Arm32/64 kernel optimizations:

- QuantizeV2
- Dequantize
- QuantizedBiasAdd
- QuantizeDownAndShrinkRange
- QuantizedRelu
- QuantizedRelu6
- QuantizedMatMul
- QuantizedConv

The optimizations are controled by three knobs:

meta::SetEnabled(bool)         -- turns codepath on/off, on by default
meta::SetUseLocalContext(bool) -- true    -- codepath will use it's own internal fine grain
                                             workers pool that offers performance improvement
                                             over the standard tensorflow worker pool. This
                                             workers pool is not compatible with other ops.
                                             Per use-case performance testing recommended.
                               -- false (default) -- use the standard tf worker pool instance
meta::SetNumThreads(int)       -- no. of compute threads when the internal worker pool is used.
                                  If 0 use intra_parallelism_count, if x > 0 then x threads.
Change: 137197327

1 files changed, 19 insertions, 8 deletions

diff --git a/tensorflow/core/kernels/quantized_conv_ops.cc b/tensorflow/core/kernels/quantized_conv_ops.cc
index fb69d770c0..2405c55c5b 100644
--- a/tensorflow/core/kernels/quantized_conv_ops.cc
+++ b/tensorflow/core/kernels/quantized_conv_ops.cc
@@ -18,12 +18,15 @@ limitations under the License.
 #include <algorithm>
 #include <vector>
 
+#define EIGEN_USE_THREADS
+
 #include "public/gemmlowp.h"
-#include "tensorflow/core/kernels/quantization_utils.h"
-#include "tensorflow/core/kernels/reference_gemm.h"
 #include "tensorflow/core/framework/op_kernel.h"
 #include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/kernels/meta_support.h"
 #include "tensorflow/core/kernels/ops_util.h"
+#include "tensorflow/core/kernels/quantization_utils.h"
+#include "tensorflow/core/kernels/reference_gemm.h"
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/util/padding.h"
 
@@ -338,12 +341,20 @@ class Im2ColConvFunctor {
     const int lda = filter_value_count;
     const int ldb = filter_count;
     const int ldc = filter_count;
-    // The gemmlowp optimized library only works for a particular set of data
-    // types, so check if we meet those requirements and
-    // fall back to a slower reference implementation if not.
-    if (std::is_same<T1, quint8>() && std::is_same<T2, quint8>() &&
-        std::is_same<T3, qint32>() && (output_offset == 0) &&
-        (output_mult == 1) && (output_shift == 0)) {
+
+    if (meta::IsSupportedAndEnabled() && std::is_same<T1, quint8>() &&
+        std::is_same<T2, quint8>() && std::is_same<T3, qint32>() &&
+        (output_offset == 0) && (output_mult == 1) && (output_shift == 0) &&
+        (transpose_c == false)) {
+      meta::QuantizedGemm(op_context, transpose_a, transpose_b,
+                          im2col_buffer.get(), filter_data, output_data, m, n,
+                          k, -input_offset, -filter_offset, lda, ldb, ldc);
+    } else if (std::is_same<T1, quint8>() && std::is_same<T2, quint8>() &&
+               std::is_same<T3, qint32>() && (output_offset == 0) &&
+               (output_mult == 1) && (output_shift == 0)) {
+      // The gemmlowp optimized library only works for a particular set of data
+      // types, so check if we meet those requirements and
+      // fall back to a slower reference implementation if not.
       const uint8* im2col_data_as_uint8 = &(im2col_buffer.get()->value);
       const uint8* filter_data_as_uint8 = &(filter_data->value);
       int32* output_data_as_int32 = &(output_data->value);