Openmp transformation using llvm-openmp runtime routines - 2
Table of Contents
Introduction
In continuation to last blog, in this blog, we’ll discuss array initialisation and reduction example.
Array Initialisation
openmp
#include <stdio.h>
#include <omp.h>
void initialize_array( int n, float *a, float val ) {
int i;
#pragma omp parallel
{
#pragma omp for
for (i = 0; i < n; i++) {
a[i] = val;
}
}
}
int main() {
float a[1000000];
initialize_array(1000000, a, 12.91);
for (int i = 0; i < 10; i++) {
printf("%f\n", a[i]);
}
}
% time clang++ c.cpp -fopenmp && time ./a.out
clang++ c.cpp -fopenmp 0.07s user 0.12s system 41% cpu 0.454 total
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
./a.out 0.00s user 0.00s system 103% cpu 0.005 total
transformed code
At first, we will generate llvm:
llvm
; ModuleID = 'c.cpp'
source_filename = "c.cpp"
target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
target triple = "arm64-apple-macosx14.0.0"
%struct.ident_t = type { i32, i32, i32, i32, ptr }
@0 = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
@1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
@2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 66, i32 0, i32 22, ptr @0 }, align 8
@3 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
@.str = private unnamed_addr constant [4 x i8] c"%f\0A\00", align 1
; Function Attrs: mustprogress noinline nounwind optnone ssp uwtable(sync)
define void @_Z16initialize_arrayiPff(i32 noundef %0, ptr noundef %1, float noundef %2) #0 {
%4 = alloca i32, align 4
%5 = alloca ptr, align 8
%6 = alloca float, align 4
%7 = alloca i32, align 4
store i32 %0, ptr %4, align 4
store ptr %1, ptr %5, align 8
store float %2, ptr %6, align 4
call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @3, i32 3, ptr @_Z16initialize_arrayiPff.omp_outlined, ptr %4, ptr %5, ptr %6)
ret void
}
; Function Attrs: noinline norecurse nounwind optnone ssp uwtable(sync)
define internal void @_Z16initialize_arrayiPff.omp_outlined(ptr noalias noundef %0, ptr noalias noundef %1, ptr noundef nonnull align 4 dereferenceable(4) %2, ptr noundef nonnull align 8 dereferenceable(8) %3, ptr noundef nonnull align 4 dereferenceable(4) %4) #1 {
%6 = alloca ptr, align 8
%7 = alloca ptr, align 8
%8 = alloca ptr, align 8
%9 = alloca ptr, align 8
%10 = alloca ptr, align 8
%11 = alloca i32, align 4
%12 = alloca i32, align 4
%13 = alloca i32, align 4
%14 = alloca i32, align 4
%15 = alloca i32, align 4
%16 = alloca i32, align 4
%17 = alloca i32, align 4
%18 = alloca i32, align 4
%19 = alloca i32, align 4
%20 = alloca i32, align 4
store ptr %0, ptr %6, align 8
store ptr %1, ptr %7, align 8
store ptr %2, ptr %8, align 8
store ptr %3, ptr %9, align 8
store ptr %4, ptr %10, align 8
%21 = load ptr, ptr %8, align 8
%22 = load ptr, ptr %9, align 8
%23 = load ptr, ptr %10, align 8
%24 = load i32, ptr %21, align 4
store i32 %24, ptr %13, align 4
%25 = load i32, ptr %13, align 4
%26 = sub nsw i32 %25, 0
%27 = sdiv i32 %26, 1
%28 = sub nsw i32 %27, 1
store i32 %28, ptr %14, align 4
store i32 0, ptr %15, align 4
%29 = load i32, ptr %13, align 4
%30 = icmp slt i32 0, %29
br i1 %30, label %31, label %66
31: ; preds = %5
store i32 0, ptr %16, align 4
%32 = load i32, ptr %14, align 4
store i32 %32, ptr %17, align 4
store i32 1, ptr %18, align 4
store i32 0, ptr %19, align 4
%33 = load ptr, ptr %6, align 8
%34 = load i32, ptr %33, align 4
call void @__kmpc_for_static_init_4(ptr @1, i32 %34, i32 34, ptr %19, ptr %16, ptr %17, ptr %18, i32 1, i32 1)
%35 = load i32, ptr %17, align 4
%36 = load i32, ptr %14, align 4
%37 = icmp sgt i32 %35, %36
br i1 %37, label %38, label %40
38: ; preds = %31
%39 = load i32, ptr %14, align 4
br label %42
40: ; preds = %31
%41 = load i32, ptr %17, align 4
br label %42
42: ; preds = %40, %38
%43 = phi i32 [ %39, %38 ], [ %41, %40 ]
store i32 %43, ptr %17, align 4
%44 = load i32, ptr %16, align 4
store i32 %44, ptr %11, align 4
br label %45
45: ; preds = %59, %42
%46 = load i32, ptr %11, align 4
%47 = load i32, ptr %17, align 4
%48 = icmp sle i32 %46, %47
br i1 %48, label %49, label %62
49: ; preds = %45
%50 = load i32, ptr %11, align 4
%51 = mul nsw i32 %50, 1
%52 = add nsw i32 0, %51
store i32 %52, ptr %20, align 4
%53 = load float, ptr %23, align 4
%54 = load ptr, ptr %22, align 8
%55 = load i32, ptr %20, align 4
%56 = sext i32 %55 to i64
%57 = getelementptr inbounds float, ptr %54, i64 %56
store float %53, ptr %57, align 4
br label %58
58: ; preds = %49
br label %59
59: ; preds = %58
%60 = load i32, ptr %11, align 4
%61 = add nsw i32 %60, 1
store i32 %61, ptr %11, align 4
br label %45
62: ; preds = %45
br label %63
63: ; preds = %62
%64 = load ptr, ptr %6, align 8
%65 = load i32, ptr %64, align 4
call void @__kmpc_for_static_fini(ptr @1, i32 %65)
br label %66
