Cadence HiFi 4 Neural Network Library v2.4.1: (#330)

Updated the script to download HiFi 4 Neural Network Library v2.4.1.
Removed the hifi4 patch
Updates in v2.4.1 library:
+ Fixed the issue related to special case handling in xa_nn_dot_prod_16x16_asym8s.
  Moved the special cases handling for vec_length=8/32 to the start of the function.
+ Removed legacy condition checks for depthwise_conv kernels.
+ Fixed the non-bit-exact issue related to xa_nn_conv2d_std_per_chan_sym8sxasym8s.
  updated to select correct vec values in case of vec_align_val is 1 and

tensorflow/lite/micro/tools/make/ext_libs/xtensa_download.sh

@@ -47,9 +47,9 @@ if [ ! -d ${DOWNLOADS_DIR} ]; then
 fi
 
 if [[ ${2} == "hifi4" ]]; then
-  LIBRARY_URL="http://github.com/foss-xtensa/nnlib-hifi4/raw/master/archive/xa_nnlib_hifi4_02_11_2021.zip"
+  LIBRARY_URL="http://github.com/foss-xtensa/nnlib-hifi4/raw/master/archive/xa_nnlib_hifi4_07_27_2021.zip"
   LIBRARY_DIRNAME="xa_nnlib_hifi4"
-  LIBRARY_MD5="8b934f61ffe0a966644849602810fb1b"
+  LIBRARY_MD5="24b8844f8e0c53c1ed8561b09968bb98"
 elif [[ ${2} == "hifi5" ]]; then
   LIBRARY_URL="http://github.com/foss-xtensa/nnlib-hifi5/raw/master/archive/xa_nnlib_hifi5_06_30.zip"
   LIBRARY_DIRNAME="xa_nnlib_hifi5"
@@ -76,10 +76,6 @@ else
 
   unzip -qo /tmp/${TMP_ZIP_ARCHIVE_NAME} -d ${DOWNLOADS_DIR} >&2
 
-  if [[ ${2} == "hifi4" ]]; then
-    apply_patch_to_folder ${DOWNLOADS_DIR}/xa_nnlib_hifi4/ ../../ext_libs/xtensa_patch.patch
-  fi
-
 fi
 
 echo "SUCCESS"

