Skip to content

A
aml
Commit d3558a6d authored by Swann Perarnau's avatar Swann Perarnau
Browse files
Options

[refactor] remove decl/init/fini from library 

Removes static allocations of all objects from the library.
While this will make our code less compact in theory,
this part of the API is too confusing for new users.
parent c9d8ed8c
Hide whitespace changes
Inline Side-by-side
Showing with 613 additions and 1123 deletions
benchmarks/dgemm_noprefetch.c
View file @ d3558a6d
......@@ -21,8 +21,8 @@
#include <math.h>
#include <stdlib.h>
AML_TILING_2D_ROWMAJOR_DECL(tiling_row);
AML_TILING_2D_COLMAJOR_DECL(tiling_col);
struct aml_tiling *tiling_row;
struct aml_tiling *tiling_col;
struct aml_area *slow, *fast;
size_t memsize, tilesize, N, T;
double *a, *b, *c;
......@@ -34,7 +34,7 @@ void do_work()
ldb = lda;
ldc = lda;
size_t ndims[2];
aml_tiling_ndims(&tiling_row, &ndims[0], &ndims[1]);
aml_tiling_ndims(tiling_row, &ndims[0], &ndims[1]);
for(int k = 0; k < ndims[1]; k++)
{
......@@ -45,12 +45,12 @@ void do_work()
{
size_t aoff, boff, coff;
double *ap, *bp, *cp;
aoff = aml_tiling_tileid(&tiling_col, i, k);
boff = aml_tiling_tileid(&tiling_row, k, j);
coff = aml_tiling_tileid(&tiling_row, i, j);
ap = aml_tiling_tilestart(&tiling_col, a, aoff);
bp = aml_tiling_tilestart(&tiling_row, b, boff);
cp = aml_tiling_tilestart(&tiling_row, c, coff);
aoff = aml_tiling_tileid(tiling_col, i, k);
boff = aml_tiling_tileid(tiling_row, k, j);
coff = aml_tiling_tileid(tiling_row, i, j);
ap = aml_tiling_tilestart(tiling_col, a, aoff);
bp = aml_tiling_tilestart(tiling_row, b, boff);
cp = aml_tiling_tilestart(tiling_row, c, coff);
cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, ldc, lda, ldb, 1.0, ap, lda, bp, ldb, 1.0, cp, ldc);
}
}
......@@ -72,10 +72,10 @@ int main(int argc, char* argv[])
tilesize = sizeof(double)*T*T;
/* the initial tiling, of 2D square tiles */
assert(!aml_tiling_2d_init(&tiling_row, AML_TILING_TYPE_2D_ROWMAJOR,
tilesize, memsize, N/T , N/T));
assert(!aml_tiling_2d_init(&tiling_col, AML_TILING_TYPE_2D_COLMAJOR,
tilesize, memsize, N/T , N/T));
assert(!aml_tiling_2d_create(&tiling_row, AML_TILING_TYPE_2D_ROWMAJOR,
tilesize, memsize, N/T , N/T));
assert(!aml_tiling_2d_create(&tiling_col, AML_TILING_TYPE_2D_COLMAJOR,
tilesize, memsize, N/T , N/T));
aml_area_linux_create(&slow, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&slowb, AML_AREA_LINUX_BINDING_FLAG_BIND);
......@@ -156,8 +156,8 @@ int main(int argc, char* argv[])
aml_area_munmap(fast, c, memsize);
aml_area_linux_destroy(&slow);
aml_area_linux_destroy(&fast);
aml_tiling_2d_fini(&tiling_row);
aml_tiling_2d_fini(&tiling_col);
aml_tiling_2d_destroy(&tiling_row);
aml_tiling_2d_destroy(&tiling_col);
aml_finalize();
return 0;
}
benchmarks/dgemm_prefetch.c
View file @ d3558a6d
......@@ -24,11 +24,11 @@
#include <math.h>
