/cmsis-nn-3.7.0/Source/NNSupportFunctions/ |
D | arm_nn_mat_mult_nt_t_s8_s32.c | 52 const int32_t rhs_rows, in arm_nn_mat_mult_nt_t_s8_s32() argument 57 int32_t rhs_rows_idx = rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 70 lhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 72 lhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 74 lhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 76 lhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 102 rhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 122 rhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 124 lhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() 141 lhs_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8_s32() [all …]
|
D | arm_nn_mat_mult_nt_t_s16.c | 55 const int32_t rhs_rows, in arm_nn_mat_mult_nt_t_s16() argument 62 const uint32_t rhs_rows_offset = (uint32_t)rhs_rows * sizeof(int16_t); in arm_nn_mat_mult_nt_t_s16() 76 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s16() 205 dst += (3 * rhs_rows); in arm_nn_mat_mult_nt_t_s16() 216 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s16() 264 const int32_t tail_rows = rhs_rows & 0x3; in arm_nn_mat_mult_nt_t_s16() 280 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s16() 350 (void)rhs_rows; in arm_nn_mat_mult_nt_t_s16()
|
D | arm_nn_mat_mult_nt_t_s8.c | 55 const int32_t rhs_rows, in arm_nn_mat_mult_nt_t_s8() argument 69 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s8() 150 dst += 4 * row_address_offset - rhs_rows; in arm_nn_mat_mult_nt_t_s8() 158 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s8() 212 const int32_t tail_rows = rhs_rows & 0x3; in arm_nn_mat_mult_nt_t_s8() 222 dst += row_address_offset - rhs_rows; in arm_nn_mat_mult_nt_t_s8() 230 for (int32_t rhs_rows_idx = 0; rhs_rows_idx <= (rhs_rows - 2); rhs_rows_idx += 2) in arm_nn_mat_mult_nt_t_s8() 439 dst_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8() 442 dst_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s8() 583 if (rhs_rows % 2) in arm_nn_mat_mult_nt_t_s8() [all …]
|
D | arm_nn_vec_mat_mul_result_acc_s8_s16.c | 50 const int32_t rhs_rows, in arm_nn_vec_mat_mul_result_acc_s8_s16() argument 62 for (size_t row_loop_cnt = rhs_rows / 4; row_loop_cnt != 0; --row_loop_cnt) in arm_nn_vec_mat_mul_result_acc_s8_s16() 121 for (size_t row_loop_cnt = rhs_rows % 4; row_loop_cnt != 0; --row_loop_cnt) in arm_nn_vec_mat_mul_result_acc_s8_s16() 155 for (int32_t row_loop_cnt = rhs_rows / 2; row_loop_cnt != 0; --row_loop_cnt) in arm_nn_vec_mat_mul_result_acc_s8_s16() 216 if (rhs_rows & 0x1) in arm_nn_vec_mat_mul_result_acc_s8_s16() 258 const int32_t row_loop_cnt = rhs_rows / 3; in arm_nn_vec_mat_mul_result_acc_s8_s16() 309 const int loop_cnt = rhs_rows % 3; in arm_nn_vec_mat_mul_result_acc_s8_s16()
|