tensorflow/lite/micro/tools/make/ext_libs/xtensa_patch.patch

-diff --git a/algo/kernels/basic/hifi4/xa_nn_dot_prod_16x16.c b/algo/kernels/basic/hifi4/xa_nn_dot_prod_16x16.c
-index 0c58a3c..7439d6d 100644
---- a/algo/kernels/basic/hifi4/xa_nn_dot_prod_16x16.c
-+++ b/algo/kernels/basic/hifi4/xa_nn_dot_prod_16x16.c
-@@ -90,8 +90,8 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-   ae_int32x2 d_bias;
-   int i;
- 
--  /* inp1 and inp2 8-byte aligned case */
--  if(((vec_length & 3) == 0) && (((int)p_inp1_start & 7) == 0) && (((int)p_inp2_start & 7) == 0))
-+ /* handle cases where vec_length is multiple of 8 */
-+  if(vec_length == 8)
-   {
-     /* Assumption: 
-      * p_inp1_start - memory is continuous => vec_count1 end and vect_count2 start are continuous 
-@@ -100,19 +100,21 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-     pt_inp1 = (const ae_int16x4 *)((WORD16 *)p_inp1_start);
-     pt_inp2 = (const ae_int16x4 *)((WORD16 *)p_inp2_start);
- 
-+    align_inp1 = AE_LA64_PP(pt_inp1);
-+    align_inp2 = AE_LA64_PP(pt_inp2);
-     /* TBD: multiple vec_count processing in a single loop can be done */
-     for(loopcnt = 0; loopcnt < vec_count; loopcnt++)
-     {
-       AE_L32_XP(d_bias, (ae_int32 *)p_bias_load, bias_address_increment);
- 
-       d_out64_0 = ZERO64;
-+      AE_LA16X4_IP(d_inp1_0, align_inp1, pt_inp1);
-+      AE_LA16X4_IP(d_inp2_0, align_inp2, pt_inp2);
-+      AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
-+      AE_LA16X4_IP(d_inp1_0, align_inp1, pt_inp1);
-+      AE_LA16X4_IP(d_inp2_0, align_inp2, pt_inp2);
-+      AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
- 
--      for(i = 0; i < (vec_length >> 2); i++)
--      {
--        AE_L16X4_IP(d_inp1_0, pt_inp1, 8);
--        AE_L16X4_IP(d_inp2_0, pt_inp2, 8);
--        AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
--      }
-       AE_SAT32X2_HIFI4(d_out32, d_out64_0);
-       d_out32 = AE_ADD32S(d_out32, d_bias);
- 
-@@ -122,8 +124,7 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-       *p_out++ = (WORD8)AE_MOVAD32_L(d_out32);
-     }
-   }
--  /* handle cases where vec_length is multiple of 8 */
--  else if(vec_length == 8)
-+  else if(vec_length == 32)
-   {
-     /* Assumption: 
-      * p_inp1_start - memory is continuous => vec_count1 end and vect_count2 start are continuous 
-@@ -140,13 +141,13 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-       AE_L32_XP(d_bias, (ae_int32 *)p_bias_load, bias_address_increment);
- 
-       d_out64_0 = ZERO64;
--      AE_LA16X4_IP(d_inp1_0, align_inp1, pt_inp1);
--      AE_LA16X4_IP(d_inp2_0, align_inp2, pt_inp2);
--      AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
--      AE_LA16X4_IP(d_inp1_0, align_inp1, pt_inp1);
--      AE_LA16X4_IP(d_inp2_0, align_inp2, pt_inp2);
--      AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
--
-+#pragma loop_count min=3
-+      for(i = 0; i < (vec_length >> 2); i++)
-+      {
-+        AE_LA16X4_IP(d_inp1_0, align_inp1, pt_inp1);
-+        AE_LA16X4_IP(d_inp2_0, align_inp2, pt_inp2);
-+        AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
-+      }
-       AE_SAT32X2_HIFI4(d_out32, d_out64_0);
-       d_out32 = AE_ADD32S(d_out32, d_bias);
- 
-@@ -156,7 +157,8 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-       *p_out++ = (WORD8)AE_MOVAD32_L(d_out32);
-     }
-   }
--  else if(vec_length == 32)
-+  /* inp1 and inp2 8-byte aligned case */
-+  else if(((vec_length & 3) == 0) && (((int)p_inp1_start & 7) == 0) && (((int)p_inp2_start & 7) == 0))
-   {
-     /* Assumption: 
-      * p_inp1_start - memory is continuous => vec_count1 end and vect_count2 start are continuous 
-@@ -165,19 +167,17 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-     pt_inp1 = (const ae_int16x4 *)((WORD16 *)p_inp1_start);
-     pt_inp2 = (const ae_int16x4 *)((WORD16 *)p_inp2_start);
- 
--    align_inp1 = AE_LA64_PP(pt_inp1);
--    align_inp2 = AE_LA64_PP(pt_inp2);
-     /* TBD: multiple vec_count processing in a single loop can be done */
-     for(loopcnt = 0; loopcnt < vec_count; loopcnt++)
-     {
-       AE_L32_XP(d_bias, (ae_int32 *)p_bias_load, bias_address_increment);
- 
-       d_out64_0 = ZERO64;
--#pragma loop_count min=3
-+
-       for(i = 0; i < (vec_length >> 2); i++)
-       {
--        AE_LA16X4_IP(d_inp1_0, align_inp1, pt_inp1);
--        AE_LA16X4_IP(d_inp2_0, align_inp2, pt_inp2);
-+        AE_L16X4_IP(d_inp1_0, pt_inp1, 8);
-+        AE_L16X4_IP(d_inp2_0, pt_inp2, 8);
-         AE_MULAAAAQ16(d_out64_0, d_inp1_0, d_inp2_0);
-       }
-       AE_SAT32X2_HIFI4(d_out32, d_out64_0);
-@@ -189,7 +189,7 @@ WORD32 xa_nn_dot_prod_16x16_asym8s(
-       *p_out++ = (WORD8)AE_MOVAD32_L(d_out32);
-     }
-   }
--  else if(((vec_length & 3) == 0) && (((int)p_inp1_start & 7) == 0))
-+   else if(((vec_length & 3) == 0) && (((int)p_inp1_start & 7) == 0))
-   {
-     /* Assumption: 
-      * p_inp1_start - memory is continuous => vec_count1 end and vect_count2 start are continuous 
-diff --git a/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise.c b/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise.c
-index 3e29856..320987b 100644
---- a/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise.c
-+++ b/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise.c
-@@ -249,8 +249,6 @@ WORD32 xa_nn_conv2d_depthwise_getsize
-     XA_NNLIB_CHK_COND((kernel_height <= 0), -1);
-     XA_NNLIB_CHK_COND((kernel_width <= 0), -1);
-     XA_NNLIB_CHK_COND((channels_multiplier <= 0), -1);
--    XA_NNLIB_CHK_COND((x_stride <= 0 || x_stride > kernel_width), -1);
--    XA_NNLIB_CHK_COND((y_stride <= 0 || y_stride > kernel_height), -1);
-     XA_NNLIB_CHK_COND((x_padding < 0), -1);
-     XA_NNLIB_CHK_COND((y_padding < 0), -1);
-     XA_NNLIB_CHK_COND((output_height <= 0), -1);
-diff --git a/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise_sym8sxasym8s.c b/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise_sym8sxasym8s.c
-index e719da1..5b7390f 100644
---- a/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise_sym8sxasym8s.c
-+++ b/algo/kernels/cnn/hifi4/xa_nn_conv2d_depthwise_sym8sxasym8s.c
-@@ -659,7 +659,6 @@ WORD32 xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s
-     XA_NNLIB_ARG_CHK_COND((input_channels <= 0), -1);
-     XA_NNLIB_ARG_CHK_COND((kernel_height <= 0 || kernel_width <= 0), -1);
-     XA_NNLIB_ARG_CHK_COND((kernel_height > input_height), -1);
--    XA_NNLIB_ARG_CHK_COND((kernel_width > input_width), -1);
-     XA_NNLIB_ARG_CHK_COND((channels_multiplier <= 0), -1);
-     XA_NNLIB_ARG_CHK_COND((y_stride <= 0 || x_stride <= 0), -1);
-     XA_NNLIB_ARG_CHK_COND((y_padding < 0 || x_padding < 0), -1);
-@@ -671,8 +670,6 @@ WORD32 xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s
-     XA_NNLIB_ARG_CHK_COND((inp_data_format != 0 && inp_data_format != 1), -1);
-     XA_NNLIB_ARG_CHK_COND((out_data_format != 0), -1);
-     /* Implementation dependent checks */
--    XA_NNLIB_ARG_CHK_COND((y_stride > kernel_height), -1);
--    XA_NNLIB_ARG_CHK_COND((x_stride > kernel_width), -1);
- 
-     if(inp_data_format == 0)
-     {