64_i10_unroll2(double *dest, double *src1, int sstr1, double *src2,
     int sstr2, int n)
 {