summaryrefslogtreecommitdiffstats
path: root/Drivers/CMSIS/NN/Source/SoftmaxFunctions
diff options
context:
space:
mode:
Diffstat (limited to 'Drivers/CMSIS/NN/Source/SoftmaxFunctions')
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/CMakeLists.txt22
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_nn_softmax_common_s8.c141
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c118
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c107
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s16.c122
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c215
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8_s16.c55
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c103
-rw-r--r--Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c74
9 files changed, 957 insertions, 0 deletions
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/CMakeLists.txt b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/CMakeLists.txt
new file mode 100644
index 0000000..622b990
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/CMakeLists.txt
@@ -0,0 +1,22 @@
+#
+# Copyright (c) 2019-2022 Arm Limited.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+file(GLOB SRC "./*_s8.c")
+target_sources(cmsis-nn PRIVATE ${SRC} arm_softmax_s8_s16.c
+ arm_softmax_s16.c
+ arm_nn_softmax_common_s8.c)
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_nn_softmax_common_s8.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_nn_softmax_common_s8.c
new file mode 100644
index 0000000..84d1ac8
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_nn_softmax_common_s8.c
@@ -0,0 +1,141 @@
+/*
+ * Copyright (C) 2022 Arm Limited or its affiliates.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_nn_softmax_common_s8.c
+ * Description: Softmax with s8 input and output of s8 or s16.
+ *
+ * $Date: 17 March 2022
+ * $Revision: V.1.0.1
+ *
+ * Target Processor: Cortex-M processors
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnsupportfunctions.h"
+
+#define ACCUM_BITS 12
+
+/**
+ * @ingroup groupSupport
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+/*
+ * Softmax function with s8 input and output of s8 or s16.
+ *
+ * Refer header file for details.
+ *
+ */
+void arm_nn_softmax_common_s8(const int8_t *input,
+ const int32_t num_rows,
+ const int32_t row_size,
+ const int32_t mult,
+ const int32_t shift,
+ const int32_t diff_min,
+ const bool int16_output,
+ void *output)
+{
+ const int32_t mask = (1 << shift);
+
+ int32_t col = 0;
+ int32_t row_idx;
+
+ for (row_idx = 0; row_idx < num_rows; ++row_idx)
+ {
+ // Find the maximum value in order to ensure numerical stability
+ int8_t max = *input;
+
+ for (col = 1; col < row_size; ++col)
+ {
+ max = MAX(max, input[col]);
+ }
+
+ int32_t diff = 0;
+ int32_t sum = 0;
+
+ for (col = 0; col < row_size; ++col)
+ {
+ diff = input[col] - max;
+ if (diff >= diff_min)
+ {
+ sum += DIV_POW2(EXP_ON_NEG(MUL_SAT(diff * mask, mult)), ACCUM_BITS);
+ }
+ }
+
+ const int32_t headroom = __CLZ(sum);
+ const int32_t shifted_scale = ONE_OVER1((sum > 0 ? sum << headroom : 0) - (1 << 31));
+ int32_t bits_over_unit;
+
+ if (int16_output)
+ {
+ int16_t *output_s16 = (int16_t *)output + row_idx * row_size;
+
+ bits_over_unit = ACCUM_BITS - headroom + 15;
+
+ for (col = 0; col < row_size; ++col)
+ {
+ diff = input[col] - max;
+
+ if (diff >= diff_min)
+ {
+ const int32_t res =
+ DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit) +
+ NN_Q15_MIN;
+ output_s16[col] = (int16_t)CLAMP(res, (int32_t)NN_Q15_MAX, (int32_t)NN_Q15_MIN);
+ }
+ else
+ {
+ output_s16[col] = NN_Q15_MIN;
+ }
+ }
+ }
+ else
+ {
+ int8_t *output_s8 = (int8_t *)output + row_idx * row_size;
+
+ bits_over_unit = ACCUM_BITS - headroom + 23;
+
+ for (col = 0; col < row_size; ++col)
+ {
+ diff = input[col] - max;
+ if (diff >= diff_min)
+ {
+ const int32_t res =
+ DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit) +
+ NN_Q7_MIN;
+ output_s8[col] = (int8_t)CLAMP(res, (int32_t)NN_Q7_MAX, (int32_t)NN_Q7_MIN);
+ }
+ else
+ {
+ output_s8[col] = NN_Q7_MIN;
+ }
+ }
+ }
+
+ input += row_size;
+ }
+}
+
+/**
+ * @} end of NNBasicMath group
+ */
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c
new file mode 100644
index 0000000..18f3e83
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c
@@ -0,0 +1,118 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_q15.c
+ * Description: Q15 softmax function
+ *
+ * $Date: 09. October 2020
+ * $Revision: V.1.0.1
+ *
+ * Target Processor: Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+/**
+ * @brief Q15 softmax function
+ * @param[in] vec_in pointer to input vector
+ * @param[in] dim_vec input vector dimention
+ * @param[out] p_out pointer to output vector
+ *
+ * @details
+ *
+ * Here, instead of typical e based softmax, we use
+ * 2-based softmax, i.e.,:
+ *
+ * y_i = 2^(x_i) / sum(2^x_j)
+ *
+ * The relative output will be different here.
