#include <algorithm>
#include <cmath>
#include <cstdint>
#include <cstring>
#include <limits>
#include <utility>
#include "fixedpoint/fixedpoint.h"
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/compatibility.h"
#include "tensorflow/lite/kernels/internal/cppmath.h"
#include "tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h"
#if defined(_MSC_VER)
#define __restrict__ __restrict
#endif
namespace tflite {
namespace tensor_utils {
namespace {
const int32_t kInt16Max = …;
const int32_t kInt16Min = …;
}
void PortableSymmetricQuantizeFloats(const float* values, const int size,
int8_t* quantized_values, float* min_value,
float* max_value, float* scaling_factor) { … }
void PortableSymmetricQuantizeFloats(const float* values, const int size,
int8_t* quantized_values, float min_value,
float max_value, float* scaling_factor) { … }
void PortableAsymmetricQuantizeFloats(const float* values, const int size,
int8_t* quantized_values,
float* scaling_factor, int32_t* offset) { … }
void PortableMatrixBatchVectorMultiplyAccumulate(const float* matrix,
int m_rows, int m_cols,
const float* vector,
int n_batch, float* result) { … }
void PortableMatrixBatchVectorMultiplyAccumulate(
const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
const int8_t* __restrict__ vectors, const float* scaling_factors,
int n_batch, float* __restrict__ result) { … }
void PortableMatrixBatchVectorMultiplyAccumulate(
const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
const int8_t* __restrict__ vectors, const float* scaling_factors,
int n_batch, float* __restrict__ result, const float* per_channel_scale,
const int32_t* input_offset, int32_t* scratch, int32_t* row_sums,
bool* compute_row_sums, CpuBackendContext* context) { … }
void PortableSparseMatrixBatchVectorMultiplyAccumulate1x4(
const float* __restrict__ matrix, const int32_t* __restrict__ segments,
const int32_t* __restrict__ indices, int m_rows, int m_cols,
const float* __restrict__ vector, int n_batch, float* __restrict__ result) { … }
void PortableSparseMatrixBatchVectorMultiplyAccumulate1x16(
const int8_t* __restrict__ matrix, const int32_t* __restrict__ segments,
const int32_t* __restrict__ indices, int m_rows, int m_cols,
const int8_t* __restrict__ vector, const int32_t* __restrict__ bias_vector,
int n_batch, const int32_t input_offset, const int32_t output_multiplier,
const int32_t output_shift, const int32_t* per_channel_scale,
const int32_t* per_channel_shift, const int32_t output_offset,
const int32_t output_activation_min, const int32_t output_activation_max,
int8_t* __restrict__ result) { … }
void PortableSparseMatrixBatchVectorMultiplyAccumulate(
const float* __restrict__ matrix, const uint8_t* __restrict__ ledger,
int m_rows, int m_cols, const float* __restrict__ vector, int n_batch,
float* __restrict__ result) { … }
void PortableSparseMatrixBatchVectorMultiplyAccumulate(
const int8_t* __restrict__ matrix, const uint8_t* ledger, const int m_rows,
const int m_cols, const int8_t* __restrict__ vectors,
const float* scaling_factors, int n_batch, float* __restrict__ result,
const float* per_channel_scale) { … }
template <typename T>
void PortableMatrixBatchVectorMultiplyAccumulateImpl(
const int8_t* input, const int32_t* bias,
const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
T* output) { … }
void PortableMatrixBatchVectorMultiplyAccumulate(
const int8_t* input, const int32_t* bias,
const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
int32_t* scratch, int16_t* output, CpuBackendContext* context) { … }
void PortableMatrixBatchVectorMultiplyAccumulate(
const int8_t* input, const int32_t* bias,
