Home
last modified time | relevance | path

Searched refs:out_shift (Results 1 – 25 of 30) sorted by relevance

12

/hal_nxp-latest/mcux/mcux-sdk/CMSIS/NN/Source/FullyConnectedFunctions/
Darm_fully_connected_q15_opt.c101 const uint16_t out_shift, in arm_fully_connected_q15_opt() argument
119 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15_opt()
120 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15_opt()
121 q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15_opt()
122 q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15_opt()
197 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16)); in arm_fully_connected_q15_opt()
198 *pO++ = (q15_t)(__SSAT((sum2 >> out_shift), 16)); in arm_fully_connected_q15_opt()
199 *pO++ = (q15_t)(__SSAT((sum3 >> out_shift), 16)); in arm_fully_connected_q15_opt()
200 *pO++ = (q15_t)(__SSAT((sum4 >> out_shift), 16)); in arm_fully_connected_q15_opt()
211 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15_opt()
[all …]
Darm_fully_connected_mat_q7_vec_q15_opt.c121 const uint16_t out_shift, in arm_fully_connected_mat_q7_vec_q15_opt() argument
140 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15_opt()
141 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15_opt()
142 q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15_opt()
143 q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15_opt()
275 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15_opt()
276 *pO++ = (q15_t)(__SSAT((sum2 >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15_opt()
277 *pO++ = (q15_t)(__SSAT((sum3 >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15_opt()
278 *pO++ = (q15_t)(__SSAT((sum4 >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15_opt()
289 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15_opt()
[all …]
Darm_fully_connected_q7_opt.c133 const uint16_t out_shift, in arm_fully_connected_q7_opt() argument
153 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7_opt()
154 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7_opt()
155 q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7_opt()
156 q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7_opt()
334 *pO++ = (q7_t)(__SSAT((sum >> out_shift), 8)); in arm_fully_connected_q7_opt()
335 *pO++ = (q7_t)(__SSAT((sum2 >> out_shift), 8)); in arm_fully_connected_q7_opt()
336 *pO++ = (q7_t)(__SSAT((sum3 >> out_shift), 8)); in arm_fully_connected_q7_opt()
337 *pO++ = (q7_t)(__SSAT((sum4 >> out_shift), 8)); in arm_fully_connected_q7_opt()
348 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7_opt()
[all …]
Darm_fully_connected_mat_q7_vec_q15.c73 const uint16_t out_shift, in arm_fully_connected_mat_q7_vec_q15() argument
92 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15()
93 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15()
128 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15()
129 *pO++ = (q15_t)(__SSAT((sum2 >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15()
141 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15()
171 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16)); in arm_fully_connected_mat_q7_vec_q15()
181 int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_mat_q7_vec_q15()
186 pOut[i] = (q15_t)__SSAT((ip_out >> out_shift), 16); in arm_fully_connected_mat_q7_vec_q15()
Darm_fully_connected_q7.c72 const uint16_t out_shift, in arm_fully_connected_q7() argument
93 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7()
94 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7()
129 *pO++ = (q7_t)(__SSAT((sum >> out_shift), 8)); in arm_fully_connected_q7()
130 *pO++ = (q7_t)(__SSAT((sum2 >> out_shift), 8)); in arm_fully_connected_q7()
143 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7()
172 *pO++ = (q7_t)(__SSAT((sum >> out_shift), 8)); in arm_fully_connected_q7()
184 int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q7()
189 pOut[i] = (q7_t)__SSAT((ip_out >> out_shift), 8); in arm_fully_connected_q7()
Darm_fully_connected_q15.c70 const uint16_t out_shift, in arm_fully_connected_q15() argument
89 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15()
90 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15()
125 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16)); in arm_fully_connected_q15()
126 *pO++ = (q15_t)(__SSAT((sum2 >> out_shift), 16)); in arm_fully_connected_q15()
137 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15()
169 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16)); in arm_fully_connected_q15()
179 int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_fully_connected_q15()