+ * But mathematically, the gradient will be the same
+ * with a log(2) scaling factor.
+ *
+ */
+
+void arm_softmax_q15(const q15_t *vec_in, const uint16_t dim_vec, q15_t *p_out)
+{
+ q31_t sum;
+ int16_t i;
+ uint8_t shift;
+ q31_t base;
+ base = -1 * 0x100000;
+ for (i = 0; i < dim_vec; i++)
+ {
+ if (vec_in[i] > base)
+ {
+ base = vec_in[i];
+ }
+ }
+
+ /* we ignore really small values
+ * anyway, they will be 0 after shrinking
+ * to q15_t
+ */
+ base = base - 16;
+
+ sum = 0;
+
+ for (i = 0; i < dim_vec; i++)
+ {
+ if (vec_in[i] > base)
+ {
+ shift = (uint8_t)__USAT(vec_in[i] - base, 5);
+ sum += 0x1 << shift;
+ }
+ }
+
+ /* This is effectively (0x1 << 32) / sum */
+ int64_t div_base = 0x100000000LL;
+ int output_base = (int32_t)(div_base / sum);
+
+ /* Final confidence will be output_base >> ( 17 - (vec_in[i] - base) )
+ * so 32768 (0x1<<15) -> 100% confidence when sum = 0x1 << 16, output_base = 0x1 << 16
+ * and vec_in[i]-base = 16
+ */
+ for (i = 0; i < dim_vec; i++)
+ {
+ if (vec_in[i] > base)
+ {
+ /* Here minimum value of 17+base-vec[i] will be 1 */
+ shift = (uint8_t)__USAT(17 + base - vec_in[i], 5);
+ p_out[i] = (q15_t)__SSAT((output_base >> shift), 16);
+ }
+ else
+ {
+ p_out[i] = 0;
+ }
+ }
+}
+
+/**
+ * @} end of Softmax group
+ */
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c
new file mode 100644
index 0000000..58eb990
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c
@@ -0,0 +1,107 @@
+/*
+ * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_q7.c
+ * Description: Q7 softmax function
+ *
+ * $Date: 09. October 2020
+ * $Revision: V.1.0.2
+ *
+ * Target Processor: Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+/**
+ * @brief Q7 softmax function
+ * @param[in] vec_in pointer to input vector
+ * @param[in] dim_vec input vector dimention
+ * @param[out] p_out pointer to output vector
+ *
+ * @details
+ *
+ * Here, instead of typical natural logarithm e based softmax, we use
+ * 2-based softmax here, i.e.,:
+ *
+ * y_i = 2^(x_i) / sum(2^x_j)
+ *
+ * The relative output will be different here.
+ * But mathematically, the gradient will be the same
+ * with a log(2) scaling factor.
+ *
+ */
+
+void arm_softmax_q7(const q7_t *vec_in, const uint16_t dim_vec, q7_t *p_out)
+{
+ q31_t sum;
+ int16_t i;
+ uint8_t shift;
+ q15_t base;
+ base = -128;
+
+ /* We first search for the maximum */
+ for (i = 0; i < dim_vec; i++)
+ {
+ if (vec_in[i] > base)
+ {
+ base = vec_in[i];
+ }
+ }
+
+ /*
+ * So the base is set to max-8, meaning
+ * that we ignore really small values.
+ * anyway, they will be 0 after shrinking to q7_t.
