[Libclc-dev] [PATCH 2/3] remquo: Port from amd builtins

Jan Vesely via Libclc-dev libclc-dev at lists.llvm.org
Tue May 1 15:01:26 PDT 2018


double version passes on carrizo. float version fails on denormals.
Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
---
denormal failure is fixed in the next patch

 generic/include/clc/clc.h           |   1 +
 generic/include/clc/math/remquo.h   |  18 +++
 generic/include/clc/math/remquo.inc |   1 +
 generic/include/math/clc_remquo.h   |   8 ++
 generic/lib/SOURCES                 |   2 +
 generic/lib/math/clc_remquo.cl      | 254 ++++++++++++++++++++++++++++++++++++
 generic/lib/math/remquo.cl          |  17 +++
 generic/lib/math/remquo.inc         |   6 +
 8 files changed, 307 insertions(+)
 create mode 100644 generic/include/clc/math/remquo.h
 create mode 100644 generic/include/clc/math/remquo.inc
 create mode 100644 generic/include/math/clc_remquo.h
 create mode 100644 generic/lib/math/clc_remquo.cl
 create mode 100644 generic/lib/math/remquo.cl
 create mode 100644 generic/lib/math/remquo.inc

diff --git a/generic/include/clc/clc.h b/generic/include/clc/clc.h
index 75ca7d0..171b06a 100644
--- a/generic/include/clc/clc.h
+++ b/generic/include/clc/clc.h
@@ -104,6 +104,7 @@
 #include <clc/math/pown.h>
 #include <clc/math/powr.h>
 #include <clc/math/remainder.h>
+#include <clc/math/remquo.h>
 #include <clc/math/rint.h>
 #include <clc/math/rootn.h>
 #include <clc/math/round.h>
diff --git a/generic/include/clc/math/remquo.h b/generic/include/clc/math/remquo.h
new file mode 100644
index 0000000..7daf82f
--- /dev/null
+++ b/generic/include/clc/math/remquo.h
@@ -0,0 +1,18 @@
+#define __CLC_FUNCTION remquo
+
+#define __CLC_BODY <clc/math/remquo.inc>
+#define __CLC_ADDRESS_SPACE global
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
+
+#define __CLC_BODY <clc/math/remquo.inc>
+#define __CLC_ADDRESS_SPACE local
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
+
+#define __CLC_BODY <clc/math/remquo.inc>
+#define __CLC_ADDRESS_SPACE private
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
+
+#undef __CLC_FUNCTION
diff --git a/generic/include/clc/math/remquo.inc b/generic/include/clc/math/remquo.inc
new file mode 100644
index 0000000..42c7b67
--- /dev/null
+++ b/generic/include/clc/math/remquo.inc
@@ -0,0 +1 @@
+_CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION(__CLC_GENTYPE x, __CLC_GENTYPE y, __CLC_ADDRESS_SPACE __CLC_INTN *q);
diff --git a/generic/include/math/clc_remquo.h b/generic/include/math/clc_remquo.h
new file mode 100644
index 0000000..ed57ec9
--- /dev/null
+++ b/generic/include/math/clc_remquo.h
@@ -0,0 +1,8 @@
+#define __CLC_FUNCTION __clc_remquo
+
+#define __CLC_BODY <clc/math/remquo.inc>
+#define __CLC_ADDRESS_SPACE private
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
+
+#undef __CLC_FUNCTION
diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
index 8598fa5..1f1a427 100644
--- a/generic/lib/SOURCES
+++ b/generic/lib/SOURCES
@@ -163,6 +163,8 @@ math/clc_powr.cl
 math/powr.cl
 math/clc_remainder.cl
 math/remainder.cl
+math/clc_remquo.cl
+math/remquo.cl
 math/clc_rootn.cl
 math/rootn.cl
 math/sin.cl
diff --git a/generic/lib/math/clc_remquo.cl b/generic/lib/math/clc_remquo.cl
new file mode 100644
index 0000000..ff741f6
--- /dev/null
+++ b/generic/lib/math/clc_remquo.cl
@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2014 Advanced Micro Devices, Inc.
+ *
+ * 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 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 <clc/clc.h>
+
+#include <math/clc_remainder.h>
+#include "../clcmacro.h"
+#include "config.h"
+#include "math.h"
+
+_CLC_DEF _CLC_OVERLOAD float __clc_remquo(float x, float y, __private int *quo)
