static int ga_extcopy(GpuArray *dst, const GpuArray *src) {
struct extcopy_args a, *aa;
gpucontext *ctx = gpudata_context(dst->data);
GpuElemwise *k = NULL;
void *args[2];
if (ctx != gpudata_context(src->data))
return GA_INVALID_ERROR;
a.itype = src->typecode;
a.otype = dst->typecode;
if (ctx->extcopy_cache != NULL)
k = cache_get(ctx->extcopy_cache, &a);
if (k == NULL) {
gpuelemwise_arg gargs[2];
gargs[0].name = "src";
gargs[0].typecode = src->typecode;
gargs[0].flags = GE_READ;
gargs[1].name = "dst";
gargs[1].typecode = dst->typecode;
gargs[1].flags = GE_WRITE;
k = GpuElemwise_new(ctx, "", "dst = src", 2, gargs, 0, 0);
if (k == NULL)
return GA_MISC_ERROR;
aa = memdup(&a, sizeof(a));
if (aa == NULL) {
GpuElemwise_free(k);
return GA_MEMORY_ERROR;
}
if (ctx->extcopy_cache == NULL)
ctx->extcopy_cache = cache_twoq(4, 8, 8, 2, extcopy_eq, extcopy_hash,
extcopy_free,
(cache_freev_fn)GpuElemwise_free,
ctx->err);
if (ctx->extcopy_cache == NULL)
return GA_MISC_ERROR;
if (cache_add(ctx->extcopy_cache, aa, k) != 0)
return GA_MISC_ERROR;
}
args[0] = (void *)src;
args[1] = (void *)dst;
return GpuElemwise_call(k, args, GE_BROADCAST);
}
static gpukernel *cuda_newkernel(void *c, unsigned int count,
const char **strings, const size_t *lengths,
const char *fname, unsigned int argcount,
const int *types, int flags, int *ret,
char **err_str) {
cuda_context *ctx = (cuda_context *)c;
strb sb = STRB_STATIC_INIT;
char *bin, *log = NULL;
srckey k, *ak;
binval *av;
gpukernel *res;
size_t bin_len = 0, log_len = 0;
CUdevice dev;
unsigned int i;
int ptx_mode = 0;
int binary_mode = 0;
int major, minor;
if (count == 0) FAIL(NULL, GA_VALUE_ERROR);
if (flags & GA_USE_OPENCL)
FAIL(NULL, GA_DEVSUP_ERROR);
if (flags & GA_USE_BINARY) {
// GA_USE_BINARY is exclusive
if (flags & ~GA_USE_BINARY)
FAIL(NULL, GA_INVALID_ERROR);
// We need the length for binary data and there is only one blob.
if (count != 1 || lengths == NULL || lengths[0] == 0)
FAIL(NULL, GA_VALUE_ERROR);
}
cuda_enter(ctx);
ctx->err = cuCtxGetDevice(&dev);
if (ctx->err != CUDA_SUCCESS) {
cuda_exit(ctx);
FAIL(NULL, GA_IMPL_ERROR);
}
ctx->err = cuDeviceComputeCapability(&major, &minor, dev);
if (ctx->err != CUDA_SUCCESS) {
cuda_exit(ctx);
FAIL(NULL, GA_IMPL_ERROR);
}
// GA_USE_CLUDA is done later
// GA_USE_SMALL will always work
if (flags & GA_USE_DOUBLE) {
if (major < 1 || (major == 1 && minor < 3)) {
cuda_exit(ctx);
FAIL(NULL, GA_DEVSUP_ERROR);
}
}
if (flags & GA_USE_COMPLEX) {
// just for now since it is most likely broken
cuda_exit(ctx);
FAIL(NULL, GA_DEVSUP_ERROR);
}
// GA_USE_HALF should always work
if (flags & GA_USE_PTX) {
ptx_mode = 1;
} else if (flags & GA_USE_BINARY) {
binary_mode = 1;
}
if (binary_mode) {
bin = memdup(strings[0], lengths[0]);
bin_len = lengths[0];
if (bin == NULL) {
cuda_exit(ctx);
FAIL(NULL, GA_MEMORY_ERROR);