#include <stdlib.h>
AML_TILING_2D_ROWMAJOR_DECL(tiling_row);
AML_TILING_2D_COLMAJOR_DECL(tiling_col);
AML_TILING_1D_DECL(tiling_prefetch);
AML_SCRATCH_PAR_DECL(sa);
AML_SCRATCH_PAR_DECL(sb);
struct aml_tiling *tiling_row;
struct aml_tiling *tiling_col;
struct aml_tiling *tiling_prefetch;
struct aml_scratch *sa;
struct aml_scratch *sb;
struct aml_area *slow, *fast;
size_t memsize, tilesize, N, T;
......@@ -45,19 +45,19 @@ void do_work()
int ai, bi, oldai, oldbi;
void *abaseptr, *bbaseptr;
struct aml_scratch_request *ar, *br;
aml_tiling_ndims(&tiling_row, &ndims[0], &ndims[1]);
abaseptr = aml_scratch_baseptr(&sa);
bbaseptr = aml_scratch_baseptr(&sb);
prea = aml_tiling_tilestart(&tiling_prefetch, a, 0);
preb = aml_tiling_tilestart(&tiling_prefetch, b, 0);
aml_tiling_ndims(tiling_row, &ndims[0], &ndims[1]);
abaseptr = aml_scratch_baseptr(sa);
bbaseptr = aml_scratch_baseptr(sb);
prea = aml_tiling_tilestart(tiling_prefetch, a, 0);
preb = aml_tiling_tilestart(tiling_prefetch, b, 0);
ai = -1; bi = -1;
for(int k = 0; k < ndims[1]; k++)
{
oldbi = bi;
oldai = ai;
aml_scratch_async_pull(&sa, &ar, abaseptr, &ai, a, k + 1);
aml_scratch_async_pull(&sb, &br, bbaseptr, &bi, b, k + 1);
aml_scratch_async_pull(sa, &ar, abaseptr, &ai, a, k + 1);
aml_scratch_async_pull(sb, &br, bbaseptr, &bi, b, k + 1);
#pragma omp parallel for
for(int i = 0; i < ndims[0]; i++)
{
......@@ -65,25 +65,25 @@ void do_work()
{
size_t coff;
double *ap, *bp, *cp;
ap = aml_tiling_tilestart(&tiling_row, prea, i);
bp = aml_tiling_tilestart(&tiling_row, preb, j);
coff = aml_tiling_tileid(&tiling_row, i, j);
cp = aml_tiling_tilestart(&tiling_row, c, coff);
ap = aml_tiling_tilestart(tiling_row, prea, i);
bp = aml_tiling_tilestart(tiling_row, preb, j);
coff = aml_tiling_tileid(tiling_row, i, j);
cp = aml_tiling_tilestart(tiling_row, c, coff);
cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, ldc, lda, ldb, 1.0, ap, lda, bp, ldb, 1.0, cp, ldc);
}
}
aml_scratch_wait(&sa, ar);
aml_scratch_wait(&sb, br);
prea = aml_tiling_tilestart(&tiling_prefetch, abaseptr, ai);
preb = aml_tiling_tilestart(&tiling_prefetch, bbaseptr, bi);
aml_scratch_release(&sa, oldai);
aml_scratch_release(&sb, oldbi);
aml_scratch_wait(sa, ar);
aml_scratch_wait(sb, br);
prea = aml_tiling_tilestart(tiling_prefetch, abaseptr, ai);
preb = aml_tiling_tilestart(tiling_prefetch, bbaseptr, bi);
aml_scratch_release(sa, oldai);
aml_scratch_release(sb, oldbi);
}
}
int main(int argc, char* argv[])
{
AML_DMA_LINUX_SEQ_DECL(dma);
struct aml_dma *dma;
struct aml_bitmap slowb, fastb;
aml_init(&argc, &argv);
assert(argc == 5);
......@@ -97,13 +97,13 @@ int main(int argc, char* argv[])
tilesize = sizeof(double)*T*T;
/* the initial tiling, 2d grid of tiles */
assert(!aml_tiling_2d_init(&tiling_row, AML_TILING_TYPE_2D_ROWMAJOR,
tilesize, memsize, N/T , N/T));