+{
+    int ux = as_int(x);
+    int ax = ux & EXSIGNBIT_SP32;
+    float xa = as_float(ax);
+    int sx = ux ^ ax;
+    int ex = ax >> EXPSHIFTBITS_SP32;
+
+    int uy = as_int(y);
+    int ay = uy & EXSIGNBIT_SP32;
+    float ya = as_float(ay);
+    int sy = uy ^ ay;
+    int ey = ay >> EXPSHIFTBITS_SP32;
+
+    float xr = as_float(0x3f800000 | (ax & 0x007fffff));
+    float yr = as_float(0x3f800000 | (ay & 0x007fffff));
+    int c;
+    int k = ex - ey;
+
+    uint q = 0;
+
+    while (k > 0) {
+        c = xr >= yr;
+        q = (q << 1) | c;
+        xr -= c ? yr : 0.0f;
+        xr += xr;
+	--k;
+    }
+
+    c = xr > yr;
+    q = (q << 1) | c;
+    xr -= c ? yr : 0.0f;
+
+    int lt = ex < ey;
+
+    q = lt ? 0 : q;
+    xr = lt ? xa : xr;
+    yr = lt ? ya : yr;
+
+    c = (yr < 2.0f * xr) | ((yr == 2.0f * xr) & ((q & 0x1) == 0x1));
+    xr -= c ? yr : 0.0f;
+    q += c;
+
+    float s = as_float(ey << EXPSHIFTBITS_SP32);
+    xr *= lt ? 1.0f : s;
+
+    int qsgn = sx == sy ? 1 : -1;
+    int quot = (q & 0x7f) * qsgn;
+
+    c = ax == ay;
+    quot = c ? qsgn : quot;
+    xr = c ? 0.0f : xr;
+
+    xr = as_float(sx ^ as_int(xr));
+
+    c = ax > PINFBITPATT_SP32 | ay > PINFBITPATT_SP32 | ax == PINFBITPATT_SP32 | ay == 0;
+    quot = c ? 0 : quot;
+    xr = c ? as_float(QNANBITPATT_SP32) : xr;
+
+    *quo = quot;
+
+    return xr;
+}
+// remquo singature is special, we don't have macro for this
+#define __VEC_REMQUO(TYPE, VEC_SIZE, HALF_VEC_SIZE) \
+_CLC_DEF _CLC_OVERLOAD TYPE##VEC_SIZE __clc_remquo(TYPE##VEC_SIZE x, TYPE##VEC_SIZE y, __private int##VEC_SIZE *quo) \
+{ \
+	int##HALF_VEC_SIZE lo, hi; \
+	TYPE##VEC_SIZE ret; \
+	ret.lo = __clc_remquo(x.lo, y.lo, &lo); \
+	ret.hi = __clc_remquo(x.hi, y.hi, &hi); \
+	(*quo).lo = lo; \
+	(*quo).hi = hi; \
+	return ret; \
+}
+__VEC_REMQUO(float, 2,)
+__VEC_REMQUO(float, 3, 2)
+__VEC_REMQUO(float, 4, 2)
+__VEC_REMQUO(float, 8, 4)
+__VEC_REMQUO(float, 16, 8)
+
+#ifdef cl_khr_fp64
+_CLC_DEF _CLC_OVERLOAD double __clc_remquo(double x, double y, __private int *pquo)
+{
+    ulong ux = as_ulong(x);
+    ulong ax = ux & ~SIGNBIT_DP64;
+    ulong xsgn = ux ^ ax;
+    double dx = as_double(ax);
+    int xexp = convert_int(ax >> EXPSHIFTBITS_DP64);
+    int xexp1 = 11 - (int) clz(ax & MANTBITS_DP64);
+    xexp1 = xexp < 1 ? xexp1 : xexp;
+
+    ulong uy = as_ulong(y);
+    ulong ay = uy & ~SIGNBIT_DP64;
+    double dy = as_double(ay);
+    int yexp = convert_int(ay >> EXPSHIFTBITS_DP64);
+    int yexp1 = 11 - (int) clz(ay & MANTBITS_DP64);
+    yexp1 = yexp < 1 ? yexp1 : yexp;
+
+    int qsgn = ((ux ^ uy) & SIGNBIT_DP64) == 0UL ? 1 : -1;
+
+    // First assume |x| > |y|
+
+    // Set ntimes to the number of times we need to do a
+    // partial remainder. If the exponent of x is an exact multiple
+    // of 53 larger than the exponent of y, and the mantissa of x is
+    // less than the mantissa of y, ntimes will be one too large
+    // but it doesn't matter - it just means that we'll go round
+    // the loop below one extra time.
+    int ntimes = max(0, (xexp1 - yexp1) / 53);
+    double w =  ldexp(dy, ntimes * 53);
+    w = ntimes == 0 ? dy : w;
+    double scale = ntimes == 0 ? 1.0 : 0x1.0p-53;
+
+    // Each time round the loop we compute a partial remainder.
+    // This is done by subtracting a large multiple of w
+    // from x each time, where w is a scaled up version of y.
+    // The subtraction must be performed exactly in quad
+    // precision, though the result at each stage can
+    // fit exactly in a double precision number.
+    int i;
+    double t, v, p, pp;
+
+    for (i = 0; i < ntimes; i++) {
+        // Compute integral multiplier
+        t = trunc(dx / w);