66: ; preds = %63, %5
%67 = load ptr, ptr %6, align 8
%68 = load i32, ptr %67, align 4
call void @__kmpc_barrier(ptr @2, i32 %68)
ret void
}
; Function Attrs: nounwind
declare void @__kmpc_for_static_init_4(ptr, i32, i32, ptr, ptr, ptr, ptr, i32, i32) #2
; Function Attrs: nounwind
declare void @__kmpc_for_static_fini(ptr, i32) #2
; Function Attrs: convergent nounwind
declare void @__kmpc_barrier(ptr, i32) #3
; Function Attrs: nounwind
declare !callback !6 void @__kmpc_fork_call(ptr, i32, ptr, ...) #2
; Function Attrs: mustprogress noinline norecurse optnone ssp uwtable(sync)
define noundef i32 @main() #4 {
%1 = alloca i32, align 4
%2 = alloca [1000000 x float], align 4
%3 = alloca i32, align 4
store i32 0, ptr %1, align 4
%4 = getelementptr inbounds [1000000 x float], ptr %2, i64 0, i64 0
call void @_Z16initialize_arrayiPff(i32 noundef 1000000, ptr noundef %4, float noundef 0x4029D1EB80000000)
store i32 0, ptr %3, align 4
br label %5
5: ; preds = %15, %0
%6 = load i32, ptr %3, align 4
%7 = icmp slt i32 %6, 10
br i1 %7, label %8, label %18
8: ; preds = %5
%9 = load i32, ptr %3, align 4
%10 = sext i32 %9 to i64
%11 = getelementptr inbounds [1000000 x float], ptr %2, i64 0, i64 %10
%12 = load float, ptr %11, align 4
%13 = fpext float %12 to double
%14 = call i32 (ptr, ...) @printf(ptr noundef @.str, double noundef %13)
br label %15
15: ; preds = %8
%16 = load i32, ptr %3, align 4
%17 = add nsw i32 %16, 1
store i32 %17, ptr %3, align 4
br label %5, !llvm.loop !8
18: ; preds = %5
%19 = load i32, ptr %1, align 4
ret i32 %19
}
declare i32 @printf(ptr noundef, ...) #5
attributes #0 = { mustprogress noinline nounwind optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #1 = { noinline norecurse nounwind optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #2 = { nounwind }
attributes #3 = { convergent nounwind }
attributes #4 = { mustprogress noinline norecurse optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #5 = { "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
!llvm.module.flags = !{!0, !1, !2, !3, !4}
!llvm.ident = !{!5}
!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 7, !"openmp", i32 51}
!2 = !{i32 8, !"PIC Level", i32 2}
!3 = !{i32 7, !"uwtable", i32 1}
!4 = !{i32 7, !"frame-pointer", i32 1}
!5 = !{!"Homebrew clang version 18.1.4"}
!6 = !{!7}
!7 = !{i64 2, i64 -1, i64 -1, i1 true}
!8 = distinct !{!8, !9}
!9 = !{!"llvm.loop.mustprogress"}
Here you can see use of __kmpc_for_static_init_4, __kmpc_for_static_fini and __kmpc_barrier. So, we’ll have to port it into kmp.h file and thus it looks like:
typedef int kmp_int32;
typedef struct ident {
kmp_int32 reserved_1; /**< might be used in Fortran; see above */
kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
identifies this union member */
kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */
#if USE_ITT_BUILD
/* but currently used for storing region-specific ITT */
/* contextual information. */
#endif /* USE_ITT_BUILD */
kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */
char const *psource; /**< String describing the source location.
The string is composed of semi-colon separated fields
which describe the source file, the function and a pair
of line numbers that delimit the construct. */
// Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
// kmp_int32 get_openmp_version() {
// return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
// }
} ident_t;
typedef struct ident ident_t;
typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
extern "C" {
/*!
@ingroup PARALLEL
@param loc source location information
@param argc total number of arguments in the ellipsis
@param microtask pointer to callback routine consisting of outlined parallel
construct
@param ... pointers to shared variables that aren't global
Do the actual fork and call the microtask in the relevant number of threads.
*/
void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
kmpc_micro microtask, ...);
/*
@param loc Source code location
@param gtid Global thread id of this thread
@param schedtype Scheduling type
@param plastiter Pointer to the "last iteration" flag
@param plower Pointer to the lower bound
@param pupper Pointer to the upper bound
@param pstride Pointer to the stride
@param incr Loop increment
@param chunk The chunk size
The functions compute the upper and lower bounds and stride to be used for the
set of iterations to be executed by the current thread from the statically
scheduled loop that is described by the initial values of the bounds, stride,
increment and chunk size.
*/
void __kmpc_for_static_init_4(ident_t *loc, kmp_int32 gtid, kmp_int32 schedtype,
kmp_int32 *plastiter, kmp_int32 *plower,
kmp_int32 *pupper, kmp_int32 *pstride,
kmp_int32 incr, kmp_int32 chunk);
/*
Mark the end of a statically scheduled loop.
