tree 65bab43d0feddaa66b160ac7dc746651dc7c48de
parent ec320d9fc418e2d95a3a38ce87233397535f467d
author Gunes Bayir <gunes.bayir@arm.com> 1674998664 +0000
committer Gunes Bayir <gunes.bayir@arm.com> 1675245570 +0000

Add new operator AddMulAdd for Neon™ backend for Float/Quantized types

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 <gunes.bayir@arm.com>
Change-Id: I8ef577aa623b036e9a9f655cc088493fd19a6109
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9055
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>