[Mesa-dev] [RFC PATCH] mesa: Replace _mesa_round_to_even() with roundeven().
Matt Turner
mattst88 at gmail.com
Wed Mar 11 14:52:18 PDT 2015
Eric's initial patch adding constant expression evaluation for
ir_unop_round_even used nearbyint. The open-coded _mesa_round_to_even
implementation came about without much explanation after a reviewer
asked whether nearbyint depended on the application not modifying the
rounding mode. Of course (as Eric commented) we rely on the application
not changing the rounding mode from its default (round-to-nearest) in
many other places, including the IROUND function used by
_mesa_round_to_even!
Worse, IROUND() is implemented using the trunc(x + 0.5) trick which
fails for x = nextafterf(0.5, 0.0).
Still worse, _mesa_round_to_even unexpectedly returns an int. I suspect
that could cause problems when rounding large integral values not
representable as an int in ir_constant_expression.cpp's ir_unop_round_even
evaluation. Its use of _mesa_round_to_even is clearly broken for doubles
(as noted during review).
The constant expression evaluation code for the packing built-in
functions also mistakenly assumed that _mesa_round_to_even returned a
float, as can be seen by the cast through a signed integer type to an
unsigned (since negative float -> unsigned conversions are undefined).
rint() and nearbyint() implement the round-half-to-even behavior we want
when the rounding mode is set to the default round-to-nearest. The only
difference between them is that nearbyint() raises the inexact
exception.
This patch implements roundeven{f,}, a function added by a yet
unimplemented technical specification (ISO/IEC TS 18661-1:2014), with a
small difference in behavior -- we don't bother raising the inexact
exception, which I don't think we care about anyway.
At least recent Intel CPUs can quickly change a subset of the bits in
the x87 floating-point control register, but the exception mask bits are
not included. rint() does not need to change these bits, but nearbyint()
does (twice: save old, set new, and restore) in order to raise the inexact
exception, which would incur some penalty.
The SSE 4.1 ROUND instructions let us implement roundeven directly.
Otherwise we assume that the rounding mode has not been modified (as we
do in the rest of Mesa) and use rint().
---
v1 of Eric's previously mentioned patch is
http://lists.freedesktop.org/archives/mesa-dev/2011-October/012878.html
Comments on v1 start with
http://lists.freedesktop.org/archives/mesa-dev/2011-October/012900.html
v2 is
http://lists.freedesktop.org/archives/mesa-dev/2011-October/013400.html
As far as I can tell, v2 came about because Eric didn't know that the
default rounding mode was what we wanted for the libc functions to work,
and that we didn't need to change it.
If we do indeed want the don't-raise-the-inexact-exception behavior,
maybe we should just keep the _mesa_round_to_even name?
src/glsl/ir_constant_expression.cpp | 18 ++++----
src/glsl/nir/nir_constant_expressions.py | 15 ++++---
src/glsl/nir/nir_opcodes.py | 2 +-
src/mesa/main/imports.c | 25 ++---------
src/mesa/main/imports.h | 3 --
src/util/Makefile.sources | 1 +
src/util/rounding.h | 73 ++++++++++++++++++++++++++++++++
7 files changed, 97 insertions(+), 40 deletions(-)
create mode 100644 src/util/rounding.h
diff --git a/src/glsl/ir_constant_expression.cpp b/src/glsl/ir_constant_expression.cpp
index 388c4c2..b5938bf 100644
--- a/src/glsl/ir_constant_expression.cpp
+++ b/src/glsl/ir_constant_expression.cpp
@@ -35,6 +35,7 @@
#include <math.h>
#include "main/core.h" /* for MAX2, MIN2, CLAMP */
+#include "util/rounding.h" /* for roundeven */
#include "ir.h"
#include "glsl_types.h"
#include "program/hash_table.h"
@@ -245,8 +246,8 @@ pack_snorm_1x8(float x)
* We must first cast the float to an int, because casting a negative
* float to a uint is undefined.
*/
- return (uint8_t) (int8_t)
- _mesa_round_to_even(CLAMP(x, -1.0f, +1.0f) * 127.0f);
+ return (uint8_t) (int)
+ roundevenf(CLAMP(x, -1.0f, +1.0f) * 127.0f);
}
/**
@@ -267,8 +268,8 @@ pack_snorm_1x16(float x)
* We must first cast the float to an int, because casting a negative
* float to a uint is undefined.