*/
void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
/*
Execute a barrier.
*/
void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
}
Now, translating llvm code to cpp we get:
c-transformed.cpp
#include <stdio.h>
#include <omp.h>
#include "kmp.h"
void kmp_initialize_array( kmp_int32 *global_tid, kmp_int32 *bound_tid, int *n, float *arr, float *value ) {
// @1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 514; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
int *lastiter = new int(0);
int *lower = new int(0);
int *upper = new int(*n);
int *stride = new int(1);
__kmpc_for_static_init_4(&loc, *global_tid, 34, lastiter, lower, upper, stride, 1, 1);
while ( *lower < *upper ) {
// set the value of arr[*lower] to *value
arr[*lower] = *value;
*lower += 1;
}
__kmpc_for_static_fini(&loc, *global_tid);
__kmpc_barrier(&loc, *global_tid);
return;
}
void initialize_array( int n, float *a, float val ) {
ident_t loc;
/*
@0 = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
@1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
@2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 66, i32 0, i32 22, ptr @0 }, align 8
@3 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
ident_t for this case is @3
*/
loc.reserved_1 = 0; loc.flags = 2; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;";
printf("value here = %f\n", val);
__kmpc_fork_call(&loc, 3, (kmpc_micro)kmp_initialize_array, &n, a, &val);
}
int main() {
float a[1000000];
float val = 12.91;
initialize_array(1000000, a, val);
for (int i = 0; i < 10; i++) {
printf("%f\n", a[i]);
}
}
% time clang++ c-transformed.cpp -fopenmp && time ./a.out
clang++ c-transformed.cpp -fopenmp 0.08s user 0.03s system 119% cpu 0.093 total
value here = 12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
12.910000
./a.out 0.00s user 0.00s system 188% cpu 0.003 total
Reduction
openmp
#include <stdio.h>
#include <omp.h>
#include "kmp.h"
void initialize_array( int n, float *a, float val ) {
int i;
#pragma omp parallel
{
#pragma omp for
for (i = 0; i < n; i++) {
a[i] = val;
}
}
#pragma omp barrier
return;
}
void parallel_Sum_created( kmp_int32 *global_tid, kmp_int32 *bound_tid, float *total_Sum, int *n, float *a ) {
// @1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 514; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
// @2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 66, i32 0, i32 22, ptr @0 }, align 8
ident_t loc2; loc2.reserved_1 = 0; loc2.flags = 66; loc2.reserved_2 = 0; loc2.reserved_3 = 22; loc2.psource = ";unknown;unknown;0;0;;\00";
float partial_sum = 0;
int *lastiter = new int(0);
int *lower = new int(0);
int *upper = new int(*n);
int *stride = new int(1);
__kmpc_for_static_init_4(&loc, *global_tid, 34, lastiter, lower, upper, stride, 1, 1);
while ( *lower < *upper ) {
partial_sum += a[*lower];
*lower += 1;
}
__kmpc_for_static_fini(&loc, *global_tid);
__kmpc_barrier(&loc2, *global_tid);
*total_Sum += partial_sum;
printf("Thread %d: partial sum: %f total sum: %f\n", omp_get_thread_num(), partial_sum, *total_Sum);
return;
}
void parallel_Sum( int n, float *a ) {
float total_Sum;
// @3 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 2; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
__kmpc_fork_call(&loc, 3, (kmpc_micro)parallel_Sum_created, &total_Sum, &n, a);
printf("Total sum: %f\n", total_Sum);
}
int main() {
float a[1000000];
initialize_array(1000000, a, 1.0);
parallel_Sum(1000000, a);
}
% time clang++ e.cpp -fopenmp && time ./a.out
clang++ e.cpp -fopenmp 0.09s user 0.03s system 118% cpu 0.095 total
Total sum: 1000000.000000
./a.out 0.01s user 0.00s system 253% cpu 0.003 total
transformed code
Beginning with llvm:
llvm
; ModuleID = 'e.cpp'
source_filename = "e.cpp"
target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
target triple = "arm64-apple-macosx14.0.0"
%struct.ident_t = type { i32, i32, i32, i32, ptr }
@0 = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
@1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
@2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 66, i32 0, i32 22, ptr @0 }, align 8
@3 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
