-
Notifications
You must be signed in to change notification settings - Fork 57
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Minimum and Maximum s8 operator support (#148)
* Adds Minimum and Maximum functions for s8 * Adds Refactored Unit Tests for Minimum and Maximum s8 * Fix small issue with build_and_run_tests.sh Change-Id: I38333a14888b59293dcafa633105ec65c2d582a1 --------- Signed-off-by: Ryan O'Shea <[email protected]>
- Loading branch information
Showing
86 changed files
with
2,802 additions
and
6 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,263 @@ | ||
| /* | ||
| * SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <[email protected]> | ||
| * | ||
| * 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_maximum_s8 | ||
| * Description: Minimum and Maximum | ||
| * | ||
| * $Date: 08 October 2024 | ||
| * $Revision: V.1.0.0 | ||
| * | ||
| * Target : Arm(R) M-Profile Architecture | ||
| * | ||
| * -------------------------------------------------------------------- */ | ||
|
|
||
| #include "arm_nnfunctions.h" | ||
| #include "arm_nnsupportfunctions.h" | ||
|
|
||
| /** | ||
| * @ingroup Public | ||
| */ | ||
|
|
||
| /** | ||
| * @addtogroup minimumMaximum | ||
| * @{ | ||
| */ | ||
|
|
||
| static arm_cmsis_nn_status | ||
| arm_max_no_broadcast_s8(const int8_t *input_1, const int8_t *input_2, int8_t *output, int32_t flat_size) | ||
| { | ||
| #if defined(ARM_MATH_MVEI) | ||
| while (flat_size > 0) | ||
| { | ||
| mve_pred16_t p = vctp8q(flat_size); | ||
|
|
||
| int8x16_t vec1 = vldrbq_z_s8(input_1, p); | ||
| input_1 += 16; | ||
| int8x16_t vec2 = vldrbq_z_s8(input_2, p); | ||
| input_2 += 16; | ||
|
|
||
| vstrbq_p_s8(output, vmaxq_s8(vec1, vec2), p); | ||
| output += 16; | ||
| flat_size -= 16; | ||
| } | ||
| #else | ||
| while (flat_size > 0) | ||
| { | ||
| int8_t in1 = *input_1++; | ||
| int8_t in2 = *input_2++; | ||
| *output++ = in1 >= in2 ? in1 : in2; | ||
| --flat_size; | ||
| } | ||
| #endif | ||
|
|
||
| return ARM_CMSIS_NN_SUCCESS; | ||
| } | ||
|
|
||
| static arm_cmsis_nn_status | ||
| arm_max_scalar_s8(const int8_t *input_1, const int8_t *input_2, int8_t *output, int32_t flat_size) | ||
| { | ||
| #if defined(ARM_MATH_MVEI) | ||
| int8x16_t scalar_vec = vdupq_n_s8(*input_1); | ||
|
|
||
| while (flat_size > 0) | ||
| { | ||
| mve_pred16_t p = vctp8q(flat_size); | ||
| int8x16_t vec = vldrbq_z_s8(input_2, p); | ||
| input_2 += 16; | ||
|
|
||
| vstrbq_p_s8(output, vmaxq_s8(scalar_vec, vec), p); | ||
| output += 16; | ||
| flat_size -= 16; | ||
| } | ||
| #else | ||
| int8_t in1 = *input_1; | ||
| while (flat_size > 0) | ||
| { | ||
| int8_t in2 = *input_2++; | ||
| *output++ = in1 >= in2 ? in1 : in2; | ||
| --flat_size; | ||
| } | ||
| #endif | ||
| return ARM_CMSIS_NN_SUCCESS; | ||
| } | ||
|
|
||
| /* | ||
| * s8 maximum | ||
| * | ||
| * Refer header file for details. | ||
| * | ||
| */ | ||
| arm_cmsis_nn_status arm_maximum_s8(const cmsis_nn_context *ctx, | ||
| const int8_t *input_1_data, | ||
| const cmsis_nn_dims *input_1_dims, | ||
| const int8_t *input_2_data, | ||
| const cmsis_nn_dims *input_2_dims, | ||
| int8_t *output_data, | ||
| const cmsis_nn_dims *output_dims) | ||
| { | ||
| (void)ctx; | ||
| const int32_t output_batch = output_dims->n; | ||
| const int32_t output_height = output_dims->h; | ||
| const int32_t output_width = output_dims->w; | ||
|
|
||
| const int32_t input_1_batch = input_1_dims->n; | ||
| const int32_t input_1_height = input_1_dims->h; | ||
| const int32_t input_1_width = input_1_dims->w; | ||
| const int32_t input_1_channels = input_1_dims->c; | ||
|
|
||
| const int32_t input_2_batch = input_2_dims->n; | ||
| const int32_t input_2_height = input_2_dims->h; | ||
| const int32_t input_2_width = input_2_dims->w; | ||
| const int32_t input_2_channels = input_2_dims->c; | ||
|
|
||
| int32_t flat_size_1 = input_1_batch * input_1_height * input_1_width * input_1_channels; | ||
| int32_t flat_size_2 = input_2_batch * input_2_height * input_2_width * input_2_channels; | ||
|
|
||
| if (arm_check_broadcast_required(input_1_dims, input_2_dims)) | ||
| { | ||
| if (flat_size_1 == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_1_data, input_2_data, output_data, flat_size_2); | ||
| } | ||
