Return to patch-aj CVS log

Up to [cvs.NetBSD.org] / pkgsrc / audio / mpg123 / patches

Integrate ARM 64bit fixedpoint patch submitted by Tim Deegan in pkg/16479.
This feature is enabled by MPG123_ARM_FIXED64.

mpg123/Makefile.common contains common stuff of mpg123 and mpg123-esound.

$NetBSD: patch-aj,v 1.2 2002/06/23 08:45:09 kent Exp $

--- mpg123.h.orig	Fri May 28 23:51:58 1999
+++ mpg123.h
@@ -56,8 +56,115 @@
 #  define real float
 #elif defined(REAL_IS_LONG_DOUBLE)
 #  define real long double
+#elif defined(REAL_IS_FIXED)
+# define real long
+
+# define REAL_RADIX		15
+# define REAL_FACTOR		(32.0 * 1024.0)
+
+# define REAL_PLUS_32767	( 32767 << REAL_RADIX )
+# define REAL_MINUS_32768	( -32768 << REAL_RADIX )
+
+# define DOUBLE_TO_REAL(x)	((int)((x) * REAL_FACTOR))
+# define REAL_TO_SHORT(x)	((x) >> REAL_RADIX)
+# define REAL_MUL(x, y)		(((long long)(x) * (long long)(y)) >> REAL_RADIX)
+
+#elif defined(REAL_IS_FIXED_ARMV4)
+# define REAL_IS_FIXED
+# define real long long
+
+# define REAL_RADIX		32
+# define REAL_FACTOR		(4294967296.0)
+
+# define REAL_PLUS_32767	( 32767LL << REAL_RADIX )
+# define REAL_MINUS_32768	( -32768LL << REAL_RADIX )
+
+# define DOUBLE_TO_REAL(x)	((long long)((x) * REAL_FACTOR))
+# define REAL_TO_SHORT(x)	((x) >> REAL_RADIX)
+
+/*
+ * "Reals" are signed, 15 bits before the point, 32 bits after.
+ * Multiplication is done like so:
+ *
+ *   if |x| < 2^8, shift x >> 7, else shift x >>15
+ *   if |y| < 2^8, shift y >> 7, else shift y >>15
+ *   multiply x * y (just avoids clipping MSB)
+ *   shift the result >> 2, 10 or 18 bits, to make a total shift of 32 bits
+ *
+ * I'd do it in C, but gcc optimises it rather poorly (2K stack
+ * frames, anyone?).  Hence the bit of (ARMv4) asm magic below.
+ * (Unfortunately, the CLZ opcode, which would make for a much neater
+ * solution to this, doesn't appear until ARMv5)
+ *
+ * (tjd April 2002)
+ */
+
+# define REAL_MUL(x, y)		real_mul((x),(y))
+
+static __inline__ long long real_mul(long long x, long long y) {
+    const register long long _x asm("r0") = (x);
+    const register long long _y asm("r2") = (y);
+    register long long rv asm("r0");
+    __asm__ __volatile__ (
+"@ we'll need to shift the result at least >> 2\n"
+"	mov	r4, #2\n"
+"@ both args >> 7,\n"
+"	mov	r0, r0, lsr #7\n"
+"	mov	r2, r2, lsr #7\n"
+"	orr	r0, r0, r1, lsl #25\n"
+"	orr	r2, r2, r3, lsl #25\n"
+"	mov	r1, r1, asr #7\n"
+"	mov	r3, r3, asr #7\n"
+"@ check if x needs further shiftage\n"
+"	cmp	r1, #0\n"
+"	cmnne	r1, #1\n"
+"	addeq	r4, r4, #8\n"
+"@ shift x if needed\n"
+"	movne	r0, r0, lsr #8\n"
+"	orrne	r0, r0, r1, lsl #24\n"
+"	movne	r1, r1, asr #8\n"
+"@ check if y needs further shiftage\n"
+"	cmp	r3, #0\n"
+"	cmnne	r3, #1\n"
+"	addeq	r4, r4, #8\n"
+"@ shift y if needed\n"
+"	movne	r2, r2, lsr #8\n"
+"	orrne	r2, r2, r3, lsl #24\n"
+"	movne	r3, r3, asr #8\n"
+"@ multiply (ignoring overflow)\n"
+"	mul	r3, r0, r3\n"
+"	mla	r3, r1, r2, r3\n"
+"	umull	r0, r1, r2, r0\n"
+"	add	r1, r1, r3\n"
+"@ shift the result the appropriate amount\n"
+"	rsb	r3, r4, #32\n"
+"	mov	r0, r0, lsr r4\n"
+"	orr	r0, r0, r1, lsl r3\n"
+"	mov	r1, r1, asr r4\n"
+	: "=r" (rv)
+	: "0" (_x), "r" (_y)
+	: "r1", "r3", "r4", "cc");
+    return rv;
+}
+
 #else
 #  define real double
+#endif
+
+#ifndef DOUBLE_TO_REAL
+# define DOUBLE_TO_REAL(x)	(x)
+#endif
+#ifndef REAL_TO_SHORT
+# define REAL_TO_SHORT(x)	(x)
+#endif
+#ifndef REAL_PLUS_32767
+# define REAL_PLUS_32767	32767.0
+#endif
+#ifndef REAL_MINUS_32768
+# define REAL_MINUS_32768	-32768.0
+#endif
+#ifndef REAL_MUL
+# define REAL_MUL(x, y)		((x) * (y))
 #endif
 
 #ifdef __GNUC__