Add int16 support to RELU #2727
Conversation
| TF_LITE_MICRO_TEST(SimpleReluTestInt16) { | ||
| const int elements_count = 10; | ||
|
|
||
| int input_shape[] = {2, 1, 5}; |
| int input_shape[] = {2, 1, 5}; | ||
| const float input_data[] = {1, 2, 3, 4, 5, -1, -2, -3, -4, -5}; | ||
| int16_t input_quantized[elements_count]; | ||
| int output_shape[] = {2, 1, 5}; |
There was a problem hiding this comment.
In ReluPrepare the following code should be used to conform with TfLite (LiteRT):
if (input->type == kTfLiteInt8) {
CalculateReluOpData<int8_t>(input, output, data);
}
else if (input->type == kTfLiteInt16) {
TF_LITE_ENSURE_EQ(context, input->params.zero_point, 0);
TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
CalculateReluOpData<int16_t>(input, output, data);
}
I am very surprised (and concerned) that the unit test passed, even though CalculateReluIOpData<int16_t> was not being called.
There was a problem hiding this comment.
It might have had something to do with the chosen input data values. I changed them so that they need to are actually int16 values and then the test failed. Could be just a coincidence that it happened to work or there might be something else going elsewhere
|
Thank you for submitting this PR. I have flagged the errors in the kernel and unit test above. On to @suleshahid for approval. |
|
"This PR is being marked as stale due to inactivity. Remove label or comment to prevent closure in 5 days." |
|
Waiting for approval |
This PR adds int16 support to RELU
bug=#2726