*/
- return (uint16_t) (int16_t)
- _mesa_round_to_even(CLAMP(x, -1.0f, +1.0f) * 32767.0f);
+ return (uint16_t) (int)
+ roundevenf(CLAMP(x, -1.0f, +1.0f) * 32767.0f);
}
/**
@@ -322,7 +323,7 @@ pack_unorm_1x8(float x)
*
* packUnorm4x8: round(clamp(c, 0, +1) * 255.0)
*/
- return (uint8_t) _mesa_round_to_even(CLAMP(x, 0.0f, 1.0f) * 255.0f);
+ return (uint8_t) (int) roundevenf(CLAMP(x, 0.0f, 1.0f) * 255.0f);
}
/**
@@ -340,7 +341,8 @@ pack_unorm_1x16(float x)
*
* packUnorm2x16: round(clamp(c, 0, +1) * 65535.0)
*/
- return (uint16_t) _mesa_round_to_even(CLAMP(x, 0.0f, 1.0f) * 65535.0f);
+ return (uint16_t) (int)
+ roundevenf(CLAMP(x, 0.0f, 1.0f) * 65535.0f);
}
/**
@@ -733,9 +735,9 @@ ir_expression::constant_expression_value(struct hash_table *variable_context)
case ir_unop_round_even:
for (unsigned c = 0; c < op[0]->type->components(); c++) {
if (op[0]->type->base_type == GLSL_TYPE_DOUBLE)
- data.d[c] = _mesa_round_to_even(op[0]->value.d[c]);
+ data.d[c] = roundeven(op[0]->value.d[c]);
else
- data.f[c] = _mesa_round_to_even(op[0]->value.f[c]);
+ data.f[c] = roundevenf(op[0]->value.f[c]);
}
break;
diff --git a/src/glsl/nir/nir_constant_expressions.py b/src/glsl/nir/nir_constant_expressions.py
index 22bc4f0..cceb422 100644
--- a/src/glsl/nir/nir_constant_expressions.py
+++ b/src/glsl/nir/nir_constant_expressions.py
@@ -28,6 +28,7 @@ template = """\
#include <math.h>
#include "main/core.h"
+#include "util/rounding.h" /* for roundeven */
#include "nir_constant_expressions.h"
#if defined(_MSC_VER) && (_MSC_VER < 1800)
@@ -68,8 +69,8 @@ pack_snorm_1x8(float x)
* We must first cast the float to an int, because casting a negative
* float to a uint is undefined.
*/
- return (uint8_t) (int8_t)
- _mesa_round_to_even(CLAMP(x, -1.0f, +1.0f) * 127.0f);
+ return (uint8_t) (int)
+ roundevenf(CLAMP(x, -1.0f, +1.0f) * 127.0f);
}
/**
@@ -90,8 +91,8 @@ pack_snorm_1x16(float x)
* We must first cast the float to an int, because casting a negative
* float to a uint is undefined.
*/
- return (uint16_t) (int16_t)
- _mesa_round_to_even(CLAMP(x, -1.0f, +1.0f) * 32767.0f);
+ return (uint16_t) (int)
+ roundevenf(CLAMP(x, -1.0f, +1.0f) * 32767.0f);
}
/**
@@ -145,7 +146,8 @@ pack_unorm_1x8(float x)
*
* packUnorm4x8: round(clamp(c, 0, +1) * 255.0)
*/
- return (uint8_t) _mesa_round_to_even(CLAMP(x, 0.0f, 1.0f) * 255.0f);
+ return (uint8_t) (int)
+ roundevenf(CLAMP(x, 0.0f, 1.0f) * 255.0f);
}
/**
@@ -163,7 +165,8 @@ pack_unorm_1x16(float x)
*
* packUnorm2x16: round(clamp(c, 0, +1) * 65535.0)
*/
- return (uint16_t) _mesa_round_to_even(CLAMP(x, 0.0f, 1.0f) * 65535.0f);
+ return (uint16_t) (int)
+ roundevenf(CLAMP(x, 0.0f, 1.0f) * 65535.0f);
}
/**
diff --git a/src/glsl/nir/nir_opcodes.py b/src/glsl/nir/nir_opcodes.py
index 77f3bb8..ad03eef 100644
--- a/src/glsl/nir/nir_opcodes.py
+++ b/src/glsl/nir/nir_opcodes.py
@@ -183,7 +183,7 @@ unop("ftrunc", tfloat, "truncf(src0)")
unop("fceil", tfloat, "ceilf(src0)")
unop("ffloor", tfloat, "floorf(src0)")
unop("ffract", tfloat, "src0 - floorf(src0)")
-unop("fround_even", tfloat, "_mesa_round_to_even(src0)")
+unop("fround_even", tfloat, "roundevenf(src0)")