assert(!aml_tiling_2d_init(&tiling_col, AML_TILING_TYPE_2D_COLMAJOR,
tilesize, memsize, N/T , N/T));
assert(!aml_tiling_2d_create(&tiling_row, AML_TILING_TYPE_2D_ROWMAJOR,
tilesize, memsize, N/T , N/T));
assert(!aml_tiling_2d_create(&tiling_col, AML_TILING_TYPE_2D_COLMAJOR,
tilesize, memsize, N/T , N/T));
/* the prefetch tiling, 1D sequence of columns of tiles */
assert(!aml_tiling_1d_init(&tiling_prefetch,
tilesize*(N/T), memsize));
assert(!aml_tiling_1d_create(&tiling_prefetch,
tilesize*(N/T), memsize));
aml_area_linux_create(&slow, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&slowb, AML_AREA_LINUX_BINDING_FLAG_BIND);
......@@ -112,9 +112,9 @@ int main(int argc, char* argv[])
&fastb, AML_AREA_LINUX_BINDING_FLAG_BIND);
assert(fast != NULL);
assert(!aml_dma_linux_seq_init(&dma, 2));
assert(!aml_scratch_par_init(&sa, &fast, &slow, &dma, &tiling_prefetch, (size_t)2, (size_t)2));
assert(!aml_scratch_par_init(&sb, &fast, &slow, &dma, &tiling_prefetch, (size_t)2, (size_t)2));
assert(!aml_dma_linux_seq_create(&dma, 2));
assert(!aml_scratch_par_create(&sa, fast, slow, dma, tiling_prefetch, (size_t)2, (size_t)2));
assert(!aml_scratch_par_create(&sb, fast, slow, dma, tiling_prefetch, (size_t)2, (size_t)2));
/* allocation */
a = aml_area_mmap(slow, NULL, memsize);
b = aml_area_mmap(slow, NULL, memsize);
......@@ -182,17 +182,17 @@ int main(int argc, char* argv[])
/* print the flops in GFLOPS */
printf("dgemm-prefetch: %llu %lld %lld %f\n", N, memsize, time,
flops/1e9);
aml_scratch_par_fini(&sa);
aml_scratch_par_fini(&sb);
aml_dma_linux_seq_fini(&dma);
aml_scratch_par_destroy(&sa);
aml_scratch_par_destroy(&sb);
aml_dma_linux_seq_destroy(&dma);
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);
aml_tiling_2d_fini(&tiling_row);
aml_tiling_2d_fini(&tiling_col);
aml_tiling_1d_fini(&tiling_prefetch);
aml_tiling_2d_destroy(&tiling_row);
aml_tiling_2d_destroy(&tiling_col);
aml_tiling_1d_destroy(&tiling_prefetch);
aml_finalize();
return 0;
}
benchmarks/stream_add_omp.c
View file @ d3558a6d
......@@ -38,8 +38,8 @@ int main(int argc, char *argv[])
*/
struct aml_area *slow = &aml_area_linux, *fast = aml_area_linux;
struct aml_dma dma;
assert(!aml_dma_init(&dma, 0));
struct aml_dma *dma;
assert(!aml_dma_create(&dma, 0));
void *a, *b, *c;
......@@ -75,19 +75,19 @@ int main(int argc, char *argv[])
{
for(unsigned long i = 0; i < numthreads*CHUNKING; i++) {
#pragma omp task depend(inout: wa[i*esz:esz])
assert(!aml_dma_move(&dma, &fast, &slow, &wa[i*esz], esz));
assert(!aml_dma_move(dma, &fast, &slow, &wa[i*esz], esz));
#pragma omp task depend(inout: wb[i*esz:esz])
assert(!aml_dma_move(&dma, &fast, &slow, &wb[i*esz], esz));
assert(!aml_dma_move(dma, &fast, &slow, &wb[i*esz], esz));
#pragma omp task depend(inout: wc[i*esz:esz])
assert(!aml_dma_move(&dma, &fast, &slow, &wc[i*esz], esz));
