diff options
| author |  A. Unique TensorFlower <gardener@tensorflow.org> | 2017-07-18 08:47:06 -0700 |
|---|---|---|
| committer | TensorFlower Gardener <gardener@tensorflow.org> | 2017-07-18 08:51:16 -0700 |
| commit | bdb2967a298236e24011405907cd19737386934e (patch) | |
| tree | dcff69670374af7fd3c30cee2161d791153c7654 | |
| parent | 0a9ce966098fd2e1039daefd2960a9135a567cf9 (diff) | |
Fix metagemm calls in quantized ops. Only use metagemm multiplication for
k <= 2048.
PiperOrigin-RevId: 162356878
| -rw-r--r-- | tensorflow/core/kernels/quantized_conv_ops.cc | 2 | ||||
| -rw-r--r-- | tensorflow/core/kernels/quantized_matmul_op.cc | 2 |
2 files changed, 2 insertions, 2 deletions
diff --git a/tensorflow/core/kernels/quantized_conv_ops.cc b/tensorflow/core/kernels/quantized_conv_ops.cc index 5658dcc069..3b0764bb9b 100644 --- a/tensorflow/core/kernels/quantized_conv_ops.cc +++ b/tensorflow/core/kernels/quantized_conv_ops.cc @@ -381,7 +381,7 @@ class Im2ColConvFunctor { 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)) { + (transpose_c == false) && (k <= 2048)) { meta::QuantizedGemm(context, transpose_a, transpose_b, im2col_buffer, filter_data, chunk_output_data, m, n, k, -input_offset, -filter_offset, lda, ldb, ldc); diff --git a/tensorflow/core/kernels/quantized_matmul_op.cc b/tensorflow/core/kernels/quantized_matmul_op.cc index 86c722e587..afb30d5f62 100644 --- a/tensorflow/core/kernels/quantized_matmul_op.cc +++ b/tensorflow/core/kernels/quantized_matmul_op.cc @@ -135,7 +135,7 @@ class QuantizedMatMulOp : public OpKernel { if (meta::IsSupportedAndEnabled() && std::is_same<T1, quint8>() && std::is_same<T2, quint8>() && std::is_same<Toutput, qint32>() && (offset_c == 0) && (mult_c == 1) && (shift_c == 0) && - (transpose_c == false)) { + (transpose_c == false) && (k <= 2048)) { // Gemmlowp/meta code path works on 32 & 64 bit Arm with NEON Simd and // allows optimized quantized 8bit to 32bit gemm. meta::QuantizedGemm(context, transpose_a_, transpose_b_, a_data, b_data, |