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

As we completed the move from gemmlowp direct Gemm calls to cpu_backend_gemm,

Browse Source 
the GemmlowpOutputPipelines are no longer used outside of legacy code, so move them there.

PiperOrigin-RevId: 246650849
This commit is contained in:
Benoit Jacob 2019-05-04 07:24:48 -07:00 committed by TensorFlower Gardener
parent 8fa823c345
commit f9cdb743b4
3 changed files with 56 additions and 84 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
tensorflow/lite/kernels/internal/optimized/integer_ops/fully_connected.h
View File

@ -365,62 +365,6 @@ inline void FullyConnectedAsGEMV(
}
#endif // USE_NEON
struct GemmlowpOutputPipeline {
typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
ColVectorMap;
typedef std::tuple<gemmlowp::OutputStageBiasAddition<ColVectorMap>,
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent,
gemmlowp::OutputStageClamp,
gemmlowp::OutputStageSaturatingCastToInt8>
Pipeline;
static Pipeline MakeExp(const int32* bias_data, int output_rows,
int32 output_offset, int32 output_multiplier,
int output_left_shift, int32 output_activation_min,
int32 output_activation_max) {
ColVectorMap bias_vector(bias_data, output_rows);
gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
bias_addition_stage.bias_vector = bias_vector;
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent quantize_down_stage;
quantize_down_stage.result_offset_after_shift = output_offset;
quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
quantize_down_stage.result_exponent = output_left_shift;
gemmlowp::OutputStageClamp clamp_stage;
clamp_stage.min = output_activation_min;
clamp_stage.max = output_activation_max;
gemmlowp::OutputStageSaturatingCastToInt8 saturating_cast_stage;
return std::make_tuple(bias_addition_stage, quantize_down_stage,
clamp_stage, saturating_cast_stage);
}
};
struct GemmlowpOutputPipelineInt8 {
typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
ColVectorMap;
typedef std::tuple<gemmlowp::OutputStageBiasAddition<ColVectorMap>,
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent,
gemmlowp::OutputStageClamp,
gemmlowp::OutputStageSaturatingCastToInt8>
Pipeline;
static Pipeline MakeExp(const int32* bias_data, int output_rows,
int32 output_offset, int32 output_multiplier,
int output_left_shift, int32 output_activation_min,
int32 output_activation_max) {
ColVectorMap bias_vector(bias_data, output_rows);
gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
bias_addition_stage.bias_vector = bias_vector;
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent quantize_down_stage;
quantize_down_stage.result_offset_after_shift = output_offset;
quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
quantize_down_stage.result_exponent = output_left_shift;
gemmlowp::OutputStageClamp clamp_stage;
clamp_stage.min = output_activation_min;
clamp_stage.max = output_activation_max;
gemmlowp::OutputStageSaturatingCastToInt8 saturating_cast_stage;
return std::make_tuple(bias_addition_stage, quantize_down_stage,
clamp_stage, saturating_cast_stage);
}
};
inline void FullyConnected(
const FullyConnectedParams& params, const RuntimeShape& input_shape,
const int8* input_data, const RuntimeShape& filter_shape,

56
tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h
View File

@ -598,6 +598,62 @@ void FullyConnected(const float* input_data, const Dims<4>& input_dims,
output_data, output_dims);
}
struct GemmlowpOutputPipeline {
typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
ColVectorMap;
typedef std::tuple<gemmlowp::OutputStageBiasAddition<ColVectorMap>,
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent,
gemmlowp::OutputStageClamp,
gemmlowp::OutputStageSaturatingCastToUint8>
Pipeline;
static Pipeline MakeExp(const int32* bias_data, int output_rows,
int32 output_offset, int32 output_multiplier,
int output_left_shift, int32 output_activation_min,
int32 output_activation_max) {
ColVectorMap bias_vector(bias_data, output_rows);
gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
bias_addition_stage.bias_vector = bias_vector;
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent quantize_down_stage;
quantize_down_stage.result_offset_after_shift = output_offset;
quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
quantize_down_stage.result_exponent = output_left_shift;
gemmlowp::OutputStageClamp clamp_stage;
clamp_stage.min = output_activation_min;
clamp_stage.max = output_activation_max;
gemmlowp::OutputStageSaturatingCastToUint8 saturating_cast_stage;
return std::make_tuple(bias_addition_stage, quantize_down_stage,
clamp_stage, saturating_cast_stage);
}
};
struct GemmlowpOutputPipelineInt8 {
typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
ColVectorMap;
typedef std::tuple<gemmlowp::OutputStageBiasAddition<ColVectorMap>,
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent,
gemmlowp::OutputStageClamp,
gemmlowp::OutputStageSaturatingCastToInt8>
Pipeline;
static Pipeline MakeExp(const int32* bias_data, int output_rows,
int32 output_offset, int32 output_multiplier,
int output_left_shift, int32 output_activation_min,
int32 output_activation_max) {
ColVectorMap bias_vector(bias_data, output_rows);
gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
bias_addition_stage.bias_vector = bias_vector;
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent quantize_down_stage;
quantize_down_stage.result_offset_after_shift = output_offset;
quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
quantize_down_stage.result_exponent = output_left_shift;
gemmlowp::OutputStageClamp clamp_stage;
clamp_stage.min = output_activation_min;
clamp_stage.max = output_activation_max;
gemmlowp::OutputStageSaturatingCastToInt8 saturating_cast_stage;
return std::make_tuple(bias_addition_stage, quantize_down_stage,
clamp_stage, saturating_cast_stage);
}
};
#ifdef USE_NEON
struct LegacyFullyConnectedAsGEMVWorkerTask : public gemmlowp::Task {
LegacyFullyConnectedAsGEMVWorkerTask(

28
tensorflow/lite/kernels/internal/optimized/optimized_ops.h
View File

@ -1148,34 +1148,6 @@ inline void FullyConnectedAsGEMV(
}
#endif // USE_NEON
struct GemmlowpOutputPipeline {
typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
ColVectorMap;
typedef std::tuple<gemmlowp::OutputStageBiasAddition<ColVectorMap>,
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent,
gemmlowp::OutputStageClamp,
gemmlowp::OutputStageSaturatingCastToUint8>
Pipeline;
static Pipeline MakeExp(const int32* bias_data, int output_rows,
int32 output_offset, int32 output_multiplier,
int output_left_shift, int32 output_activation_min,
int32 output_activation_max) {
ColVectorMap bias_vector(bias_data, output_rows);
gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
bias_addition_stage.bias_vector = bias_vector;
gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent quantize_down_stage;
quantize_down_stage.result_offset_after_shift = output_offset;
quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
quantize_down_stage.result_exponent = output_left_shift;
gemmlowp::OutputStageClamp clamp_stage;
clamp_stage.min = output_activation_min;
clamp_stage.max = output_activation_max;
gemmlowp::OutputStageSaturatingCastToUint8 saturating_cast_stage;
return std::make_tuple(bias_addition_stage, quantize_down_stage,
clamp_stage, saturating_cast_stage);
}
};
inline void FullyConnected(
const FullyConnectedParams& params, const RuntimeShape& input_shape,
const uint8* input_data, const RuntimeShape& filter_shape,