# Trigonometric operations.
diff --git a/src/mesa/main/imports.c b/src/mesa/main/imports.c
index ac8deeb..8da5372 100644
--- a/src/mesa/main/imports.c
+++ b/src/mesa/main/imports.c
@@ -45,6 +45,7 @@
#include <stdio.h>
#include <stdarg.h>
#include "c99_math.h"
+#include "util/rounding.h" /* for roundeven */
#include "imports.h"
#include "context.h"
#include "mtypes.h"
@@ -307,26 +308,6 @@ _mesa_bitcount_64(uint64_t n)
#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.
*
@@ -388,7 +369,7 @@ _mesa_float_to_half(float val)
* or normal.
*/
e = 0;
- m = _mesa_round_to_even((1 << 24) * fabsf(fi.f));
+ m = (int) roundevenf((1 << 24) * fabsf(fi.f));
}
else if (new_exp > 15) {
/* map this value to infinity */
@@ -402,7 +383,7 @@ _mesa_float_to_half(float val)
* either normal or infinite.
*/
e = new_exp + 15;
- m = _mesa_round_to_even(flt_m / (float) (1 << 13));
+ m = (int) roundevenf(flt_m / (float) (1 << 13));
}
}
diff --git a/src/mesa/main/imports.h b/src/mesa/main/imports.h
index ee6b399..29f2499 100644
--- a/src/mesa/main/imports.h
+++ b/src/mesa/main/imports.h
@@ -433,9 +433,6 @@ _mesa_fls(unsigned int n)
#endif
}
-extern int
-_mesa_round_to_even(float val);
-
extern GLhalfARB
_mesa_float_to_half(float f);
diff --git a/src/util/Makefile.sources b/src/util/Makefile.sources
index 560ea83..3e0d02b 100644
--- a/src/util/Makefile.sources
+++ b/src/util/Makefile.sources
@@ -14,6 +14,7 @@ MESA_UTIL_FILES := \
register_allocate.h \
rgtc.c \
rgtc.h \
+ rounding.h \
set.c \
set.h \
simple_list.h \
diff --git a/src/util/rounding.h b/src/util/rounding.h
new file mode 100644
index 0000000..328909b
--- /dev/null
+++ b/src/util/rounding.h
@@ -0,0 +1,73 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * 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 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.
+ */
+
+#include <math.h>
+
+#ifdef __SSE4_1__
+#include <smmintrin.h>
+#endif
+
+/* The C standard library has functions round()/rint()/nearbyint() that round
+ * their arguments according to the rounding mode set in the floating-point
+ * control register. While there are trunc()/ceil()/floor() functions that do
+ * a specific operation without modifying the rounding mode, there is no
+ * roundeven() in any version of C.
+ *
+ * Technical Specification 18661 (ISO/IEC TS 18661-1:2014) adds roundeven(),
+ * but it's unfortunately not implemented by glibc.
+ *
+ * This implementation differs in that it does not raise the inexact exception.
+ */
+static inline float
+roundevenf(float x)
+{
+ float ret;
+#ifdef __SSE4_1__
+ __m128 m = _mm_load_ss(&x);
+ m = _mm_round_ss(m, m, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
+ _mm_store_ss(&ret, m);
+#else
+ /* Assume that the floating-point rounding mode has not been changed from
+ * the default (Round to nearest).
+ */
+ ret = rintf(x);
+#endif
+ return ret;
+}
+
+static inline double
+roundeven(double x)
+{
+ double ret;
+#ifdef __SSE4_1__
+ __m128d m = _mm_load_sd(&x);
+ m = _mm_round_sd(m, m, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
+ _mm_store_sd(&ret, m);
+#else
+ /* Assume that the floating-point rounding mode has not been changed from
+ * the default (Round to nearest).
+ */
+ ret = rint(x);
+#endif
+ return ret;
+}
--
2.0.5
More information about the mesa-dev
mailing list