| /cmsis-nn-3.6.0-3.5.0/Source/NNSupportFunctions/ |
| D | arm_nn_mat_mul_core_4x_s8.c | 114 int32x4_t res = {acc_n0, acc_n1, acc_n2, acc_n3}; in arm_nn_mat_mul_core_4x_s8() local
120 res = vaddq_n_s32(res, sum_tmp); in arm_nn_mat_mul_core_4x_s8()
122 res = arm_requantize_mve(res, quant_params->multiplier[i], quant_params->shift[i]); in arm_nn_mat_mul_core_4x_s8()
123 res = vaddq_n_s32(res, conv_params->output_offset); in arm_nn_mat_mul_core_4x_s8()
125 res = vmaxq_s32(res, vdupq_n_s32(conv_params->activation.min)); in arm_nn_mat_mul_core_4x_s8()
126 res = vminq_s32(res, vdupq_n_s32(conv_params->activation.max)); in arm_nn_mat_mul_core_4x_s8()
129 vstrbq_scatter_offset_s32(output, scatter_offset, res); in arm_nn_mat_mul_core_4x_s8()
|
| D | arm_nn_mat_mul_core_1x_s8.c | 104 int32x4_t res = vldrwq_s32(acc); in arm_nn_mat_mul_core_1x_s8() local
107 res = vaddq_s32(res, vldrwq_s32(bias)); in arm_nn_mat_mul_core_1x_s8()
110 res = arm_requantize_mve_32x4(res, vldrwq_s32(output_mult), vldrwq_s32(output_shift)); in arm_nn_mat_mul_core_1x_s8()
113 res = vaddq_n_s32(res, out_offset); in arm_nn_mat_mul_core_1x_s8()
114 res = vmaxq_s32(res, vdupq_n_s32(out_activation_min)); in arm_nn_mat_mul_core_1x_s8()
115 res = vminq_s32(res, vdupq_n_s32(out_activation_max)); in arm_nn_mat_mul_core_1x_s8()
116 vstrbq_s32(output, res); in arm_nn_mat_mul_core_1x_s8()
|
| D | arm_nn_mat_mult_nt_t_s8.c | 129 int32x4_t res = {acc_n0, acc_n1, acc_n2, acc_n3}; in arm_nn_mat_mult_nt_t_s8() local
135 res = vaddq_n_s32(res, sum_tmp); in arm_nn_mat_mult_nt_t_s8()
137 res = arm_requantize_mve(res, dst_multipliers[i], dst_shifts[i]); in arm_nn_mat_mult_nt_t_s8()
138 res = vaddq_n_s32(res, dst_offset); in arm_nn_mat_mult_nt_t_s8()
140 res = vmaxq_s32(res, vdupq_n_s32(activation_min)); in arm_nn_mat_mult_nt_t_s8()
141 res = vminq_s32(res, vdupq_n_s32(activation_max)); in arm_nn_mat_mult_nt_t_s8()
144 vstrbq_scatter_offset_s32(dst, scatter_offset, res); in arm_nn_mat_mult_nt_t_s8()
198 int32x4_t res = vldrwq_s32(acc); in arm_nn_mat_mult_nt_t_s8() local
199 res = arm_requantize_mve_32x4(res, vldrwq_s32(multipliers), vldrwq_s32(shifts)); in arm_nn_mat_mult_nt_t_s8()
202 res = vaddq_n_s32(res, dst_offset); in arm_nn_mat_mult_nt_t_s8()
[all …]
|
| /cmsis-nn-3.6.0-3.5.0/Source/ConvolutionFunctions/ |
| D | arm_nn_mat_mult_s8.c | 104 int32x4_t res = {acc_0, acc_1, acc_2, acc_3}; in arm_nn_mat_mult_s8() local
107 res = vaddq_n_s32(res, bias[i_out_ch]); in arm_nn_mat_mult_s8()
109 res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]); in arm_nn_mat_mult_s8()
110 res = vaddq_n_s32(res, out_offset); in arm_nn_mat_mult_s8()
112 res = vmaxq_s32(res, vdupq_n_s32(activation_min)); in arm_nn_mat_mult_s8()
