static void cuda_free_ctx(cuda_context *ctx) {
gpuarray_blas_ops *blas_ops;
gpudata *next, *curr;
ASSERT_CTX(ctx);
ctx->refcnt--;
if (ctx->refcnt == 0) {
assert(ctx->enter == 0 && "Context was active when freed!");
if (ctx->blas_handle != NULL) {
ctx->err = cuda_property(ctx, NULL, NULL, GA_CTX_PROP_BLAS_OPS,
&blas_ops);
blas_ops->teardown(ctx);
}
cuMemFreeHost((void *)ctx->errbuf->ptr);
deallocate(ctx->errbuf);
cuStreamDestroy(ctx->s);
/* Clear out the freelist */
for (curr = ctx->freeblocks; curr != NULL; curr = next) {
next = curr->next;
cuMemFree(curr->ptr);
deallocate(curr);
}
if (!(ctx->flags & DONTFREE))
cuCtxDestroy(ctx->ctx);
cache_destroy(ctx->extcopy_cache);
CLEAR(ctx);
free(ctx);
}
}
static void cuda_free_ctx(cuda_context *ctx) {
gpuarray_blas_ops *blas_ops;
ASSERT_CTX(ctx);
ctx->refcnt--;
if (ctx->refcnt == 0) {
if (ctx->blas_handle != NULL) {
ctx->err = cuda_property(ctx, NULL, NULL, GA_CTX_PROP_BLAS_OPS, &blas_ops);
blas_ops->teardown(ctx);
}
cuStreamDestroy(ctx->s);
if (!(ctx->flags & DONTFREE))
cuCtxDestroy(ctx->ctx);
cache_free(ctx->extcopy_cache);
CLEAR(ctx);
free(ctx);
}
}
static void cuda_free_ctx(cuda_context *ctx) {
gpuarray_blas_ops *blas_ops;
ASSERT_CTX(ctx);
ctx->refcnt--;
if (ctx->refcnt == 0) {
assert(ctx->enter == 0 && "Context was active when freed!");
if (ctx->blas_handle != NULL) {
ctx->err = cuda_property(ctx, NULL, NULL, GA_CTX_PROP_BLAS_OPS,
&blas_ops);
blas_ops->teardown(ctx);
}
ctx->refcnt = 2; /* Prevent recursive calls */
cuda_free(ctx->errbuf);
cuStreamDestroy(ctx->s);
if (!(ctx->flags & DONTFREE))
cuCtxDestroy(ctx->ctx);
cache_free(ctx->extcopy_cache);
CLEAR(ctx);
free(ctx);
}
}
static int cuda_extcopy(gpudata *input, size_t ioff, gpudata *output,
size_t ooff, int intype, int outtype,
unsigned int a_nd, const size_t *a_dims,
const ssize_t *a_str, unsigned int b_nd,
const size_t *b_dims, const ssize_t *b_str) {
cuda_context *ctx = input->ctx;
void *args[2];
int res = GA_SYS_ERROR;
unsigned int i;
size_t nEls = 1, ls, gs;
gpukernel *k;
extcopy_args a, *aa;
ASSERT_BUF(input);
ASSERT_BUF(output);
if (input->ctx != output->ctx)
return GA_INVALID_ERROR;
for (i = 0; i < a_nd; i++) {
nEls *= a_dims[i];
}
if (nEls == 0) return GA_NO_ERROR;
a.ind = a_nd;
a.ond = b_nd;
a.itype = intype;
a.otype = outtype;
a.ioff = ioff;
a.ooff = ooff;
a.idims = a_dims;
a.odims = b_dims;
a.istr = a_str;
a.ostr = b_str;
k = cache_get(ctx->extcopy_cache, &a);
if (k == NULL) {
res = gen_extcopy_kernel(&a, input->ctx, &k, nEls);
if (res != GA_NO_ERROR)
return res;
/* Cache the kernel */
aa = memdup(&a, sizeof(a));
if (aa == NULL) goto done;
aa->idims = memdup(a_dims, a_nd*sizeof(size_t));
aa->odims = memdup(b_dims, b_nd*sizeof(size_t));
aa->istr = memdup(a_str, a_nd*sizeof(ssize_t));
aa->ostr = memdup(b_str, b_nd*sizeof(ssize_t));
if (aa->idims == NULL || aa->odims == NULL ||
aa->istr == NULL || aa->ostr == NULL) {
extcopy_free(aa);
goto done;
}
/* One ref is given to the cache, we manage the other */
cuda_retainkernel(k);
cache_add(ctx->extcopy_cache, aa, k);
} else {
/* This is our reference */
cuda_retainkernel(k);
}
done:
/* Cheap kernel scheduling */
res = cuda_property(NULL, NULL, k, GA_KERNEL_PROP_MAXLSIZE, &ls);
if (res != GA_NO_ERROR) goto fail;
gs = ((nEls-1) / ls) + 1;
args[0] = input;
args[1] = output;
res = cuda_callkernel(k, 1, &ls, &gs, 0, args);
fail:
/* We free our reference here */
cuda_freekernel(k);
return res;
}
