From d3dbe347fdd5c82fdd643fd231877832d4ec940c Mon Sep 17 00:00:00 2001
From: sshiddib <sharada.shiddibhavi@intel.com>
Date: Tue, 26 May 2020 15:03:42 -0700
Subject: [PATCH 1/7] Removing OpenMP dependency from Mkl-dnn supporting
threadpool
---
.../mkl_threadpool_device_test.cc | 4 +--
tensorflow/core/kernels/mkl_conv_ops.cc | 5 ++--
tensorflow/core/kernels/mkl_conv_ops_test.cc | 2 +-
tensorflow/core/kernels/mkl_qmatmul_op.cc | 27 +++++++++++++++++--
.../core/kernels/mkl_quantized_conv_ops.h | 21 +++++++++++++--
...mkl_requantization_range_per_channel_op.cc | 22 +++++++++++++++
tensorflow/tensorflow.bzl | 4 +--
7 files changed, 73 insertions(+), 12 deletions(-)
diff --git a/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc b/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc
index 5d583a8360b..c29752d3c2c 100644
--- a/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc
+++ b/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc
@@ -25,7 +25,7 @@ limitations under the License.
namespace tensorflow {
-#ifdef _OPENMP
+#if defined(_OPENMP) && !defined(ENABLE_MKLDNN_THREADPOOL)
TEST(MKLThreadPoolDeviceTest, TestOmpDefaults) {
SessionOptions options;
unsetenv("OMP_NUM_THREADS");
@@ -46,7 +46,7 @@ TEST(MKLThreadPoolDeviceTest, TestOmpPreSets) {
EXPECT_EQ(omp_get_max_threads(), 314);
}
-#endif // _OPENMP
+#endif // defined(_OPENMP) && !defined(ENABLE_MKLDNN_THREADPOOL)
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/mkl_conv_ops.cc b/tensorflow/core/kernels/mkl_conv_ops.cc
index 59de3229211..2208401c7b3 100644
--- a/tensorflow/core/kernels/mkl_conv_ops.cc
+++ b/tensorflow/core/kernels/mkl_conv_ops.cc
@@ -1493,7 +1493,8 @@ class MklQuantizedConv2DOp
float max_output_value;
MklQuantizationRangeForMultiplication<Tinput, qint8, qint32>(
min_input, max_input, min_filter.flat<float>()(0),
- max_filter.flat<float>()(0), &min_output_value, &max_output_value);
+ max_filter.flat<float>()(0), &min_output_value, &max_output_value,
+ context);
AllocateOutputSetMklShape(context, 1, &output_min, {},
output_min_mkl_shape);
AllocateOutputSetMklShape(context, 2, &output_max, {},
@@ -1510,7 +1511,7 @@ class MklQuantizedConv2DOp
output_max_mkl_shape);
MklQuantizationRangeForMultiplication<Tinput, qint8, qint32>(
min_input, max_input, min_filter, max_filter, &output_min,
- &output_max);
+ &output_max, context);
}
}
}
diff --git a/tensorflow/core/kernels/mkl_conv_ops_test.cc b/tensorflow/core/kernels/mkl_conv_ops_test.cc
index a055351337c..9d11b0fb006 100644
--- a/tensorflow/core/kernels/mkl_conv_ops_test.cc
+++ b/tensorflow/core/kernels/mkl_conv_ops_test.cc
@@ -28,7 +28,7 @@ limitations under the License.
#include "tensorflow/core/public/session.h"
#if defined(INTEL_MKL_DNN_ONLY)
-#include "third_party/intel_mkl_dnn/include/mkldnn.h"
+#include "mkldnn.hpp"
#include "tensorflow/core/util/mkl_util.h"
#endif
diff --git a/tensorflow/core/kernels/mkl_qmatmul_op.cc b/tensorflow/core/kernels/mkl_qmatmul_op.cc
index cc7127e0559..382566c4ab5 100644
--- a/tensorflow/core/kernels/mkl_qmatmul_op.cc
+++ b/tensorflow/core/kernels/mkl_qmatmul_op.cc
@@ -91,12 +91,15 @@ limitations under the License.