@.str = private unnamed_addr constant [15 x i8] c"Total sum: %f\0A\00", align 1
@.str.1 = private unnamed_addr constant [11 x i8] c"a[i] = %f\0A\00", align 1
; Function Attrs: mustprogress noinline nounwind optnone ssp uwtable(sync)
define void @_Z16initialize_arrayiPff(i32 noundef %0, ptr noundef %1, float noundef %2) #0 {
%4 = alloca i32, align 4
%5 = alloca ptr, align 8
%6 = alloca float, align 4
%7 = alloca i32, align 4
store i32 %0, ptr %4, align 4
store ptr %1, ptr %5, align 8
store float %2, ptr %6, align 4
call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @3, i32 3, ptr @_Z16initialize_arrayiPff.omp_outlined, ptr %4, ptr %5, ptr %6)
ret void
}
; Function Attrs: noinline norecurse nounwind optnone ssp uwtable(sync)
define internal void @_Z16initialize_arrayiPff.omp_outlined(ptr noalias noundef %0, ptr noalias noundef %1, ptr noundef nonnull align 4 dereferenceable(4) %2, ptr noundef nonnull align 8 dereferenceable(8) %3, ptr noundef nonnull align 4 dereferenceable(4) %4) #1 {
%6 = alloca ptr, align 8
%7 = alloca ptr, align 8
%8 = alloca ptr, align 8
%9 = alloca ptr, align 8
%10 = alloca ptr, align 8
%11 = alloca i32, align 4
%12 = alloca i32, align 4
%13 = alloca i32, align 4
%14 = alloca i32, align 4
%15 = alloca i32, align 4
%16 = alloca i32, align 4
%17 = alloca i32, align 4
%18 = alloca i32, align 4
%19 = alloca i32, align 4
%20 = alloca i32, align 4
store ptr %0, ptr %6, align 8
store ptr %1, ptr %7, align 8
store ptr %2, ptr %8, align 8
store ptr %3, ptr %9, align 8
store ptr %4, ptr %10, align 8
%21 = load ptr, ptr %8, align 8
%22 = load ptr, ptr %9, align 8
%23 = load ptr, ptr %10, align 8
%24 = load i32, ptr %21, align 4
store i32 %24, ptr %13, align 4
%25 = load i32, ptr %13, align 4
%26 = sub nsw i32 %25, 0
%27 = sdiv i32 %26, 1
%28 = sub nsw i32 %27, 1
store i32 %28, ptr %14, align 4
store i32 0, ptr %15, align 4
%29 = load i32, ptr %13, align 4
%30 = icmp slt i32 0, %29
br i1 %30, label %31, label %66
31: ; preds = %5
store i32 0, ptr %16, align 4
%32 = load i32, ptr %14, align 4
store i32 %32, ptr %17, align 4
store i32 1, ptr %18, align 4
store i32 0, ptr %19, align 4
%33 = load ptr, ptr %6, align 8
%34 = load i32, ptr %33, align 4
call void @__kmpc_for_static_init_4(ptr @1, i32 %34, i32 34, ptr %19, ptr %16, ptr %17, ptr %18, i32 1, i32 1)
%35 = load i32, ptr %17, align 4
%36 = load i32, ptr %14, align 4
%37 = icmp sgt i32 %35, %36
br i1 %37, label %38, label %40
38: ; preds = %31
%39 = load i32, ptr %14, align 4
br label %42
40: ; preds = %31
%41 = load i32, ptr %17, align 4
br label %42
42: ; preds = %40, %38
%43 = phi i32 [ %39, %38 ], [ %41, %40 ]
store i32 %43, ptr %17, align 4
%44 = load i32, ptr %16, align 4
store i32 %44, ptr %11, align 4
br label %45
45: ; preds = %59, %42
%46 = load i32, ptr %11, align 4
%47 = load i32, ptr %17, align 4
%48 = icmp sle i32 %46, %47
br i1 %48, label %49, label %62
49: ; preds = %45
%50 = load i32, ptr %11, align 4
%51 = mul nsw i32 %50, 1
%52 = add nsw i32 0, %51
store i32 %52, ptr %20, align 4
%53 = load float, ptr %23, align 4
%54 = load ptr, ptr %22, align 8
%55 = load i32, ptr %20, align 4
%56 = sext i32 %55 to i64
%57 = getelementptr inbounds float, ptr %54, i64 %56
store float %53, ptr %57, align 4
br label %58
58: ; preds = %49
br label %59
59: ; preds = %58
%60 = load i32, ptr %11, align 4
%61 = add nsw i32 %60, 1
store i32 %61, ptr %11, align 4
br label %45
62: ; preds = %45
br label %63
63: ; preds = %62
%64 = load ptr, ptr %6, align 8
%65 = load i32, ptr %64, align 4
call void @__kmpc_for_static_fini(ptr @1, i32 %65)
br label %66
66: ; preds = %63, %5
%67 = load ptr, ptr %6, align 8
%68 = load i32, ptr %67, align 4
call void @__kmpc_barrier(ptr @2, i32 %68)
ret void
}
; Function Attrs: nounwind
declare void @__kmpc_for_static_init_4(ptr, i32, i32, ptr, ptr, ptr, ptr, i32, i32) #2
; Function Attrs: nounwind
declare void @__kmpc_for_static_fini(ptr, i32) #2
; Function Attrs: convergent nounwind
declare void @__kmpc_barrier(ptr, i32) #3
; Function Attrs: nounwind
declare !callback !6 void @__kmpc_fork_call(ptr, i32, ptr, ...) #2
; Function Attrs: mustprogress noinline optnone ssp uwtable(sync)
