stream_add_pth_st.c 4.15 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12
#include <aml.h>
#include <assert.h>
#include <errno.h>
#include <omp.h>
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>

#define ITER 10
#define MEMSIZE (1UL<<20)
#define CHUNKING 4

13 14 15 16 17 18 19
size_t numthreads, tilesz, esz;
unsigned long *a, *b, *c;
AML_TILING_1D_DECL(tiling);
AML_AREA_LINUX_DECL(slow);
AML_AREA_LINUX_DECL(fast);
AML_SCRATCH_PAR_DECL(sa);
AML_SCRATCH_PAR_DECL(sb);
20 21 22 23 24 25 26 27 28 29 30

int kernel(unsigned long *a, unsigned long *b, unsigned long *c, size_t n)
{
	size_t i;
	printf("%p = %p + %p [%zi]\n",c,a,b,n);
	for(i = 0; i < n; i++)
		c[i] = a[i] + b[i];
	return 0;
}

struct winfo {
31 32
	int tid;
	pthread_t th;
33 34 35 36
};

void *th_work(void *arg)
{
37 38

	int offset, i, ai, bi, oldai, oldbi;
39
	unsigned long *ap, *bp, *cp;
40
	void *abaseptr, *bbaseptr;
41
	struct winfo *wi = arg;
42 43 44
	offset = wi->tid*CHUNKING;
	ap = aml_tiling_tilestart(&tiling, a, offset);
	bp = aml_tiling_tilestart(&tiling, b, offset);
45
	cp = aml_tiling_tilestart(&tiling, c, offset);
46 47 48 49 50 51 52 53
	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);
54
		kernel(ap, bp, cp, esz);
55 56 57 58
		aml_scratch_wait(&sa, ar);
		aml_scratch_wait(&sb, br);
		ap = aml_tiling_tilestart(&tiling, abaseptr, ai);
		bp = aml_tiling_tilestart(&tiling, bbaseptr, bi);
59
		cp = aml_tiling_tilestart(&tiling, c, offset+i+1);
60 61 62
		aml_scratch_release(&sa, oldai);
		aml_scratch_release(&sb, oldbi);
	}
63
	kernel(ap, bp, cp, esz);
64 65 66 67 68

	return arg;
}
int main(int argc, char *argv[])
{
69 70 71 72
	AML_BINDING_SINGLE_DECL(binding);
	AML_ARENA_JEMALLOC_DECL(arena);
	AML_DMA_LINUX_SEQ_DECL(dma);
	unsigned long nodemask[AML_NODEMASK_SZ];
73
	aml_init(&argc, &argv);
74
	assert(argc == 1);
75 76 77 78 79 80 81

	/* 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();
82 83
		tilesz = MEMSIZE/(numthreads*CHUNKING);
		esz = tilesz/sizeof(unsigned long);
84
	}
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102

	/* initialize all the supporting struct */
	assert(!aml_binding_init(&binding, AML_BINDING_TYPE_SINGLE, 0));
	assert(!aml_tiling_init(&tiling, AML_TILING_TYPE_1D, tilesz, MEMSIZE));
	AML_NODEMASK_ZERO(nodemask);
	AML_NODEMASK_SET(nodemask, 0);
	assert(!aml_arena_jemalloc_init(&arena, AML_ARENA_JEMALLOC_TYPE_REGULAR));

	assert(!aml_area_linux_init(&slow,
				    AML_AREA_LINUX_MANAGER_TYPE_SINGLE,
				    AML_AREA_LINUX_MBIND_TYPE_REGULAR,
				    AML_AREA_LINUX_MMAP_TYPE_ANONYMOUS,
				    &arena, MPOL_BIND, nodemask));
	assert(!aml_area_linux_init(&fast,
				    AML_AREA_LINUX_MANAGER_TYPE_SINGLE,
				    AML_AREA_LINUX_MBIND_TYPE_REGULAR,
				    AML_AREA_LINUX_MMAP_TYPE_ANONYMOUS,
				    &arena, MPOL_BIND, nodemask));
103
	assert(!aml_dma_linux_seq_init(&dma, (size_t)numthreads*4));
104
	assert(!aml_scratch_par_init(&sa, &fast, &slow, &dma, &tiling,
105
				     (size_t)2*numthreads, (size_t)numthreads));
106
	assert(!aml_scratch_par_init(&sb, &fast, &slow, &dma, &tiling,
107
				     (size_t)2*numthreads, (size_t)numthreads));
108 109

	/* allocation */
110 111 112 113
	a = aml_area_malloc(&slow, MEMSIZE);
	b = aml_area_malloc(&slow, MEMSIZE);
	c = aml_area_malloc(&fast, MEMSIZE);
	assert(a != NULL && b != NULL && c != NULL);
114 115 116

	unsigned long esize = MEMSIZE/sizeof(unsigned long);
	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 */
123 124 125 126 127 128 129
	struct winfo *wis = aml_area_calloc(&slow, numthreads, sizeof(struct winfo));
	for(unsigned long i = 0; i < numthreads; i++) {
		wis[i].tid = i;
		pthread_create(&wis[i].th, NULL, &th_work, (void*)&wis[i]);
	}
	for(unsigned long j = 0; j < numthreads; j++) {
		pthread_join(wis[j].th, NULL);
130
	}
131
	aml_area_free(&slow, wis);
132 133 134

	/* validate */
	for(unsigned long i = 0; i < esize; i++) {
135
		assert(c[i] == esize);
136 137
	}

138 139 140
	aml_scratch_par_destroy(&sa);
	aml_scratch_par_destroy(&sb);
	aml_dma_linux_seq_destroy(&dma);
141 142
	aml_area_free(&slow, a);
	aml_area_free(&slow, b);
143 144 145 146 147
	aml_area_free(&slow, c);
	aml_area_linux_destroy(&slow);
	aml_area_linux_destroy(&fast);
	aml_tiling_destroy(&tiling, AML_TILING_TYPE_1D);
	aml_binding_destroy(&binding, AML_BINDING_TYPE_SINGLE);
148 149 150
	aml_finalize();
	return 0;
}