CMSIS-NN: Add int32 bias support int16xint8 conv (#2557) BUG=CMSIS-NN glue is updated to support int32bias for int16x8
diff --git a/tensorflow/lite/micro/kernels/cmsis_nn/conv.cc b/tensorflow/lite/micro/kernels/cmsis_nn/conv.cc index 6691b59..4c35970 100644 --- a/tensorflow/lite/micro/kernels/cmsis_nn/conv.cc +++ b/tensorflow/lite/micro/kernels/cmsis_nn/conv.cc
@@ -21,7 +21,6 @@ #include "tensorflow/lite/kernels/internal/common.h" #include "tensorflow/lite/kernels/internal/quantization_util.h" #include "tensorflow/lite/kernels/internal/reference/conv.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/conv.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" #include "tensorflow/lite/kernels/kernel_util.h" #include "tensorflow/lite/kernels/padding.h" @@ -214,9 +213,26 @@ const cmsis_nn_dims* filter_dims, const int8_t* filter, const cmsis_nn_dims* bias_dims, const int64_t* bias, const cmsis_nn_dims* output_dims, int16_t* output, TfLiteType weightsT) { + const cmsis_nn_bias_data bias_data = {bias, false}; + return arm_convolve_wrapper_s16(ctx, conv_params, quant_params, input_dims, - input, filter_dims, filter, bias_dims, bias, - output_dims, output); + input, filter_dims, filter, bias_dims, + &bias_data, output_dims, output); +} + +template <> +arm_cmsis_nn_status convolve_wrapper( + const cmsis_nn_context* ctx, const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, const int16_t* input, + const cmsis_nn_dims* filter_dims, const int8_t* filter, + const cmsis_nn_dims* bias_dims, const int32_t* bias, + const cmsis_nn_dims* output_dims, int16_t* output, TfLiteType weightsT) { + const cmsis_nn_bias_data bias_data = {bias, true}; + + return arm_convolve_wrapper_s16(ctx, conv_params, quant_params, input_dims, + input, filter_dims, filter, bias_dims, + &bias_data, output_dims, output); } template <typename ActType, typename BiasType, TfLiteType type> @@ -362,25 +378,17 @@ TFLITE_DCHECK(node->user_data != nullptr); const OpData& data = *(static_cast<const OpData*>(node->user_data)); - if (bias == nullptr || bias->type == kTfLiteInt64) { + if (bias == nullptr || bias->type == kTfLiteInt32) { + return EvalQuantizedPerChannel<int16_t, int32_t, kTfLiteInt16>( + context, node, params, data, input, filter, bias, output); + } else if (bias->type == kTfLiteInt64) { return EvalQuantizedPerChannel<int16_t, int64_t, kTfLiteInt16>( context, node, params, data, input, filter, bias, output); } else { - reference_integer_ops::ConvPerChannel( - ConvParamsQuantized(params, data.reference_op_data), - data.reference_op_data.per_channel_output_multiplier, - data.reference_op_data.per_channel_output_shift, - tflite::micro::GetTensorShape(input), - tflite::micro::GetTensorData<int16_t>(input), - tflite::micro::GetTensorShape(filter), - tflite::micro::GetTensorData<int8_t>(filter), - tflite::micro::GetTensorShape(bias), - tflite::micro::GetOptionalTensorData<std::int32_t>(bias), - tflite::micro::GetTensorShape(output), - tflite::micro::GetTensorData<int16_t>(output)); + MicroPrintf("Bias type %s (%d) not supported.", + TfLiteTypeGetName(bias->type), bias->type); + return kTfLiteError; } - - return kTfLiteOk; } TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { @@ -443,22 +451,12 @@ break; } case kTfLiteInt16: { - if (bias == nullptr || bias->type == kTfLiteInt64) { + if (bias == nullptr || bias->type == kTfLiteInt32) { + return EvalQuantizedPerChannel<int16_t, int32_t, kTfLiteInt16>( + context, node, params, data, input, filter, bias, output); + } else if (bias->type == kTfLiteInt64) { return EvalQuantizedPerChannel<int16_t, int64_t, kTfLiteInt16>( context, node, params, data, input, filter, bias, output); - } else if (bias->type == kTfLiteInt32) { - reference_integer_ops::ConvPerChannel( - ConvParamsQuantized(params, data.reference_op_data), - data.reference_op_data.per_channel_output_multiplier, - data.reference_op_data.per_channel_output_shift, - tflite::micro::GetTensorShape(input), - tflite::micro::GetTensorData<int16_t>(input), - tflite::micro::GetTensorShape(filter), - tflite::micro::GetTensorData<int8_t>(filter), - tflite::micro::GetTensorShape(bias), - tflite::micro::GetOptionalTensorData<std::int32_t>(bias), - tflite::micro::GetTensorShape(output), - tflite::micro::GetTensorData<int16_t>(output)); } else { MicroPrintf("Bias type %s (%d) not supported.", TfLiteTypeGetName(bias->type), bias->type);
diff --git a/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh b/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh index 482766c..195c0ca 100755 --- a/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh +++ b/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh
@@ -47,9 +47,9 @@ echo >&2 "${DOWNLOADED_CMSIS_NN_PATH} already exists, skipping the download." else - ZIP_PREFIX_NN="8492d82a1a81651977c5f5128492b4a0f0cf6715" + ZIP_PREFIX_NN="15dbe7cda7130766777f18a69388238bc6540cef" CMSIS_NN_URL="http://github.com/ARM-software/CMSIS-NN/archive/${ZIP_PREFIX_NN}.zip" - CMSIS_NN_MD5="2cb03e4f044b78af6751009cd53247a8" + CMSIS_NN_MD5="6121fa707d4dac17acb4e222453a38a6" # wget is much faster than git clone of the entire repo. So we wget a specific # version and can then apply a patch, as needed.