define void @_Z12parallel_SumiPf(i32 noundef %0, ptr noundef %1) #4 {
%3 = alloca i32, align 4
%4 = alloca ptr, align 8
%5 = alloca i32, align 4
%6 = alloca float, align 4
%7 = alloca float, align 4
store i32 %0, ptr %3, align 4
store ptr %1, ptr %4, align 8
call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @3, i32 3, ptr @_Z12parallel_SumiPf.omp_outlined, ptr %7, ptr %3, ptr %4)
%8 = load float, ptr %7, align 4
%9 = fpext float %8 to double
%10 = call i32 (ptr, ...) @printf(ptr noundef @.str, double noundef %9)
ret void
}
; Function Attrs: noinline norecurse nounwind optnone ssp uwtable(sync)
define internal void @_Z12parallel_SumiPf.omp_outlined(ptr noalias noundef %0, ptr noalias noundef %1, ptr noundef nonnull align 4 dereferenceable(4) %2, ptr noundef nonnull align 4 dereferenceable(4) %3, ptr noundef nonnull align 8 dereferenceable(8) %4) #1 {
%6 = alloca ptr, align 8
%7 = alloca ptr, align 8
%8 = alloca ptr, align 8 // total_Sum
%9 = alloca ptr, align 8 // n
%10 = alloca ptr, align 8 // a
%11 = alloca float, align 4
%12 = alloca i32, align 4
%13 = alloca i32, align 4
%14 = alloca i32, align 4
%15 = alloca i32, align 4
%16 = alloca i32, align 4
%17 = alloca i32, align 4
%18 = alloca i32, align 4
%19 = alloca i32, align 4
%20 = alloca i32, align 4
%21 = alloca i32, align 4
store ptr %0, ptr %6, align 8
store ptr %1, ptr %7, align 8
store ptr %2, ptr %8, align 8 // total_Sum
store ptr %3, ptr %9, align 8 // n
store ptr %4, ptr %10, align 8 // a
%22 = load ptr, ptr %8, align 8
%23 = load ptr, ptr %9, align 8
%24 = load ptr, ptr %10, align 8
store float 0.000000e+00, ptr %11, align 4 // partial_Sum = 0.0
store float 0.000000e+00, ptr %22, align 4 // total_Sum = 0.0
%25 = load i32, ptr %23, align 4
store i32 %25, ptr %14, align 4
%26 = load i32, ptr %14, align 4
%27 = sub nsw i32 %26, 0
%28 = sdiv i32 %27, 1
%29 = sub nsw i32 %28, 1
store i32 %29, ptr %15, align 4
store i32 0, ptr %16, align 4
%30 = load i32, ptr %14, align 4
%31 = icmp slt i32 0, %30
br i1 %31, label %32, label %69
32: ; preds = %5
store i32 0, ptr %17, align 4
%33 = load i32, ptr %15, align 4
store i32 %33, ptr %18, align 4
store i32 1, ptr %19, align 4
store i32 0, ptr %20, align 4
%34 = load ptr, ptr %6, align 8
%35 = load i32, ptr %34, align 4
call void @__kmpc_for_static_init_4(ptr @1, i32 %35, i32 34, ptr %20, ptr %17, ptr %18, ptr %19, i32 1, i32 1)
%36 = load i32, ptr %18, align 4
%37 = load i32, ptr %15, align 4
%38 = icmp sgt i32 %36, %37
br i1 %38, label %39, label %41
39: ; preds = %32
%40 = load i32, ptr %15, align 4
br label %43
41: ; preds = %32
%42 = load i32, ptr %18, align 4
br label %43
43: ; preds = %41, %39
%44 = phi i32 [ %40, %39 ], [ %42, %41 ]
store i32 %44, ptr %18, align 4
%45 = load i32, ptr %17, align 4
store i32 %45, ptr %12, align 4
br label %46
46: ; preds = %62, %43
%47 = load i32, ptr %12, align 4
%48 = load i32, ptr %18, align 4
%49 = icmp sle i32 %47, %48
br i1 %49, label %50, label %65
50: ; preds = %46
%51 = load i32, ptr %12, align 4
%52 = mul nsw i32 %51, 1
%53 = add nsw i32 0, %52
store i32 %53, ptr %21, align 4
%54 = load ptr, ptr %24, align 8
%55 = load i32, ptr %21, align 4
%56 = sext i32 %55 to i64
%57 = getelementptr inbounds float, ptr %54, i64 %56
%58 = load float, ptr %57, align 4
%59 = load float, ptr %11, align 4
%60 = fadd float %59, %58
store float %60, ptr %11, align 4
br label %61
61: ; preds = %50
br label %62
62: ; preds = %61
%63 = load i32, ptr %12, align 4
%64 = add nsw i32 %63, 1
store i32 %64, ptr %12, align 4
br label %46
65: ; preds = %46
br label %66
66: ; preds = %65
%67 = load ptr, ptr %6, align 8
%68 = load i32, ptr %67, align 4
call void @__kmpc_for_static_fini(ptr @1, i32 %68)
br label %69
69: ; preds = %66, %5
%70 = load ptr, ptr %6, align 8
%71 = load i32, ptr %70, align 4
call void @__kmpc_barrier(ptr @2, i32 %71)
%72 = load float, ptr %11, align 4
%73 = load float, ptr %22, align 4
%74 = fadd float %73, %72
store float %74, ptr %22, align 4
ret void
}
declare i32 @printf(ptr noundef, ...) #5
; Function Attrs: mustprogress noinline nounwind optnone ssp uwtable(sync)
