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 --- arm_compute/runtime/NEON/NEFunctions.h | 3 +- arm_compute/runtime/NEON/functions/NEAddMulAdd.h | 105 +++++++++++++++++++++++ 2 files changed, 107 insertions(+), 1 deletion(-) create mode 100644 arm_compute/runtime/NEON/functions/NEAddMulAdd.h (limited to 'arm_compute') diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h index a679e8c04e..8dbe6c59c3 100644 --- a/arm_compute/runtime/NEON/NEFunctions.h +++ b/arm_compute/runtime/NEON/NEFunctions.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2022 Arm Limited. + * Copyright (c) 2016-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,6 +25,7 @@ #define ARM_COMPUTE_NEFUNCTIONS_H #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" +#include "arm_compute/runtime/NEON/functions/NEAddMulAdd.h" #include "arm_compute/runtime/NEON/functions/NEArgMinMaxLayer.h" #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h" #include "arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h" diff --git a/arm_compute/runtime/NEON/functions/NEAddMulAdd.h b/arm_compute/runtime/NEON/functions/NEAddMulAdd.h new file mode 100644 index 0000000000..28185f338f --- /dev/null +++ b/arm_compute/runtime/NEON/functions/NEAddMulAdd.h @@ -0,0 +1,105 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEADDMULADD +#define ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEADDMULADD + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/IFunction.h" +#include "arm_compute/runtime/IMemoryManager.h" + +#include + +namespace arm_compute +{ +class ITensor; +class ITensorInfo; + +/** Function to compute Add+Mul+Add fused operation */ +class NEAddMulAdd : public IFunction +{ +public: + /** Constructor */ + NEAddMulAdd(std::shared_ptr memory_manager = nullptr); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEAddMulAdd(const NEAddMulAdd &) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + NEAddMulAdd(NEAddMulAdd &&) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEAddMulAdd &operator=(const NEAddMulAdd &) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + NEAddMulAdd &operator=(NEAddMulAdd &&) = delete; + /** Destructor */ + ~NEAddMulAdd(); + /** Initialize the function's inputs and outputs. + * + * Valid data layouts: + * - Any + * + * Valid data type configurations: + * |input1 |input2 |bn_mul |bn_add |add_output |final_output | + * |:--------------|:--------------|:--------------|:--------------|:--------------|:--------------| + * |QASYMM8 |QASYMM8 |QASYMM8 |QASYMM8 |QASYMM8 |QASYMM8 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED |QASYMM8_SIGNED |QASYMM8_SIGNED |QASYMM8_SIGNED |QASYMM8_SIGNED | + * |F16 |F16 |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 |F32 |F32 | + * + * This is what this composite function (tailored for add followed by a batch norm operation) does: + * add_output <- input1 + input2 (add) + * final_output <- add_output * bn_mul + bn_add (batch norm = mul+add) + * + * @param[in] input1 First tensor input. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] input2 Second tensor input. Data types supported: Same as @p input. + * @param[in] bn_mul The multiplication coefficient on the feature dimension. Data types supported: Same as @p input. + * It's one dimensional tensor with size equal to the feature maps [FM] + * @param[in] bn_add The addition coefficient on the feature dimension. Data types supported: Same as @p input. + * It's one dimensional tensor with size equal to the feature maps [FM] + * @param[out] add_output Output of the first add. Data type supported: Same as @p input. + * @param[out] final_output Output of the add+mul+add+act composite operation. Data type supported: Same as @p input. + * @param[in] policy Policy to handle overflow + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + * + */ + void configure(ITensor *input1, ITensor *input2, ITensor *bn_mul, ITensor *bn_add, + ITensor *add_output, ITensor *final_output, + ConvertPolicy policy, const ActivationLayerInfo &act_info); + /** Static function to check if given info will lead to a valid configuration of @ref NEAddMulAdd + * + * Similar to @ref NEAddMulAdd::configure() except the arguments are @ref ITensorInfo * instead of @ref ITensor * + * + * @return a status + */ + static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, + const ITensorInfo *bn_mul, const ITensorInfo *bn_add, + const ITensorInfo *add_output, const ITensorInfo *final_output, + ConvertPolicy policy, const ActivationLayerInfo &act_info); + + // Inherited methods overridden: + void run() override; + +private: + struct Impl; + std::unique_ptr _impl; +}; +} // namespace arm_compute +#endif /* ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEADDMULADD */ -- cgit v1.2.1