From ae72a46e495742863dba44fcf5fdc673c9d2afbc Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Sun, 29 Jan 2023 13:24:24 +0000 Subject: =?UTF-8?q?Add=20new=20operator=20AddMulAdd=20for=20Neon=E2=84=A2?= =?UTF-8?q?=20backend=20for=20Float/Quantized=20types?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This is a fused operator that merges Add + Mul + Add [+ Relu-based-Activation] layers and have an intermediate output after the first Add. It's supported for FP16/32/QASYMM8/QASYMM8_SIGNED data types. The subsequent Add and Mul are intended for scaling and the coefficients only have one dimension (per channel). The inputs are - input1 : nD tensor [X, Y, Z, W, ..] - input2 : nD tensor [X, Y, Z, W, ..] - add_coef : 1D tensor [X] - mul_coef : 1D tensor [X] The outputs are - out1 : nD tensor (intermediate output) [X, Y, Z, W, ..] - out2 : nD tensor (final output) [X, Y, Z, W, ..] The operation can be summarized as follows: out1 <- input1 + input2 out2 <- Act(out1 * mul_coef + add_coef) The activation function can be Identity, Relu, Bounded Relu or Lower/Upper Bounded Relu. The intermediate output can be skipped by providing a nullptr. The reason of providing this operator is to be able to fuse in case of Residual network patterns and save computations by reducing memory back and forward. Resolves: COMPMID-5463 Signed-off-by: Gunes Bayir Change-Id: I8ef577aa623b036e9a9f655cc088493fd19a6109 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9055 Comments-Addressed: Arm Jenkins Reviewed-by: Jakub Sujak Reviewed-by: Viet-Hoa Do Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- src/BUILD.bazel | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'src/BUILD.bazel') diff --git a/src/BUILD.bazel b/src/BUILD.bazel index 93f7d1acb9..d8b0127179 100644 --- a/src/BUILD.bazel +++ b/src/BUILD.bazel @@ -653,6 +653,7 @@ filegroup( "cpu/CpuTensor.cpp", "cpu/kernels/CpuActivationKernel.cpp", "cpu/kernels/CpuAddKernel.cpp", + "cpu/kernels/CpuAddMulAddKernel.cpp", "cpu/kernels/CpuCastKernel.cpp", "cpu/kernels/CpuCol2ImKernel.cpp", "cpu/kernels/CpuConcatenateBatchKernel.cpp", @@ -710,6 +711,10 @@ filegroup( "cpu/kernels/add/generic/neon/qasymm8.cpp", "cpu/kernels/add/generic/neon/qasymm8_signed.cpp", "cpu/kernels/add/generic/neon/qsymm16.cpp", + "cpu/kernels/addmuladd/generic/neon/fp16.cpp", + "cpu/kernels/addmuladd/generic/neon/fp32.cpp", + "cpu/kernels/addmuladd/generic/neon/qasymm8.cpp", + "cpu/kernels/addmuladd/generic/neon/qasymm8_signed.cpp", "cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp", "cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp", "cpu/kernels/boundingboxtransform/generic/neon/impl.cpp", @@ -810,6 +815,7 @@ filegroup( "cpu/kernels/sub/neon/qsymm16.cpp", "cpu/operators/CpuActivation.cpp", "cpu/operators/CpuAdd.cpp", + "cpu/operators/CpuAddMulAdd.cpp", "cpu/operators/CpuCast.cpp", "cpu/operators/CpuConcatenate.cpp", "cpu/operators/CpuConv2d.cpp", @@ -867,6 +873,7 @@ filegroup( "runtime/NEON/INESimpleFunction.cpp", "runtime/NEON/INESimpleFunctionNoBorder.cpp", "runtime/NEON/functions/NEActivationLayer.cpp", + "runtime/NEON/functions/NEAddMulAdd.cpp", "runtime/NEON/functions/NEArgMinMaxLayer.cpp", "runtime/NEON/functions/NEArithmeticAddition.cpp", "runtime/NEON/functions/NEArithmeticSubtraction.cpp", -- cgit v1.2.1