define void @_Z5checkiPff(i32 noundef %0, ptr noundef %1, float noundef %2) #0 {
%4 = alloca i32, align 4
%5 = alloca ptr, align 8
%6 = alloca float, align 4
%7 = alloca i32, align 4
store i32 %0, ptr %4, align 4
store ptr %1, ptr %5, align 8
store float %2, ptr %6, align 4
call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @3, i32 3, ptr @_Z5checkiPff.omp_outlined, ptr %4, ptr %5, ptr %6)
ret void
}
; Function Attrs: noinline norecurse nounwind optnone ssp uwtable(sync)
define internal void @_Z5checkiPff.omp_outlined(ptr noalias noundef %0, ptr noalias noundef %1, ptr noundef nonnull align 4 dereferenceable(4) %2, ptr noundef nonnull align 8 dereferenceable(8) %3, ptr noundef nonnull align 4 dereferenceable(4) %4) #1 personality ptr @__gxx_personality_v0 {
%6 = alloca ptr, align 8
%7 = alloca ptr, align 8
%8 = alloca ptr, align 8
%9 = alloca ptr, align 8
%10 = alloca ptr, align 8
%11 = alloca i32, align 4
%12 = alloca i32, align 4
%13 = alloca i32, align 4
%14 = alloca i32, align 4
%15 = alloca i32, align 4
%16 = alloca i32, align 4
%17 = alloca i32, align 4
%18 = alloca i32, align 4
%19 = alloca i32, align 4
%20 = alloca i32, align 4
store ptr %0, ptr %6, align 8
store ptr %1, ptr %7, align 8
store ptr %2, ptr %8, align 8
store ptr %3, ptr %9, align 8
store ptr %4, ptr %10, align 8
%21 = load ptr, ptr %8, align 8
%22 = load ptr, ptr %9, align 8
%23 = load ptr, ptr %10, align 8
%24 = load i32, ptr %21, align 4
store i32 %24, ptr %13, align 4
%25 = load i32, ptr %13, align 4
%26 = sub nsw i32 %25, 0
%27 = sdiv i32 %26, 1
%28 = sub nsw i32 %27, 1
store i32 %28, ptr %14, align 4
store i32 0, ptr %15, align 4
%29 = load i32, ptr %13, align 4
%30 = icmp slt i32 0, %29
br i1 %30, label %31, label %78
31: ; preds = %5
store i32 0, ptr %16, align 4
%32 = load i32, ptr %14, align 4
store i32 %32, ptr %17, align 4
store i32 1, ptr %18, align 4
store i32 0, ptr %19, align 4
%33 = load ptr, ptr %6, align 8
%34 = load i32, ptr %33, align 4
call void @__kmpc_for_static_init_4(ptr @1, i32 %34, i32 34, ptr %19, ptr %16, ptr %17, ptr %18, i32 1, i32 1)
%35 = load i32, ptr %17, align 4
%36 = load i32, ptr %14, align 4
%37 = icmp sgt i32 %35, %36
br i1 %37, label %38, label %40
38: ; preds = %31
%39 = load i32, ptr %14, align 4
br label %42
40: ; preds = %31
%41 = load i32, ptr %17, align 4
br label %42
42: ; preds = %40, %38
%43 = phi i32 [ %39, %38 ], [ %41, %40 ]
store i32 %43, ptr %17, align 4
%44 = load i32, ptr %16, align 4
store i32 %44, ptr %11, align 4
br label %45
45: ; preds = %71, %42
%46 = load i32, ptr %11, align 4
%47 = load i32, ptr %17, align 4
%48 = icmp sle i32 %46, %47
br i1 %48, label %49, label %74
49: ; preds = %45
%50 = load i32, ptr %11, align 4
%51 = mul nsw i32 %50, 1
%52 = add nsw i32 0, %51
store i32 %52, ptr %20, align 4
%53 = load ptr, ptr %22, align 8
%54 = load i32, ptr %20, align 4
%55 = sext i32 %54 to i64
%56 = getelementptr inbounds float, ptr %53, i64 %55
%57 = load float, ptr %56, align 4
%58 = load float, ptr %23, align 4
%59 = fcmp une float %57, %58
br i1 %59, label %60, label %69
60: ; preds = %49
%61 = load ptr, ptr %22, align 8
%62 = load i32, ptr %20, align 4
%63 = sext i32 %62 to i64
%64 = getelementptr inbounds float, ptr %61, i64 %63
%65 = load float, ptr %64, align 4
%66 = fpext float %65 to double
%67 = invoke i32 (ptr, ...) @printf(ptr noundef @.str.1, double noundef %66)
to label %68 unwind label %81
68: ; preds = %60
br label %69
69: ; preds = %68, %49
br label %70
70: ; preds = %69
br label %71
71: ; preds = %70
%72 = load i32, ptr %11, align 4
%73 = add nsw i32 %72, 1
store i32 %73, ptr %11, align 4
br label %45
74: ; preds = %45
br label %75
75: ; preds = %74
%76 = load ptr, ptr %6, align 8
%77 = load i32, ptr %76, align 4
call void @__kmpc_for_static_fini(ptr @1, i32 %77)
br label %78
78: ; preds = %75, %5
%79 = load ptr, ptr %6, align 8
%80 = load i32, ptr %79, align 4
call void @__kmpc_barrier(ptr @2, i32 %80)
ret void
81: ; preds = %60
%82 = landingpad { ptr, i32 }
catch ptr null
%83 = extractvalue { ptr, i32 } %82, 0
call void @__clang_call_terminate(ptr %83) #8
unreachable
}
declare i32 @__gxx_personality_v0(...)
