/*
* Mesa 3-D graphics library
- * Version: 7.1
*
* Copyright (C) 1999-2007 Brian Paul All Rights Reserved.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
- * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
- * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
- * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
*/
#endif
-#ifdef WIN32
+#ifdef _WIN32
#define vsnprintf _vsnprintf
-#elif defined(__IBMC__) || defined(__IBMCPP__) || ( defined(__VMS) && __CRTL_VER < 70312000 )
+#elif defined(__IBMC__) || defined(__IBMCPP__)
extern int vsnprintf(char *str, size_t count, const char *fmt, va_list arg);
-#ifdef __VMS
-#include "vsnprintf.c"
-#endif
#endif
/**********************************************************************/
ASSERT( alignment > 0 );
- ptr = (uintptr_t) malloc(bytes + alignment + sizeof(void *));
+ ptr = (uintptr_t)malloc(bytes + alignment + sizeof(void *));
if (!ptr)
return NULL;
ASSERT( alignment > 0 );
- ptr = (uintptr_t) calloc(1, bytes + alignment + sizeof(void *));
+ ptr = (uintptr_t)calloc(1, bytes + alignment + sizeof(void *));
if (!ptr)
return NULL;
* \param ptr pointer to the memory to be freed.
* The actual address to free is stored in the word immediately before the
* address the client sees.
+ * Note that it is legal to pass NULL pointer to this function and will be
+ * handled accordingly.
*/
void
_mesa_align_free(void *ptr)
#elif defined(_WIN32) && defined(_MSC_VER)
_aligned_free(ptr);
#else
- void **cubbyHole = (void **) ((char *) ptr - sizeof(void *));
- void *realAddr = *cubbyHole;
- free(realAddr);
+ if (ptr) {
+ void **cubbyHole = (void **) ((char *) ptr - sizeof(void *));
+ void *realAddr = *cubbyHole;
+ free(realAddr);
+ }
#endif /* defined(HAVE_POSIX_MEMALIGN) */
}
if (newBuf && oldBuffer && copySize > 0) {
memcpy(newBuf, oldBuffer, copySize);
}
- if (oldBuffer)
- _mesa_align_free(oldBuffer);
+
+ _mesa_align_free(oldBuffer);
return newBuf;
#endif
}
void *newBuffer = malloc(newSize);
if (newBuffer && oldBuffer && copySize > 0)
memcpy(newBuffer, oldBuffer, copySize);
- if (oldBuffer)
- free(oldBuffer);
+ free(oldBuffer);
return newBuffer;
}
-/**
- * Fill memory with a constant 16bit word.
- * \param dst destination pointer.
- * \param val value.
- * \param n number of words.
- */
-void
-_mesa_memset16( unsigned short *dst, unsigned short val, size_t n )
-{
- while (n-- > 0)
- *dst++ = val;
-}
-
/*@}*/
/** \name Math */
/*@{*/
-/** Wrapper around sqrt() */
-double
-_mesa_sqrtd(double x)
-{
- return sqrt(x);
-}
-
-
-/*
- * A High Speed, Low Precision Square Root
- * by Paul Lalonde and Robert Dawson
- * from "Graphics Gems", Academic Press, 1990
- *
- * SPARC implementation of a fast square root by table
- * lookup.
- * SPARC floating point format is as follows:
- *
- * BIT 31 30 23 22 0
- * sign exponent mantissa
- */
-static short sqrttab[0x100]; /* declare table of square roots */
-
-void
-_mesa_init_sqrt_table(void)
-{
-#if defined(USE_IEEE) && !defined(DEBUG)
- unsigned short i;
- fi_type fi; /* to access the bits of a float in C quickly */
- /* we use a union defined in glheader.h */
-
- for(i=0; i<= 0x7f; i++) {
- fi.i = 0;
-
- /*
- * Build a float with the bit pattern i as mantissa
- * and an exponent of 0, stored as 127
- */
-
- fi.i = (i << 16) | (127 << 23);
- fi.f = _mesa_sqrtd(fi.f);
-
- /*
- * Take the square root then strip the first 7 bits of
- * the mantissa into the table
- */
-
- sqrttab[i] = (fi.i & 0x7fffff) >> 16;
-
- /*
- * Repeat the process, this time with an exponent of
- * 1, stored as 128
- */
-
- fi.i = 0;
- fi.i = (i << 16) | (128 << 23);
- fi.f = sqrt(fi.f);
- sqrttab[i+0x80] = (fi.i & 0x7fffff) >> 16;
- }
-#else
- (void) sqrttab; /* silence compiler warnings */
-#endif /*HAVE_FAST_MATH*/
-}
-
-
-/**
- * Single precision square root.