184 pOut[i] = (q15_t)__SSAT((ip_out >> out_shift), 16); in arm_fully_connected_q15()
/hal_nxp-latest/mcux/mcux-sdk/CMSIS/NN/Source/ConvolutionFunctions/
Darm_nn_mat_mult_kernel_q7_q15.c45 const uint16_t out_shift, in arm_nn_mat_mult_kernel_q7_q15() argument
66 q31_t sum = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15()
67 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15()
68 q31_t sum3 = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15()
69 q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15()
112 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15()
113 *pOut++ = (q7_t)__SSAT((sum3 >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15()
114 *pOut2++ = (q7_t)__SSAT((sum2 >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15()
115 *pOut2++ = (q7_t)__SSAT((sum4 >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15()
130 q31_t sum = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15()
[all …]
Darm_nn_mat_mult_kernel_q7_q15_reordered.c45 const uint16_t out_shift, in arm_nn_mat_mult_kernel_q7_q15_reordered() argument
66 q31_t sum = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15_reordered()
67 q31_t sum2 = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15_reordered()
68 q31_t sum3 = ((q31_t)(bias[i + 1]) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15_reordered()
69 q31_t sum4 = ((q31_t)(bias[i + 1]) << bias_shift) + NN_ROUND(out_shift); in arm_nn_mat_mult_kernel_q7_q15_reordered()
112 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15_reordered()
113 *pOut++ = (q7_t)__SSAT((sum3 >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15_reordered()
114 *pOut2++ = (q7_t)__SSAT((sum2 >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15_reordered()
115 *pOut2++ = (q7_t)__SSAT((sum4 >> out_shift), 8); in arm_nn_mat_mult_kernel_q7_q15_reordered()
131 (void)out_shift; in arm_nn_mat_mult_kernel_q7_q15_reordered()
Darm_convolve_HWC_q15_fast.c89 const uint16_t out_shift, in arm_convolve_HWC_q15_fast() argument
157 q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast()
158 q31_t sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast()
159 q31_t sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast()
160 q31_t sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast()
192 *pOut++ = (q15_t)__SSAT(sum >> out_shift, 16); in arm_convolve_HWC_q15_fast()
193 *pOut++ = (q15_t)__SSAT(sum3 >> out_shift, 16); in arm_convolve_HWC_q15_fast()
194 *pOut2++ = (q15_t)__SSAT(sum2 >> out_shift, 16); in arm_convolve_HWC_q15_fast()
195 *pOut2++ = (q15_t)__SSAT(sum4 >> out_shift, 16); in arm_convolve_HWC_q15_fast()
227 conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast()
[all …]
Darm_convolve_HWC_q15_fast_nonsquare.c98 const uint16_t out_shift, in arm_convolve_HWC_q15_fast_nonsquare() argument
169 q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast_nonsquare()
170 q31_t sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast_nonsquare()
171 q31_t sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast_nonsquare()
172 q31_t sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast_nonsquare()
204 *pOut++ = (q15_t)__SSAT(sum >> out_shift, 16); in arm_convolve_HWC_q15_fast_nonsquare()
205 *pOut++ = (q15_t)__SSAT(sum3 >> out_shift, 16); in arm_convolve_HWC_q15_fast_nonsquare()
206 *pOut2++ = (q15_t)__SSAT(sum2 >> out_shift, 16); in arm_convolve_HWC_q15_fast_nonsquare()
207 *pOut2++ = (q15_t)__SSAT(sum4 >> out_shift, 16); in arm_convolve_HWC_q15_fast_nonsquare()
239 conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q15_fast_nonsquare()
[all …]
Darm_depthwise_separable_conv_HWC_q7.c92 const uint16_t out_shift, in arm_depthwise_separable_conv_HWC_q7() argument
148 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7()
149 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7()
150 q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7()
151 q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7()
338 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7()
339 *pOut++ = (q7_t)__SSAT((sum2 >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7()
340 *pOut++ = (q7_t)__SSAT((sum3 >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7()
341 *pOut++ = (q7_t)__SSAT((sum4 >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7()
351 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7()
[all …]
Darm_depthwise_separable_conv_HWC_q7_nonsquare.c88 const uint16_t out_shift, in arm_depthwise_separable_conv_HWC_q7_nonsquare() argument
158 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
159 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
160 q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