// https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training
#ifdef INTEL_MKL
+#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/kernels/fill_functor.h"
#include "tensorflow/core/kernels/mkl_matmul_ops_common.h"
#include "tensorflow/core/kernels/mkl_quantized_conv_ops.h"
#include "tensorflow/core/kernels/no_op.h"
#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/util/mkl_threadpool.h"
+#include "tensorflow/core/util/work_sharder.h"
namespace {
enum {
@@ -342,7 +345,7 @@ class MklDnnQuantizedMatMulOp : public MklDnnMatMulOpBase<Tweight, Toutput> {
const float max_weight = context->input(6).flat<float>()(0);
MklQuantizationRangeForMultiplication<quint8, qint8, qint32>(
min_input, max_input, min_weight, max_weight, min_output_value,
- max_output_value);
+ max_output_value, context);
}
virtual void ExtendMklDnnMatMulFwdParams(OpKernelContext* context,
@@ -428,6 +431,26 @@ class MklDnnQuantizedMatMulOp : public MklDnnMatMulOpBase<Tweight, Toutput> {
((max_input - min_input) *
std::max(std::abs(max_weight), std::abs(min_weight)));
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ auto parallel_func = [&](int64 start, int64 end) {
+ for (int64 j = start ; j < end; j++) {
+ int x = 0;
+ for (int64 i = 0; i < k; ++i) {
+ x += wt_buf[i * n + j];
+ }
+ comp_bias[j] =
+ ((bias_buf[j] * out_scale) + static_cast<float>(x * qa_amin));
+ }
+ };
+
+ const float kArithCost = 2.5f;
+ const float kMovCost = 1.0f;
+ float shard_cost = 4*kArithCost + kMovCost;
+ const DeviceBase::CpuWorkerThreads& worker_threads =
+ *(context->device()->tensorflow_cpu_worker_threads());
+ Shard(worker_threads.num_threads, worker_threads.workers, n, shard_cost,
+ parallel_func);
+#else
#pragma omp parallel for schedule(static)
for (int j = 0; j < n; ++j) {
int x = 0;
@@ -437,7 +460,7 @@ class MklDnnQuantizedMatMulOp : public MklDnnMatMulOpBase<Tweight, Toutput> {
comp_bias[j] =
((bias_buf[j] * out_scale) + static_cast<float>(x * qa_amin));
}
-
+#endif // ENABLE_MKLDNN_THREADPOOL
return reinterpret_cast<Tbias*>(comp_bias_);
} else if (mode_ == QUANTIZE_MODE_SCALED) {
diff --git a/tensorflow/core/kernels/mkl_quantized_conv_ops.h b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
index fef2d837cf2..a95f8d29bfc 100644
--- a/tensorflow/core/kernels/mkl_quantized_conv_ops.h
+++ b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
@@ -39,7 +39,8 @@ float MklFloatForOneQuantizedLevel(float range_min, float range_max) {
template <class T1, class T2, class T3>
void MklQuantizationRangeForMultiplication(float min_a, float max_a,
float min_b, float max_b,
- float* min_c, float* max_c) {
+ float* min_c, float* max_c,
+ OpKernelContext* context) {
const float a_float_for_one_quant_level =
MklFloatForOneQuantizedLevel<T1>(min_a, max_a);
const float b_float_for_one_quant_level =
@@ -59,7 +60,8 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
const Tensor& min_b_vector,
const Tensor& max_b_vector,
Tensor** min_c_vector,
- Tensor** max_c_vector) {
+ Tensor** max_c_vector,
+ OpKernelContext* context) {
DCHECK(min_b_vector.NumElements() == (*min_c_vector)->NumElements());
DCHECK(max_b_vector.NumElements() == (*max_c_vector)->NumElements());
size_t n_channel = min_b_vector.NumElements();
@@ -69,6 +71,20 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
const float* max_b = max_b_vector.flat<float>().data();
float* min_c = (*min_c_vector)->flat<float>().data();
float* max_c = (*max_c_vector)->flat<float>().data();
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ auto eigen_tp =
+ MklDnnThreadPoolWrapper::GetInstance().CreateThreadPoolPtr(context);
+ eigen_tp->parallel_for(n_channel, [&](int n, int n_channel) {
+ float a_float_for_one_quant_level =
+ MklFloatForOneQuantizedLevel<T1>(min_a, max_a);
+ float b_float_for_one_quant_level =
+ MklFloatForOneQuantizedLevel<T2>(min_b[n], max_b[n]);
+ float c_float_for_one_quant_level =
+ a_float_for_one_quant_level * b_float_for_one_quant_level;
+ min_c[n] = c_float_for_one_quant_level * c_lowest;
+ max_c[n] = c_float_for_one_quant_level * c_highest;
+ });
+#else
#pragma omp parallel for
for (size_t n = 0; n < n_channel; ++n) {
float a_float_for_one_quant_level =
@@ -80,6 +96,7 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
min_c[n] = c_float_for_one_quant_level * c_lowest;
max_c[n] = c_float_for_one_quant_level * c_highest;
}
+#endif // ENABLE_MKLDNN_THREADPOOL
}
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
index 767a6f1c397..0a19573d901 100644
--- a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
+++ b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
@@ -28,6 +28,7 @@ limitations under the License.
