Updated ARM CMSIS NN Lib

This commit is contained in:
Kwabena W. Agyeman 2018-06-09 17:53:52 -04:00
parent 107019a357
commit e4f238e794
2 changed files with 64 additions and 1 deletions

View File

@ -427,6 +427,69 @@ extern "C"
q15_t * bufferA,
q7_t * bufferB);
/**
* @brief Fast Q15 convolution function (non-sqaure shape)
* @param[in] Im_in pointer to input tensor
* @param[in] dim_im_in_x input tensor dimention x
* @param[in] dim_im_in_y input tensor dimention y
* @param[in] ch_im_in number of input tensor channels
* @param[in] wt pointer to kernel weights
* @param[in] ch_im_out number of filters, i.e., output tensor channels
* @param[in] dim_kernel_x filter kernel size x
* @param[in] dim_kernel_y filter kernel size y
* @param[in] padding_x padding size x
* @param[in] padding_y padding size y
* @param[in] stride_x convolution stride x
* @param[in] stride_y convolution stride y
* @param[in] bias pointer to bias
* @param[in] bias_shift amount of left-shift for bias
* @param[in] out_shift amount of right-shift for output
* @param[in,out] Im_out pointer to output tensor
* @param[in] dim_im_out_x output tensor dimension x
* @param[in] dim_im_out_y output tensor dimension y
* @param[in,out] bufferA pointer to buffer space for input
* @param[in,out] bufferB pointer to buffer space for output
* @return The function returns either
* <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
*
* @details
*
* <b>Buffer size:</b>
*
* bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
*
* bufferB size: 0
*
* <b>Input dimension constraints:</b>
*
* ch_im_in is multiple of 2
*
* ch_im_out is multipe of 2
*
*/
arm_status
arm_convolve_HWC_q15_fast_nonsquare(const q15_t * Im_in,
const uint16_t dim_im_in_x,
const uint16_t dim_im_in_y,
const uint16_t ch_im_in,
const q15_t * wt,
const uint16_t ch_im_out,
const uint16_t dim_kernel_x,
const uint16_t dim_kernel_y,
const uint16_t padding_x,
const uint16_t padding_y,
const uint16_t stride_x,
const uint16_t stride_y,
const q15_t * bias,
const uint16_t bias_shift,
const uint16_t out_shift,
q15_t * Im_out,
const uint16_t dim_im_out_x,
const uint16_t dim_im_out_y,
q15_t * bufferA,
q7_t * bufferB);
/**
* @brief Q7 depthwise separable convolution function
* @param[in] Im_in pointer to input tensor

View File

@ -80,7 +80,7 @@ void arm_nn_activations_direct_q7(q7_t * data, uint16_t size, uint16_t int_width
while (i)
{
in = *pIn++;
out = lookup_table[(uint8_t) in >> shift_size];
out = lookup_table[(uint8_t) (in >> shift_size)];
*pOut++ = out;
i--;
}