161 q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
342 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
343 *pOut++ = (q7_t)__SSAT((sum2 >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
344 *pOut++ = (q7_t)__SSAT((sum3 >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
345 *pOut++ = (q7_t)__SSAT((sum4 >> out_shift), 8); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
355 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); in arm_depthwise_separable_conv_HWC_q7_nonsquare()
[all …]
Darm_nn_mat_mult_kernel_s8_s16_reordered.c46 const int32_t *out_shift, in arm_nn_mat_mult_kernel_s8_s16_reordered() argument
105 ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16_reordered()
111 ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16_reordered()
117 out_shift++; in arm_nn_mat_mult_kernel_s8_s16_reordered()
119 ch_1_out_0 = arm_nn_requantize(ch_1_out_0, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16_reordered()
125 ch_1_out_1 = arm_nn_requantize(ch_1_out_1, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16_reordered()
131 out_shift++; in arm_nn_mat_mult_kernel_s8_s16_reordered()
169 ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16_reordered()
175 ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16_reordered()
190 (void)out_shift; in arm_nn_mat_mult_kernel_s8_s16_reordered()
Darm_nn_mat_mult_kernel_s8_s16.c43 const int32_t *out_shift, in arm_nn_mat_mult_kernel_s8_s16() argument
149 int32x4_t shift = vldrwq_s32(out_shift); in arm_nn_mat_mult_kernel_s8_s16()
151 out_shift += ROW_PER_LOOP; in arm_nn_mat_mult_kernel_s8_s16()
199 shift_tail[i_ch] = out_shift[i_ch]; in arm_nn_mat_mult_kernel_s8_s16()
283 ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16()
289 ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16()
295 out_shift++; in arm_nn_mat_mult_kernel_s8_s16()
297 ch_1_out_0 = arm_nn_requantize(ch_1_out_0, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16()
303 ch_1_out_1 = arm_nn_requantize(ch_1_out_1, *out_mult, *out_shift); in arm_nn_mat_mult_kernel_s8_s16()
309 out_shift++; in arm_nn_mat_mult_kernel_s8_s16()
[all …]
Darm_convolve_HWC_q7_fast.c101 const uint16_t out_shift, in arm_convolve_HWC_q7_fast() argument
161 … wt, bufferA, ch_im_out, ch_im_in * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast()
197 … wt, bufferA, ch_im_out, ch_im_in * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast()
219 … wt, bufferA, ch_im_out, ch_im_in * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast()
250 … wt, bufferA, ch_im_out, ch_im_in * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast()
283 … wt, bufferA, ch_im_out, ch_im_in * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast()
298 q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_fast()
326 *pOut = (q7_t)__SSAT((sum >> out_shift), 8); in arm_convolve_HWC_q7_fast()
349 conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_fast()
367 … Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8); in arm_convolve_HWC_q7_fast()
Darm_convolve_HWC_q7_fast_nonsquare.c88 const uint16_t out_shift, in arm_convolve_HWC_q7_fast_nonsquare() argument
151 …wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast_nonsquare()
189 …wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast_nonsquare()
212 …wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast_nonsquare()
245 …wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast_nonsquare()
280 …wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_fast_nonsquare()
294 q31_t sum = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_fast_nonsquare()
322 *pOut = (q7_t)__SSAT((sum >> out_shift), 8); in arm_convolve_HWC_q7_fast_nonsquare()
346 conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_fast_nonsquare()
365 … Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8); in arm_convolve_HWC_q7_fast_nonsquare()
Darm_convolve_1x1_HWC_q7_fast_nonsquare.c95 const uint16_t out_shift, in arm_convolve_1x1_HWC_q7_fast_nonsquare() argument
136 wt, bufferA, ch_im_out, ch_im_in, bias_shift, out_shift, bias, pOut); in arm_convolve_1x1_HWC_q7_fast_nonsquare()
149 q31_t sum = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift); in arm_convolve_1x1_HWC_q7_fast_nonsquare()
178 *pOut = (q7_t)__SSAT((sum >> out_shift), 8); in arm_convolve_1x1_HWC_q7_fast_nonsquare()
203 conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); in arm_convolve_1x1_HWC_q7_fast_nonsquare()
222 … Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8); in arm_convolve_1x1_HWC_q7_fast_nonsquare()
Darm_convolve_HWC_q7_basic.c84 const uint16_t out_shift, in arm_convolve_HWC_q7_basic() argument