const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
int32_t* scratch, int8_t* output, CpuBackendContext* context) { … }
void PortableMatrixBatchVectorMultiply(const int8_t* input,
int32_t input_zeropoint,
const int8_t* input_to_gate_weights,
int32_t input_to_gate_effective_scale_a,
int32_t input_to_gate_effective_scale_b,
int32_t n_batch, int32_t n_input,
int32_t n_cell, int8_t* gate_output,
int8_t gate_output_zp) { … }
void PortableMatrixBatchVectorMultiply(
const int16_t* hidden, const int8_t* hidden_to_output_weights,
int32_t proj_effective_scale_a, int32_t proj_effective_scale_b,
const int32_t* gate_bias, int32_t n_batch, int32_t n_hidden,
int32_t n_output, int32_t output_zp, int8_t* proj_output) { … }
void PortableApplyLayerNorm(const int16_t* input,
const int16_t* layer_norm_weights,
const int32_t* bias, int32_t layer_norm_scale_a,
int32_t layer_norm_scale_b, int32_t variance_limit,
int n_batch, int n_input, int16_t* output) { … }
void PortableApplyLayerNormFloat(const int16_t* input,
const int16_t* layer_norm_weights,
int32_t layer_norm_scale_a,
int32_t layer_norm_scale_b,
const int32_t* bias, int n_batch, int n_input,
int16_t* output) { … }
void PortableMatrixScalarMultiplyAccumulate(const int8_t* matrix,
int32_t scalar, int32_t n_row,
int32_t n_col, int32_t* output) { … }
void PortableApplySigmoid(const int16_t* input, int32_t n_batch,
int32_t n_input, int16_t* output) { … }
void PortableApplySigmoidFloat(const int16_t* input, int32_t n_batch,
int32_t n_input, int16_t* output) { … }
template <int IntegerBits>
void PortableApplyTanhImpl(const int16_t* input, int32_t n_batch,
int32_t n_input, int16_t* output) { … }
void PortableApplyTanh(int32_t integer_bits, const int16_t* input,
int32_t n_batch, int32_t n_input, int16_t* output) { … }
void PortableApplyTanhFloat(const int16_t* input, int32_t n_batch,
int32_t n_input, int32_t integer_bits,
int16_t* output) { … }
void PortableCwiseMul(const int16_t* input_1, const int16_t* input_2,
int n_batch, int n_input, int shift, int16_t* output) { … }
void PortableCwiseMul(const int16_t* input_1, const int16_t* input_2,
int32_t multiplier, int32_t shift, int32_t n_batch,
int32_t n_input, int32_t output_zp, int8_t* output) { … }
void PortableCwiseAdd(const int16_t* input_1, const int16_t* input_2,
int n_batch, int n_input, int16_t* output) { … }
float PortableVectorVectorDotProduct(const float* vector1, const float* vector2,
int v_size) { … }
namespace {
inline int32_t VectorVectorDotProduct(const int16_t* vector1,
const int16_t* vector2, int v_size) { … }
}
void PortableBatchVectorBatchVectorDotProduct(const int16_t* vector1,
const int16_t* vector2,
int v_size, int n_batch,
int32_t* result) { … }
void PortableVectorBatchVectorCwiseProductAccumulate(
const int16_t* vector, int v_size, const int16_t* batch_vector, int n_batch,
int32_t multiplier, int shift, int16_t* result) { … }
void PortableSub1Vector(const float* vector, int v_size, float* result) { … }
void PortableSub1Vector(const int16_t* vector, int v_size, int16_t* result) { … }
void PortableVectorScalarMultiply(const int8_t* vector, const int v_size,
const float scale, float* result) { … }
void PortableMeanStddevNormalization(const float* __restrict__ input_vector,
float* __restrict__ output_vector,
int v_size, int n_batch) { … }
void PortableTwoGateSaturatingAdd(const int8_t* input, int8_t input_zp,
const int8_t* recurrent, int8_t recurrent_zp,
int32_t input_effective_scale_a,
int32_t input_effective_scale_b,
int32_t recurrent_effective_scale_a,
int32_t recurrent_effective_scale_b,
int32_t n_batch, int32_t n_cell,
int16_t* output) { … }
}
}