util: Reimplement half <-> float conversions.

Removed u_half.py used to generate the table for previous method.

Previous implementation of float to half conversion was faulty for
denormalised and NaNs and would require extra logic to fix,
thus making the speedup of using tables irrelevant.

Reviewed-by: Jose Fonseca <[email protected]>

3 files changed, 67 insertions, 214 deletions

diff --git a/src/gallium/auxiliary/util/.gitignore b/src/gallium/auxiliary/util/.gitignore
index 5dd0408effb..da74de623d7 100644
--- a/src/gallium/auxiliary/util/.gitignore
+++ b/src/gallium/auxiliary/util/.gitignore
@@ -1,3 +1,2 @@
 u_format_srgb.c
 u_format_table.c
-u_half.c
diff --git a/src/gallium/auxiliary/util/u_half.h b/src/gallium/auxiliary/util/u_half.h
index ad030e90c67..f7009f54844 100644
--- a/src/gallium/auxiliary/util/u_half.h
+++ b/src/gallium/auxiliary/util/u_half.h
@@ -35,51 +35,84 @@
 extern "C" {
 #endif
 
-extern const uint32_t util_half_to_float_mantissa_table[2048];
-extern const uint32_t util_half_to_float_exponent_table[64];
-extern const uint32_t util_half_to_float_offset_table[64];
-extern const uint16_t util_float_to_half_base_table[512];
-extern const uint8_t util_float_to_half_shift_table[512];
-
 /*
- * Note that if the half float is a signaling NaN, the x87 FPU will turn
- * it into a quiet NaN immediately upon loading into a float.
- *
- * Additionally, denormals may be flushed to zero.
+ * References for float <-> half conversions
  *
- * To avoid this, use the floatui functions instead of the float ones
- * when just doing conversion rather than computation on the resulting
- * floats.
+ *  http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
+ *  https://gist.github.com/2156668
+ *  https://gist.github.com/2144712
  */
 
-static INLINE uint32_t
-util_half_to_floatui(uint16_t h)
+static INLINE uint16_t
+util_float_to_half(float f)
 {
-   unsigned exp = h >> 10;
-   return util_half_to_float_mantissa_table[util_half_to_float_offset_table[exp] + (h & 0x3ff)] + util_half_to_float_exponent_table[exp];
+   uint32_t sign_mask  = 0x80000000;
+   uint32_t round_mask = ~0xfff;
+   uint32_t f32inf = 0xff << 23;
+   uint32_t f16inf = 0x1f << 23;
+   uint32_t sign;
+   union fi magic;
+   union fi f32;
+   uint16_t f16;
+
+   magic.ui = 0xf << 23;
+
+   f32.f = f;
+
+   /* Sign */
+   sign = f32.ui & sign_mask;
+   f32.ui ^= sign;
+
+   if (f32.ui == f32inf) {
+      /* Inf */
+      f16 = 0x7c00;
+   } else if (f32.ui > f32inf) {
+      /* NaN */
+      f16 = 0x7e00;
+   } else {
+      /* Number */
+      f32.ui &= round_mask;
+      f32.f  *= magic.f;
+      f32.ui -= round_mask;
+
+      /* Clamp to infinity if overflowed */
+      if (f32.ui > f16inf)
+         f32.ui = f16inf;
+
+      f16 = f32.ui >> 13;
+   }
+
+   /* Sign */
+   f16 |= sign >> 16;
+
+   return f16;
 }
 
 static INLINE float
-util_half_to_float(uint16_t h)
+util_half_to_float(uint16_t f16)
 {
-   union fi r;
-   r.ui = util_half_to_floatui(h);
-   return r.f;
-}
+   union fi infnan;
+   union fi magic;
+   union fi f32;
 
-static INLINE uint16_t
-util_floatui_to_half(uint32_t v)
-{
-   unsigned signexp = v >> 23;
-   return util_float_to_half_base_table[signexp] + ((v & 0x007fffff) >> util_float_to_half_shift_table[signexp]);
-}
+   infnan.ui = 0x8f << 23;
+   infnan.f = 65536.0f;
+   magic.ui  = 0xef << 23;
 
-static INLINE uint16_t
-util_float_to_half(float f)
-{
-   union fi i;
-   i.f = f;
-   return util_floatui_to_half(i.ui);
+   /* Exponent / Mantissa */
+   f32.ui = (f16 & 0x7fff) << 13;
+
+   /* Adjust */
+   f32.f *= magic.f;
+
+   /* Inf / NaN */
+   if (f32.f >= infnan.f)
+      f32.ui |= 0xff << 23;
+
+   /* Sign */
+   f32.ui |= (f16 & 0x8000) << 16;
+
+   return f32.f;
 }
 
