在 2019/12/16 17:49, Liu Hao 写道:
> 在 2019/12/16 下午4:43, Martin Storsjö 写道:
> The former. It is exactly why `fmaf()` uses three `+=` instead of adding
> all parts together in a single expression. The `+=` operator, being an
> assignment operator, forces rounding to `float` if `FLT_EVAL_METHOD` is
> zero. Probably I should have added `#pragma STDC FP_CONTRACT OFF`. I
> will try that later today.
> 

'if `FLT_EVAL_METHOD` is zero' should be 'if `FLT_EVAL_METHOD` is _not_
zero'

Online compilers show that `#pragma STDC FP_CONTRACT OFF` only has
effects on GCC when compiling C [1], and has no effect on GCC when
compiling C++ [2], nor does it have any effect on Clang, no matter when
compiling C or C++ [1] [2].

So I don't recommend it, as it does not have consistent behavior, albeit
being a standard pragma.


[1] https://gcc.godbolt.org/z/I4Ff5T
[2] https://gcc.godbolt.org/z/yLF5wd

>> If the latter, should we build that part of mingw-w64-crt with specific
>> compiler flags, or perhaps with the implementation converted to assembly
>> (either inline or separate) to enforce the right form?
>>
>>> +static inline void break_down(iec559_double *restrict lo, iec559_double
>>> *restrict hi, double x) {
>>> +  hi->f = x;
>>> +  /* Erase low-order significant bits. `hi->f` now has only 32
>>> significant bits. */
>>
>> 32 bits? I can't get that number to add up properly, regardless if
>> including the sign/exponent or omitting it...
>>
> 
> Oops sorry it was copied from the `long double` variant. Perhaps all
> such comments are wrong. Will fix them later.
> 

I have attached a new patch that has those comments updated. In 'fmaf.c'
the lower 13 bits are erased, leaving 24 - 13 = 11 bits in the upper
half. In 'fma.c' the lower 27 bits are erased, leaving 53 - 27 = 26 bits
in the upper half.



> Assuming `double`:
> 
>      |---- Z -----|                          ; 53 bits
>   |-- X1 Y1 --|                              ; 52 bits
>               |-- X1 Y2 ---|                 ; 53 bits
>               |--- X2 Y1 --|                 ; 53 bits
>                            |--- X2 Y2 ---|   ; 54 bits <inexact>
>   ^MSB         ^ULP
> 
> I coundn't proof that there will be no rounding errors in this case. The
> only thing I know for sure at the moment is that if no number is
> denormal then the MSB of both `X1` and `Y1` are ones.
> 
> 

This chart was wrong. It should look like follows, where each character
represents a bit:

       |---------Z---------|                    ; 53 bits
    |------X1--Y1------|                        ; 52 bits
              |-------X2--Y1------|             ; 53 bits
              |------X1--Y2-------|             ; 53 bits
                        |-------X2--Y2-------|  ; 54 bits <inexact>
    ^                  ^
    MSB                ULP
  (Bit 52)           (Bit 0)

Summing the result from top to bottom might cause a roundoff error at ULP.


-- 
Best regards,
LH_Mouse
From 7c01abfcc1e3f86280e0dd16cfe71e878ec079f7 Mon Sep 17 00:00:00 2001
From: Liu Hao <[email protected]>
Date: Thu, 12 Dec 2019 21:04:04 +0800
Subject: [PATCH] crt/fma{,f}.c: Implement FMA for `double` and `float`
 properly

FMA of `double` shall not be implemented using `long double`, as
the result has 106 bits and cannot be stored accurately in a
`long double`, whose mantissa only has 64 bits.

Let's calculate `FMA(0x1.0000000000001p0, 1, 0x0.00000000000007FFp0)`:
First, we multiply `0x1.0000000000001p0` and `1`, which yields
`0x1.0000000000001p0`. Then we add `0x0.00000000000007FFp0` to it:

      0x1.0000 0000 0000 1     p0
  +)  0x0.0000 0000 0000 07FF  p0
  ---------------------------------
      0x1.0000 0000 0000 17FF  p0   (long double)

This result has 65 significant bits. When it is stored into a
`long double`, the last bit is rounded to even, as follows:

      0x1.0000 0000 0000 17FF  p0
                            1       <= rounded UP as this bit is set
  ---------------------------------
      0x1.0000 0000 0000 1800  p0   (long double)

If we attempt to round this value into `double` again, we get:

      0x1.0000 0000 0000 1800  p0
                          8         <= rounded UP as this bit is set
  ---------------------------------
      0x1.0000 0000 0000 2000  p0   (double)

This is wrong. Because FMA shall round the result exactly once. If we
round the first result to double we get a different result:

      0x0.0000 0000 0000 17FF  p0
                          8         <= rounded DOWN as this bit is clear
  ---------------------------------
      0x1.0000 0000 0000 1000  p0   (double)

`float` suffers from a similar problem.

