WEAK int halide_dev_malloc(void *user_context, buffer_t *buf) {
DEBUG_PRINTF( user_context, "CUDA: halide_dev_malloc (user_context: %p, buf: %p)\n", user_context, buf );
CudaContext ctx(user_context);
if (ctx.error != CUDA_SUCCESS) {
return ctx.error;
}
size_t size = _buf_size(user_context, buf);
if (buf->dev) {
// This buffer already has a device allocation
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf, size));
return 0;
}
halide_assert(user_context, buf->stride[0] >= 0 && buf->stride[1] >= 0 &&
buf->stride[2] >= 0 && buf->stride[3] >= 0);
DEBUG_PRINTF(user_context, " allocating buffer of %lld bytes, "
"extents: %lldx%lldx%lldx%lld strides: %lldx%lldx%lldx%lld (%d bytes per element)\n",
(long long)size,
(long long)buf->extent[0], (long long)buf->extent[1],
(long long)buf->extent[2], (long long)buf->extent[3],
(long long)buf->stride[0], (long long)buf->stride[1],
(long long)buf->stride[2], (long long)buf->stride[3],
buf->elem_size);
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
CUdeviceptr p;
DEBUG_PRINTF( user_context, " cuMemAlloc %lld -> ", size );
CUresult err = cuMemAlloc(&p, size);
if (err != CUDA_SUCCESS) {
DEBUG_PRINTF( user_context, "%s\n", _get_error_name(err));
halide_error_varargs(user_context, "CUDA: cuMemAlloc failed (%s)",
_get_error_name(err));
return err;
} else {
DEBUG_PRINTF( user_context, "%p\n", p );
}
halide_assert(user_context, p);
buf->dev = (uint64_t)p;
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
return 0;
}
WEAK int halide_dev_malloc(void *user_context, buffer_t* buf) {
DEBUG_PRINTF( user_context, "CL: halide_dev_malloc (user_context: %p, buf: %p)\n", user_context, buf );
ClContext ctx(user_context);
if (ctx.error != CL_SUCCESS) {
return ctx.error;
}
size_t size = _buf_size(user_context, buf);
if (buf->dev) {
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf, size));
return 0;
}
halide_assert(user_context, buf->stride[0] >= 0 && buf->stride[1] >= 0 &&
buf->stride[2] >= 0 && buf->stride[3] >= 0);
DEBUG_PRINTF(user_context, " Allocating buffer of %lld bytes, "
"extents: %lldx%lldx%lldx%lld strides: %lldx%lldx%lldx%lld (%d bytes per element)\n",
(long long)size,
(long long)buf->extent[0], (long long)buf->extent[1],
(long long)buf->extent[2], (long long)buf->extent[3],
(long long)buf->stride[0], (long long)buf->stride[1],
(long long)buf->stride[2], (long long)buf->stride[3],
buf->elem_size);
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
cl_int err;
DEBUG_PRINTF( user_context, " clCreateBuffer -> ", size );
buf->dev = (uint64_t)clCreateBuffer(ctx.context, CL_MEM_READ_WRITE, size, NULL, &err);
if (err != CL_SUCCESS || buf->dev == 0) {
DEBUG_PRINTF( user_context, "%d\n", err);
halide_error_varargs(user_context, "CL: clCreateBuffer failed (%d)\n", err);
return err;
} else {
DEBUG_PRINTF( user_context, "%p\n", (cl_mem)buf->dev );
}
DEBUG_PRINTF(user_context, " Allocated device buffer %p for buffer %p\n",
(void *)buf->dev, buf);
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
return CL_SUCCESS;
}
