10 void ApplyLayerNorm(const int16_t* input, const int16_t* layer_norm_weights, const int32_t* bias,  in ApplyLayerNorm()
11                     int32_t layer_norm_scale_a, int32_t layer_norm_scale_b, int32_t variance_limit,  in ApplyLayerNorm()
18             const int32_t index = i * n_input + j;  in ApplyLayerNorm()
19             int32_t val = static_cast<int32_t>(input[index]);  in ApplyLayerNorm()
23         int32_t mean = static_cast<int32_t>(static_cast<int64_t>(sum) * 1024 / n_input);  in ApplyLayerNorm()
25         int32_t temp = kOverflowGuard / n_input;  in ApplyLayerNorm()
27         int32_t variance2 = static_cast<int32_t>(variance / kOverflowGuard);  in ApplyLayerNorm()
31         int32_t stddev_inverse_a;  in ApplyLayerNorm()
37             const int32_t index = i * n_input + j;  in ApplyLayerNorm()
38             int32_t val = static_cast<int32_t>(input[index]);  in ApplyLayerNorm()
39             int32_t shifted = 1024 * val - mean;  in ApplyLayerNorm()
40             int32_t rescaled =  in ApplyLayerNorm()
44             int32_t val4 = static_cast<int32_t>((val3 > 0 ? val3 + 512 : val3 - 512) / 1024);  in ApplyLayerNorm()
45             int32_t val5 = MultiplyByQuantizedMultiplier(val4, layer_norm_scale_a,  in ApplyLayerNorm()
53 void MatrixScalarMultiplyAccumulate(const int8_t* matrix, int32_t scalar, int32_t n_row,  in MatrixScalarMultiplyAccumulate()
54                                     int32_t n_col, int32_t* output) {  in MatrixScalarMultiplyAccumulate()
56         int32_t row_sum = 0;  in MatrixScalarMultiplyAccumulate()
64 bool PrecomputeZeroPointTimesWeightWithBias(int32_t zero_point, const int8_t* weight_tensor,  in PrecomputeZeroPointTimesWeightWithBias()
65                                             const Shape& weight_shape, const int32_t* bias_tensor,  in PrecomputeZeroPointTimesWeightWithBias()
66                                             std::unique_ptr<int32_t[]>* output) {  in PrecomputeZeroPointTimesWeightWithBias()
74     *output = std::make_unique<int32_t[]>(row);  in PrecomputeZeroPointTimesWeightWithBias()
76         memset(output->get(), 0, row * sizeof(int32_t));  in PrecomputeZeroPointTimesWeightWithBias()
78         memcpy(output->get(), bias_tensor, row * sizeof(int32_t));  in PrecomputeZeroPointTimesWeightWithBias()
86 void ApplySigmoid(const int16_t* input, int32_t n_batch, int32_t n_input, int16_t* output) {  in ApplySigmoid()
106             const int32_t value = static_cast<int32_t>(a) * static_cast<int32_t>(b);  in CwiseMul()
112 void CwiseMul(const int16_t* input_1, const int16_t* input_2, int32_t multiplier, int32_t shift,  in CwiseMul()
113               int32_t n_batch, int32_t n_input, int32_t output_zp, int8_t* output) {  in CwiseMul()
119             int32_t value = static_cast<int32_t>(a) * static_cast<int32_t>(b);  in CwiseMul()
143             int32_t sum = input_1[index] + input_2[index];  in CwiseAdd()
144             const int32_t sum_clamped = std::min(INT16_MAX, std::max(INT16_MIN, sum));  in CwiseAdd()
150 void CwiseClipping(int16_t* input, const int16_t clipping_value, int32_t n_batch, int32_t n_input) {  in CwiseClipping()
164 void CwiseClipping(int8_t* input, const int8_t clipping_value, int32_t n_batch, int32_t n_input) {  in CwiseClipping()
180                                              int32_t multiplier, int shift, int16_t* result) {  in VectorBatchVectorCwiseProductAccumulate()
183             int32_t prod = vector[v] * *batch_vector++;  in VectorBatchVectorCwiseProductAccumulate()
185             int32_t output = prod + *result;  in VectorBatchVectorCwiseProductAccumulate()