experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Update README regarding Metal delegate instructions.

Browse Source 
PiperOrigin-RevId: 320880762
Change-Id: I6b36930f9b946b1f798c335326af182be49fedb2
This commit is contained in:
Taehee Jeong 2020-07-12 19:16:08 -07:00 committed by TensorFlower Gardener
parent d91193d641
commit 7b14d48d8a
1 changed files with 15 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
tensorflow/lite/delegates/gpu/README.md
View File

@ -51,6 +51,10 @@ TFLite on GPU supports the following ops in 16-bit and 32-bit float precision:
## Basic Usage
**Note:** Following section describes the example usage for Android GPU delegate
with C++. For other languages and platforms, please see
[the documentation](https://www.tensorflow.org/lite/performance/gpu).
Using TFLite on GPU is as simple as getting the GPU delegate via
`TfLiteGpuDelegateV2Create()` and then passing it to
`Interpreter::ModifyGraphWithDelegate()` instead of calling
@ -99,13 +103,13 @@ Metal shaders are used for iOS, which were introduced with iOS 8. Thus,
compilation flags should look like:
```sh
bazel build --config ios_arm64 //path/to/your:project
bazel build --config ios_fat //path/to/your:project
```
## Advanced Usage: Delegate Options
There are GPU options that can be set and passed on to
`TfLiteGpuDelegateCreate()`. When option is set to `nullptr` as shown in the
`TfLiteGpuDelegateV2Create()`. When option is set to `nullptr` as shown in the
Basic Usage, it translates to:
```c++
@ -113,12 +117,13 @@ const TfLiteGpuDelegateOptionsV2 kDefaultOptions =
TfLiteGpuDelegateOptionsV2Default();
```
Similar for `NewTfLiteMetalDelegate()`:
Similar for `TFLGpuDelegateCreate()`:
```c++
const TfLiteMetalDelegateOptions kDefaultOptions = {
.precision_loss_allowed = 0, // false
.wait_type = TFLITE_METAL_WAIT_TYPE_SLEEP,
const TFLGpuDelegateOptions kDefaultOptions = {
.allow_precision_loss = false,
.wait_type = TFLGpuDelegateWaitTypePassive,
.enable_quantization = false,
};
```
@ -126,9 +131,10 @@ While it is convenient to just supply `nullptr`, it is recommended to explicitly
set the options to avoid any unexpected artifacts in case default values are
changed.
*IMPORTANT:* Note that the default option does not allow precision loss, and
thus may not be the fastest. For faster execution, you may want to set
`precision_loss_allowed` to `1` for FP16 execution.
*IMPORTANT:* Note that the default option may not be the fastest. For faster
execution, you may want to set `allow_precision_loss` to `true` so that the GPU
performs FP16 calculation internally, and set `wait_type` to
`TFLGpuDelegateWaitTypeAggressive` to avoid GPU sleep mode.
## Tips and Tricks