This patch fixes such issues.

Testcase for `double`:

  #include <assert.h>
  #include <math.h>

  int main(void)
    {
      volatile double a = 0x1.0000000000001p0;
      volatile double b = 1;
      volatile double c = 0x0.00000000000007FFp0;

      assert(a * b + c    == 0x1.0000000000001p0);
      assert(fma(a, b, c) == 0x1.0000000000001p0);
      assert((double)((long double)a * b + c)
                          == 0x1.0000000000002p0);
    }

Testcase for `float`:

  #include <assert.h>
  #include <math.h>

  int main(void)
    {
      volatile float a = 0x0.800001p0f;
      volatile float b = 0x0.5FFFFFp0f;;
      volatile float c = 0x0.000000000000FFFFFFp0f;

      assert(a * b + c     == 0x0.2FFFFFCp0);
      assert(fmaf(a, b, c) == 0x0.2FFFFFCp0);
      assert((float)((double)a * b + c)
                           == 0x0.3000000p0);
    }

Reference: https://www.exploringbinary.com/double-rounding-errors-in-floating
Signed-off-by: Liu Hao <[email protected]>
---
 mingw-w64-crt/math/fma.c  | 66 ++++++++++++++++++++++++++++++++++++---
 mingw-w64-crt/math/fmaf.c | 66 ++++++++++++++++++++++++++++++++++++---
 2 files changed, 122 insertions(+), 10 deletions(-)

diff --git a/mingw-w64-crt/math/fma.c b/mingw-w64-crt/math/fma.c
index c4ce738b..6cd6a31f 100644
--- a/mingw-w64-crt/math/fma.c
+++ b/mingw-w64-crt/math/fma.c
@@ -29,13 +29,69 @@ double fma(double x, double y, double z){
   return z;
 }
 
-#else
+#elif defined(_AMD64_) || defined(__x86_64__) || defined(_X86_) || 
defined(__i386__)
 
-long double fmal(long double x, long double y, long double z);
+#include <math.h>
+#include <stdint.h>
 
-/* For platforms that don't have hardware FMA, emulate it. */
-double fma(double x, double y, double z){
-  return (double)fmal(x, y, z);
+/* This is in accordance with the IEC 559 double-precision format.
+ * Be advised that due to the hidden bit, the higher half actually has 26 bits.
+ * Multiplying two 27-bit numbers will cause a 1-ULP error, which we cannot
+ * avoid. It is kept in the very last position.
+ */
+typedef union iec559_double_ {
+  struct __attribute__((__packed__)) {
+    uint64_t mlo : 27;
+    uint64_t mhi : 25;
+    uint64_t exp : 11;
+    uint64_t sgn :  1;
+  };
+  double f;
+} iec559_double;
+
+static inline void break_down(iec559_double *restrict lo, iec559_double 
*restrict hi, double x) {
+  hi->f = x;
+  /* Erase low-order significant bits. `hi->f` now has only 26 significant 
bits. */
+  hi->mlo = 0;
+  /* Store the low-order half. It will be normalized by the hardware. */
+  lo->f = x - hi->f;
+  /* Preserve signness in case of zero. */
+  lo->sgn = hi->sgn;
 }
 
