commit 8a164884dddf769643cf3b9f7f94e43cb4f3c20b
author ramelg01 <ramy.elgammal@arm.com> Thu Apr 07 02:42:52 2022 +0100
committer Ramy Elgammal <ramy.elgammal@arm.com> Tue Apr 26 15:51:22 2022 +0000
tree 35958dd48b6df1a851c880dad2b2ce285671b611
parent c827e99fc46521f43719b0c2d1b6f05d66abf68c

Update Neon™ depthwise kernel

- Reduce duplication and simplify overall structure.
- Improve multi-threaded performance by sharing more data
  in lower-level caches.

Partially Resolves: COMPMID-5054
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Iac747f39b21c540122fa75218762631c4d787911
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7449
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Andrew Mundy
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

Android.bp

diff --git a/Android.bp b/Android.bp
index 17c56f3..c072c0e 100644
--- a/Android.bp
+++ b/Android.bp
@@ -301,9 +301,12 @@
         "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
         "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
         "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
+        "src/core/NEON/kernels/arm_conv/depthwise/depthwise_strategies_common.cpp",
         "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp",
         "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp",
         "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp",
+        "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp",
+        "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp",
         "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp",
         "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp",
         "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp",
@@ -962,6 +965,9 @@
                 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
                 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
                 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+                "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+                "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+                "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
                 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
                 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
                 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",