; Function Attrs: noinline noreturn nounwind ssp uwtable(sync)
define linkonce_odr hidden void @__clang_call_terminate(ptr noundef %0) #6 {
%2 = call ptr @__cxa_begin_catch(ptr %0) #2
call void @_ZSt9terminatev() #8
unreachable
}
declare ptr @__cxa_begin_catch(ptr)
declare void @_ZSt9terminatev()
; Function Attrs: mustprogress noinline norecurse optnone ssp uwtable(sync)
define noundef i32 @main() #7 {
%1 = alloca i32, align 4
%2 = alloca [1000000 x float], align 4
store i32 0, ptr %1, align 4
%3 = getelementptr inbounds [1000000 x float], ptr %2, i64 0, i64 0
call void @_Z16initialize_arrayiPff(i32 noundef 1000000, ptr noundef %3, float noundef 1.000000e+00)
%4 = getelementptr inbounds [1000000 x float], ptr %2, i64 0, i64 0
call void @_Z12parallel_SumiPf(i32 noundef 1000000, ptr noundef %4)
%5 = getelementptr inbounds [1000000 x float], ptr %2, i64 0, i64 0
call void @_Z5checkiPff(i32 noundef 1000000, ptr noundef %5, float noundef 1.000000e+00)
ret i32 0
}
attributes #0 = { mustprogress noinline nounwind optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #1 = { noinline norecurse nounwind optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #2 = { nounwind }
attributes #3 = { convergent nounwind }
attributes #4 = { mustprogress noinline optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #5 = { "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #6 = { noinline noreturn nounwind ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #7 = { mustprogress noinline norecurse optnone ssp uwtable(sync) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="apple-m1" "target-features"="+aes,+complxnum,+crc,+dotprod,+fp-armv8,+fp16fml,+fullfp16,+jsconv,+lse,+neon,+pauth,+ras,+rcpc,+rdm,+sha2,+sha3,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+v8a,+zcm,+zcz" }
attributes #8 = { noreturn nounwind }
!llvm.module.flags = !{!0, !1, !2, !3, !4}
!llvm.ident = !{!5}
!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 7, !"openmp", i32 51}
!2 = !{i32 8, !"PIC Level", i32 2}
!3 = !{i32 7, !"uwtable", i32 1}
!4 = !{i32 7, !"frame-pointer", i32 1}
!5 = !{!"Homebrew clang version 18.1.4"}
!6 = !{!7}
!7 = !{i64 2, i64 -1, i64 -1, i1 true}
No use of any new llvm openmp routines and thus we can translate it as shown:
#include <stdio.h>
#include <omp.h>
#include <algorithm>
#include "kmp.h"
void initialize_array( int n, float *a, float val ) {
int i;
#pragma omp parallel
{
#pragma omp for
for (i = 0; i < n; i++) {
a[i] = val;
}
}
return;
}
void parallel_Sum_created( kmp_int32 *global_tid, kmp_int32 *bound_tid, float *total_Sum, int *n, float *a ) {
// @1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 514; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
// @2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 66, i32 0, i32 22, ptr @0 }, align 8
ident_t loc2; loc2.reserved_1 = 0; loc2.flags = 66; loc2.reserved_2 = 0; loc2.reserved_3 = 22; loc2.psource = ";unknown;unknown;0;0;;\00";
float partial_sum = 0;
int *lastiter = new int(0);
int *lower = new int(0);
int *upper = new int(*n);
int *stride = new int(1);
__kmpc_for_static_init_4(&loc, *global_tid, 34, lastiter, lower, upper, stride, 1, 1);
*lower = std::max(0, *lower - 1);
while ( *lower < *upper ) {
partial_sum += a[*lower];
*lower += 1;
}
__kmpc_for_static_fini(&loc, *global_tid);
__kmpc_barrier(&loc2, *global_tid);
*total_Sum += partial_sum;
printf("Thread %d: partial sum: %f total sum: %f\n", omp_get_thread_num(), partial_sum, *total_Sum);
return;
}
void parallel_Sum( int n, float *a ) {
float total_Sum = 0;
// @3 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 2; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
__kmpc_fork_call(&loc, 3, (kmpc_micro)parallel_Sum_created, &total_Sum, &n, a);
printf("Total sum: %f\n", total_Sum);
}
int main() {
float a[1000000];
initialize_array(1000000, a, 1.0);
parallel_Sum(1000000, a);
}
% time clang++ e-transformed.cpp -fopenmp && time ./a.out
ld: warning: reexported library with install name '/opt/homebrew/opt/llvm/lib/libunwind.1.dylib' found at '/opt/homebrew/Cellar/llvm/18.1.4/lib/libunwind.1.0.dylib' couldn't be matched with any parent library and will be linked directly
clang++ e-transformed.cpp -fopenmp 0.18s user 0.04s system 105% cpu 0.206 total
Thread 0: partial sum: 125000.000000 total sum: 375000.000000
Thread 7: partial sum: 125000.000000 total sum: 500000.000000
Thread 2: partial sum: 125000.000000 total sum: 375000.000000
Thread 3: partial sum: 125000.000000 total sum: 375000.000000
Thread 6: partial sum: 125000.000000 total sum: 750000.000000
Thread 5: partial sum: 125000.000000 total sum: 875000.000000
Thread 1: partial sum: 125000.000000 total sum: 1000000.000000
Thread 4: partial sum: 125000.000000 total sum: 750000.000000
Total sum: 1000000.000000
./a.out 0.01s user 0.00s system 258% cpu 0.003 total
We need to fully understand how is #pragma critical being handled in llvm openmp runtime routine calls.