+ */
+ base = base - (1 << 3);
+
+ sum = 0;
+
+ for (i = 0; i < dim_vec; i++)
+ {
+ shift = (uint8_t)__USAT(vec_in[i] - base, 3);
+ sum += 0x1 << shift;
+ }
+
+ /* This is effectively (0x1 << 20) / sum */
+ int output_base = (1 << 20) / sum;
+
+ for (i = 0; i < dim_vec; i++)
+ {
+
+ /* Here minimum value of 13+base-vec_in[i] will be 5 */
+ shift = (uint8_t)__USAT(13 + base - vec_in[i], 5);
+ p_out[i] = (q7_t)__SSAT((output_base >> shift), 8);
+ }
+}
+
+/**
+ * @} end of Softmax group
+ */
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s16.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s16.c
new file mode 100644
index 0000000..e840893
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s16.c
@@ -0,0 +1,122 @@
+/*
+ * Copyright (C) 2022 Arm Limited or its affiliates.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_s16.c
+ * Description: S16 softmax function
+ *
+ * $Date: 9 March 2022
+ * $Revision: V.1.0.0
+ *
+ * Target Processor: Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+#include "arm_nnsupportfunctions.h"
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+arm_status arm_softmax_s16(const int16_t *input,
+ const int32_t num_rows,
+ const int32_t row_size,
+ const int32_t mult,
+ const int32_t shift,
+ const cmsis_nn_softmax_lut_s16 *softmax_params,
+ int16_t *output)
+{
+ int32_t col = 0;
+ int32_t row_idx;
+
+ if (softmax_params->exp_lut == NULL || softmax_params->one_by_one_lut == NULL)
+ {
+ return ARM_MATH_ARGUMENT_ERROR;
+ }
+
+ for (row_idx = 0; row_idx < num_rows; ++row_idx)
+ {
+ // Find the maximum value in order to ensure numerical stability
+ int16_t max = *input;
+ for (col = 1; col < row_size; ++col)
+ {
+ max = MAX(max, input[col]);
+ }
+
+ int32_t diff = 0;
+ int32_t sum = 0;
+ int16_t *cached_exp_results = output;
+
+ for (col = 0; col < row_size; ++col)
+ {
+ diff = input[col] - max;
+ const int32_t scaled_diff = arm_nn_requantize(diff, mult, shift);
+ const int32_t symmetric_scaled_diff = scaled_diff + NN_Q15_MAX;
+ const int16_t saturated_symmetric_scaled_diff = MIN(MAX(symmetric_scaled_diff, NN_Q15_MIN), NN_Q15_MAX);
+
+ // Lookup from exp table and cache result for next step
+ const int16_t index = (256 + (saturated_symmetric_scaled_diff >> 7));
+ const int16_t offset = saturated_symmetric_scaled_diff & 0x7f;
+ const int16_t base = softmax_params->exp_lut[index];
+ const int16_t slope = softmax_params->exp_lut[index + 1] - softmax_params->exp_lut[index];
+ const int16_t delta = (slope * offset + 64) >> 7;
+ const int16_t result = (base + delta);
+ cached_exp_results[col] = result;
+
+ sum += cached_exp_results[col];
+ }
+
+ const int32_t headroom = __CLZ(sum);
+
+ // Compute the reciprocal 1/sum
+ const int32_t shifted_sum = (((sum) << (headroom - 1)) + (1 << 13)) >> 14;
+
+ // Since LUT computes 1/(1 + x), compute x = (sum - 1) => -65536
+ // Since LUT expects a symmetrical input, recenter from [UINT16_MIN, UINT16_MAX] to [INT16_MIN, INT16_MAX] =>
+ // -32768 ==> So in total -65536 -32768 => -98304
+ const int16_t symmetric_shifted_sum = shifted_sum - 98304;
+
+ // Lookup from one by one table
+ const int16_t index = (256 + (symmetric_shifted_sum >> 7));
+ const int16_t offset = symmetric_shifted_sum & 0x7f;
+ const int16_t base = softmax_params->one_by_one_lut[index];