#include "tensorflow/core/kernels/meta_support.h"
#include "tensorflow/core/kernels/no_op.h"
#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/util/mkl_threadpool.h"
#include "tensorflow/core/util/mkl_util.h"
namespace tensorflow {
@@ -73,6 +74,26 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
// Find the ranges of each channel in parallel.
float out_min_max = std::numeric_limits<float>::min();
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ // TODO: Add eigen parallel_for
+ for(size_t i = 0; i < depth; ++i) {
+ Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =
+ transposed_input.chip<0>(i).minimum();
+ Eigen::Tensor<qint32, 0, Eigen::RowMajor> max =
+ transposed_input.chip<0>(i).maximum();
+ const int32_t min_per_channel = min();
+ const int32_t max_per_channel = max();
+ const int32_t abs_max =
+ std::max(std::abs(min_per_channel), std::abs(max_per_channel));
+ float scale =
+ std::max(std::abs(input_min_data[i]), std::abs(input_max_data[i]));
+ ranges[i] =
+ scale * static_cast<float>(abs_max) / static_cast<float>(1L << 31);
+ if (min_per_channel < 0) is_non_negative = false;
+
+ out_min_max = std::max(out_min_max, ranges[i]);
+ }
+#else
#pragma omp parallel for reduction(max : out_min_max)
for (size_t i = 0; i < depth; ++i) {
Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =
@@ -92,6 +113,7 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
// Thread-local out_min_max.
out_min_max = std::max(out_min_max, ranges[i]);
}
+#endif // ENABLE_MKLDNN_THREADPOOL
// All local out_min_max gets max-reduced into one global out_min_max at
// the end of the loop by specifying reduction(max:out_min_max) along with
// omp parallel for.
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index 9a780839be3..5dc5877367b 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -354,9 +354,7 @@ def tf_copts(
)
def tf_openmp_copts():
- # TODO(intel-mkl): Remove -fopenmp for threadpool after removing all
- # omp pragmas in tensorflow/core.
- return if_mkl_lnx_x64(["-fopenmp"]) + if_mkldnn_threadpool(["-fopenmp"])
+ return (if_mkl_lnx_x64(["-fopenmp"]) + if_mkldnn_threadpool(["-fno-openmp"]))
def tfe_xla_copts():
return select({
From 4da2360572103d436f9873f50cef8c940dc963b7 Mon Sep 17 00:00:00 2001
From: sshiddib <sharada.shiddibhavi@intel.com>
Date: Tue, 26 May 2020 15:03:42 -0700
Subject: [PATCH 2/7] Removing OpenMP dependency from Mkl-dnn supporting
threadpool
---
.../mkl_threadpool_device_test.cc | 4 +--
tensorflow/core/kernels/mkl_conv_ops_test.cc | 2 +-
tensorflow/core/kernels/mkl_qmatmul_op.cc | 25 ++++++++++++++++++-
.../core/kernels/mkl_quantized_conv_ops.h | 14 +++++++++++
...mkl_requantization_range_per_channel_op.cc | 22 ++++++++++++++++
tensorflow/tensorflow.bzl | 4 +--
6 files changed, 64 insertions(+), 7 deletions(-)
diff --git a/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc b/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc
index 5d583a8360b..c29752d3c2c 100644
--- a/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc
+++ b/tensorflow/core/common_runtime/mkl_threadpool_device_test.cc
@@ -25,7 +25,7 @@ limitations under the License.
