CMSIS-NN  Version 3.1.0
CMSIS NN Software Library
 All Data Structures Files Functions Variables Enumerations Enumerator Macros Groups Pages
Revision History of CMSIS-NN
Version Description
V3.1.0
  • Added arm_convolve_fast_s16
  • Added arm_nn_mat_mult_kernel_s16 DSP implementation
  • Added conv int16 DSP implementation
  • Added unit tests for int16 DSP conv kernel
  • Added arm_fully_connected_s16 and arm_fully_connected_wrapper_s16
    with support for int16.
  • Added fully_connected int16 reference implementation
  • Added fully_connected int16 DSP implementation
  • Added unit tests for int16 fully_connected kernel
  • Added dilation support for int8 conv kernel
  • Added dilation support for int8 depthwise conv kernel
V3.0.1
  • Added arm_convolve_s16 and arm_convolve_wrapper_s16
    with support for int16.
  • Added conv int16 reference implementation
  • Added new requantization function for int16
  • Added unit tests for int16 conv kernel
V3.0.0
  • Updated arm_fully_connected_s8 to use zero weight offset
    as per the TFLM int8 quantization spec. The API is the same but,
    the weight offset parameter is expected to be zero
  • Added unit test for Softmax
  • Added MVE optimiztion for SVDF operator
  • Removed Examples folder
V2.0.0

Added the following function for int8 SVDF operator.

  • arm_svdf_s8
  • arm_depthwise_conv_wrapper_s8
  • arm_convolve_wrapper_s8
  • arm_nn_doubling_high_mult_no_sat

Deleted functions

  • arm_max_pool_s8_opt

V1.3.0 Added following functions for int8 operators with symmetric quantization.
  • arm_convolve_s8
  • arm_convolve_1x1_s8_fast
  • arm_depthwise_conv_3x3_s8
  • arm_depthwise_conv_s8
  • arm_depthwise_conv_s8_opt
  • arm_fully_connected_s8
  • arm_softmax_q7
  • arm_softmax_s8
  • arm_softmax_u8
  • arm_elementwise_mul_s8
  • arm_elementwise_add_s8
  • arm_avgpool_s8
  • arm_maxpool_s8
  • arm_maxpool_s8_opt
  • arm_relu6_s8
  • arm_reshape_s8
  • arm_fully_connected_s8
  • arm_concatenation_s8_w
  • arm_concatenation_s8_x
  • arm_concatenation_s8_y
  • arm_concatenation_s8_z
Added following support functions.
  • arm_nn_mat_mul_kernel_s8_s16
  • arm_nn_mat_mul_kernel_s8_s16_reordered
  • arm_nn_depthwise_conv_s8_core
  • arm_nn_mat_mul_core_4x_s8
  • arm_nn_mat_mul_core_1x_s8
  • arm_convolve_1_x_n_s8
  • arm_nn_vec_mat_mult_t_s8
  • arm_nn_mat_mult_nt_t_s8
  • arm_nn_depthwise_conv_nt_t_s8
  • arm_nn_depthwise_conv_nt_t_padded_s8
V1.2.0 Added depthwise convolution function with asymmetric quantization.
  • arm_depthwise_conv_u8_basic_ver1
Added support functions for requantization.
  • arm_nn_sat_doubling_high_mult
  • arm_nn_divide_by_power_of_two
V1.1.0 Added new math functions:
  • arm_convolve_HWC_q7_basic_nonsquare
  • arm_convolve_HWC_q15_fast_nonsquare
  • arm_nn_mult_q7
  • arm_nn_mult_q15
Fixed minor issues.
V1.0.0 Initial version