}
} else {
if (flags & GA_USE_CLUDA) {
strb_appends(&sb, CUDA_PREAMBLE);
}
if (lengths == NULL) {
for (i = 0; i < count; i++)
strb_appends(&sb, strings[i]);
} else {
for (i = 0; i < count; i++) {
if (lengths[i] == 0)
strb_appends(&sb, strings[i]);
else
strb_appendn(&sb, strings[i], lengths[i]);
}
}
strb_append0(&sb);
if (strb_error(&sb)) {
strb_clear(&sb);
cuda_exit(ctx);
return NULL;
}
if (ptx_mode) {
bin = sb.s;
bin_len = sb.l;
} else {
bin = NULL;
if (compile_cache != NULL) {
k.src = sb.s;
k.len = sb.l;
memcpy(k.arch, ctx->bin_id, BIN_ID_LEN);
av = cache_get(compile_cache, &k);
if (av != NULL) {
bin = memdup(av->bin, av->len);
bin_len = av->len;
}
}
if (bin == NULL) {
bin = call_compiler(sb.s, sb.l, ctx->bin_id, &bin_len,
&log, &log_len, ret);
}
if (bin == NULL) {
if (err_str != NULL) {
strb debug_msg = STRB_STATIC_INIT;
// We're substituting debug_msg for a string with this first line:
strb_appends(&debug_msg, "CUDA kernel build failure ::\n");
/* Delete the final NUL */
sb.l--;
gpukernel_source_with_line_numbers(1, (const char **)&sb.s,
&sb.l, &debug_msg);
if (log != NULL) {
strb_appends(&debug_msg, "\nCompiler log:\n");
strb_appendn(&debug_msg, log, log_len);
free(log);
}
*err_str = strb_cstr(&debug_msg);
// *err_str will be free()d by the caller (see docs in kernel.h)
}
strb_clear(&sb);
cuda_exit(ctx);
return NULL;
}
if (compile_cache == NULL)
compile_cache = cache_twoq(16, 16, 16, 8, src_eq, src_hash, src_free,
bin_free);
if (compile_cache != NULL) {
ak = malloc(sizeof(*ak));
av = malloc(sizeof(*av));
if (ak == NULL || av == NULL) {
free(ak);
free(av);
goto done;
}
ak->src = memdup(sb.s, sb.l);
if (ak->src == NULL) {
free(ak);
free(av);
goto done;
}
ak->len = sb.l;
memmove(ak->arch, ctx->bin_id, BIN_ID_LEN);
av->len = bin_len;
av->bin = memdup(bin, bin_len);
if (av->bin == NULL) {
src_free(ak);
free(av);
goto done;
}
cache_add(compile_cache, ak, av);
}
done:
strb_clear(&sb);
}
}
res = calloc(1, sizeof(*res));
if (res == NULL) {
free(bin);
cuda_exit(ctx);
FAIL(NULL, GA_SYS_ERROR);
}
res->bin_sz = bin_len;
res->bin = bin;
res->refcnt = 1;
res->argcount = argcount;
res->types = calloc(argcount, sizeof(int));
if (res->types == NULL) {
_cuda_freekernel(res);
cuda_exit(ctx);
FAIL(NULL, GA_MEMORY_ERROR);
}
memcpy(res->types, types, argcount*sizeof(int));
res->args = calloc(argcount, sizeof(void *));
if (res->args == NULL) {
_cuda_freekernel(res);
cuda_exit(ctx);
FAIL(NULL, GA_MEMORY_ERROR);
}
ctx->err = cuModuleLoadData(&res->m, bin);
if (ctx->err != CUDA_SUCCESS) {
_cuda_freekernel(res);
cuda_exit(ctx);
FAIL(NULL, GA_IMPL_ERROR);
}
ctx->err = cuModuleGetFunction(&res->k, res->m, fname);
if (ctx->err != CUDA_SUCCESS) {
_cuda_freekernel(res);
cuda_exit(ctx);
FAIL(NULL, GA_IMPL_ERROR);
}
res->ctx = ctx;
ctx->refcnt++;
cuda_exit(ctx);
TAG_KER(res);
return res;
}
