giuseros commented on a change in pull request #7785:
URL: https://github.com/apache/tvm/pull/7785#discussion_r619379316
##########
File path: src/runtime/crt/memory/stack_allocator.c
##########
@@ -16,17 +16,22 @@
* specific language governing permissions and limitations
* under the License.
*/
-
// LINT_C_FILE
-
#include <tvm/runtime/crt/stack_allocator.h>
+#ifdef TVM_CRT_DEBUG
+#include <tvm/runtime/crt/logging.h>
+#endif
void* StackMemoryManager_Allocate(tvm_workspace_t* tvm_runtime_workspace,
int32_t nbytes) {
uint32_t offset_bytes = (~nbytes + 1) & (TVM_RUNTIME_ALLOC_ALIGNMENT_BYTES -
1);
uint8_t* current_alloc = tvm_runtime_workspace->next_alloc;
uint8_t* next_alloc = tvm_runtime_workspace->next_alloc + nbytes +
offset_bytes;
uint8_t* workspace_end = tvm_runtime_workspace->workspace +
tvm_runtime_workspace->workspace_size;
-
+#ifdef TVM_CRT_DEBUG
Review comment:
So there are two main issues here:
* We are using 4 bytes per block to check memory. Given the memory
constraints on some devices, it might be too expensive.
* More importantly, we are loosing the block alignment. There are some
microcontrollers that would not allow non-aligned memory access.
For those two reasons, I would disable it by default and only enable for
specific reasons. I can use a different MACRO (e.g.,
STACK_ALLOCATOR_CHECK_ENABLED) that people can use to enable this feature, so
it would not be only for debug. What do you think? Also adding @manupa-arm and
@Mousius to the discussion
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