+ const int16_t slope = softmax_params->one_by_one_lut[index + 1] - softmax_params->one_by_one_lut[index];
+ const int16_t delta = (slope * offset + 64) >> 7;
+ const int16_t one_by_one_result = (base + delta);
+
+ for (col = 0; col < row_size; ++col)
+ {
+ const int16_t right_shift = 30 - headroom;
+ int32_t result = (cached_exp_results[col] * one_by_one_result) >> right_shift;
+ result = (result + 1) >> 1; // Last shift position and insert round
+ output[col] = (int16_t)result;
+ }
+
+ output += row_size;
+ input += row_size;
+ }
+
+ return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of Softmax group
+ */
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c
new file mode 100644
index 0000000..fd70597
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c
@@ -0,0 +1,215 @@
+/*
+ * Copyright (C) 2010-2022 Arm Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_s8.c
+ * Description: S8 softmax function
+ *
+ * $Date: 9 March 2022
+ * $Revision: V.2.1.0
+ *
+ * Target Processor: Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+#include "arm_nnsupportfunctions.h"
+
+#define ACCUM_BITS 12
+
+#ifdef ARM_MATH_MVEI
+static int32x4_t arm_exp_on_negative_values_mve_32x4(int32x4_t val)
+{
+#define SHIFT_START (24)
+ int32_t shift = SHIFT_START;
+ int32x4_t mask;
+
+ const int32x4_t val_mod_minus_quarter =
+ vandq_s32(val, vdupq_n_s32((1 << SHIFT_START) - 1)) - vdupq_n_s32(1 << SHIFT_START);
+ const int32x4_t remainder = vsubq_s32(val_mod_minus_quarter, val);
+ const int32x4_t x = vaddq_n_s32(val_mod_minus_quarter << 5, 1 << 28);
+ const int32x4_t x2 = MUL_SAT_MVE(x, x);
+ const int32x4_t op_1 = DIV_POW2_MVE(MUL_SAT_MVE(x2, x2), 2) + MUL_SAT_MVE(x2, x);
+ const int32x4_t op_2 = x + DIV_POW2_MVE(MUL_SAT_MVE(op_1, vdupq_n_s32(715827883)) + x2, 1);
+ int32x4_t result = vdupq_n_s32(1895147668) + MUL_SAT_MVE(vdupq_n_s32(1895147668), op_2);
+
+#define SELECT_IF_NON_ZERO(x) \
+ { \
+ mve_pred16_t p = vcmpneq_n_s32(remainder & vdupq_n_s32(1 << shift++), 0); \
+ mask = vmvnq_m_s32(vdupq_n_s32(0), vdupq_n_s32(0), p); \
+ result = SELECT_USING_MASK(mask, MUL_SAT_MVE(result, vdupq_n_s32(x)), result); \
+ }
+
+ SELECT_IF_NON_ZERO(1672461947)
+ SELECT_IF_NON_ZERO(1302514674)
+ SELECT_IF_NON_ZERO(790015084)
+ SELECT_IF_NON_ZERO(290630308)
+ SELECT_IF_NON_ZERO(39332535)
+ SELECT_IF_NON_ZERO(720401)
+ SELECT_IF_NON_ZERO(242)
+
+#undef SELECT_IF_NON_ZERO
+
+ mve_pred16_t p = vcmpeqq_n_s32(val, 0);
+ mask = vmvnq_m_s32(vdupq_n_s32(0), vdupq_n_s32(0), p);
+
+ result = SELECT_USING_MASK(mask, vdupq_n_s32(NN_Q31_MAX), result);
+ return result;
+}
+#endif
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+void arm_softmax_s8(const int8_t *input,
+ const int32_t num_rows,
+ const int32_t row_size,
+ const int32_t mult,
+ const int32_t shift,
+ const int32_t diff_min,
+ int8_t *output)
+{
+#ifdef ARM_MATH_MVEI
+
+#define ACT_MIN ((int8_t)NN_Q7_MIN)
+#define ACT_MAX ((int8_t)NN_Q7_MAX)
+
+ const int32_t mask = (1 << shift);
+
+ for (int i_num_rows = 0; i_num_rows < num_rows; ++i_num_rows)
+ {
+ int8_t max = ACT_MIN;
+
+ int32_t vec_count = (row_size + 15) / 16;
+ uint32_t r_count = (uint32_t)row_size;
+ for (int i = 0; i < vec_count; i++)
+ {
+ mve_pred16_t p = vctp8q(r_count);
+ const int8x16_t ip = vldrbq_z_s8(&input[i * 16], p);