namespace tensorflow {
-#ifdef _OPENMP
+#if defined(_OPENMP) && !defined(ENABLE_MKLDNN_THREADPOOL)
TEST(MKLThreadPoolDeviceTest, TestOmpDefaults) {
SessionOptions options;
unsetenv("OMP_NUM_THREADS");
@@ -46,7 +46,7 @@ TEST(MKLThreadPoolDeviceTest, TestOmpPreSets) {
EXPECT_EQ(omp_get_max_threads(), 314);
}
-#endif // _OPENMP
+#endif // defined(_OPENMP) && !defined(ENABLE_MKLDNN_THREADPOOL)
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/mkl_conv_ops_test.cc b/tensorflow/core/kernels/mkl_conv_ops_test.cc
index a055351337c..9d11b0fb006 100644
--- a/tensorflow/core/kernels/mkl_conv_ops_test.cc
+++ b/tensorflow/core/kernels/mkl_conv_ops_test.cc
@@ -28,7 +28,7 @@ limitations under the License.
#include "tensorflow/core/public/session.h"
#if defined(INTEL_MKL_DNN_ONLY)
-#include "third_party/intel_mkl_dnn/include/mkldnn.h"
+#include "mkldnn.hpp"
#include "tensorflow/core/util/mkl_util.h"
#endif
diff --git a/tensorflow/core/kernels/mkl_qmatmul_op.cc b/tensorflow/core/kernels/mkl_qmatmul_op.cc
index cc7127e0559..d8bbc130c55 100644
--- a/tensorflow/core/kernels/mkl_qmatmul_op.cc
+++ b/tensorflow/core/kernels/mkl_qmatmul_op.cc
@@ -91,12 +91,15 @@ limitations under the License.
// https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training
#ifdef INTEL_MKL
+#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/kernels/fill_functor.h"
#include "tensorflow/core/kernels/mkl_matmul_ops_common.h"
#include "tensorflow/core/kernels/mkl_quantized_conv_ops.h"
#include "tensorflow/core/kernels/no_op.h"
#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/util/mkl_threadpool.h"
+#include "tensorflow/core/util/work_sharder.h"
namespace {
enum {
@@ -428,6 +431,26 @@ class MklDnnQuantizedMatMulOp : public MklDnnMatMulOpBase<Tweight, Toutput> {
((max_input - min_input) *
std::max(std::abs(max_weight), std::abs(min_weight)));
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ auto parallel_func = [&](int64 start, int64 end) {
+ for (int64 j = start ; j < end; j++) {
+ int x = 0;
+ for (int64 i = 0; i < k; ++i) {
+ x += wt_buf[i * n + j];
+ }
+ comp_bias[j] =
+ ((bias_buf[j] * out_scale) + static_cast<float>(x * qa_amin));
+ }
+ };
+
+ const float kArithCost = 2.5f;
+ const float kMovCost = 1.0f;
+ float shard_cost = 4*kArithCost + kMovCost;
+ const DeviceBase::CpuWorkerThreads& worker_threads =
+ *(context->device()->tensorflow_cpu_worker_threads());
+ Shard(worker_threads.num_threads, worker_threads.workers, n, shard_cost,
+ parallel_func);
+#else
#pragma omp parallel for schedule(static)
for (int j = 0; j < n; ++j) {
int x = 0;
@@ -437,7 +460,7 @@ class MklDnnQuantizedMatMulOp : public MklDnnMatMulOpBase<Tweight, Toutput> {
comp_bias[j] =
((bias_buf[j] * out_scale) + static_cast<float>(x * qa_amin));
}
-
+#endif // ENABLE_MKLDNN_THREADPOOL
return reinterpret_cast<Tbias*>(comp_bias_);
} else if (mode_ == QUANTIZE_MODE_SCALED) {
diff --git a/tensorflow/core/kernels/mkl_quantized_conv_ops.h b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
index fef2d837cf2..037a3a5f3ff 100644
--- a/tensorflow/core/kernels/mkl_quantized_conv_ops.h
+++ b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
@@ -69,6 +69,19 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
