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

Disable dot-product depthwise conv when not compiling for Android with Clang.

Browse Source 
PiperOrigin-RevId: 255181633
This commit is contained in:
A. Unique TensorFlower 2019-06-26 07:11:29 -07:00 committed by TensorFlower Gardener
parent 99023249f7
commit 851b3f5a46
3 changed files with 12 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tensorflow/lite/kernels/internal/depthwiseconv_quantized_test.cc
View File

@ -170,7 +170,8 @@ inline void DispatchDepthwiseConv(
// This is compiled-in even if dot-product instructions are unavailable. // This is compiled-in even if dot-product instructions are unavailable.
// However, tests should skip dot-product testing in that case and not // However, tests should skip dot-product testing in that case and not
// call this code. // call this code.
#if defined(__aarch64__) && !defined(GOOGLE_L4T) #if defined(__aarch64__) && !defined(GOOGLE_L4T) && defined(__ANDROID__) && \
defined(__clang__)
DotProduct3x3KernelType kernel_type = DotProduct3x3KernelType kernel_type =
optimized_ops::depthwise_conv::CategorizeDotProductKernel( optimized_ops::depthwise_conv::CategorizeDotProductKernel(
input_shape, filter_shape, params); input_shape, filter_shape, params);
@ -683,7 +684,8 @@ void TestOneDepthwiseConv3x3Filter(
} }
void TestOneNeonDot3x3(const TestParam& test_param) { void TestOneNeonDot3x3(const TestParam& test_param) {
#if defined(__aarch64__) && !defined(GOOGLE_L4T) #if defined(__aarch64__) && !defined(GOOGLE_L4T) && defined(__ANDROID__) && \
defined(__clang__)
CpuBackendContext backend_context; CpuBackendContext backend_context;
ruy::Context* ruy_context = backend_context.ruy_context(); ruy::Context* ruy_context = backend_context.ruy_context();
const auto ruy_paths = ruy_context != nullptr const auto ruy_paths = ruy_context != nullptr
@ -854,7 +856,8 @@ INSTANTIATE_TEST_SUITE_P(
TestParam::TestNameSuffix); TestParam::TestNameSuffix);
#endif #endif
#if defined(__aarch64__) && !defined(GOOGLE_L4T) #if defined(__aarch64__) && !defined(GOOGLE_L4T) && defined(__ANDROID__) && \
defined(__clang__)
INSTANTIATE_TEST_SUITE_P( INSTANTIATE_TEST_SUITE_P(
NeonAsm, DepthwiseConvTest, NeonAsm, DepthwiseConvTest,
testing::Combine( testing::Combine(

5
tensorflow/lite/kernels/internal/optimized/depthwiseconv_uint8.h
View File

@ -2006,7 +2006,8 @@ inline void DepthwiseConvWithRounding(
// Enable for arm64 except for the Nvidia Linux 4 Tegra (L4T) running on // Enable for arm64 except for the Nvidia Linux 4 Tegra (L4T) running on
// Jetson TX-2. This compiler does not support the offsetof() macro. // Jetson TX-2. This compiler does not support the offsetof() macro.
#if defined(__aarch64__) && !defined(GOOGLE_L4T) #if defined(__aarch64__) && !defined(GOOGLE_L4T) && defined(__ANDROID__) && \
defined(__clang__)
// Dispatch to dot-product 3x3 kernels when supported. // Dispatch to dot-product 3x3 kernels when supported.
if (cpu_flags.neon_dotprod) { if (cpu_flags.neon_dotprod) {
using optimized_ops::depthwise_conv::DotProduct3x3KernelType; using optimized_ops::depthwise_conv::DotProduct3x3KernelType;
@ -2025,6 +2026,8 @@ inline void DepthwiseConvWithRounding(
} }
} }
#elif defined(__aarch64__) && !defined(GOOGLE_L4T)
// Dispatch to non-dot-product 3x3 kernels when supported. // Dispatch to non-dot-product 3x3 kernels when supported.
const int stride_width = params.stride_width; const int stride_width = params.stride_width;

3
tensorflow/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h
View File

@ -5786,7 +5786,8 @@ struct WorkspacePrefetchWrite<
#endif // __aarch64__ #endif // __aarch64__
#if defined(__aarch64__) && !defined(GOOGLE_L4T) #if defined(__aarch64__) && !defined(GOOGLE_L4T) && defined(__ANDROID__) && \
defined(__clang__)
// Dot product ops hard-coded // Dot product ops hard-coded
template <> template <>