- */
-float
-_mesa_sqrtf( float x )
-{
-#if defined(USE_IEEE) && !defined(DEBUG)
- fi_type num;
- /* to access the bits of a float in C
- * we use a union from glheader.h */
-
- short e; /* the exponent */
- if (x == 0.0F) return 0.0F; /* check for square root of 0 */
- num.f = x;
- e = (num.i >> 23) - 127; /* get the exponent - on a SPARC the */
- /* exponent is stored with 127 added */
- num.i &= 0x7fffff; /* leave only the mantissa */
- if (e & 0x01) num.i |= 0x800000;
- /* the exponent is odd so we have to */
- /* look it up in the second half of */
- /* the lookup table, so we set the */
- /* high bit */
- e >>= 1; /* divide the exponent by two */
- /* note that in C the shift */
- /* operators are sign preserving */
- /* for signed operands */
- /* Do the table lookup, based on the quaternary mantissa,
- * then reconstruct the result back into a float
- */
- num.i = ((sqrttab[num.i >> 16]) << 16) | ((e + 127) << 23);
-
- return num.f;
-#else
- return (float) _mesa_sqrtd((double) x);
-#endif
-}
-
-
-/**
- inv_sqrt - A single precision 1/sqrt routine for IEEE format floats.
- written by Josh Vanderhoof, based on newsgroup posts by James Van Buskirk
- and Vesa Karvonen.
-*/
-float
-_mesa_inv_sqrtf(float n)
-{
-#if defined(USE_IEEE) && !defined(DEBUG)
- float r0, x0, y0;
- float r1, x1, y1;
- float r2, x2, y2;
-#if 0 /* not used, see below -BP */
- float r3, x3, y3;
-#endif
- fi_type u;
- unsigned int magic;
-
- /*
- Exponent part of the magic number -
-
- We want to:
- 1. subtract the bias from the exponent,
- 2. negate it
- 3. divide by two (rounding towards -inf)
- 4. add the bias back
-
- Which is the same as subtracting the exponent from 381 and dividing
- by 2.
-
- floor(-(x - 127) / 2) + 127 = floor((381 - x) / 2)
- */
-
- magic = 381 << 23;
-
- /*
- Significand part of magic number -
-
- With the current magic number, "(magic - u.i) >> 1" will give you:
-
- for 1 <= u.f <= 2: 1.25 - u.f / 4
- for 2 <= u.f <= 4: 1.00 - u.f / 8
-
- This isn't a bad approximation of 1/sqrt. The maximum difference from
- 1/sqrt will be around .06. After three Newton-Raphson iterations, the
- maximum difference is less than 4.5e-8. (Which is actually close
- enough to make the following bias academic...)
-
- To get a better approximation you can add a bias to the magic
- number. For example, if you subtract 1/2 of the maximum difference in
- the first approximation (.03), you will get the following function:
-
- for 1 <= u.f <= 2: 1.22 - u.f / 4
- for 2 <= u.f <= 3.76: 0.97 - u.f / 8
- for 3.76 <= u.f <= 4: 0.72 - u.f / 16
- (The 3.76 to 4 range is where the result is < .5.)
-
- This is the closest possible initial approximation, but with a maximum
- error of 8e-11 after three NR iterations, it is still not perfect. If
- you subtract 0.0332281 instead of .03, the maximum error will be
- 2.5e-11 after three NR iterations, which should be about as close as
- is possible.
-
- for 1 <= u.f <= 2: 1.2167719 - u.f / 4
- for 2 <= u.f <= 3.73: 0.9667719 - u.f / 8
- for 3.73 <= u.f <= 4: 0.7167719 - u.f / 16
-
- */
-
- magic -= (int)(0.0332281 * (1 << 25));
-
- u.f = n;
- u.i = (magic - u.i) >> 1;
-
- /*
- Instead of Newton-Raphson, we use Goldschmidt's algorithm, which
- allows more parallelism. From what I understand, the parallelism
- comes at the cost of less precision, because it lets error
- accumulate across iterations.
- */
- x0 = 1.0f;
- y0 = 0.5f * n;
- r0 = u.f;
-
- x1 = x0 * r0;
- y1 = y0 * r0 * r0;
- r1 = 1.5f - y1;
-
- x2 = x1 * r1;
- y2 = y1 * r1 * r1;
- r2 = 1.5f - y2;
-
-#if 1
- return x2 * r2; /* we can stop here, and be conformant -BP */
-#else
- x3 = x2 * r2;
- y3 = y2 * r2 * r2;
- r3 = 1.5f - y3;
-
- return x3 * r3;
-#endif
-#else
- return (float) (1.0 / sqrt(n));
-#endif
-}
#ifndef __GNUC__
/**
#endif
+/* Using C99 rounding functions for roundToEven() implementation is
+ * difficult, because round(), rint, and nearbyint() are affected by
+ * fesetenv(), which the application may have done for its own
+ * purposes. Mesa's IROUND macro is close to what we want, but it
+ * rounds away from 0 on n + 0.5.
