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>
diff --git a/filelist.json b/filelist.json
index 84884e1..93dfdff 100644
--- a/filelist.json
+++ b/filelist.json
@@ -665,7 +665,7 @@
"Reduction": {
"deps": [ "Reshape" ],
"files": {
- "common": [
+ "common": [
"src/core/CL/kernels/CLReductionOperationKernel.cpp",
"src/runtime/CL/functions/CLReductionOperation.cpp"
]
@@ -1154,7 +1154,7 @@
"src/core/NEON/kernels/convolution/common/qsymm8.cpp",
"src/core/NEON/kernels/convolution/common/utils.cpp",
"src/core/NEON/kernels/arm_conv/addressing.cpp",
- "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.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",
@@ -1171,6 +1171,7 @@
"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_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+ "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
@@ -1208,13 +1209,18 @@
"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",
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_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/interleaves/a64_s8q_3x3_dot.cpp",
"src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.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/cpu/kernels/depthwiseconv2d/generic/neon/impl.cpp"
- ],
+ ],
"fp16":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp16.cpp"],
- "fp32":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp32.cpp"],
- "qasymm8":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8.cpp"],
+ "fp32":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp32.cpp"],
+ "qasymm8":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8.cpp"],
"qasymm8_signed":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8_signed.cpp"]
},
"sve": {
@@ -1315,7 +1321,7 @@
"fp32": ["src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp"],
"fp16": ["src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp"]
- },
+ },
"sve2":{
"qasymm8": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8.cpp"],
"qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8_signed.cpp"]
@@ -1519,7 +1525,7 @@
"src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
"src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
"src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
- "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp",
+ "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp",
"src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp",
"src/cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp"
],
@@ -1603,7 +1609,7 @@
"common": [
"src/core/NEON/kernels/NEInstanceNormalizationLayerKernel.cpp",
"src/runtime/NEON/functions/NEInstanceNormalizationLayer.cpp"
- ],
+ ],
"neon":{
"common":["src/cpu/kernels/instancenorm/generic/neon/impl.cpp"],
"fp16":["src/cpu/kernels/instancenorm/generic/neon/fp16.cpp"],
@@ -1663,7 +1669,7 @@
"files": {
"common": [
"src/cpu/kernels/CpuMaxUnpoolingLayerKernel.cpp",
- "src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp",
+ "src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp",
"src/cpu/operators/CpuMaxUnpooling.cpp"
],
"neon":{
@@ -1764,12 +1770,12 @@
"src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
"src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
"src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
- "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"
+ "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"
],
"nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ],
"fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ],
- "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ],
- "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ],
+ "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ],
+ "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ],
"qasymm8_signed":["src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"]
},
"sve": {
@@ -1969,8 +1975,8 @@
"neon":{
"common":["src/cpu/kernels/softmax/generic/neon/impl.cpp"],
"fp32": ["src/cpu/kernels/softmax/generic/neon/fp32.cpp"],
- "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"],
- "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"],
+ "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"],
+ "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"],
"qasymm8_signed":["src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp"]
},
"sve": {
@@ -1982,7 +1988,7 @@
},
"sve2":{
"common" :["src/cpu/kernels/softmax/generic/sve2/impl.cpp"],
- "qasymm8":[ "src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp"],
+ "qasymm8":[ "src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp"],
"qasymm8_signed":["src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp"]
}
}
@@ -2074,4 +2080,4 @@
"src/gpu/cl/kernels/experimental/dynamic_fusion/ClCompositeKernel.cpp"
]
}
-}
\ No newline at end of file
+}