assert(!aml_dma_move(dma, &fast, &slow, &wc[i*esz], esz));
#pragma omp task depend(in: wa[i*esz:esz], wb[i*esz:esz]) depend(out: wc[i*esz:esz])
kernel(&wa[i*esz], &wb[i*esz], &wc[i*esz], esz);
#pragma omp task depend(inout: wa[i*esz:esz])
assert(!aml_dma_move(&dma, &slow, &fast, &wa[i*esz], esz));
assert(!aml_dma_move(dma, &slow, &fast, &wa[i*esz], esz));
#pragma omp task depend(inout: wb[i*esz:esz])
assert(!aml_dma_move(&dma, &slow, &fast, &wb[i*esz], esz));
assert(!aml_dma_move(dma, &slow, &fast, &wb[i*esz], esz));
#pragma omp task depend(inout: wc[i*esz:esz])
assert(!aml_dma_move(&dma, &slow, &fast, &wc[i*esz], esz));
assert(!aml_dma_move(dma, &slow, &fast, &wc[i*esz], esz));
}
}
......
benchmarks/stream_add_omp_mt.c
View file @ d3558a6d
......@@ -27,10 +27,10 @@
size_t numthreads, tilesz, esz;
unsigned long *a, *b, *c;
AML_TILING_1D_DECL(tiling);
struct aml_tiling *tiling;
struct aml_area *slow, *fast;
AML_SCRATCH_SEQ_DECL(sa);
AML_SCRATCH_SEQ_DECL(sb);
struct aml_scratch *sa;
struct aml_scratch *sb;
int kernel(unsigned long *a, unsigned long *b, unsigned long *c, size_t n)
{
......@@ -42,7 +42,7 @@ int kernel(unsigned long *a, unsigned long *b, unsigned long *c, size_t n)
int main(int argc, char *argv[])
{
AML_DMA_LINUX_PAR_DECL(dma);
struct aml_dma *dma;
struct aml_bitmap slowb, fastb;
aml_init(&argc, &argv);
assert(argc == 4);
......@@ -63,17 +63,17 @@ int main(int argc, char *argv[])
}
/* initialize all the supporting struct */
assert(!aml_tiling_1d_init(&tiling, tilesz, memsize));
assert(!aml_tiling_1d_create(&tiling, tilesz, memsize));
aml_area_linux_create(&slow, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&slowb, AML_AREA_LINUX_BINDING_FLAG_BIND);
assert(slow != NULL);
aml_area_linux_create(&fast, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&fastb, AML_AREA_LINUX_BINDING_FLAG_BIND);
assert(fast != NULL);
assert(!aml_dma_linux_par_init(&dma, numthreads*2, numthreads));
assert(!aml_scratch_seq_init(&sa, &fast, &slow, &dma, &tiling,
assert(!aml_dma_linux_par_create(&dma, numthreads*2, numthreads));
assert(!aml_scratch_seq_create(&sa, fast, slow, dma, tiling,
(size_t)2*numthreads, (size_t)1));
assert(!aml_scratch_seq_init(&sb, &fast, &slow, &dma, &tiling,
assert(!aml_scratch_seq_create(&sb, fast, slow, dma, tiling,
(size_t)2*numthreads, (size_t)1));
/* allocation */
......@@ -93,23 +93,23 @@ int main(int argc, char *argv[])
int i, ai, bi, oldai, oldbi;
unsigned long *ap, *bp;
void *abaseptr, *bbaseptr;
ap = aml_tiling_tilestart(&tiling, a, 0);
bp = aml_tiling_tilestart(&tiling, b, 0);
abaseptr = aml_scratch_baseptr(&sa);
bbaseptr = aml_scratch_baseptr(&sb);
ap = aml_tiling_tilestart(tiling, a, 0);
bp = aml_tiling_tilestart(tiling, b, 0);
abaseptr = aml_scratch_baseptr(sa);
bbaseptr = aml_scratch_baseptr(sb);
ai = -1; bi = -1;
for(i = 0; i < (memsize/tilesz) -1; i++) {
