CMSIS-NN: Add bias null check for convolution functions

NULL check is done for optional bias parameter

Change-Id: I371aad8b900a90ab045789d8b27317b994dba71a

CMSIS/NN/Include/arm_nnfunctions.h

@@ -21,8 +21,8 @@
  * Title:        arm_nnfunctions.h
  * Description:  Public header file for CMSIS NN Library
  *
- * $Date:        June 11, 2020
- * $Revision:    V.6.0.1
+ * $Date:        July 27, 2020
+ * $Revision:    V.6.0.2
  *
  * Target Processor:  Cortex-M CPUs
  * -------------------------------------------------------------------- */
@@ -258,7 +258,7 @@ extern    "C"
    * @param[in]      filter_dims    Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions
    * @param[in]      filter_data    Filter data pointer. Data type: int8
    * @param[in]      bias_dims      Bias tensor dimensions. Format: [C_OUT]
-   * @param[in]      bias_data      Bias data pointer. Data type: int32
+   * @param[in]      bias_data      Optional bias data pointer. Data type: int32
    * @param[in]      output_dims    Output tensor dimensions. Format: [N, H, W, C_OUT]
    * @param[out]     output_data    Output data pointer. Data type: int8
 
@@ -574,7 +574,7 @@ extern    "C"
    * @param[in]      filter_dims    Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]
    * @param[in]      filter_data    Filter data pointer. Data type: int8
    * @param[in]      bias_dims      Bias tensor dimensions. Format: [C_OUT]
-   * @param[in]      bias_data      Bias data pointer. Data type: int32
+   * @param[in]      bias_data      Optional bias data pointer. Data type: int32
    * @param[in]      output_dims    Output tensor dimensions. Format: [N, H, W, C_OUT]
    * @param[out]     output_data    Output data pointer. Data type: int8
    *
@@ -626,7 +626,7 @@ extern    "C"
    * @param[in]      filter_dims    Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension
    * @param[in]      filter_data    Filter data pointer. Data type: int8
    * @param[in]      bias_dims      Bias tensor dimensions. Format: [C_OUT]
-   * @param[in]      bias_data      Bias data pointer. Data type: int32
+   * @param[in]      bias_data      Optional bias data pointer. Data type: int32
    * @param[in]      output_dims    Output tensor dimensions. Format: [N, H, W, C_OUT]
    * @param[out]     output_data    Output data pointer. Data type: int8
    *

CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c

@@ -21,8 +21,8 @@
  * Title:        arm_convolve_1_x_n_s8.c
  * Description:  s8 version of 1xN convolution using symmetric quantization.
  *
- * $Date:        May 18, 2020
- * $Revision:    V.2.0.0
+ * $Date:        July 27, 2020
+ * $Revision:    V.2.0.1
  *
  * Target Processor:  Cortex-M cores
  *
@@ -147,9 +147,11 @@ arm_status arm_convolve_1_x_n_s8(const cmsis_nn_context* ctx,
             }
             int32x4_t res = vldrwq_s32(acc);
             s_offset = vmulq_n_s32(s_offset, input_offset);
-
-            res = vaddq_n_s32(res, bias_data[i_out_ch]);
             res = vaddq_s32(res, s_offset);
+            if (bias_data)
+            {
+                res = vaddq_n_s32(res, bias_data[i_out_ch]);
+            }
             res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]);
             res = vaddq_n_s32(res, out_offset);
 

CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c

@@ -21,8 +21,8 @@
  * Title:        arm_convolve_1x1_s8_fast.c
  * Description:  Fast q7 version of 1x1 convolution (non-square shape)
  *
- * $Date:        May 29, 2020
- * $Revision:    V.2.0.1
+ * $Date:        July 27, 2020
+ * $Revision:    V.2.0.2
  *
  * Target Processor:  Cortex-M cores
  *
@@ -99,8 +99,10 @@ arm_status arm_convolve_1x1_s8_fast(const cmsis_nn_context *ctx,
                                             &sum_row,
                                             temp_out);
             int32x4_t res = vldrwq_s32(temp_out);
-
-            res = vaddq_n_s32(res, bias_data[i_out_ch]);
+            if (bias_data)
+            {
+                res = vaddq_n_s32(res, bias_data[i_out_ch]);
+            }
             sum_row = sum_row * input_offset;
             res = vaddq_n_s32(res, sum_row);
             res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]);
@@ -131,8 +133,10 @@ arm_status arm_convolve_1x1_s8_fast(const cmsis_nn_context *ctx,
                                             filter_data + i_out_ch * input_ch,
                                             &sum_row,
                                             &acc);
-
-            acc += bias_data[i_out_ch];
+            if (bias_data)
+            {
+                acc += bias_data[i_out_ch];
+            }
             sum_row = (sum_row * input_offset);
             acc += sum_row;
             acc = arm_nn_requantize(acc, output_mult[i_out_ch], output_shift[i_out_ch]);

CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c

@@ -21,8 +21,8 @@
  * Title:        arm_convolve_s8.c
  * Description:  s8 version of convolution using symmetric quantization.
  *
- * $Date:        May 29, 2020
- * $Revision:    V.2.0.1
+ * $Date:        July 27, 2020
+ * $Revision:    V.2.0.2
  *
  * Target Processor:  Cortex-M cores
  *
@@ -140,8 +140,10 @@ arm_status arm_convolve_s8(const cmsis_nn_context* ctx,
 
                         int32x4_t res = vldrwq_s32(acc);
                         s_offset = vmulq_n_s32(s_offset, input_offset);
-
-                        res = vaddq_n_s32(res, bias_data[i_out_ch]);
+                        if (bias_data)
+                        {
+                            res = vaddq_n_s32(res, bias_data[i_out_ch]);
+                        }
                         res = vaddq_s32(res, s_offset);
                         res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]);
                         res = vaddq_n_s32(res, out_offset);
@@ -260,7 +262,11 @@ arm_status arm_convolve_s8(const cmsis_nn_context* ctx,
             for (i = 0; i < output_ch; i++)
             {
                 /* Load the accumulator with bias first */
-                q31_t sum = bias_data[i];
+                q31_t sum = 0;
+                if (bias_data)
+                {
+                    sum = bias_data[i];
+                }
 
                 /* Point to the beginning of the im2col buffer where the input is available as a rearranged column */
                 const q15_t *ip_as_col = buffer_a;
@@ -311,7 +317,7 @@ arm_status arm_convolve_s8(const cmsis_nn_context* ctx,
             {
                 for (i_out_x = 0; i_out_x < output_x; i_out_x++)
                 {
-                    conv_out = bias_data[i_out_ch];
+                    conv_out = 0;
 
                     const int32_t base_idx_y = stride_y * i_out_y - pad_y;
                     const int32_t base_idx_x = stride_x * i_out_x - pad_x;
@@ -337,6 +343,10 @@ arm_status arm_convolve_s8(const cmsis_nn_context* ctx,
                             }
                         }
                     }
+                    if (bias_data)
+                    {
+                        conv_out += bias_data[i_out_ch];
+                    }
                     conv_out = arm_nn_requantize(conv_out, output_mult[i_out_ch], output_shift[i_out_ch]);
                     conv_out += out_offset;
                     conv_out = MAX(conv_out, out_activation_min);

CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c

@@ -21,8 +21,8 @@
  * Title:        arm_nn_mat_mult_s8.c
  * Description:  General Matrix-multiplication function
  *
- * $Date:        May 29, 2020
- * $Revision:    V.2.0.3
+ * $Date:        July 27, 2020
+ * $Revision:    V.2.0.4
  *
  * Target Processor:  Cortex-M cores
  * -------------------------------------------------------------------- */
@@ -65,10 +65,10 @@ q7_t *arm_nn_mat_mult_s8(const q7_t *input_row,
             const int8_t *ip_c2 = input_col + (2 * row_len);
             const int8_t *ip_c3 = input_col + (3 * row_len);
 
-            int32_t acc_0 = bias[i_out_ch];
-            int32_t acc_1 = bias[i_out_ch];
-            int32_t acc_2 = bias[i_out_ch];
-            int32_t acc_3 = bias[i_out_ch];
+            int32_t acc_0 = 0;
+            int32_t acc_1 = 0;
+            int32_t acc_2 = 0;
+            int32_t acc_3 = 0;
             const int32_t row_loop_cnt = (row_len + 7) / 8;
 
             for (int i_row_loop = 0; i_row_loop < row_loop_cnt; i_row_loop++)
@@ -103,6 +103,10 @@ q7_t *arm_nn_mat_mult_s8(const q7_t *input_row,
             }
 
             int32x4_t res = {acc_0, acc_1, acc_2, acc_3};
+            if (bias)
+            {
+                res = vaddq_n_s32(res, bias[i_out_ch]);
+            }
             res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]);
             res = vaddq_n_s32(res, out_offset);
 
