[OpenMP][libomptarget] Add math functions support in OpenMP offloading

[gtbercea]

Summary

Add kmpc function definition to libomptarget-nvptx.bc.

[ABataev]

libomptarget/deviceRTLs/nvptx/src/interface.h

570

I think you need to add __kmpc_powf(float), __kmpc_powl(long double), __kmpc_sinf(float), __kmpc_sinl(long double)

[gtbercea]

• Address comments.

[jdoerfert]

I guess here, and other places, we will later need to add more function/type combinations. Given that they all follow the same scheme, we should probably use a template for generation soon.

[gtbercea]

• Use macros.

[jdoerfert]

libomptarget/deviceRTLs/nvptx/src/interface.h

585

Shouldn't we do this with macros again?

I would even propose a separate "math_macro.inc" file that is included
in both places. In one the macro DECL_ONLY is set and we get only declarations
while in the other we expand to definitions.
The idea is we simplify the maintenance in the future and cut down code.
Finally, we probably want to reuse the "math_macro.inc" also in other deviceRTLs so
we always support the same functions across all targets. Does that makes sense?

libomptarget/deviceRTLs/nvptx/src/math.cu

21

I was thinking we have some macro for float/double generation:

#define __OPENMP_MATH_FUNC_1_FP(__fn, __kmpc_fn)   \
   __OPENMP_MATH_FUNC_1(float, __fn, __kmpc_fn) \
   __OPENMP_MATH_FUNC_1(double, __fn, __kmpc_fn) \

so we can cut down further below

29

I don't think we want the semicolon at the end.

34

Wouldn't we get the correct conversion for:

`__OPENMP_MATH_FUNC_2(long double, pow, __kmpc_powl);`

If not, should we add a two type version that casts the arguments
and define __OPENMP_MATH_FUNC_2 in terms of that one?

[gtbercea]

Replaced by: D61399

libomptarget/deviceRTLs/nvptx/CMakeLists.txt

set(cuda_src_files
       src/data_sharing.cu
       src/libcall.cu
       src/loop.cu
       src/omptarget-nvptx.cu
       src/parallel.cu
       src/reduction.cu
       src/sync.cu
       src/task.cu
+      src/math.cu
   )
 
   set(omp_data_objects src/omp_data.cu)
 
   # Get the compute capability the user requested or use SM_35 by default.
   # SM_35 is what clang uses by default.
   set(default_capabilities 35)
   if (DEFINED LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITY)

libomptarget/deviceRTLs/nvptx/src/interface.h

 
 EXTERN void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode,
                                           const void *buf, size_t size,
                                           int16_t is_shared, const void **res);
 
 EXTERN void __kmpc_restore_team_static_memory(int16_t isSPMDExecutionMode,
                                               int16_t is_shared);
 
+// POW
+EXTERN float __kmpc_powf(float a, float b);
+EXTERN double __kmpc_pow(double, double);
+EXTERN long double __kmpc_powl(long double a, long double b);
+
+// LOG
+EXTERN double __kmpc_log(double);
+EXTERN float __kmpc_logf(float);
+EXTERN double __kmpc_log10(double);
+EXTERN float __kmpc_log10f(float);
+EXTERN double __kmpc_log1p(double);
+EXTERN float __kmpc_log1pf(float);
+EXTERN double __kmpc_log2(double);
+EXTERN float __kmpc_log2f(float);
+EXTERN double __kmpc_logb(double);
+EXTERN float __kmpc_logbf(float);
+
+// SIN
+EXTERN float __kmpc_sinf(float);
+EXTERN double __kmpc_sin(double);
+EXTERN long double __kmpc_sinl(long double);
+
+// COS
+EXTERN float __kmpc_cosf(float);
+EXTERN double __kmpc_cos(double);
+EXTERN long double __kmpc_cosl(long double);
+
 #endif

libomptarget/deviceRTLs/nvptx/src/math.cu

+//===------------ math.cu - NVPTX OpenMP math constructs --------- CUDA -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains the implementation of math function not already handled.
+// Function in this file will end up as part of the .bc library of the
+// device and will call libdevice functions.
+//
+//===----------------------------------------------------------------------===//
+
+#include "omptarget-nvptx.h"
+
+// Single argument functions
+#define __OPENMP_MATH_FUNC_1(__ty, __fn, __kmpc_fn)                            \
+  EXTERN __ty __kmpc_fn(__ty __x) {                                            \
+    return __fn(__x);                                                          \
+  }
+
+// Double argument functions
+#define __OPENMP_MATH_FUNC_2(__ty, __fn, __kmpc_fn)                            \
+  EXTERN __ty __kmpc_fn(__ty __x, __ty __y) {                                  \
+    return __fn(__x, __y);                                                     \
+  }
+
+__OPENMP_MATH_FUNC_2(float, powf, __kmpc_powf);
+__OPENMP_MATH_FUNC_2(double, pow, __kmpc_pow);
+// no powl defined for the GPU device so use pow.
+EXTERN long double __kmpc_powl(long double a, long double b) {
+  return pow((double) a, (double) b);
+}
+
+// LOG
+__OPENMP_MATH_FUNC_1(double, log, __kmpc_log);
+__OPENMP_MATH_FUNC_1(float, logf, __kmpc_logf);
+__OPENMP_MATH_FUNC_1(double, log10, __kmpc_log10);
+__OPENMP_MATH_FUNC_1(float, log10f, __kmpc_log10f);
+__OPENMP_MATH_FUNC_1(double, log1p, __kmpc_log1p);
+__OPENMP_MATH_FUNC_1(float, log1pf, __kmpc_log1pf);
+__OPENMP_MATH_FUNC_1(double, log2, __kmpc_log2);
+__OPENMP_MATH_FUNC_1(float, log2f, __kmpc_log2f);
+__OPENMP_MATH_FUNC_1(double, logb, __kmpc_logb);
+__OPENMP_MATH_FUNC_1(float, logbf, __kmpc_logbf);
+
+__OPENMP_MATH_FUNC_1(float, sinf, __kmpc_sinf);
+__OPENMP_MATH_FUNC_1(double, sin, __kmpc_sin);
+// no sinl defined for the GPU device so use pow.
+EXTERN long double __kmpc_sinl(long double a) {
+  return sin((double) a);
+}
+
+// COS
+__OPENMP_MATH_FUNC_1(float, cosf, __kmpc_cosf);
+__OPENMP_MATH_FUNC_1(double, cos, __kmpc_cos);
+EXTERN long double __kmpc_cosl(long double a) {
+  return cos((double) a);
+}
 No newline at end of file