+ max = vmaxvq_p_s8(max, ip, p);
+ r_count -= 16;
+ }
+
+ vec_count = row_size / 4;
+ int32_t idx = 0;
+ int32_t sum = 0;
+
+ while (vec_count)
+ {
+ int32x4_t ip = vldrbq_s32(&input[idx * 4]);
+ ip = vsubq_n_s32(ip, max);
+ mve_pred16_t p = vcmpgeq_n_s32(ip, diff_min);
+ if (p != 0)
+ {
+ ip = vmulq_n_s32(ip, mask);
+
+ int32x4_t res = MUL_SAT_MVE(ip, vdupq_n_s32(mult));
+
+ res = arm_exp_on_negative_values_mve_32x4(res);
+ res = DIV_POW2_MVE(res, ACCUM_BITS);
+ res = vpselq_s32(res, vdupq_n_s32(0), p);
+ sum += vaddvq_s32(res);
+ }
+
+ vec_count--;
+ idx++;
+ }
+
+ const int32_t tail_idx = row_size & ~3;
+ for (int i = 0; i < (row_size & 3); i++)
+ {
+ const int32_t diff = input[tail_idx + i] - max;
+ if (diff >= diff_min)
+ {
+ sum += DIV_POW2(EXP_ON_NEG(MUL_SAT(diff * mask, mult)), ACCUM_BITS);
+ }
+ }
+
+ const int32_t headroom = __CLZ((uint32_t)sum);
+ const int32_t bits_over_unit = ACCUM_BITS - headroom + 23;
+ const int32_t shifted_scale = ONE_OVER1((sum > 0 ? sum << headroom : 0) - (1 << 31));
+
+ vec_count = row_size / 4;
+ idx = 0;
+
+ while (vec_count)
+ {
+ int32x4_t ip = vldrbq_s32(&input[idx]);
+ ip = vsubq_n_s32(ip, max);
+
+ mve_pred16_t p = vcmpgeq_n_s32(ip, diff_min);
+
+ int32x4_t tmp_res;
+
+ if (p != 0)
+ {
+ ip = vmulq_n_s32(ip, mask);
+
+ tmp_res = MUL_SAT_MVE(ip, vdupq_n_s32(mult));
+ tmp_res = arm_exp_on_negative_values_mve_32x4(tmp_res);
+ tmp_res = MUL_SAT_MVE(vdupq_n_s32(shifted_scale), tmp_res);
+ tmp_res = DIV_POW2_MVE(tmp_res, bits_over_unit);
+ tmp_res += vdupq_n_s32(ACT_MIN);
+
+ tmp_res = vmaxq_s32(tmp_res, vdupq_n_s32(ACT_MIN));
+ tmp_res = vminq_s32(tmp_res, vdupq_n_s32(ACT_MAX));
+ tmp_res = vpselq_s32(tmp_res, vdupq_n_s32(ACT_MIN), p);
+ }
+ else
+ {
+ tmp_res = vdupq_n_s32(ACT_MIN);
+ }
+ vstrbq_s32(&output[idx], tmp_res);
+ vec_count--;
+ idx += 4;
+ }
+
+ for (int i = 0; i < (row_size & 3); i++)
+ {
+ int32_t diff = input[tail_idx + i] - max;
+ if (diff >= diff_min)
+ {
+ const int32_t res =
+ DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit) +
+ NN_Q7_MIN;
+ output[tail_idx + i] = (int8_t)CLAMP(res, (int32_t)ACT_MAX, (int32_t)ACT_MIN);
+ }
+ else
+ {
+ output[tail_idx + i] = ACT_MIN;
+ }
+ }
+
+ input += row_size;
+ output += row_size;
+ }
+#else
+ arm_nn_softmax_common_s8(input, num_rows, row_size, mult, shift, diff_min, false, (void *)output);
+#endif
+}
+
+/**
+ * @} end of Softmax group
+ */
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8_s16.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8_s16.c
new file mode 100644
index 0000000..9ba0b9a
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8_s16.c
@@ -0,0 +1,55 @@
+/*
+ * Copyright (C) 2022 Arm Limited or its affiliates.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_s8_s16.c
+ * Description: S8 to s16 softmax function
+ *
+ * $Date: 7 January 2022
+ * $Revision: V.1.0.0
+ *
+ * Target Processor: Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+#include "arm_nnsupportfunctions.h"
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+void arm_softmax_s8_s16(const int8_t *input,
+ const int32_t num_rows,
+ const int32_t row_size,
+ const int32_t mult,
+ const int32_t shift,
+ const int32_t diff_min,
+ int16_t *output)
+{