@@ -124,7 +128,7 @@ q7_t *arm_nn_mat_mult_s8(const q7_t *input_row,
 
                 const int8_t *ip_r0 = input_row + (i_out_ch * row_len);
                 const int8_t *ip_c0 = input_col + (i_col_batch * row_len);
-                int32_t acc_0 = bias[i_out_ch];
+                int32_t acc_0 = 0;
                 const int32_t row_loop_cnt = (row_len + 7) / 8;
 
                 for (int i_row_loop = 0; i_row_loop < row_loop_cnt; i_row_loop++)
@@ -142,6 +146,10 @@ q7_t *arm_nn_mat_mult_s8(const q7_t *input_row,
                     acc_0 = vmladavaq_p_s16(acc_0, r0, c0, p);
                 }
 
+                if (bias)
+                {
+                    acc_0 += bias[i_out_ch];
+                }
                 acc_0 = arm_nn_requantize(acc_0, output_mult[i_out_ch], output_shift[i_out_ch]);
                 acc_0 += out_offset;
                 acc_0 = MAX(acc_0, activation_min);

CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c

@@ -21,8 +21,8 @@
  * Title:        arm_nn_mat_mult_s8_nt_t_s8
  * Description:  Matrix multiplication support function with the right-hand-side (rhs) matrix transposed
  *
- * $Date:        March 17 2020
- * $Revision:    V.1.0.1
+ * $Date:        July 27 2020
+ * $Revision:    V.1.0.2
  *
  * Target Processor:  Cortex-M
  *
@@ -80,9 +80,11 @@ arm_status arm_nn_mat_mult_nt_t_s8(const q7_t *lhs,
 
         lhs_offset_contribution0 *= lhs_offset;
         lhs_offset_contribution1 *= lhs_offset;
-
-        lhs_offset_contribution0 += bias[rhs_rows_idx];
-        lhs_offset_contribution1 += bias[rhs_rows_idx + 1];
+        if (bias)
+        {
+            lhs_offset_contribution0 += bias[rhs_rows_idx];
+            lhs_offset_contribution1 += bias[rhs_rows_idx + 1];
+        }
 
         int32_t lhs_rows_idx = lhs_rows >> 1;
 
@@ -375,7 +377,11 @@ arm_status arm_nn_mat_mult_nt_t_s8(const q7_t *lhs,
         for (int32_t lhs_rows_idx = 0; lhs_rows_idx < lhs_rows; ++lhs_rows_idx)
         {
             const q7_t *rhs_ptr = &rhs[0];
-            q31_t res00 = bias[rhs_rows - 1];
+            q31_t res00 = 0;
+            if (bias)
+            {
+                res00 = bias[rhs_rows - 1];
+            }
 
             for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx)
             {
@@ -419,9 +425,11 @@ arm_status arm_nn_mat_mult_nt_t_s8(const q7_t *lhs,
 
         lhs_offset_contribution0 *= lhs_offset;
         lhs_offset_contribution1 *= lhs_offset;
-
-        lhs_offset_contribution0 += bias[rhs_rows_idx];
-        lhs_offset_contribution1 += bias[rhs_rows_idx + 1];
+        if (bias)
+        {
+            lhs_offset_contribution0 += bias[rhs_rows_idx];
+            lhs_offset_contribution1 += bias[rhs_rows_idx + 1];
+        }
 
         int32_t lhs_rows_idx = lhs_rows >> 1;
 
@@ -536,7 +544,11 @@ arm_status arm_nn_mat_mult_nt_t_s8(const q7_t *lhs,
         for (int32_t lhs_rows_idx = 0; lhs_rows_idx < lhs_rows; ++lhs_rows_idx)
         {
             const q7_t *rhs_ptr = &rhs[0];
-            q31_t res00 = bias[rhs_rows - 1];
+            q31_t res00 = 0;
+            if (bias)
+            {
+                res00 = bias[rhs_rows - 1];
+            }
 
             for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx)
             {