struct aml_scratch_request *ar, *br;
oldai = ai; oldbi = bi;
aml_scratch_async_pull(&sa, &ar, abaseptr, &ai, a, i+1);
aml_scratch_async_pull(&sb, &br, bbaseptr, &bi, b, i+1);
aml_scratch_async_pull(sa, &ar, abaseptr, &ai, a, i+1);
aml_scratch_async_pull(sb, &br, bbaseptr, &bi, b, i+1);
kernel(ap, bp, &c[i*esz], esz);
aml_scratch_wait(&sa, ar);
aml_scratch_wait(&sb, br);
ap = aml_tiling_tilestart(&tiling, abaseptr, ai);
bp = aml_tiling_tilestart(&tiling, bbaseptr, bi);
aml_scratch_release(&sa, oldai);
aml_scratch_release(&sb, oldbi);
aml_scratch_wait(sa, ar);
aml_scratch_wait(sb, br);
ap = aml_tiling_tilestart(tiling, abaseptr, ai);
bp = aml_tiling_tilestart(tiling, bbaseptr, bi);
aml_scratch_release(sa, oldai);
aml_scratch_release(sb, oldbi);
}
kernel(ap, bp, &c[i*esz], esz);
......@@ -118,15 +118,15 @@ int main(int argc, char *argv[])
assert(c[i] == esize);
}
aml_scratch_seq_fini(&sa);
aml_scratch_seq_fini(&sb);
aml_dma_linux_par_fini(&dma);
aml_scratch_seq_destroy(&sa);
aml_scratch_seq_destroy(&sb);
aml_dma_linux_par_destroy(&dma);
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);
aml_tiling_1d_fini(&tiling);
aml_tiling_1d_destroy(&tiling);
aml_finalize();
return 0;
}
benchmarks/stream_add_omp_st.c
View file @ d3558a6d
......@@ -26,10 +26,10 @@
size_t numthreads, tilesz, esz;
unsigned long *a, *b, *c;
AML_TILING_1D_DECL(tiling);
struct aml_tiling *tiling;
struct aml_area *slow, *fast;
AML_SCRATCH_PAR_DECL(sa);
AML_SCRATCH_PAR_DECL(sb);
struct aml_scratch *sa;
struct aml_scratch *sb;
int kernel(unsigned long *a, unsigned long *b, unsigned long *c, size_t n)
{
......@@ -47,32 +47,32 @@ void do_work(unsigned long tid)
unsigned long *ap, *bp, *cp;
void *abaseptr, *bbaseptr;
offset = tid*CHUNKING;
ap = aml_tiling_tilestart(&tiling, a, offset);
bp = aml_tiling_tilestart(&tiling, b, offset);
cp = aml_tiling_tilestart(&tiling, c, offset);
abaseptr = aml_scratch_baseptr(&sa);
bbaseptr = aml_scratch_baseptr(&sb);
ap = aml_tiling_tilestart(tiling, a, offset);
bp = aml_tiling_tilestart(tiling, b, offset);
cp = aml_tiling_tilestart(tiling, c, offset);
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);
aml_scratch_async_pull(sa, &ar, abaseptr, &ai, a, offset+i+1);
aml_scratch_async_pull(sb, &br, bbaseptr, &bi, b, offset+i+1);
kernel(ap, bp, cp, esz);
aml_scratch_wait(&sa, ar);
aml_scratch_wait(&sb, br);
ap = aml_tiling_tilestart(&tiling, abaseptr, ai);
bp = aml_tiling_tilestart(&tiling, bbaseptr, bi);
cp = aml_tiling_tilestart(&tiling, c, offset+i+1);
aml_scratch_release(&sa, oldai);
aml_scratch_release(&sb, oldbi);
aml_scratch_wait(sa, ar);
aml_scratch_wait(sb, br);
ap = aml_tiling_tilestart(tiling, abaseptr, ai);
bp = aml_tiling_tilestart(tiling, bbaseptr, bi);
cp = aml_tiling_tilestart(tiling, c, offset+i+1);
aml_scratch_release(sa, oldai);
aml_scratch_release(sb, oldbi);
}