+ arm_nn_softmax_common_s8(input, num_rows, row_size, mult, shift, diff_min, true, (void *)output);
+}
+/**
+ * @} end of Softmax group
+ */
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c
new file mode 100644
index 0000000..c4df8f8
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c
@@ -0,0 +1,103 @@
+/*
+ * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_u8.c
+ * Description: U8 softmax function
+ *
+ * $Date: 09. October 2020
+ * $Revision: V.1.0.2
+ *
+ * Target Processor: Cortex-M CPUs
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+#include "arm_nnsupportfunctions.h"
+
+#define ACCUM_BITS 12
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+void arm_softmax_u8(const uint8_t *input,
+ const int32_t num_rows,
+ const int32_t row_size,
+ const int32_t mult,
+ const int32_t shift,
+ const int32_t diff_min,
+ uint8_t *output)
+{
+ const int32_t mask = (1 << shift);
+
+ int32_t col = 0;
+ int32_t row_idx;
+
+ for (row_idx = 0; row_idx < num_rows; ++row_idx)
+ {
+ // Find the maximum value in order to ensure numerical stability
+ uint8_t max = *input;
+
+ for (col = 1; col < row_size; ++col)
+ {
+ max = MAX(max, input[col]);
+ }
+
+ int32_t diff = 0;
+ int32_t sum = 0;
+
+ for (col = 0; col < row_size; ++col)
+ {
+ diff = input[col] - max;
+ if (diff >= diff_min)
+ {
+ sum += DIV_POW2(EXP_ON_NEG(MUL_SAT(diff * mask, mult)), ACCUM_BITS);
+ }
+ }
+
+ const int32_t headroom = __CLZ((uint32_t)sum);
+ const int32_t bits_over_unit = ACCUM_BITS - headroom + 23;
+ const int32_t shifted_scale = ONE_OVER1((sum << headroom) - (1 << 31));
+
+ for (col = 0; col < row_size; ++col)
+ {
+ diff = input[col] - max;
+ if (diff >= diff_min)
+ {
+ const int32_t res =
+ DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit);
+ output[col] = (uint8_t)CLAMP(res, (int32_t)255, (int32_t)0);
+ }
+ else
+ {
+ output[col] = 0;
+ }
+ }
+ input += row_size;
+ output += row_size;
+ }
+}
+/**
+ * @} end of Softmax group
+ */ \ No newline at end of file
diff --git a/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c
new file mode 100644
index 0000000..66e892e
--- /dev/null
+++ b/Drivers/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c
@@ -0,0 +1,74 @@
+/*
+ * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_softmax_with_batch_q7.c
+ * Description: Q7 softmax function
+ *
+ * $Date: 09. October 2020
+ * $Revision: V.1.0.1
+ *
+ * Target Processor: Cortex-M and Cortex-A cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+/**
+ * @brief Q7 softmax function with batch parameter
+ * @param[in] vec_in pointer to input vector
+ * @param[in] nb_batches number of batches
+ * @param[in] dim_vec input vector dimention
+ * @param[out] p_out pointer to output vector
+ *
+ * @details
+ *
+ * Here, instead of typical natural logarithm e based softmax, we use
+ * 2-based softmax here, i.e.,:
+ *
+ * y_i = 2^(x_i) / sum(2^x_j)
+ *
+ * The relative output will be different here.
+ * But mathematically, the gradient will be the same
+ * with a log(2) scaling factor.
+ *
+ */
+
+void arm_softmax_with_batch_q7(const q7_t *vec_in, const uint16_t nb_batches, const uint16_t dim_vec, q7_t *p_out)
+{
+ for (int i = 0; i < nb_batches; i++)
+ {
+ arm_softmax_q7(vec_in, dim_vec, p_out);
+ vec_in += dim_vec;
+ p_out += dim_vec;
+ }
+}
+
+/**
+ * @} end of Softmax group
+ */