+
+        // Compute w * t in quad precision
+        p = w * t;
+        pp = fma(w, t, -p);
+
+        // Subtract w * t from dx
+        v = dx - p;
+        dx = v + (((dx - v) - p) - pp);
+
+        // If t was one too large, dx will be negative. Add back one w.
+        dx += dx < 0.0 ? w : 0.0;
+
+        // Scale w down by 2^(-53) for the next iteration
+        w *= scale;
+    }
+
+    // One more time
+    // Variable todd says whether the integer t is odd or not
+    t = floor(dx / w);
+    long lt = (long)t;
+    int todd = lt & 1;
+
+    p = w * t;
+    pp = fma(w, t, -p);
+    v = dx - p;
+    dx = v + (((dx - v) - p) - pp);
+    i = dx < 0.0;
+    todd ^= i;
+    dx += i ? w : 0.0;
+
+    lt -= i;
+
+    // At this point, dx lies in the range [0,dy)
+
+    // For the remainder function, we need to adjust dx
+    // so that it lies in the range (-y/2, y/2] by carefully
+    // subtracting w (== dy == y) if necessary. The rigmarole
+    // with todd is to get the correct sign of the result
+    // when x/y lies exactly half way between two integers,
+    // when we need to choose the even integer.
+
+    int al = (2.0*dx > w) | (todd & (2.0*dx == w));
+    double dxl = dx - (al ? w : 0.0);
+
+    int ag = (dx > 0.5*w) | (todd & (dx == 0.5*w));
+    double dxg = dx - (ag ? w : 0.0);
+
+    dx = dy < 0x1.0p+1022 ? dxl : dxg;
+    lt += dy < 0x1.0p+1022 ? al : ag;
+    int quo = ((int)lt & 0x7f) * qsgn;
+
+    double ret = as_double(xsgn ^ as_ulong(dx));
+    dx = as_double(ax);
+
+    // Now handle |x| == |y|
+    int c = dx == dy;
+    t = as_double(xsgn);
+    quo = c ? qsgn : quo;
+    ret = c ? t : ret;
+
+    // Next, handle |x| < |y|
+    c = dx < dy;
+    quo = c ? 0 : quo;
+    ret = c ? x : ret;
+
+    c &= (yexp < 1023 & 2.0*dx > dy) | (dx > 0.5*dy);
+    quo = c ? qsgn : quo;
+    // we could use a conversion here instead since qsgn = +-1
+    p = qsgn == 1 ? -1.0 : 1.0;
+    t = fma(y, p, x);
+    ret = c ? t : ret;
+
+    // We don't need anything special for |x| == 0
+
+    // |y| is 0
+    c = dy == 0.0;
+    quo = c ? 0 : quo;
+    ret = c ? as_double(QNANBITPATT_DP64) : ret;
+
+    // y is +-Inf, NaN
+    c = yexp > BIASEDEMAX_DP64;
+    quo = c ? 0 : quo;
+    t = y == y ? x : y;
+    ret = c ? t : ret;
+
+    // x is +=Inf, NaN
+    c = xexp > BIASEDEMAX_DP64;
+    quo = c ? 0 : quo;
+    ret = c ? as_double(QNANBITPATT_DP64) : ret;
+
+    *pquo = quo;
+    return ret;
+}
+__VEC_REMQUO(double, 2,)
+__VEC_REMQUO(double, 3, 2)
+__VEC_REMQUO(double, 4, 2)
+__VEC_REMQUO(double, 8, 4)
+__VEC_REMQUO(double, 16, 8)
+#endif
diff --git a/generic/lib/math/remquo.cl b/generic/lib/math/remquo.cl
new file mode 100644
index 0000000..fc29b36
--- /dev/null
+++ b/generic/lib/math/remquo.cl
@@ -0,0 +1,17 @@
+#include <clc/clc.h>
+#include <math/clc_remquo.h>
+
+#define __CLC_BODY <remquo.inc>
+#define __CLC_ADDRESS_SPACE global
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
+
+#define __CLC_BODY <remquo.inc>
+#define __CLC_ADDRESS_SPACE local
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
+
+#define __CLC_BODY <remquo.inc>
+#define __CLC_ADDRESS_SPACE private
+#include <clc/math/gentype.inc>
+#undef __CLC_ADDRESS_SPACE
diff --git a/generic/lib/math/remquo.inc b/generic/lib/math/remquo.inc
new file mode 100644
index 0000000..c1de78a
--- /dev/null
+++ b/generic/lib/math/remquo.inc
@@ -0,0 +1,6 @@
+_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE remquo(__CLC_GENTYPE x, __CLC_GENTYPE y, __CLC_ADDRESS_SPACE __CLC_INTN *q) {
+  __CLC_INTN local_q;
+  __CLC_GENTYPE ret = __clc_remquo(x, y, &local_q);
+  *q = local_q;
+  return ret;
+}
-- 
2.14.3



More information about the Libclc-dev mailing list