+double fma(double x, double y, double z) {
+  /*
+    POSIX-2013:
+    1. If x or y are NaN, a NaN shall be returned.
+    2. If x multiplied by y is an exact infinity and z is also an infinity
+       but with the opposite sign, a domain error shall occur, and either a NaN
+       (if supported), or an implementation-defined value shall be returned.
+    3. If one of x and y is infinite, the other is zero, and z is not a NaN,
+       a domain error shall occur, and either a NaN (if supported), or an
+       implementation-defined value shall be returned.
+    4. If one of x and y is infinite, the other is zero, and z is a NaN, a NaN
+       shall be returned and a domain error may occur.
+    5. If x* y is not 0*Inf nor Inf*0 and z is a NaN, a NaN shall be returned.
+  */
+  /* Check whether the result is finite. */
+  double ret = x * y + z;
+  if(!isfinite(ret)) {
+    return ret; /* If this naive check doesn't yield a finite value, the FMA 
isn't
+                   likely to return one either. Forward the value as is. */
+  }
+  iec559_double xlo, xhi, ylo, yhi;
+  break_down(&xlo, &xhi, x);
+  break_down(&ylo, &yhi, y);
+  /* The order of these four statements is essential. Don't move them around. 
*/
+  ret = z;
+  ret += xhi.f * yhi.f;                 /* The most significant item comes 
first. */
+  ret += xhi.f * ylo.f + xlo.f * yhi.f; /* They are equally significant. */
+  ret += xlo.f * ylo.f;                 /* The least significant item comes 
last. */
+  return ret;
+}
+
+#else
+
+#error Please add FMA implementation for this platform.
+
 #endif
diff --git a/mingw-w64-crt/math/fmaf.c b/mingw-w64-crt/math/fmaf.c
index b3f58a84..3e00738f 100644
--- a/mingw-w64-crt/math/fmaf.c
+++ b/mingw-w64-crt/math/fmaf.c
@@ -29,13 +29,69 @@ float fmaf(float x, float y, float z){
   return z;
 }
 
-#else
+#elif defined(_AMD64_) || defined(__x86_64__) || defined(_X86_) || 
defined(__i386__)
 
-long double fmal(long double x, long double y, long double z);
+#include <math.h>
+#include <stdint.h>
 
-/* For platforms that don't have hardware FMA, emulate it. */
-float fmaf(float x, float y, float z){
-  return (float)fmal(x, y, z);
+/* This is in accordance with the IEC 559 single-precision format.
+ * Be advised that due to the hidden bit, the higher half actually has 11 bits.
+ * Multiplying two 13-bit numbers will cause a 1-ULP error, which we cannot
+ * avoid. It is kept in the very last position.
+ */
+typedef union iec559_float_ {
+  struct __attribute__((__packed__)) {
+    uint32_t mlo : 13;
+    uint32_t mhi : 10;
+    uint32_t exp :  8;
+    uint32_t sgn :  1;
+  };
+  float f;
+} iec559_float;
+
+static inline void break_down(iec559_float *restrict lo, iec559_float 
*restrict hi, float x) {
+  hi->f = x;
+  /* Erase low-order significant bits. `hi->f` now has only 11 significant 
bits. */
+  hi->mlo = 0;
+  /* Store the low-order half. It will be normalized by the hardware. */
+  lo->f = x - hi->f;
+  /* Preserve signness in case of zero. */
+  lo->sgn = hi->sgn;
 }
 
+float fmaf(float x, float y, float z) {
+  /*
+    POSIX-2013:
+    1. If x or y are NaN, a NaN shall be returned.
+    2. If x multiplied by y is an exact infinity and z is also an infinity
+       but with the opposite sign, a domain error shall occur, and either a NaN
+       (if supported), or an implementation-defined value shall be returned.
+    3. If one of x and y is infinite, the other is zero, and z is not a NaN,
+       a domain error shall occur, and either a NaN (if supported), or an
+       implementation-defined value shall be returned.
+    4. If one of x and y is infinite, the other is zero, and z is a NaN, a NaN
+       shall be returned and a domain error may occur.
+    5. If x* y is not 0*Inf nor Inf*0 and z is a NaN, a NaN shall be returned.
+  */
+  /* Check whether the result is finite. */
+  float ret = x * y + z;
+  if(!isfinite(ret)) {
+    return ret; /* If this naive check doesn't yield a finite value, the FMA 
isn't
+                   likely to return one either. Forward the value as is. */
+  }
+  iec559_float xlo, xhi, ylo, yhi;
+  break_down(&xlo, &xhi, x);
+  break_down(&ylo, &yhi, y);
+  /* The order of these four statements is essential. Don't move them around. 
*/
+  ret = z;
+  ret += xhi.f * yhi.f;                 /* The most significant item comes 
first. */
+  ret += xhi.f * ylo.f + xlo.f * yhi.f; /* They are equally significant. */
+  ret += xlo.f * ylo.f;                 /* The least significant item comes 
last. */
+  return ret;
+}
+
+#else
+
+#error Please add FMA implementation for this platform.
+
 #endif
-- 
2.24.1

Attachment: signature.asc
Description: OpenPGP digital signature

_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public

Reply via email to