kernel(ap, bp, cp, esz);
}
int main(int argc, char *argv[])
{
AML_DMA_LINUX_SEQ_DECL(dma);
struct aml_dma *dma;
struct aml_bitmap slowb, fastb;
aml_init(&argc, &argv);
assert(argc == 4);
......@@ -93,7 +93,7 @@ int main(int argc, char *argv[])
}
/* initialize all the supporting struct */
assert(!aml_tiling_1d_init(&tiling, tilesz, memsize));
assert(!aml_tiling_1d_create(&tiling, tilesz, memsize));
aml_area_linux_create(&slow, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&slowb, AML_AREA_LINUX_BINDING_FLAG_BIND);
......@@ -101,10 +101,10 @@ int main(int argc, char *argv[])
aml_area_linux_create(&fast, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&fastb, AML_AREA_LINUX_BINDING_FLAG_BIND);
assert(fast != NULL);
assert(!aml_dma_linux_seq_init(&dma, numthreads*2));
assert(!aml_scratch_par_init(&sa, &fast, &slow, &dma, &tiling,
assert(!aml_dma_linux_seq_create(&dma, numthreads*2));
assert(!aml_scratch_par_create(&sa, fast, slow, dma, tiling,
2*numthreads, numthreads));
assert(!aml_scratch_par_init(&sb, &fast, &slow, &dma, &tiling,
assert(!aml_scratch_par_create(&sb, fast, slow, dma, tiling,
2*numthreads, numthreads));
/* allocation */
......@@ -131,15 +131,15 @@ int main(int argc, char *argv[])
assert(c[i] == esize);
}
aml_scratch_par_fini(&sa);
aml_scratch_par_fini(&sb);
aml_dma_linux_seq_fini(&dma);
aml_scratch_par_destroy(&sa);
aml_scratch_par_destroy(&sb);
aml_dma_linux_seq_destroy(&dma);
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);
aml_tiling_1d_fini(&tiling);
aml_tiling_1d_destroy(&tiling);
aml_finalize();
return 0;
}
benchmarks/stream_add_pth_st.c
View file @ d3558a6d
......@@ -27,9 +27,9 @@
size_t numthreads, tilesz, esz;
unsigned long *a, *b, *c;
struct aml_area *slow, *fast;
AML_TILING_1D_DECL(tiling);
AML_SCRATCH_PAR_DECL(sa);
AML_SCRATCH_PAR_DECL(sb);
struct aml_tiling *tiling;
struct aml_scratch *sa;
struct aml_scratch *sb;
int kernel(unsigned long *a, unsigned long *b, unsigned long *c, size_t n)
{
......@@ -53,25 +53,25 @@ void *th_work(void *arg)
void *abaseptr, *bbaseptr;
struct winfo *wi = arg;
offset = wi->tid*CHUNKING;
ap = aml_tiling_tilestart(&tiling, a, offset);
bp = aml_tiling_tilestart(&tiling, b, offset);
cp = aml_tiling_tilestart(&tiling, c, offset);
abaseptr = aml_scratch_baseptr(&sa);
bbaseptr = aml_scratch_baseptr(&sb);
ap = aml_tiling_tilestart(tiling, a, offset);
bp = aml_tiling_tilestart(tiling, b, offset);
cp = aml_tiling_tilestart(tiling, c, offset);
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);
aml_scratch_async_pull(sa, &ar, abaseptr, &ai, a, offset+i+1);
aml_scratch_async_pull(sb, &br, bbaseptr, &bi, b, offset+i+1);
kernel(ap, bp, cp, esz);
aml_scratch_wait(&sa, ar);
aml_scratch_wait(&sb, br);
ap = aml_tiling_tilestart(&tiling, abaseptr, ai);
bp = aml_tiling_tilestart(&tiling, bbaseptr, bi);
cp = aml_tiling_tilestart(&tiling, c, offset+i+1);
