Add int4 MVE support for FC (#134)

Unit tests are updated as well.
ARM-software · May 6, 2024 · 01dee38 · 01dee38
1 parent 429fb5c
commit 01dee38
Show file tree

Hide file tree

Showing 73 changed files with 1,368 additions and 396 deletions.
diff --git a/Include/arm_nnsupportfunctions.h b/Include/arm_nnsupportfunctions.h
@@ -22,7 +22,7 @@
  * Description:  Public header file of support functions for CMSIS NN Library
  *
  * $Date:        30 April 2024
- * $Revision:    V.21.1.0
+ * $Revision:    V.22.0.0
  *
  * Target :  Arm(R) M-Profile Architecture
  * -------------------------------------------------------------------- */
@@ -596,8 +596,6 @@ arm_cmsis_nn_status arm_nn_mat_mult_nt_t_s8_s32(const int8_t *lhs,
  * @param[in]      rhs_rows        Number of rows in the right-hand side input matrix
  * @param[in]      activation_min  Minimum value to clamp the output to. Range: int8
  * @param[in]      activation_max  Maximum value to clamp the output to. Range: int8
- * @param[in]      address_offset  Memory position offset for dst. First output is stored at 'dst', the
- *                                 second at 'dst + address_offset' and so on. Default value is typically 1.
  *
  * @return         The function returns <code>ARM_CMSIS_NN_SUCCESS</code>
  *
@@ -613,8 +611,7 @@ arm_cmsis_nn_status arm_nn_vec_mat_mult_t_s4(const int8_t *lhs,
                                              const int32_t rhs_cols,
                                              const int32_t rhs_rows,
                                              const int32_t activation_min,
-                                             const int32_t activation_max,
-                                             const int32_t address_offset);
+                                             const int32_t activation_max);
 
 /**
  * @brief s8 Vector by Matrix (transposed) multiplication

diff --git a/README.md b/README.md
@@ -29,7 +29,7 @@ Examples are Cortex-M55 or Cortex-M85 configured with MVE.
 | Conv2D          | Yes         | Yes        | Yes        | Yes         | Yes          | Yes          | Yes         | Yes          | Yes          |
 | DepthwiseConv2D | Yes         | Yes        | Yes        | Yes         | Yes          | Yes          | Yes         | Yes          | Yes          |
 | TransposeConv2D | Yes         | No         | No         | Yes         | No           | No           | Yes         | No           | No           |
-| Fully Connected | Yes         | Yes        | Yes        | Yes         | Yes          | Yes          | Yes         | Yes          | No           |
+| Fully Connected | Yes         | Yes        | Yes        | Yes         | Yes          | Yes          | Yes         | Yes          | Yes          |
 | Add             | Yes         | Yes        | N/A        | Yes         | Yes          | N/A          | Yes         | Yes          | N/A          |
 | Mul             | Yes         | Yes        | N/A        | Yes         | Yes          | N/A          | Yes         | Yes          | N/A          |
 | MaxPooling      | Yes         | Yes        | N/A        | Yes         | Yes          | N/A          | Yes         | Yes          | N/A          |

diff --git a/Source/FullyConnectedFunctions/arm_fully_connected_s4.c b/Source/FullyConnectedFunctions/arm_fully_connected_s4.c
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright 2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
+ * SPDX-FileCopyrightText: Copyright 2023-2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
  *
  * SPDX-License-Identifier: Apache-2.0
  *
@@ -21,8 +21,8 @@
  * Title:        arm_fully_connected_s4
  * Description:  Fully connected function compatible with TF Lite.
  *
- * $Date:        10 October 2023
- * $Revision:    V.1.0.0
+ * $Date:        22 April 2024
+ * $Revision:    V.1.1.0
  *
  * Target :  Arm(R) M-Profile Architecture
  *
@@ -79,8 +79,8 @@ arm_cmsis_nn_status arm_fully_connected_s4(const cmsis_nn_context *ctx,
                                  filter_dims->n, /* col_dim or accum_depth */
                                  output_dims->c, /* row_dim or output_depth */
                                  fc_params->activation.min,
-                                 fc_params->activation.max,
-                                 1L);
+                                 fc_params->activation.max);
+
         input += filter_dims->n;
         output += output_dims->c;
         batch_cnt--;