const float* max_b = max_b_vector.flat<float>().data();
float* min_c = (*min_c_vector)->flat<float>().data();
float* max_c = (*max_c_vector)->flat<float>().data();
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ // TODO: Add eigen parallel_for
+ for(size_t n = 0; n < n_channel; ++n) {
+ float a_float_for_one_quant_level =
+ MklFloatForOneQuantizedLevel<T1>(min_a, max_a);
+ float b_float_for_one_quant_level =
+ MklFloatForOneQuantizedLevel<T2>(min_b[n], max_b[n]);
+ float c_float_for_one_quant_level =
+ a_float_for_one_quant_level * b_float_for_one_quant_level;
+ min_c[n] = c_float_for_one_quant_level * c_lowest;
+ max_c[n] = c_float_for_one_quant_level * c_highest;
+ }
+#else
#pragma omp parallel for
for (size_t n = 0; n < n_channel; ++n) {
float a_float_for_one_quant_level =
@@ -80,6 +93,7 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
min_c[n] = c_float_for_one_quant_level * c_lowest;
max_c[n] = c_float_for_one_quant_level * c_highest;
}
+#endif // ENABLE_MKLDNN_THREADPOOL
}
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
index 767a6f1c397..0a19573d901 100644
--- a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
+++ b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
@@ -28,6 +28,7 @@ limitations under the License.
#include "tensorflow/core/kernels/meta_support.h"
#include "tensorflow/core/kernels/no_op.h"
#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/util/mkl_threadpool.h"
#include "tensorflow/core/util/mkl_util.h"
namespace tensorflow {
@@ -73,6 +74,26 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
// Find the ranges of each channel in parallel.
float out_min_max = std::numeric_limits<float>::min();
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ // TODO: Add eigen parallel_for
+ for(size_t i = 0; i < depth; ++i) {
+ Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =
+ transposed_input.chip<0>(i).minimum();
+ Eigen::Tensor<qint32, 0, Eigen::RowMajor> max =
+ transposed_input.chip<0>(i).maximum();
+ const int32_t min_per_channel = min();
+ const int32_t max_per_channel = max();
+ const int32_t abs_max =
+ std::max(std::abs(min_per_channel), std::abs(max_per_channel));
+ float scale =
+ std::max(std::abs(input_min_data[i]), std::abs(input_max_data[i]));
+ ranges[i] =
+ scale * static_cast<float>(abs_max) / static_cast<float>(1L << 31);
+ if (min_per_channel < 0) is_non_negative = false;
+
+ out_min_max = std::max(out_min_max, ranges[i]);
+ }
+#else
#pragma omp parallel for reduction(max : out_min_max)
for (size_t i = 0; i < depth; ++i) {
Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =
@@ -92,6 +113,7 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
// Thread-local out_min_max.
out_min_max = std::max(out_min_max, ranges[i]);
}
+#endif // ENABLE_MKLDNN_THREADPOOL
// All local out_min_max gets max-reduced into one global out_min_max at
// the end of the loop by specifying reduction(max:out_min_max) along with
// omp parallel for.
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index 9a780839be3..5dc5877367b 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -354,9 +354,7 @@ def tf_copts(
)
def tf_openmp_copts():
- # TODO(intel-mkl): Remove -fopenmp for threadpool after removing all
- # omp pragmas in tensorflow/core.
