[Libclc-dev] [PATCH] Add mul_hi implementation [v2]

Aaron Watry awatry at gmail.com
Thu Aug 15 14:35:01 PDT 2013


Everything except long/ulong is handled by just casting to the next larger type,
doing the math and then shifting/casting the result.

For 64-bit types, we break the high/low parts of each operand apart, and do
a FOIL-based multiplication.

v2:
  Discard the stack-overflow implementation due to copyright concerns.
  - The implementation is still FOIL-based, but discards the previous code.
---
 generic/include/clc/clc.h              |   1 +
 generic/include/clc/integer/mul_hi.h   |   2 +
 generic/include/clc/integer/mul_hi.inc |   1 +
 generic/lib/SOURCES                    |   1 +
 generic/lib/integer/mul_hi.cl          | 109 +++++++++++++++++++++++++++++++++
 5 files changed, 114 insertions(+)
 create mode 100644 generic/include/clc/integer/mul_hi.h
 create mode 100644 generic/include/clc/integer/mul_hi.inc
 create mode 100644 generic/lib/integer/mul_hi.cl

diff --git a/generic/include/clc/clc.h b/generic/include/clc/clc.h
index ae212ed..305f058 100644
--- a/generic/include/clc/clc.h
+++ b/generic/include/clc/clc.h
@@ -68,6 +68,7 @@
 #include <clc/integer/hadd.h>
 #include <clc/integer/mad24.h>
 #include <clc/integer/mul24.h>
+#include <clc/integer/mul_hi.h>
 #include <clc/integer/rhadd.h>
 #include <clc/integer/rotate.h>
 #include <clc/integer/sub_sat.h>
diff --git a/generic/include/clc/integer/mul_hi.h b/generic/include/clc/integer/mul_hi.h
new file mode 100644
index 0000000..27b95d8
--- /dev/null
+++ b/generic/include/clc/integer/mul_hi.h
@@ -0,0 +1,2 @@
+#define __CLC_BODY <clc/integer/mul_hi.inc>
+#include <clc/integer/gentype.inc>
diff --git a/generic/include/clc/integer/mul_hi.inc b/generic/include/clc/integer/mul_hi.inc
new file mode 100644
index 0000000..ce9e5c0
--- /dev/null
+++ b/generic/include/clc/integer/mul_hi.inc
@@ -0,0 +1 @@
+_CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE mul_hi(__CLC_GENTYPE x, __CLC_GENTYPE y);
diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
index 5a68d07..d6e2d8c 100644
--- a/generic/lib/SOURCES
+++ b/generic/lib/SOURCES
@@ -14,6 +14,7 @@ integer/clz_impl.ll
 integer/hadd.cl
 integer/mad24.cl
 integer/mul24.cl
+integer/mul_hi.cl
 integer/rhadd.cl
 integer/rotate.cl
 integer/sub_sat.cl
diff --git a/generic/lib/integer/mul_hi.cl b/generic/lib/integer/mul_hi.cl
new file mode 100644
index 0000000..4c02efb
--- /dev/null
+++ b/generic/lib/integer/mul_hi.cl
@@ -0,0 +1,109 @@
+#include <clc/clc.h>
+
+//For all types EXCEPT long, which is implemented separately
+#define __CLC_MUL_HI_IMPL(BGENTYPE, GENTYPE, GENSIZE) \
+    _CLC_OVERLOAD _CLC_DECL GENTYPE mul_hi(GENTYPE x, GENTYPE y){ \
+        return (GENTYPE)(((BGENTYPE)x * (BGENTYPE)y) >> GENSIZE); \
+    } \
+
+//FOIL-based long mul_hi
+//
+// Summary: Treat mul_hi(long x, long y) as:
+// (a+b) * (c+d) where a and c are the high-order parts of x and y respectively
+// and b and d are the low-order parts of x and y.
+// Thinking back to algebra, we use FOIL to do the work.
+
+_CLC_OVERLOAD _CLC_DECL long mul_hi(long x, long y){
+    long f, o, i;
+    ulong l;
+
+    //Move the high/low halves of x/y into the lower 32-bits of variables so
+    //that we can multiply them without worrying about overflow.
+    long x_hi = x >> 32;
+    long x_lo = x & UINT_MAX;