+ */
+int
+_mesa_round_to_even(float val)
+{
+ int rounded = IROUND(val);
+
+ if (val - floor(val) == 0.5) {
+ if (rounded % 2 != 0)
+ rounded += val > 0 ? -1 : 1;
+ }
+
+ return rounded;
+}
+
+
/**
* Convert a 4-byte float to a 2-byte half float.
- * Based on code from:
- * http://www.opengl.org/discussion_boards/ubb/Forum3/HTML/008786.html
+ *
+ * Not all float32 values can be represented exactly as a float16 value. We
+ * round such intermediate float32 values to the nearest float16. When the
+ * float32 lies exactly between to float16 values, we round to the one with
+ * an even mantissa.
+ *
+ * This rounding behavior has several benefits:
+ * - It has no sign bias.
+ *
+ * - It reproduces the behavior of real hardware: opcode F32TO16 in Intel's
+ * GPU ISA.
+ *
+ * - By reproducing the behavior of the GPU (at least on Intel hardware),
+ * compile-time evaluation of constant packHalf2x16 GLSL expressions will
+ * result in the same value as if the expression were executed on the GPU.
*/
GLhalfARB
_mesa_float_to_half(float val)
else {
/* regular number */
const int new_exp = flt_e - 127;
- if (new_exp < -24) {
- /* this maps to 0 */
- /* m = 0; - already set */
+ if (new_exp < -14) {
+ /* The float32 lies in the range (0.0, min_normal16) and is rounded
+ * to a nearby float16 value. The result will be either zero, subnormal,
+ * or normal.
+ */
e = 0;
- }
- else if (new_exp < -14) {
- /* this maps to a denorm */
- unsigned int exp_val = (unsigned int) (-14 - new_exp); /* 2^-exp_val*/
- e = 0;
- switch (exp_val) {
- case 0:
- _mesa_warning(NULL,
- "float_to_half: logical error in denorm creation!\n");
- /* m = 0; - already set */
- break;
- case 1: m = 512 + (flt_m >> 14); break;
- case 2: m = 256 + (flt_m >> 15); break;
- case 3: m = 128 + (flt_m >> 16); break;
- case 4: m = 64 + (flt_m >> 17); break;
- case 5: m = 32 + (flt_m >> 18); break;
- case 6: m = 16 + (flt_m >> 19); break;
- case 7: m = 8 + (flt_m >> 20); break;
- case 8: m = 4 + (flt_m >> 21); break;
- case 9: m = 2 + (flt_m >> 22); break;
- case 10: m = 1; break;
- }
+ m = _mesa_round_to_even((1 << 24) * fabsf(fi.f));
}
else if (new_exp > 15) {
/* map this value to infinity */
e = 31;
}
else {
- /* regular */
+ /* The float32 lies in the range
+ * [min_normal16, max_normal16 + max_step16)
+ * and is rounded to a nearby float16 value. The result will be
+ * either normal or infinite.
+ */
e = new_exp + 15;
- m = flt_m >> 13;
+ m = _mesa_round_to_even(flt_m / (float) (1 << 13));
}
}
+ assert(0 <= m && m <= 1024);
+ if (m == 1024) {
+ /* The float32 was rounded upwards into the range of the next exponent,
+ * so bump the exponent. This correctly handles the case where f32
+ * should be rounded up to float16 infinity.
+ */
+ ++e;
+ m = 0;
+ }
+
result = (s << 15) | (e << 10) | m;
return result;
}
{
if (s) {
size_t l = strlen(s);
- char *s2 = (char *) malloc(l + 1);
+ char *s2 = malloc(l + 1);
if (s2)
strcpy(s2, s);
return s2;
_mesa_strtof( const char *s, char **end )
{
#if defined(_GNU_SOURCE) && !defined(__CYGWIN__) && !defined(__FreeBSD__) && \
- !defined(ANDROID) && !defined(__HAIKU__)
+ !defined(ANDROID) && !defined(__HAIKU__) && !defined(__UCLIBC__) && \
+ !defined(__NetBSD__)
static locale_t loc = NULL;
if (!loc) {
loc = newlocale(LC_CTYPE_MASK, "C", NULL);
projects / mesa.git / blobdiff