- return if_mkl_lnx_x64(["-fopenmp"]) + if_mkldnn_threadpool(["-fopenmp"])
+ return (if_mkl_lnx_x64(["-fopenmp"]) + if_mkldnn_threadpool(["-fno-openmp"]))
def tfe_xla_copts():
return select({
From 6ae6ef636d608822696508e9535cdb39d547a0c6 Mon Sep 17 00:00:00 2001
From: Sharada Shiddibhavi <sharada.shiddibhavi@intel.com>
Date: Mon, 8 Jun 2020 10:02:22 -0700
Subject: [PATCH 3/7] Addressing review comments
---
.../core/kernels/mkl_quantized_conv_ops.h | 20 +++++--------------
1 file changed, 5 insertions(+), 15 deletions(-)
diff --git a/tensorflow/core/kernels/mkl_quantized_conv_ops.h b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
index 037a3a5f3ff..37022f46113 100644
--- a/tensorflow/core/kernels/mkl_quantized_conv_ops.h
+++ b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
@@ -69,21 +69,12 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
const float* max_b = max_b_vector.flat<float>().data();
float* min_c = (*min_c_vector)->flat<float>().data();
float* max_c = (*max_c_vector)->flat<float>().data();
-#ifdef ENABLE_MKLDNN_THREADPOOL
- // TODO: Add eigen parallel_for
- for(size_t n = 0; n < n_channel; ++n) {
- float a_float_for_one_quant_level =
- MklFloatForOneQuantizedLevel<T1>(min_a, max_a);
- float b_float_for_one_quant_level =
- MklFloatForOneQuantizedLevel<T2>(min_b[n], max_b[n]);
- float c_float_for_one_quant_level =
- a_float_for_one_quant_level * b_float_for_one_quant_level;
- min_c[n] = c_float_for_one_quant_level * c_lowest;
- max_c[n] = c_float_for_one_quant_level * c_highest;
- }
-#else
+
+#ifndef ENABLE_MKLDNN_THREADPOOL
#pragma omp parallel for
- for (size_t n = 0; n < n_channel; ++n) {
+#endif // ENABLE_MKLDNN_THREADPOOL
+ // TODO: Add eigen parallel_for
+ for (size_t n = 0; n < n_channel; ++n) {
float a_float_for_one_quant_level =
MklFloatForOneQuantizedLevel<T1>(min_a, max_a);
float b_float_for_one_quant_level =
@@ -93,7 +84,6 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
min_c[n] = c_float_for_one_quant_level * c_lowest;
max_c[n] = c_float_for_one_quant_level * c_highest;
}
-#endif // ENABLE_MKLDNN_THREADPOOL
}
} // namespace tensorflow
From f32fcb56ffd2f6cbd362e5accab39b264a7d169f Mon Sep 17 00:00:00 2001
From: Sharada Shiddibhavi <sharada.shiddibhavi@intel.com>
Date: Mon, 8 Jun 2020 10:05:49 -0700
Subject: [PATCH 4/7] Addressing review comments
---
...mkl_requantization_range_per_channel_op.cc | 24 ++++---------------
1 file changed, 4 insertions(+), 20 deletions(-)
diff --git a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
index 0a19573d901..3aa744c0a5e 100644
--- a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
+++ b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
@@ -74,27 +74,11 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
// Find the ranges of each channel in parallel.
float out_min_max = std::numeric_limits<float>::min();
-#ifdef ENABLE_MKLDNN_THREADPOOL
- // TODO: Add eigen parallel_for
- for(size_t i = 0; i < depth; ++i) {
- Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =
- transposed_input.chip<0>(i).minimum();
- Eigen::Tensor<qint32, 0, Eigen::RowMajor> max =
- transposed_input.chip<0>(i).maximum();
- const int32_t min_per_channel = min();
- const int32_t max_per_channel = max();
- const int32_t abs_max =
- std::max(std::abs(min_per_channel), std::abs(max_per_channel));
- float scale =
- std::max(std::abs(input_min_data[i]), std::abs(input_max_data[i]));
- ranges[i] =
- scale * static_cast<float>(abs_max) / static_cast<float>(1L << 31);
- if (min_per_channel < 0) is_non_negative = false;
- out_min_max = std::max(out_min_max, ranges[i]);
- }
-#else
+#ifndef ENABLE_MKLDNN_THREADPOOL
#pragma omp parallel for reduction(max : out_min_max)
+#endif // ENABLE_MKLDNN_THREADPOOL
+ // TODO: Add eigen parallel_for
for (size_t i = 0; i < depth; ++i) {
Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =
transposed_input.chip<0>(i).minimum();
@@ -113,7 +97,7 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
// Thread-local out_min_max.
out_min_max = std::max(out_min_max, ranges[i]);
}
-#endif // ENABLE_MKLDNN_THREADPOOL
+
// All local out_min_max gets max-reduced into one global out_min_max at
// the end of the loop by specifying reduction(max:out_min_max) along with
// omp parallel for.
