diff options
| author | James Ward <james.ward@arm.com> | 2022-11-15 11:36:47 +0000 |
|---|---|---|
| committer | Eric Kunze <eric.kunze@arm.com> | 2022-11-29 15:55:26 +0000 |
| commit | ee2566914d3476b8103b88915f3b81bda8490b44 (patch) | |
| tree | 0d0dd56adafb3a65d896a192eeb736200eec8f06 /reference_model/src/tensor.cc | |
| parent | 542dd3b8da39440026fa9e809eebd0a3b79cf95d (diff) | |
| download | reference_model-ee2566914d3476b8103b88915f3b81bda8490b44.tar.gz | |
FP16 improvements
* Update FP16 resize to newest spec version
* Correct casting to fp16 for graphs of >1 ops
Change-Id: Iedff9a71eb7f72948b3c00a635bb0fd07d414bcd
Signed-off-by: James Ward <james.ward@arm.com>
Diffstat (limited to 'reference_model/src/tensor.cc')
| -rw-r--r-- | reference_model/src/tensor.cc | 5 |
1 files changed, 2 insertions, 3 deletions
diff --git a/reference_model/src/tensor.cc b/reference_model/src/tensor.cc index 4eaf21d..e9598c4 100644 --- a/reference_model/src/tensor.cc +++ b/reference_model/src/tensor.cc @@ -159,8 +159,7 @@ int TosaReference::Tensor::readFromNpyFile(const char* filename) switch (dtype) { case DType_FP16: - // Convert from fp16 to fp32 - //TODO(jw): remove this once we cast to fp16 in register_fcn/eval + // Convert from fp16 to fp32 so that fp16 values can be manipulated as float for (uint32_t i=0; i < elements; i++) { fdatabuf[i] = half_float::half_cast<float, half_float::half>(f16databuf[i]); } @@ -277,7 +276,7 @@ int TosaReference::Tensor::writeToNpyFile(const char* filename) const free(f16databuf); return 1; } - // Convert fp32 to fp16 + // Convert fp32 to fp16 so that output file contains valid fp16 data for (uint32_t i=0; i < elements; i++) { f16databuf[i] = half_float::half_cast<half_float::half, float>(fdatabuf[i]); } |