D | arm_nn_vec_mat_mult_t_s16.c | 55 const int32_t rhs_rows, in arm_nn_vec_mat_mult_t_s16() argument 69 int32_t row_loop_cnt = rhs_rows / 4; in arm_nn_vec_mat_mult_t_s16() 165 for (int8_t rows_left = rhs_rows & 0x3; rows_left > 0; rows_left--) in arm_nn_vec_mat_mult_t_s16() 218 const int32_t row_loop_cnt = rhs_rows / 2; in arm_nn_vec_mat_mult_t_s16() 284 if (rhs_rows & 0x1) in arm_nn_vec_mat_mult_t_s16() 328 for (int i_row_loop_cnt = 0; i_row_loop_cnt < rhs_rows; i_row_loop_cnt++) in arm_nn_vec_mat_mult_t_s16()
|
D | arm_nn_vec_mat_mul_result_acc_s16.c | 55 const int32_t rhs_rows, in arm_nn_vec_mat_mul_result_acc_s16() argument 78 int32_t row_loop_cnt = rhs_rows / 4; in arm_nn_vec_mat_mul_result_acc_s16() 170 for (int8_t rows_left = rhs_rows & 0x3; rows_left > 0; rows_left--) in arm_nn_vec_mat_mul_result_acc_s16() 217 const int32_t row_loop_cnt = rhs_rows / 2; in arm_nn_vec_mat_mul_result_acc_s16() 283 if (rhs_rows & 0x1) in arm_nn_vec_mat_mul_result_acc_s16() 326 for (int i_row_loop_cnt = 0; i_row_loop_cnt < rhs_rows; i_row_loop_cnt++) in arm_nn_vec_mat_mul_result_acc_s16()
|
D | arm_nn_vec_mat_mult_t_s8.c | 68 const int32_t rhs_rows, in arm_nn_vec_mat_mult_t_s8() argument 77 const int32_t row_loop_cnt = rhs_rows / 4; in arm_nn_vec_mat_mult_t_s8() 154 const int loop_cnt = rhs_rows % 4; in arm_nn_vec_mat_mult_t_s8() 201 const int32_t row_loop_cnt = rhs_rows / 2; in arm_nn_vec_mat_mult_t_s8() 273 if (rhs_rows & 0x1) in arm_nn_vec_mat_mult_t_s8() 321 const int32_t row_loop_cnt = rhs_rows / 3; in arm_nn_vec_mat_mult_t_s8() 382 const int loop_cnt = rhs_rows % 3; in arm_nn_vec_mat_mult_t_s8() 427 const int32_t row_loop_cnt = rhs_rows / 4; in arm_nn_vec_mat_mult_t_s8() 498 const int loop_cnt = rhs_rows % 4; in arm_nn_vec_mat_mult_t_s8() 541 const int32_t row_loop_cnt = rhs_rows / 2; in arm_nn_vec_mat_mult_t_s8() [all …]
|
D | arm_nn_vec_mat_mult_t_svdf_s8.c | 57 const int32_t rhs_rows, in arm_nn_vec_mat_mult_t_svdf_s8() argument 67 int32_t row_loop_cnt = rhs_rows / 3; in arm_nn_vec_mat_mult_t_svdf_s8() 127 const int loop_cnt = rhs_rows % 3; in arm_nn_vec_mat_mult_t_svdf_s8() 163 int32_t row_loop_cnt = rhs_rows / 2; in arm_nn_vec_mat_mult_t_svdf_s8() 298 if (rhs_rows & 0x1) in arm_nn_vec_mat_mult_t_svdf_s8() 335 int32_t row_loop_cnt = rhs_rows / 3; in arm_nn_vec_mat_mult_t_svdf_s8() 383 const int loop_cnt = rhs_rows % 3; in arm_nn_vec_mat_mult_t_svdf_s8()
|
D | arm_nn_mat_mult_nt_t_s4.c | 55 const int32_t rhs_rows, in arm_nn_mat_mult_nt_t_s4() argument 69 …const uint32x4_t scatter_offset = {0, (uint32_t)rhs_rows, (uint32_t)rhs_rows * 2, (uint32_t)rhs_ro… in arm_nn_mat_mult_nt_t_s4() 74 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s4() 174 dst += (3 * rhs_rows); in arm_nn_mat_mult_nt_t_s4() 180 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s4() 270 dst += (2 * rhs_rows); in arm_nn_mat_mult_nt_t_s4() 282 for (int i = 0; i < rhs_rows; i++) in arm_nn_mat_mult_nt_t_s4() 351 const int32_t tail_rows = rhs_rows & 0x3; in arm_nn_mat_mult_nt_t_s4() 369 for (int32_t rhs_rows_idx = 0; rhs_rows_idx <= (rhs_rows - 4); rhs_rows_idx += 4) in arm_nn_mat_mult_nt_t_s4() 614 dst_ptr += rhs_rows; in arm_nn_mat_mult_nt_t_s4() [all …]