From e2b4994f450f5e08c3718bcd8fc0e8c3cf984780 Mon Sep 17 00:00:00 2001
From: Sharada Shiddibhavi <sharada.shiddibhavi@intel.com>
Date: Mon, 8 Jun 2020 10:07:11 -0700
Subject: [PATCH 5/7] Update mkl_quantized_conv_ops.h
---
tensorflow/core/kernels/mkl_quantized_conv_ops.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/tensorflow/core/kernels/mkl_quantized_conv_ops.h b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
index 37022f46113..442f6a53047 100644
--- a/tensorflow/core/kernels/mkl_quantized_conv_ops.h
+++ b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
@@ -73,8 +73,8 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
#ifndef ENABLE_MKLDNN_THREADPOOL
#pragma omp parallel for
#endif // ENABLE_MKLDNN_THREADPOOL
- // TODO: Add eigen parallel_for
- for (size_t n = 0; n < n_channel; ++n) {
+ // TODO: Add eigen parallel_for
+ for (size_t n = 0; n < n_channel; ++n) {
float a_float_for_one_quant_level =
MklFloatForOneQuantizedLevel<T1>(min_a, max_a);
float b_float_for_one_quant_level =
From 7fc6bf6726cee3f6178caaafeedb7a98f776cb02 Mon Sep 17 00:00:00 2001
From: Sharada Shiddibhavi <sharada.shiddibhavi@intel.com>
Date: Mon, 8 Jun 2020 10:30:03 -0700
Subject: [PATCH 6/7] Update tensorflow/core/kernels/mkl_quantized_conv_ops.h
Co-authored-by: Penporn Koanantakool <38085909+penpornk@users.noreply.github.com>
---
tensorflow/core/kernels/mkl_quantized_conv_ops.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tensorflow/core/kernels/mkl_quantized_conv_ops.h b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
index 442f6a53047..4121c88fb83 100644
--- a/tensorflow/core/kernels/mkl_quantized_conv_ops.h
+++ b/tensorflow/core/kernels/mkl_quantized_conv_ops.h
@@ -72,7 +72,7 @@ void MklQuantizationRangeForMultiplication(float min_a, float max_a,
#ifndef ENABLE_MKLDNN_THREADPOOL
#pragma omp parallel for
-#endif // ENABLE_MKLDNN_THREADPOOL
+#endif // !ENABLE_MKLDNN_THREADPOOL
// TODO: Add eigen parallel_for
for (size_t n = 0; n < n_channel; ++n) {
float a_float_for_one_quant_level =
From 7b4b07c098170bd891f2426ee9c043249ac41983 Mon Sep 17 00:00:00 2001
From: Sharada Shiddibhavi <sharada.shiddibhavi@intel.com>
Date: Mon, 8 Jun 2020 10:31:00 -0700
Subject: [PATCH 7/7] Update mkl_requantization_range_per_channel_op.cc
---
.../core/kernels/mkl_requantization_range_per_channel_op.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
index 3aa744c0a5e..a43f6a71acf 100644
--- a/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
+++ b/tensorflow/core/kernels/mkl_requantization_range_per_channel_op.cc
@@ -77,7 +77,7 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
#ifndef ENABLE_MKLDNN_THREADPOOL
#pragma omp parallel for reduction(max : out_min_max)
-#endif // ENABLE_MKLDNN_THREADPOOL
+#endif // !ENABLE_MKLDNN_THREADPOOL
// TODO: Add eigen parallel_for
for (size_t i = 0; i < depth; ++i) {
Eigen::Tensor<qint32, 0, Eigen::RowMajor> min =