132 … wt, bufferA, ch_im_out, ch_im_in * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_basic()
149 q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_basic()
180 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_convolve_HWC_q7_basic()
196 conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_basic()
214 … Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8); in arm_convolve_HWC_q7_basic()
Darm_convolve_HWC_q7_basic_nonsquare.c82 const uint16_t out_shift, in arm_convolve_HWC_q7_basic_nonsquare() argument
133 …wt, bufferA, ch_im_out, ch_im_in * dim_kernel_y * dim_kernel_x, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_basic_nonsquare()
150 q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_basic_nonsquare()
181 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_convolve_HWC_q7_basic_nonsquare()
197 conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_basic_nonsquare()
216 … Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8); in arm_convolve_HWC_q7_basic_nonsquare()
Darm_convolve_HWC_q7_RGB.c90 const uint16_t out_shift, in arm_convolve_HWC_q7_RGB() argument
182 … wt, bufferA, ch_im_out, 3 * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); in arm_convolve_HWC_q7_RGB()
198 q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_RGB()
226 *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8); in arm_convolve_HWC_q7_RGB()
248 conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift); in arm_convolve_HWC_q7_RGB()
266 … Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8); in arm_convolve_HWC_q7_RGB()
/hal_nxp-latest/mcux/mcux-sdk/CMSIS/NN/Source/NNSupportFunctions/
Darm_nn_mult_q7.c56 void arm_nn_mult_q7(q7_t *pSrcA, q7_t *pSrcB, q7_t *pDst, const uint16_t out_shift, uint32_t blockS… in arm_nn_mult_q7() argument
74 …t1 = (q7_t)__SSAT(((q15_t)((q15_t)(*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); in arm_nn_mult_q7()
75 …t2 = (q7_t)__SSAT(((q15_t)((q15_t)(*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); in arm_nn_mult_q7()
76 …t3 = (q7_t)__SSAT(((q15_t)((q15_t)(*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); in arm_nn_mult_q7()
77 …t4 = (q7_t)__SSAT(((q15_t)((q15_t)(*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); in arm_nn_mult_q7()
103 …++ = (q7_t)__SSAT(((q15_t)((q15_t)(*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); in arm_nn_mult_q7()
Darm_nn_mult_q15.c56 void arm_nn_mult_q15(q15_t *pSrcA, q15_t *pSrcB, q15_t *pDst, const uint16_t out_shift, uint32_t bl… in arm_nn_mult_q15() argument
90 out1 = (q15_t)__SSAT((q31_t)(mul1 + NN_ROUND(out_shift)) >> out_shift, 16); in arm_nn_mult_q15()
91 out2 = (q15_t)__SSAT((q31_t)(mul2 + NN_ROUND(out_shift)) >> out_shift, 16); in arm_nn_mult_q15()
92 out3 = (q15_t)__SSAT((q31_t)(mul3 + NN_ROUND(out_shift)) >> out_shift, 16); in arm_nn_mult_q15()
93 out4 = (q15_t)__SSAT((q31_t)(mul4 + NN_ROUND(out_shift)) >> out_shift, 16); in arm_nn_mult_q15()
129 … = (q15_t)__SSAT(((q31_t)((q31_t)(*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 16); in arm_nn_mult_q15()
/hal_nxp-latest/mcux/mcux-sdk/CMSIS/NN/Include/
Darm_nnfunctions.h305 const uint16_t out_shift,
349 const uint16_t out_shift,
386 const uint16_t out_shift,
427 const uint16_t out_shift,
478 const uint16_t out_shift,
534 const uint16_t out_shift,
687 const uint16_t out_shift,
729 const uint16_t out_shift,
790 const uint16_t out_shift,
833 const uint16_t out_shift,
[all …]
/hal_nxp-latest/mcux/mcux-sdk/CMSIS/NN/Source/BasicMathFunctions/
Darm_elementwise_mul_s8.c57 const int32_t out_shift, in arm_elementwise_mul_s8() argument
81 res_0 = arm_requantize_mve_32x4(res_0, vdupq_n_s32(out_mult), vdupq_n_s32(out_shift)); in arm_elementwise_mul_s8()
130 mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; in arm_elementwise_mul_s8()
141 mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; in arm_elementwise_mul_s8()
151 mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; in arm_elementwise_mul_s8()
161 mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; in arm_elementwise_mul_s8()
184 mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; in arm_elementwise_mul_s8()
Darm_elementwise_add_s8.c82 const int32_t out_shift, in arm_elementwise_add_s8() argument
110 SAT_INPUT_VECT(vect_1, out_mult, out_shift); in arm_elementwise_add_s8()
164 SAT_INPUT(sum, out_mult, out_shift); in arm_elementwise_add_s8()
178 SAT_INPUT(sum, out_mult, out_shift); in arm_elementwise_add_s8()
192 SAT_INPUT(sum, out_mult, out_shift); in arm_elementwise_add_s8()
206 SAT_INPUT(sum, out_mult, out_shift); in arm_elementwise_add_s8()
236 SAT_INPUT(sum, out_mult, out_shift); in arm_elementwise_add_s8()

12