aml_scratch_release(&sa, oldai);
aml_scratch_release(&sb, oldbi);
aml_scratch_wait(sa, ar);
aml_scratch_wait(sb, br);
ap = aml_tiling_tilestart(tiling, abaseptr, ai);
bp = aml_tiling_tilestart(tiling, bbaseptr, bi);
cp = aml_tiling_tilestart(tiling, c, offset+i+1);
aml_scratch_release(sa, oldai);
aml_scratch_release(sb, oldbi);
}
kernel(ap, bp, cp, esz);
......@@ -79,7 +79,7 @@ void *th_work(void *arg)
}
int main(int argc, char *argv[])
{
AML_DMA_LINUX_SEQ_DECL(dma);
struct aml_dma *dma;
struct aml_bitmap slowb, fastb;
aml_init(&argc, &argv);
assert(argc == 4);
......@@ -100,17 +100,17 @@ int main(int argc, char *argv[])
}
/* initialize all the supporting struct */
assert(!aml_tiling_1d_init(&tiling, tilesz, memsize));
assert(!aml_tiling_1d_create(&tiling, tilesz, memsize));
aml_area_linux_create(&slow, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&slowb, AML_AREA_LINUX_BINDING_FLAG_BIND);
assert(slow != NULL);
aml_area_linux_create(&fast, AML_AREA_LINUX_MMAP_FLAG_PRIVATE,
&fastb, AML_AREA_LINUX_BINDING_FLAG_BIND);
assert(fast != NULL);
assert(!aml_dma_linux_seq_init(&dma, (size_t)numthreads*4));
assert(!aml_scratch_par_init(&sa, &fast, &slow, &dma, &tiling,
assert(!aml_dma_linux_seq_create(dma, (size_t)numthreads*4));
assert(!aml_scratch_par_create(&sa, fast, slow, dma, tiling,
(size_t)2*numthreads, (size_t)numthreads));
assert(!aml_scratch_par_init(&sb, &fast, &slow, &dma, &tiling,
assert(!aml_scratch_par_create(&sb, fast, slow, dma, tiling,
(size_t)2*numthreads, (size_t)numthreads));
/* allocation */
......@@ -142,15 +142,15 @@ int main(int argc, char *argv[])
assert(c[i] == esize);
}
aml_scratch_par_fini(&sa);
aml_scratch_par_fini(&sb);
aml_dma_linux_seq_fini(&dma);
aml_scratch_par_destroy(&sa);
aml_scratch_par_destroy(&sb);
aml_dma_linux_seq_destroy(&dma);
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);
aml_tiling_1d_fini(&tiling);
aml_tiling_1d_destroy(&tiling);
aml_finalize();
return 0;
}
include/aml.h
View file @ d3558a6d
......@@ -265,16 +265,6 @@ struct aml_tiling_ops {
/**
* \todo Doc
**/
int (*init_iterator)(struct aml_tiling_data *tiling,
struct aml_tiling_iterator *iterator, int flags);
/**
* \todo Doc
**/
int (*fini_iterator)(struct aml_tiling_data *tiling,
struct aml_tiling_iterator *iterator);
/**
* \todo Doc
**/
int (*destroy_iterator)(struct aml_tiling_data *tiling,
struct aml_tiling_iterator **iterator);
/**
......@@ -396,23 +386,6 @@ struct aml_tiling_iterator {
int aml_tiling_create_iterator(struct aml_tiling *tiling,
struct aml_tiling_iterator **iterator,
int flags);
/**
* Initializes a tiling iterator.
* @param tiling: an initialized tiling structure.
* @param iterator: an allocated tiling iterator structure.
* @param flags: reserved for future use; pass 0 for now.
* @return 0 if successful; an error code otherwise.
