stream_add_omp_st.c 3.97 KB
Newer Older
Swann Perarnau's avatar
Swann Perarnau committed
1 2 3 4 5 6 7 8 9 10
/*******************************************************************************
 * Copyright 2019 UChicago Argonne, LLC.
 * (c.f. AUTHORS, LICENSE)
 *
 * This file is part of the AML project.
 * For more info, see https://xgitlab.cels.anl.gov/argo/aml
 *
 * SPDX-License-Identifier: BSD-3-Clause
*******************************************************************************/

11
#include "aml.h"
Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
12
#include "aml/area/linux.h"
13
#include "aml/dma/linux-seq.h"
14
#include "aml/scratch/par.h"
15 16 17 18 19 20
#include <assert.h>
#include <errno.h>
#include <omp.h>
#include <pthread.h>
#include <stdlib.h>

21 22
#include "utils.h"

23 24 25
#define ITER 10
#define CHUNKING 4

26 27 28
size_t numthreads, tilesz, esz;
unsigned long *a, *b, *c;
AML_TILING_1D_DECL(tiling);
Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
29
struct aml_area *slow, *fast;
30 31
AML_SCRATCH_PAR_DECL(sa);
AML_SCRATCH_PAR_DECL(sb);
32 33 34 35

int kernel(unsigned long *a, unsigned long *b, unsigned long *c, size_t n)
{
	size_t i;
36
	debug("%p = %p + %p [%zi]\n",c,a,b,n);
37 38 39 40 41
	for(i = 0; i < n; i++)
		c[i] = a[i] + b[i];
	return 0;
}

42
void do_work(unsigned long tid)
43
{
44 45

	int offset, i, ai, bi, oldai, oldbi;
46
	unsigned long *ap, *bp, *cp;
47
	void *abaseptr, *bbaseptr;
48
	offset = tid*CHUNKING;
49 50
	ap = aml_tiling_tilestart(&tiling, a, offset);
	bp = aml_tiling_tilestart(&tiling, b, offset);
51
	cp = aml_tiling_tilestart(&tiling, c, offset);
52 53 54 55 56 57 58 59
	abaseptr = aml_scratch_baseptr(&sa);
	bbaseptr = aml_scratch_baseptr(&sb);
	ai = -1; bi = -1;
	for(i = 0; i < CHUNKING-1; i++) {
		struct aml_scratch_request *ar, *br;
		oldai = ai; oldbi = bi;
		aml_scratch_async_pull(&sa, &ar, abaseptr, &ai, a, offset+i+1);
		aml_scratch_async_pull(&sb, &br, bbaseptr, &bi, b, offset+i+1);
60
		kernel(ap, bp, cp, esz);
61 62 63 64
		aml_scratch_wait(&sa, ar);
		aml_scratch_wait(&sb, br);
		ap = aml_tiling_tilestart(&tiling, abaseptr, ai);
		bp = aml_tiling_tilestart(&tiling, bbaseptr, bi);
65
		cp = aml_tiling_tilestart(&tiling, c, offset+i+1);
66 67 68
		aml_scratch_release(&sa, oldai);
		aml_scratch_release(&sb, oldbi);
	}
69
	kernel(ap, bp, cp, esz);
70
}
71

72 73
int main(int argc, char *argv[])
{
74
	AML_DMA_LINUX_SEQ_DECL(dma);
Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
75
	struct aml_bitmap slowb, fastb;
76
	aml_init(&argc, &argv);
77 78 79
	assert(argc == 4);

	log_init(argv[0]);
Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
80 81
	assert(aml_bitmap_from_string(&fastb, argv[1]) == 0);
	assert(aml_bitmap_from_string(&slowb, argv[2]) == 0);
82
	unsigned long memsize = 1UL << atoi(argv[3]);
83 84 85 86 87 88 89

	/* use openmp env to figure out how many threads we want
	 * (we actually use 3x as much)
	 */
	#pragma omp parallel
	{
		numthreads = omp_get_num_threads();
90
		tilesz = memsize/(numthreads*CHUNKING);
91
		esz = tilesz/sizeof(unsigned long);
92
	}
93 94

	/* initialize all the supporting struct */
95
	assert(!aml_tiling_init(&tiling, AML_TILING_TYPE_1D, tilesz, memsize));
96

Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
97 98 99 100 101 102
	slow = aml_area_linux_create(AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
				     &slowb, AML_AREA_LINUX_BINDING_FLAG_BIND);
	assert(slow != NULL);
	fast = aml_area_linux_create(AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
				     &fastb, AML_AREA_LINUX_BINDING_FLAG_BIND);
	assert(fast != NULL);
103 104 105 106 107 108 109
	assert(!aml_dma_linux_seq_init(&dma, numthreads*2));
	assert(!aml_scratch_par_init(&sa, &fast, &slow, &dma, &tiling,
				     2*numthreads, numthreads));
	assert(!aml_scratch_par_init(&sb, &fast, &slow, &dma, &tiling,
				     2*numthreads, numthreads));

	/* allocation */
Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
110 111 112
	a = aml_area_mmap(slow, NULL, memsize);
	b = aml_area_mmap(slow, NULL, memsize);
	c = aml_area_mmap(fast, NULL, memsize);
113
	assert(a != NULL && b != NULL && c != NULL);
114

115
	unsigned long esize = memsize/sizeof(unsigned long);
116
	for(unsigned long i = 0; i < esize; i++) {
117 118 119
		a[i] = i;
		b[i] = esize - i;
		c[i] = 0;
120 121 122
	}

	/* run kernel */
123
	#pragma omp parallel for
124
	for(unsigned long i = 0; i < numthreads; i++) {
125
		do_work(i);
126 127 128 129
	}

	/* validate */
	for(unsigned long i = 0; i < esize; i++) {
130
		assert(c[i] == esize);
131 132
	}

133 134 135
	aml_scratch_par_destroy(&sa);
	aml_scratch_par_destroy(&sb);
	aml_dma_linux_seq_destroy(&dma);
Nicolas Denoyelle's avatar
Nicolas Denoyelle committed
136 137 138 139 140
	aml_area_munmap(slow, a, memsize);
	aml_area_munmap(slow, b, memsize);
	aml_area_munmap(fast, c, memsize);
	aml_area_linux_destroy(slow);
	aml_area_linux_destroy(fast);
141
	aml_tiling_destroy(&tiling, AML_TILING_TYPE_1D);
142 143 144
	aml_finalize();
	return 0;
}