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

DepthwiseConv, dot-product kernel asm code, initial code copied from intrinsics.

Browse Source 
PiperOrigin-RevId: 238124479
This commit is contained in:
A. Unique TensorFlower 2019-03-12 16:40:02 -07:00 committed by TensorFlower Gardener
parent 130d901289
commit f23ef54d0d
2 changed files with 3222 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
tensorflow/lite/kernels/internal/depthwiseconv_quantized_test.cc
View File

@ -145,9 +145,29 @@ inline void DispatchDepthwiseConv(
break; break;
#endif #endif
} }
case DepthwiseConvImplementation::kUseNeon3x3DotProduct: case DepthwiseConvImplementation::kUseNeon3x3DotProduct: {
// TODO(b/118426582) Placeholder for future dispatches. #if defined(__ARM_FEATURE_DOTPROD) && defined(__aarch64__)
DotProduct3x3KernelType kernel_type =
optimized_ops::depthwise_conv::CategorizeDotProductKernel(
input_shape, filter_shape, params);
ASSERT_TRUE(
kernel_type == DotProduct3x3KernelType::kPlain ||
kernel_type == DotProduct3x3KernelType::kStride2 ||
kernel_type ==
DotProduct3x3KernelType::kWithDepthMultiplicationStride1 ||
kernel_type ==
DotProduct3x3KernelType::kWithDepthMultiplicationStride2)
<< "Kernel type = " << static_cast<int>(kernel_type);
optimized_ops::depthwise_conv::DepthwiseConvDotProduct3x3<
DepthwiseConvImplementation::kUseNeon3x3DotProduct>(
params, input_shape, input_data, filter_shape, filter_data,
bias_shape, bias_data, output_shape, output_data);
return;
#endif
break; break;
}
case DepthwiseConvImplementation::kUseCModel3x3DotProduct: { case DepthwiseConvImplementation::kUseCModel3x3DotProduct: {
DotProduct3x3KernelType kernel_type = DotProduct3x3KernelType kernel_type =
optimized_ops::depthwise_conv::CategorizeDotProductKernel( optimized_ops::depthwise_conv::CategorizeDotProductKernel(
@ -181,7 +201,6 @@ inline void DispatchDepthwiseConv(
return; return;
} }
case DepthwiseConvImplementation::kUseUnwound3x3DotProduct: { case DepthwiseConvImplementation::kUseUnwound3x3DotProduct: {
using optimized_ops::depthwise_conv::DotProduct3x3KernelType;
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);
@ -200,7 +219,6 @@ inline void DispatchDepthwiseConv(
} }
case DepthwiseConvImplementation::kUseIntrinsics3x3DotProduct: { case DepthwiseConvImplementation::kUseIntrinsics3x3DotProduct: {
#if defined(USE_NEON) #if defined(USE_NEON)
using optimized_ops::depthwise_conv::DotProduct3x3KernelType;
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);
@ -794,5 +812,21 @@ INSTANTIATE_TEST_SUITE_P(
TestParam::TestNameSuffix); TestParam::TestNameSuffix);
#endif #endif
#if defined(__ARM_FEATURE_DOTPROD) && defined(__aarch64__)
INSTANTIATE_TEST_SUITE_P(
NeonAsm, DepthwiseConvTest,
testing::Combine(
Values(DepthwiseConvImplementation::
kUseNeon3x3DotProduct), // forced_invocation
Values(1000), // tests_to_run
Bool(), // test_stride
Bool(), // test_pad
Bool(), // test_depth_multiplier
Values(DepthwiseConvOutputRounding::kUpward), // output_rounding
Values(false) // loose_tolerance
),
TestParam::TestNameSuffix);
#endif
} // namespace } // namespace
} // namespace tflite } // namespace tflite

3189
tensorflow/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h
View File

File diff suppressed because it is too large Load Diff