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

Using DeviceInfo instead of CLDevice.

Browse Source 
PiperOrigin-RevId: 335772168
Change-Id: I02d334a87c2d8495ed7ed96dbdc31f4945b2cd3b
This commit is contained in:
Raman Sarokin 2020-10-06 19:17:21 -07:00 committed by TensorFlower Gardener
parent abce641853
commit 31e4a59c18
6 changed files with 36 additions and 32 deletions

1
tensorflow/lite/delegates/gpu/cl/BUILD
View File

@ -286,6 +286,7 @@ cc_library(
":cl_command_queue",
":cl_context",
":cl_device",
":device_info",
":precision",
":program_cache",
":tensor",

9
tensorflow/lite/delegates/gpu/cl/api.cc
View File

@ -677,12 +677,13 @@ class InferenceBuilderImpl : public InferenceBuilder {
if (GetRelativeImportance(options, InferencePriority::MIN_LATENCY,
InferencePriority::MIN_MEMORY_USAGE) ==
PriorityImportance::HIGHER) {
preferred_storage_types = {GetFastestStorageType(environment_->device()),
TensorStorageType::BUFFER};
} else {
preferred_storage_types = {
GetStorageTypeWithMinimalMemoryConsumption(environment_->device()),
GetFastestStorageType(environment_->device().GetInfo()),
TensorStorageType::BUFFER};
} else {
preferred_storage_types = {GetStorageTypeWithMinimalMemoryConsumption(
environment_->device().GetInfo()),
TensorStorageType::BUFFER};
}
for (TensorStorageType storage_type : preferred_storage_types) {

1
tensorflow/lite/delegates/gpu/cl/cl_device.h
View File

@ -73,6 +73,7 @@ class CLDevice {
bool SupportsOneLayerTextureArray() const;
void DisableOneLayerTextureArray();
const DeviceInfo& GetInfo() const { return info_; }
// We update device info during context creation, so as supported texture
// formats can be requested from context only.
mutable DeviceInfo info_;

50
tensorflow/lite/delegates/gpu/cl/environment.cc
View File

@ -171,54 +171,54 @@ bool Environment::IsSupported(TensorStorageType storage_type) const {
return false;
}
TensorStorageType GetFastestStorageType(const CLDevice& gpu) {
if (gpu.IsAdreno()) {
if (gpu.IsAdreno6xxOrHigher()) {
TensorStorageType GetFastestStorageType(const DeviceInfo& gpu_info) {
if (gpu_info.IsAdreno()) {
if (gpu_info.IsAdreno6xxOrHigher()) {
return TensorStorageType::TEXTURE_ARRAY;
} else {
return TensorStorageType::TEXTURE_2D;
}
} else if (gpu.IsPowerVR()) {
} else if (gpu_info.IsPowerVR()) {
return TensorStorageType::TEXTURE_2D;
} else if (gpu.IsMali()) {
const MaliInfo mali_info = gpu.info_.mali_info;
} else if (gpu_info.IsMali()) {
const MaliInfo mali_info = gpu_info.mali_info;
if (mali_info.IsMaliT8xx() || mali_info.IsBifrostGen3() ||
mali_info.IsValhall()) {
return TensorStorageType::TEXTURE_2D;
} else {
return TensorStorageType::BUFFER;
}
} else if (gpu.IsNvidia()) {
return gpu.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu.IsAMD()) {
return gpu.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu.IsIntel()) {
} else if (gpu_info.IsNvidia()) {
return gpu_info.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu_info.IsAMD()) {
return gpu_info.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu_info.IsIntel()) {
return TensorStorageType::BUFFER;
}
return TensorStorageType::BUFFER;
}
TensorStorageType GetStorageTypeWithMinimalMemoryConsumption(
const CLDevice& gpu) {
if (gpu.IsAdreno()) {
if (gpu.IsAdreno3xx() || gpu.IsAdreno4xx()) {
const DeviceInfo& gpu_info) {
if (gpu_info.IsAdreno()) {
if (gpu_info.IsAdreno3xx() || gpu_info.IsAdreno4xx()) {
return TensorStorageType::BUFFER;
} else {
return TensorStorageType::IMAGE_BUFFER;
}
} else if (gpu.IsPowerVR()) {
} else if (gpu_info.IsPowerVR()) {
return TensorStorageType::BUFFER;
} else if (gpu.IsMali()) {
} else if (gpu_info.IsMali()) {
return TensorStorageType::BUFFER;
} else if (gpu.IsNvidia()) {
return gpu.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu.IsAMD()) {
return gpu.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu.IsIntel()) {
} else if (gpu_info.IsNvidia()) {
return gpu_info.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu_info.IsAMD()) {
return gpu_info.SupportsImageBuffer() ? TensorStorageType::IMAGE_BUFFER
: TensorStorageType::BUFFER;
} else if (gpu_info.IsIntel()) {
return TensorStorageType::BUFFER;
}
return TensorStorageType::BUFFER;

5
tensorflow/lite/delegates/gpu/cl/environment.h
View File

@ -19,6 +19,7 @@ limitations under the License.
#include "tensorflow/lite/delegates/gpu/cl/cl_command_queue.h"
#include "tensorflow/lite/delegates/gpu/cl/cl_context.h"
#include "tensorflow/lite/delegates/gpu/cl/cl_device.h"
#include "tensorflow/lite/delegates/gpu/cl/device_info.h"
#include "tensorflow/lite/delegates/gpu/cl/precision.h"
#include "tensorflow/lite/delegates/gpu/cl/program_cache.h"
#include "tensorflow/lite/delegates/gpu/cl/tensor_type.h"
@ -74,9 +75,9 @@ class Environment {
ProgramCache program_cache_;
};
TensorStorageType GetFastestStorageType(const CLDevice& gpu);
TensorStorageType GetFastestStorageType(const DeviceInfo& gpu_info);
TensorStorageType GetStorageTypeWithMinimalMemoryConsumption(
const CLDevice& gpu);
const DeviceInfo& gpu_info);
absl::Status CreateEnvironment(Environment* result);

2
tensorflow/lite/delegates/gpu/cl/testing/performance_profiling.cc
View File

@ -43,7 +43,7 @@ absl::Status RunModelSample(const std::string& model_name) {
create_info.precision = env.IsSupported(CalculationsPrecision::F16)
? CalculationsPrecision::F16
: CalculationsPrecision::F32;
create_info.storage_type = GetFastestStorageType(env.device());
create_info.storage_type = GetFastestStorageType(env.device().GetInfo());
create_info.hints.Add(ModelHints::kAllowSpecialKernels);
std::cout << "Precision: " << ToString(create_info.precision) << std::endl;
std::cout << "Storage type: " << ToString(create_info.storage_type)