113 res = vminq_s32(res, vdupq_n_s32(activation_max)); in arm_nn_mat_mult_s8()
116 vstrbq_scatter_offset_s32(&out[i_out_ch], scatter_offset, res); in arm_nn_mat_mult_s8()
|
| D | arm_depthwise_conv_s8.c | 116 int32x4_t res = vldrwq_s32(out_buff); in depthwise_conv_s8_mult_4() local
117 … res = arm_requantize_mve_32x4(res, vldrwq_s32(output_mult), vldrwq_s32(output_shift)); in depthwise_conv_s8_mult_4()
120 res = vaddq_n_s32(res, output_offset); in depthwise_conv_s8_mult_4()
122 res = vmaxq_s32(res, vdupq_n_s32(output_activation_min)); in depthwise_conv_s8_mult_4()
123 res = vminq_s32(res, vdupq_n_s32(output_activation_max)); in depthwise_conv_s8_mult_4()
124 vstrbq_s32(output, res); in depthwise_conv_s8_mult_4()
|
| /cmsis-nn-3.6.0-3.5.0/Source/SoftmaxFunctions/ |
| D | arm_softmax_s8.c | 128 int32x4_t res = MUL_SAT_MVE(ip, vdupq_n_s32(mult)); in arm_softmax_s8() local
130 res = arm_exp_on_negative_values_mve_32x4(res); in arm_softmax_s8()
131 res = DIV_POW2_MVE(res, ACCUM_BITS); in arm_softmax_s8()
132 res = vpselq_s32(res, vdupq_n_s32(0), p); in arm_softmax_s8()
133 sum += vaddvq_s32(res); in arm_softmax_s8()
194 const int32_t res = in arm_softmax_s8() local
197 output[tail_idx + i] = (int8_t)CLAMP(res, (int32_t)ACT_MAX, (int32_t)ACT_MIN); in arm_softmax_s8()
|
| D | arm_nn_softmax_common_s8.c | 108 const int32_t res = in arm_nn_softmax_common_s8() local
111 output_s16[col] = (int16_t)CLAMP(res, (int32_t)NN_Q15_MAX, (int32_t)NN_Q15_MIN); in arm_nn_softmax_common_s8()
130 const int32_t res = in arm_nn_softmax_common_s8() local
133 output_s8[col] = (int8_t)CLAMP(res, (int32_t)NN_Q7_MAX, (int32_t)NN_Q7_MIN); in arm_nn_softmax_common_s8()
|
| D | arm_softmax_u8.c | 88 const int32_t res = in arm_softmax_u8() local
90 output[col] = (uint8_t)CLAMP(res, (int32_t)255, (int32_t)0); in arm_softmax_u8()
|
| /cmsis-nn-3.6.0-3.5.0/Source/PoolingFunctions/ |
| D | arm_max_pool_s16.c | 97 int16x8_t res = vmaxq_x_s16(src, min, p); in clamp_output() local
98 res = vminq_x_s16(res, max, p); in clamp_output()
99 vstrhq_p_s16(source, res, p); in clamp_output()
|
| D | arm_max_pool_s8.c | 109 int8x16_t res = vmaxq_x_s8(src, vmin, p); in clamp_output() local
110 res = vminq_x_s8(res, vmax, p); in clamp_output()
111 vstrbq_p_s8(source, res, p); in clamp_output()
|
| /cmsis-nn-3.6.0-3.5.0/Source/SVDFunctions/ |
| D | arm_svdf_state_s16_s8.c | 114 arm_cmsis_nn_status res = arm_nn_vec_mat_mult_t_svdf_s8(input, in arm_svdf_state_s16_s8() local
126 if (res != ARM_CMSIS_NN_SUCCESS) in arm_svdf_state_s16_s8()
128 return res; in arm_svdf_state_s16_s8()
|
| D | arm_svdf_s8.c | 114 arm_cmsis_nn_status res = arm_nn_vec_mat_mult_t_s8(input, in arm_svdf_s8() local
128 if (res != ARM_CMSIS_NN_SUCCESS) in arm_svdf_s8()
130 return res; in arm_svdf_s8()
|