| else if (flat_size_2 == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_2_data, input_1_data, output_data, flat_size_1); | ||
| } | ||
| else | ||
| { | ||
| int32_t width_1_diff = input_1_width >= input_2_width ? 0 : input_1_channels; | ||
| int32_t width_2_diff = input_2_width >= input_1_width ? 0 : input_2_channels; | ||
|
|
||
| int32_t height_1_diff = | ||
| input_1_height >= input_2_height ? width_1_diff : -input_1_width * (input_1_channels - width_1_diff); | ||
| int32_t height_2_diff = | ||
| input_2_height >= input_1_height ? width_2_diff : -input_2_width * (input_2_channels - width_2_diff); | ||
|
|
||
| int32_t batch_1_diff = | ||
| input_1_batch >= input_2_batch ? input_1_channels * input_1_width * input_1_height : 0; | ||
| int32_t batch_2_diff = | ||
| input_2_batch >= input_1_batch ? input_2_channels * input_2_width * input_2_height : 0; | ||
|
|
||
| for (int32_t i_out_batch = 0; i_out_batch < output_batch; i_out_batch++) | ||
| { | ||
| const int8_t *input_1_ptr = input_1_data; | ||
| const int8_t *input_2_ptr = input_2_data; | ||
| flat_size_1 = input_1_height * input_1_width * input_1_channels; | ||
| flat_size_2 = input_2_height * input_2_width * input_2_channels; | ||
| if (input_1_height == input_2_height && input_1_width == input_2_width && | ||
| input_1_channels == input_2_channels) | ||
| { | ||
| arm_max_no_broadcast_s8(input_1_ptr, input_2_ptr, output_data, flat_size_1); | ||
| output_data += flat_size_1; | ||
| } | ||
| else if (flat_size_1 == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_1_ptr, input_2_ptr, output_data, flat_size_2); | ||
| output_data += flat_size_2; | ||
| } | ||
| else if (flat_size_2 == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_2_ptr, input_1_ptr, output_data, flat_size_1); | ||
| output_data += flat_size_1; | ||
| } | ||
| else | ||
| { | ||
| flat_size_1 = input_1_width * input_1_channels; | ||
| flat_size_2 = input_2_width * input_2_channels; | ||
| for (int32_t i_out_height = 0; i_out_height < output_height; i_out_height++) | ||
| { | ||
| if (input_1_width == input_2_width && input_1_channels == input_2_channels) | ||
| { | ||
| arm_max_no_broadcast_s8(input_1_ptr, input_2_ptr, output_data, flat_size_1); | ||
| output_data += flat_size_1; | ||
| input_1_ptr += flat_size_1; | ||
| input_2_ptr += flat_size_1; | ||
| } | ||
| else if (flat_size_1 == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_1_ptr, input_2_ptr, output_data, flat_size_2); | ||
| output_data += flat_size_2; | ||
| ++input_1_ptr; | ||
| input_2_ptr += flat_size_2; | ||
| } | ||
| else if (flat_size_2 == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_2_ptr, input_1_ptr, output_data, flat_size_1); | ||
| output_data += flat_size_1; | ||
| ++input_2_ptr; | ||
| input_1_ptr += flat_size_1; | ||
| } | ||
| else | ||
| { | ||
| for (int32_t i_out_width = 0; i_out_width < output_width; i_out_width++) | ||
| { | ||
| if (input_1_channels == input_2_channels) | ||
| { | ||
| arm_max_no_broadcast_s8(input_1_ptr, input_2_ptr, output_data, input_1_channels); | ||
| output_data += input_1_channels; | ||
| input_1_ptr += input_1_channels; | ||
| input_2_ptr += input_1_channels; | ||
| } | ||
| else if (input_1_channels == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_1_ptr, input_2_ptr, output_data, input_2_channels); | ||
| output_data += input_2_channels; | ||
| input_1_ptr++; | ||
| input_2_ptr += input_2_channels; | ||
| } | ||
| else if (input_2_channels == 1) | ||
| { | ||
| // arm_max_scalar expects the tensor with the scalar value to be provided first | ||
| arm_max_scalar_s8(input_2_ptr, input_1_ptr, output_data, input_1_channels); | ||
| output_data += input_1_channels; | ||
| input_1_ptr += input_1_channels; | ||
| input_2_ptr++; | ||
| } | ||
| input_1_ptr -= width_1_diff; | ||
| input_2_ptr -= width_2_diff; | ||
| } | ||
| } | ||
| input_1_ptr += height_1_diff; | ||
| input_2_ptr += height_2_diff; | ||
| } | ||
| } | ||
| input_1_data += batch_1_diff; | ||
| input_2_data += batch_2_diff; | ||
| } | ||
| } | ||
| } | ||
| else | ||
| { | ||
| arm_max_no_broadcast_s8(input_1_data, input_2_data, output_data, flat_size_1); | ||
| } | ||
|
|
||
| return (ARM_CMSIS_NN_SUCCESS); | ||
| } | ||
|
|
||
| /** | ||
| * @} end of Doxygen group | ||
| */ |
Oops, something went wrong.