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

Disable padded MatrixBatchVectorMultiply with sdot

Browse Source 
PiperOrigin-RevId: 358324561
Change-Id: I2ba23bf11c7b200e49cee1cdff096c3521f12e51
This commit is contained in:
David Rim 2021-02-18 20:25:36 -08:00 committed by TensorFlower Gardener
parent 3314df4945
commit 2abdad6ab1
1 changed files with 8 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
tensorflow/lite/kernels/internal/optimized/neon_tensor_utils.cc
View File

@ -60,6 +60,10 @@ limitations under the License.
#define TFLITE_UNLIKELY(x) (x)
#endif
// TODO(b/180650471): Add back padded version of
// MatrixBatchVectorMultiplyAccumulate with sdot instruction.
#define ENABLE_PADDED_DOT_PROD false
namespace tflite {
namespace tensor_utils {
namespace {
@ -68,7 +72,6 @@ constexpr int kFloatValuesPerNeonVector = 4;
constexpr int kInt16ValuesPerNeonVector = 8;
constexpr int kInt8ValuesPerNeonVector = 16;
constexpr int kNeonVectorAlignment = 4;
template <int PerNeonSize>
inline int RoundDownVectors(int size) {
return size & ~(PerNeonSize - 1);
@ -1054,7 +1057,8 @@ void NeonMatrixBatchVectorMultiplyAccumulate(const int8_t* __restrict__ matrix,
DotprodMatrixBatchFourVectorMultiplyAccumulate(
matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result);
return;
} else if (n_batch >= 2 && m_rows * m_cols >= 128 * 128) {
} else if (ENABLE_PADDED_DOT_PROD && n_batch >= 2 &&
m_rows * m_cols >= 128 * 128) {
DotprodMatrixBatchPaddedFourVectorMultiplyAccumulate(
matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result);
return;
@ -1252,7 +1256,8 @@ void NeonMatrixBatchVectorMultiplyAccumulateImpl(
matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result,
per_channel_scale, input_offset, row_sums);
return;
} else if (n_batch >= 2 && m_rows * m_cols >= 128 * 128) {
} else if (ENABLE_PADDED_DOT_PROD && n_batch >= 2 &&
m_rows * m_cols >= 128 * 128) {
DotprodMatrixBatchPaddedFourVectorMultiplyAccumulate(
matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result,
per_channel_scale, input_offset, row_sums);