|
D | arm_nn_vec_mat_mult_t_s4.c | 62 const int32_t rhs_rows, in arm_nn_vec_mat_mult_t_s4() argument 66 const int32_t row_loop_cnt = rhs_rows / 4; in arm_nn_vec_mat_mult_t_s4() 603 for (int32_t i_row_loop_cnt = 0; i_row_loop_cnt < rhs_rows % 4; ++i_row_loop_cnt) in arm_nn_vec_mat_mult_t_s4()
|
/cmsis-nn-3.7.0/Source/ConvolutionFunctions/ |
D | arm_convolve_1_x_n_s8.c | 96 const int32_t rhs_rows = output_dims->c; in arm_convolve_1_x_n_s8() local 134 rhs_rows, in arm_convolve_1_x_n_s8() 140 rhs_rows, in arm_convolve_1_x_n_s8() 143 output_data += lhs_rows * rhs_rows; in arm_convolve_1_x_n_s8() 164 rhs_rows, in arm_convolve_1_x_n_s8() 170 rhs_rows, in arm_convolve_1_x_n_s8() 173 output_data += lhs_rows * rhs_rows; in arm_convolve_1_x_n_s8() 198 rhs_rows, in arm_convolve_1_x_n_s8() 204 rhs_rows, in arm_convolve_1_x_n_s8() 207 output_data += lhs_rows * rhs_rows; in arm_convolve_1_x_n_s8()
|
D | arm_convolve_1x1_s8_fast.c | 73 const int32_t rhs_rows = output_dims->c; in arm_convolve_1x1_s8_fast() local 83 rhs_rows, in arm_convolve_1x1_s8_fast() 89 rhs_rows, in arm_convolve_1x1_s8_fast()
|
D | arm_convolve_1x1_s8.c | 71 const int32_t rhs_rows = output_dims->c; in arm_convolve_1x1_s8() local 75 const int32_t output_inc = output_dims->w * rhs_rows; in arm_convolve_1x1_s8() 93 rhs_rows, in arm_convolve_1x1_s8() 99 rhs_rows, in arm_convolve_1x1_s8()
|
D | arm_convolve_1x1_s4.c | 71 const int32_t rhs_rows = output_dims->c; in arm_convolve_1x1_s4() local 75 const int32_t output_inc = output_dims->w * rhs_rows; in arm_convolve_1x1_s4() 93 rhs_rows, in arm_convolve_1x1_s4()
|
D | arm_convolve_1x1_s4_fast.c | 73 const int32_t rhs_rows = output_dims->c; in arm_convolve_1x1_s4_fast() local 83 rhs_rows, in arm_convolve_1x1_s4_fast()
|
D | arm_convolve_s4.c | 101 const int32_t rhs_rows = output_dims->c; in arm_convolve_s4() local 143 rhs_rows, in arm_convolve_s4() 150 out += lhs_rows * rhs_rows; in arm_convolve_s4() 173 rhs_rows, in arm_convolve_s4() 180 out += lhs_rows * rhs_rows; in arm_convolve_s4()
|
D | arm_convolve_s16.c | 94 const int32_t rhs_rows = output_dims->c; in arm_convolve_s16() local 146 rhs_rows, in arm_convolve_s16() 194 rhs_rows, in arm_convolve_s16() 198 out += lhs_rows * rhs_rows; in arm_convolve_s16()
|
D | arm_convolve_1_x_n_s4.c | 104 const int32_t rhs_rows = output_dims->c; in arm_convolve_1_x_n_s4() local 143 rhs_rows, in arm_convolve_1_x_n_s4() 151 output_data += lhs_rows * rhs_rows; in arm_convolve_1_x_n_s4()
|
/cmsis-nn-3.7.0/Include/ |
D | arm_nnsupportfunctions.h | 456 const int32_t rhs_rows, 502 const int32_t rhs_rows, 549 const int32_t rhs_rows, 578 const int32_t rhs_rows, 612 const int32_t rhs_rows, 651 const int32_t rhs_rows, 681 const int32_t rhs_rows, 713 const int32_t rhs_rows, 1843 const int32_t rhs_rows, 1873 const int32_t rhs_rows,
|