**/
int aml_tiling_init_iterator(struct aml_tiling *tiling,
struct aml_tiling_iterator *iterator, int flags);
/**
* Finalize an initialized tiling iterator.
* @param tiling: an initialized tiling structure.
* @param iterator: an initialized tiling iterator structure.
**/
void aml_tiling_fini_iterator(struct aml_tiling *tiling,
struct aml_tiling_iterator *iterator);
/**
* Tears down an initialized tiling iterator.
* @param tiling: an initialized tiling structure.
......
include/aml/area/linux.h
View file @ d3558a6d
......@@ -88,23 +88,6 @@ struct aml_area_linux_data {
int mmap_flags;
};
/**
* Static declaration of an aml area with linux ops.
**/
#define AML_AREA_LINUX_DECL(name) \
struct aml_area_linux_data __ ##name## _inner_data; \
struct aml_area name = { \
&aml_area_linux_ops, \
(struct aml_area_data *)&__ ## name ## _inner_data, \
}
/**
* Static declaration of the size of a linux aml area.
**/
#define AML_AREA_LINUX_ALLOCSIZE \
(sizeof(struct aml_area_linux_data) + \
sizeof(struct aml_area))
/**
* \brief Linux area creation.
*
......@@ -137,19 +120,6 @@ int aml_area_linux_create(struct aml_area **area, const int mmap_flags,
**/
void aml_area_linux_destroy(struct aml_area **area);
/**
* Initialize a struct aml_area declared using the AML_AREA_LINUX_DECL macro.
* @see aml_area_linux_create() for details on arguments.
*/
int aml_area_linux_init(struct aml_area *area, const int mmap_flags,
const struct aml_bitmap *nodemask,
const int binding_flags);
/**
* Finalize a struct aml_area initialized with aml_area_linux_init.
*/
void aml_area_linux_fini(struct aml_area *area);
/**
* Bind memory of size "size" pointed by "ptr" to binding set in "bind".
* If mbind call was not successfull, i.e AML_FAILURE is returned, then errno
......
include/aml/dma/linux-par.h
View file @ d3558a6d
......@@ -62,7 +62,7 @@ struct aml_dma_request_linux_par {
/** Inside of a parallel request for linux movement. **/
struct aml_dma_linux_par_data {
size_t nbthreads;
struct aml_vector requests;
struct aml_vector *requests;
pthread_mutex_t lock;
};
......@@ -75,7 +75,7 @@ struct aml_dma_linux_par_ops {
/**
* aml_dma structure for linux based, parallel dma movement
* Needs to be initialized with aml_dma_linux_par_init().
* Needs to be initialized with aml_dma_linux_par_create().
* Can be passed to generic aml_dma_*() functions.
**/
struct aml_dma_linux_par {
......@@ -83,19 +83,6 @@ struct aml_dma_linux_par {
struct aml_dma_linux_par_data data;
};
/** Static declaration of aml_dma_linux_par structure. **/
#define AML_DMA_LINUX_PAR_DECL(name) \
struct aml_dma_linux_par __ ##name## _inner_data; \
struct aml_dma name = { \
&aml_dma_linux_par_ops, \
(struct aml_dma_data *)&__ ## name ## _inner_data, \
}
/** Static declaration of aml_dma_linux_par structure size. **/
#define AML_DMA_LINUX_PAR_ALLOCSIZE \
(sizeof(struct aml_dma_linux_par) + \
sizeof(struct aml_dma))
/**
* Allocates and initializes a new parallel DMA.
*
......@@ -110,24 +97,6 @@ struct aml_dma_linux_par {
int aml_dma_linux_par_create(struct aml_dma **dma, size_t nbreqs,
size_t nbthreads);
/**
* Initializes a new parallel DMA.
*
* @param dma a pointer to a dma declared with the AML_DMA_LINUX_PAR_DECL macro
* @param nbreqs the initial number of slots for asynchronous requests that are
* in-flight (will be increased automatically if necessary).
* @param nbthreads the number of threads to launch for each request.
*