diff options
Diffstat (limited to 'compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h')
| -rw-r--r-- | compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h | 190 |
1 files changed, 190 insertions, 0 deletions
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h b/compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h new file mode 100644 index 000000000..a4a5b2a78 --- /dev/null +++ b/compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h @@ -0,0 +1,190 @@ +/* + * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved + * Copyright 2020 The TensorFlow Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef LUCI_INTERPRETER_PAL_SVDF_H +#define LUCI_INTERPRETER_PAL_SVDF_H + +#include <arm_nn_types.h> +#include <arm_nnfunctions.h> + +namespace luci_interpreter_pal +{ +static inline void +IntegerSVDF(const TfLiteSVDFParams ¶ms, const tflite::RuntimeShape &input_shape, + const int8_t *input_data, const tflite::RuntimeShape &weight_feature_shape, + const int8_t *weight_feature_data, const tflite::RuntimeShape &weight_time_shape, + const int16_t *weight_time_data, const tflite::RuntimeShape &bias_shape, + const int32_t *bias_data, int16_t *activation_state_data, + const tflite::RuntimeShape &output_shape, int8_t *output_data, int32_t *scratchpad_data, + int32_t *output_temp_data, int32_t scale_1_a, int scale_1_b, int32_t scale_2_a, + int scale_2_b, int32_t input_zp, int32_t output_zp) +{ + const int32_t rank = params.rank; + const int32_t batch_size = input_shape.Dims(0); + const int32_t num_filters = weight_feature_shape.Dims(0); + const int32_t memory_size = weight_time_shape.Dims(1); + + cmsis_nn_dims input_dims; + input_dims.n = input_shape.Dims(0); + input_dims.h = input_shape.Dims(1); + + cmsis_nn_dims weights_feature_dims; + weights_feature_dims.n = weight_feature_shape.Dims(0); + weights_feature_dims.h = weight_feature_shape.Dims(1); + + cmsis_nn_dims weights_time_dims; + weights_time_dims.n = weight_time_shape.Dims(0); + weights_time_dims.h = weight_time_shape.Dims(1); + + cmsis_nn_dims bias_dims; + bias_dims.n = bias_shape.Dims(0); + + cmsis_nn_dims state_dims; + state_dims.n = batch_size; + state_dims.h = memory_size * num_filters; + + cmsis_nn_dims output_dims; + output_dims.n = output_shape.Dims(0); + output_dims.h = output_shape.Dims(1); + + cmsis_nn_svdf_params svdf_params; + svdf_params.rank = params.rank; + svdf_params.input_offset = input_zp; + svdf_params.output_offset = output_zp; + + svdf_params.input_activation.min = INT16_MIN; + svdf_params.input_activation.max = INT16_MAX; + + svdf_params.output_activation.min = INT8_MIN; + svdf_params.output_activation.max = INT8_MAX; + + cmsis_nn_per_tensor_quant_params in_quant_params; + in_quant_params.multiplier = scale_1_a; + in_quant_params.shift = scale_1_b; + + cmsis_nn_per_tensor_quant_params out_quant_params; + out_quant_params.multiplier = scale_2_a; + out_quant_params.shift = scale_2_b; + + cmsis_nn_context scratch_ctx; + scratch_ctx.buf = scratchpad_data; + + cmsis_nn_context scratch_output_ctx; + scratch_output_ctx.buf = output_temp_data; + + arm_svdf_s8(&scratch_ctx, &scratch_output_ctx, &svdf_params, &in_quant_params, &out_quant_params, + &input_dims, input_data, &state_dims, activation_state_data, &weights_feature_dims, + weight_feature_data, &weights_time_dims, weight_time_data, &bias_dims, bias_data, + &output_dims, output_data); +} +static inline void +FloatSVDF(const TfLiteSVDFParams ¶ms, const tflite::RuntimeShape &input_shape, + const float *input_data, const tflite::RuntimeShape &weight_feature_shape, + const float *weight_feature_data, const tflite::RuntimeShape &weight_time_shape, + const float *weight_time_data, const tflite::RuntimeShape &bias_shape, + const float *bias_data, float *scratchpad_data, float *activation_state_data, + const tflite::RuntimeShape &output_shape, float *output_data) +{ + const int32_t rank = params.rank; + const int32_t batch_size = input_shape.Dims(0); + const int32_t input_size = input_shape.Dims(1); + const int32_t num_filters = weight_feature_shape.Dims(0); + const int32_t num_units = num_filters / rank; + const int32_t memory_size = weight_time_shape.Dims(1); + + // Left shift the activation_state. + { + float *new_state_start = activation_state_data; + const float *old_state_start = activation_state_data + 1; + const float *old_state_end = activation_state_data + batch_size * num_filters * memory_size; + while (old_state_start != old_state_end) + { + *new_state_start++ = *old_state_start++; + } + } + + // Note: no need to clear the latest activation, matmul is not accumulative. + + // Compute conv1d(inputs, weights_feature). + // The activation_state's rightmost column is used to save current cycle + // activation. This is achieved by starting at state_ptr[memory_size - 1] and + // having the stride equal to memory_size. + + // Perform batched matrix vector multiply operation: + { + const float *matrix = weight_feature_data; + const float *vector = input_data; + float *result = &activation_state_data[memory_size - 1]; + float *result_in_batch = result; + for (int i = 0; i < batch_size; ++i) + { + const float *matrix_ptr = matrix; + for (int j = 0; j < num_filters; ++j) + { + float dot_prod = 0.0f; + const float *vector_in_batch = vector + i * input_size; + for (int k = 0; k < input_size; ++k) + { + dot_prod += *matrix_ptr++ * *vector_in_batch++; + } + *result_in_batch = dot_prod; + result_in_batch += memory_size; + } + } + } + + tflite::reference_ops::ApplyTimeWeightsBiasAndActivation( + batch_size, memory_size, num_filters, num_units, rank, weight_time_data, bias_data, + params.activation, activation_state_data, scratchpad_data, output_data); +} + +static inline void SetupScratchpadTensor( + const luci_interpreter::DataType &input_data_type, + const luci_interpreter::DataType &weight_feature_data_type, + luci_interpreter::Tensor *scratchpad_1, luci_interpreter::Tensor *scratchpad_2, + luci_interpreter::Tensor *scratchpad_3, luci_interpreter::Tensor *scratchpad_4, + luci_interpreter::Tensor *scratchpad_5, luci_interpreter::Tensor *scratchpad_6, + const luci_interpreter::Shape input_shape, const luci_interpreter::Shape weight_time_shape, + const int32_t batch_size, const int32_t num_filters, const int32_t num_units) +{ + if (input_data_type == loco::DataType::FLOAT32 && + (weight_feature_data_type == loco::DataType::S8 || + weight_feature_data_type == loco::DataType::U8)) + { + (void)input_shape; + (void)weight_time_shape; + (void)scratchpad_3; + (void)scratchpad_4; + (void)scratchpad_5; + (void)scratchpad_6; + + throw std::runtime_error("Hybrid type is not supported for cmsisnn"); + } + + // Resize scratchpad_1 tensor + scratchpad_1->resize({batch_size, num_filters}); + + if (input_data_type == loco::DataType::S8) + { + // Resize scratchpad_2 for full_integer op + scratchpad_2->resize({batch_size, num_units}); + } +} + +} // namespace luci_interpreter_pal + +#endif // LUCI_INTERPRETER_PAL_SVDF_H |