 #ifdef __cplusplus
diff --git a/src/gallium/auxiliary/util/u_half.py b/src/gallium/auxiliary/util/u_half.py
deleted file mode 100644
index 915cf3b9273..00000000000
--- a/src/gallium/auxiliary/util/u_half.py
+++ /dev/null
@@ -1,179 +0,0 @@
-# Copyright 2010 Luca Barbieri
-#
-# Permission is hereby granted, free of charge, to any person obtaining
-# a copy of this software and associated documentation files (the
-# "Software"), to deal in the Software without restriction, including
-# without limitation the rights to use, copy, modify, merge, publish,
-# distribute, sublicense, and/or sell copies of the Software, and to
-# permit persons to whom the Software is furnished to do so, subject to
-# the following conditions:
-#
-# The above copyright notice and this permission notice (including the
-# next paragraph) shall be included in all copies or substantial
-# portions of the Software.
-#
-# 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 THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS 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 code is a reimplementation of the algorithm in
-#  www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
-# "Fast Half Float Conversions" by Jeroen van der Zijp, Nov 2008
-#
-# The table contents have been slightly changed so that the exponent
-# bias is now in the exponent table instead of the mantissa table (mostly
-# for cosmetic reasons, and because it theoretically allows a variant
-# that flushes denormal to zero but uses a mantissa table with 24-bit
-# entries).
-#
-# The tables are also constructed slightly differently.
-#
-
-# Note that using a 64K * 4 table is a terrible idea since it will not fit
-# in the L1 cache and will massively pollute the L2 cache as well
-#
-# These should instead fit in the L1 cache.
-#
-# TODO: we could use a denormal bias table instead of the mantissa/offset
-# tables: this would reduce the L1 cache usage from 8704 to 2304 bytes
-# but would involve more computation
-#
-# Note however that if denormals are never encountered, the L1 cache usage
-# is only about 4608 bytes anyway.
-
-table_index = None
-table_length = None
-
-def begin(t, n, l):
-	global table_length
-	global table_index
-	table_index = 0
-	table_length = l
-	print
-	print "const " + t + " " + n + "[" + str(l) + "] = {"
-
-def value(v):
-	global table_index
-	table_index += 1
-	print "\t" + hex(v) + ","
-
-def end():
-	global table_length
-	global table_index
-	print "};"
-	assert table_index == table_length
-
-print "/* This file is autogenerated by u_half.py. Do not edit directly. */"
-print "#include \"util/u_half.h\""
-
-begin("uint32_t", "util_half_to_float_mantissa_table", 2048)
-# zero
-value(0)
-
-# denormals
-for i in xrange(1, 1024):
-	m = i << 13
-	e = 0
-
-	# normalize number
-	while (m & 0x00800000) == 0:
-		e -= 0x00800000
-		m <<= 1
-
-	m &= ~0x00800000
-	e += 0x38800000
-	value(m | e)
-
-# normals
-for i in xrange(1024, 2048):
-	value((i - 1024) << 13)
-end()
-
-begin("uint32_t", "util_half_to_float_exponent_table", 64)
-# positive zero or denormals
-value(0)
-
-# positive numbers
-for i in xrange(1, 31):
-	value(0x38000000 + (i << 23))
-
-# positive infinity/NaN
-value(0x7f800000)
-
-# negative zero or denormals
-value(0x80000000)
-
-# negative numbers
-for i in range(33, 63):
-	value(0xb8000000 + ((i - 32) << 23))
-
-# negative infinity/NaN
-value(0xff800000)
-end()
-
-begin("uint32_t", "util_half_to_float_offset_table", 64)
-# positive zero or denormals
-value(0)
-
-# positive normals
-for i in range(1, 32):
-	value(1024)
-
-# negative zero or denormals
-value(0)
-
-# negative normals
-for i in xrange(33, 64):
-	value(1024)
-end()
-
-begin("uint16_t", "util_float_to_half_base_table", 512)
-for sign in (0, 0x8000):
-	# very small numbers mapping to zero
-	for i in xrange(-127, -24):
-		value(sign | 0)
-
-	# small numbers mapping to denormals
-	for i in xrange(-24, -14):
-		value(sign | (0x400 >> (-14 -i)))
-
-	# normal numbers
-	for i in xrange(-14, 16):
-		value(sign | ((i + 15) << 10))
-
-	# large numbers mapping to infinity
-	for i in xrange(16, 128):
-		value(sign | 0x7c00)
-
-	# infinity and NaNs
-	value(sign | 0x7c00)
-end()
-
-begin("uint8_t", "util_float_to_half_shift_table", 512)
-for sign in (0, 0x8000):
-	# very small numbers mapping to zero
-	for i in xrange(-127, -24):
-		value(24)
-
-	# small numbers mapping to denormals
-	for i in xrange(-24, -14):
-		value(-1 - i)
-
-	# normal numbers
-	for i in xrange(-14, 16):
-		value(13)
-
-	# large numbers mapping to infinity
-	for i in xrange(16, 128):
-		value(24)
-
-	# infinity and NaNs
-	value(13)
-end()
-