That’s it for this blog, we’ll continue to push and get LFortran support openmp code :)
race free reduction
Umm, no, not yet. Upon executing reduction code multiple times, I got this:
% time clang++ e.cpp -fopenmp && time ./a.out
clang++ e.cpp -fopenmp 0.08s user 0.03s system 117% cpu 0.091 total
Total sum: 875000.000000
./a.out 0.01s user 0.00s system 266% cpu 0.003 total
Something is wrong, and then we get to know that reduction code is not free from race conditions and thus needs to be rewritten as:
e-reduction.cpp
#include <stdio.h>
#include <omp.h>
void initialize_array( int n, float *a, float val ) {
int i;
#pragma omp parallel
{
#pragma omp for
for (i = 0; i < n; i++) {
a[i] = val;
}
}
}
void parallel_Sum( int n, float *a ) {
int i;
float total_sum;
#pragma omp parallel for reduction(+:total_sum)
for (i = 0; i < n; i++) {
total_sum += a[i];
}
printf("Total sum: %f\n", total_sum);
}
int main() {
float a[1000000];
initialize_array(1000000, a, 1.0);
parallel_Sum(1000000, a);
return 0;
}
Which can then on understanding llvm transformed to:
#include <stdio.h>
#include <omp.h>
#include <algorithm>
#include "kmp.h"
void initialize_array( int n, float *a, float val ) {
int i;
#pragma omp parallel
{
#pragma omp for
for (i = 0; i < n; i++) {
a[i] = val;
}
}
return;
}
void parallel_Sum_reduction( float *lhs_data, float *rhs_data ) {
*lhs_data += *rhs_data;
return;
}
void parallel_Sum_created( kmp_int32 *global_tid, kmp_int32 *bound_tid, float *total_Sum, int *n, float *a ) {
// @1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 514, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 514; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
// @2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 66, i32 0, i32 22, ptr @0 }, align 8
ident_t loc2; loc2.reserved_1 = 0; loc2.flags = 66; loc2.reserved_2 = 0; loc2.reserved_3 = 22; loc2.psource = ";unknown;unknown;0;0;;\00";
float partial_sum = 0;
int *lastiter = new int(0);
int *lower = new int(0);
int *upper = new int(*n);
int *stride = new int(1);
__kmpc_for_static_init_4(&loc, *global_tid, 34, lastiter, lower, upper, stride, 1, 1);
*lower = std::max(0, *lower - 1);
while ( *lower < *upper ) {
partial_sum += a[*lower];
*lower += 1;
}
__kmpc_for_static_fini(&loc, *global_tid);
// @4 = private unnamed_addr constant %struct.ident_t { i32 0, i32 18, i32 0, i32 22, ptr @0 }, align 8
ident_t loc3; loc3.reserved_1 = 0; loc3.flags = 18; loc3.reserved_2 = 0; loc3.reserved_3 = 22; loc3.psource = ";unknown;unknown;0;0;;\00";
// %73 = call i32 @__kmpc_reduce_nowait(ptr @4, i32 %72, i32 1, i64 8, ptr %22, ptr @_Z12parallel_SumiPf.omp_outlined.omp.reduction.reduction_func, ptr @.gomp_critical_user_.reduction.var)
kmp_critical_name *lck;
kmp_int32 res = __kmpc_reduce_nowait(&loc3, *global_tid, 1, 8, &partial_sum, (void (*)(void *, void *))parallel_Sum_reduction, lck);
if ( res == 1 ) {
// @result 1 for the primary thread, 0 for all other team threads, 2 for all team
// threads if atomic reduction needed
*total_Sum += partial_sum;
__kmpc_end_reduce_nowait(&loc3, *global_tid, lck);
} else if ( res == 2 ) {
// %80 = atomicrmw fadd ptr %24, float %79 monotonic, align 4
__kmpc_atomic_float4_add(&loc2, *global_tid, total_Sum, partial_sum);
}
return;
}
void parallel_Sum( int n, float *a ) {
float total_Sum = 0;
// @3 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
ident_t loc; loc.reserved_1 = 0; loc.flags = 2; loc.reserved_2 = 0; loc.reserved_3 = 22; loc.psource = ";unknown;unknown;0;0;;\00";
__kmpc_fork_call(&loc, 3, (kmpc_micro)parallel_Sum_created, &total_Sum, &n, a);
printf("Total sum: %f\n", total_Sum);
}
int main() {
float a[1000000];
initialize_array(1000000, a, 1.0);
parallel_Sum(1000000, a);
}
Here you can notice __kmpc_reduce_nowait, __kmpc_end_reduce_nowait and __kmpc_atomic_float4_add are used, we ported corresponding functions in kmp.h.
kmp.h
typedef float kmp_real32;
typedef kmp_int32 kmp_critical_name[8];
....
/* 2.a.i. Reduce Block without a terminating barrier */
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid global thread number
@param num_vars number of items (variables) to be reduced
@param reduce_size size of data in bytes to be reduced
@param reduce_data pointer to data to be reduced
@param reduce_func callback function providing reduction operation on two
operands and returning result of reduction in lhs_data
@param lck pointer to the unique lock data structure
@result 1 for the primary thread, 0 for all other team threads, 2 for all team
threads if atomic reduction needed
The nowait version is used for a reduce clause with the nowait argument.
*/
kmp_int32
__kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size, void *reduce_data,
void (*reduce_func)(void *lhs_data, void *rhs_data),
kmp_critical_name *lck);
/*!
@ingroup SYNCHRONIZATION
@param loc source location information
@param global_tid global thread id.
@param lck pointer to the unique lock data structure
Finish the execution of a reduce nowait.
*/
void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
kmp_critical_name *lck);
// 4-byte add float
void __kmpc_atomic_float4_add(ident_t *id_ref, int gtid, kmp_real32 *lhs,
kmp_real32 rhs);
And this can be executed using:
% time clang++ e-transformed-reduction.cpp -fopenmp && time ./a.out
clang++ e-transformed-reduction.cpp -fopenmp 0.18s user 0.04s system 107% cpu 0.207 total
Total sum: 1000000.000000
./a.out 0.01s user 0.00s system 227% cpu 0.004 total
Now, I think this blog can be wrapped, thanks for following!