+    long y_hi = y >> 32;
+    long y_lo = y & UINT_MAX;
+
+    //Multiply all of the components according to FOIL method
+    f = x_hi * y_hi;
+    o = x_hi * y_lo;
+    i = x_lo * y_hi;
+    l = x_lo * y_lo;
+
+    //Now add the components back together in the following steps:
+    //F: doesn't need to be modified
+    //O/I: Need to be added together.
+    //L: Shift right by 32-bits, then add into the sum of O and I
+    //Once O/I/L are summed up, then shift the sum by 32-bits and add to F.
+    //
+    //We use hadd to give us a bit of extra precision for the intermediate sums
+    //but as a result, we shift by 31 bits instead of 32
+    return (long)(f + (hadd(o, (i + (long)((ulong)l>>32))) >> 31));
+}
+
+_CLC_OVERLOAD _CLC_DECL ulong mul_hi(ulong x, ulong y){
+    ulong f, o, i;
+    ulong l;
+
+    //Move the high/low halves of x/y into the lower 32-bits of variables so
+    //that we can multiply them without worrying about overflow.
+    ulong x_hi = x >> 32;
+    ulong x_lo = x & UINT_MAX;
+    ulong y_hi = y >> 32;
+    ulong y_lo = y & UINT_MAX;
+
+    //Multiply all of the components according to FOIL method
+    f = x_hi * y_hi;
+    o = x_hi * y_lo;
+    i = x_lo * y_hi;
+    l = x_lo * y_lo;
+
+    //Now add the components back together, taking care to respect the fact that:
+    //F: doesn't need to be modified
+    //O/I: Need to be added together.
+    //L: Shift right by 32-bits, then add into the sum of O and I
+    //Once O/I/L are summed up, then shift the sum by 32-bits and add to F.
+    //
+    //We use hadd to give us a bit of extra precision for the intermediate sums
+    //but as a result, we shift by 31 bits instead of 32
+    return (f + (hadd(o, (i + (l>>32))) >> 31));
+}
+
+#define __CLC_MUL_HI_VEC(GENTYPE) \
+    _CLC_OVERLOAD _CLC_DECL GENTYPE##2 mul_hi(GENTYPE##2 x, GENTYPE##2 y){ \
+        return (GENTYPE##2){mul_hi(x.s0, y.s0), mul_hi(x.s1, y.s1)}; \
+    } \
+    _CLC_OVERLOAD _CLC_DECL GENTYPE##3 mul_hi(GENTYPE##3 x, GENTYPE##3 y){ \
+        return (GENTYPE##3){mul_hi(x.s0, y.s0), mul_hi(x.s1, y.s1), mul_hi(x.s2, y.s2)}; \
+    } \
+    _CLC_OVERLOAD _CLC_DECL GENTYPE##4 mul_hi(GENTYPE##4 x, GENTYPE##4 y){ \
+        return (GENTYPE##4){mul_hi(x.lo, y.lo), mul_hi(x.hi, y.hi)}; \
+    } \
+    _CLC_OVERLOAD _CLC_DECL GENTYPE##8 mul_hi(GENTYPE##8 x, GENTYPE##8 y){ \
+        return (GENTYPE##8){mul_hi(x.lo, y.lo), mul_hi(x.hi, y.hi)}; \
+    } \
+    _CLC_OVERLOAD _CLC_DECL GENTYPE##16 mul_hi(GENTYPE##16 x, GENTYPE##16 y){ \
+        return (GENTYPE##16){mul_hi(x.lo, y.lo), mul_hi(x.hi, y.hi)}; \
+    } \
+
+#define __CLC_MUL_HI_DEC_IMPL(BTYPE, TYPE, BITS) \
+    __CLC_MUL_HI_IMPL(BTYPE, TYPE, BITS) \
+    __CLC_MUL_HI_VEC(TYPE)
+
+#define __CLC_MUL_HI_TYPES() \
+    __CLC_MUL_HI_DEC_IMPL(short, char, 8) \
+    __CLC_MUL_HI_DEC_IMPL(ushort, uchar, 8) \
+    __CLC_MUL_HI_DEC_IMPL(int, short, 16) \
+    __CLC_MUL_HI_DEC_IMPL(uint, ushort, 16) \
+    __CLC_MUL_HI_DEC_IMPL(long, int, 32) \
+    __CLC_MUL_HI_DEC_IMPL(ulong, uint, 32) \
+    __CLC_MUL_HI_VEC(long) \
+    __CLC_MUL_HI_VEC(ulong)
+
+__CLC_MUL_HI_TYPES()
+
+#undef __CLC_MUL_HI_TYPES
+#undef __CLC_MUL_HI_DEC_IMPL
+#undef __CLC_MUL_HI_IMPL
+#undef __CLC_MUL_HI_VEC
+#undef __CLC_B32
-- 
1.8.1.2





More information about the Libclc-dev mailing list