COMPMID-748 - Integrating optimized SGEMM for bifrost
This patch introduces a new GEMM capable to improve the mac utilisation
of 10% compared to the GEMM without reshape. However this implementation
is not faster in all cases as we need to take into account the time for
reshaping the matrices. For this reason an heuristic solution to select
the optimal GEMM to use has been added to the function. More information
about the heuristic implementation can be found at COMPMID-852.
With this new patch, GoogleNet, MobileNet, VGG16 and SqueezeNet can
improved the performance of 1.5x.
More information about the performance uplift can be found here:
https://confluence.arm.com/display/MLENG/GEMM+FP32+performance%3A+ACL+18.02
Change-Id: I024563c06b9aed02a211a974e452bae5c233b04c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/117140
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h b/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h
index 8721643..9a3069e 100644
--- a/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -62,7 +62,7 @@
* \end{array} \right)
* @f]
*
- * @note The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor)
+ * @note The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor) * mult_transpose1xW_width
*
*/
class CLGEMMTranspose1xWKernel : public ICLSimple2DKernel
@@ -70,18 +70,20 @@
public:
/** Initialise the kernel's input and output.
*
- * @param[in] input Input tensor. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
- * @param[out] output Output tensor. Data type supported: same as @p input
+ * @param[in] input Input tensor. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+ * @param[out] output Output tensor. Data type supported: same as @p input
+ * @param[in] mult_transpose1xW_width (Optional) Multiplication factor for the width of the 1xW transposed block
*/
- void configure(const ICLTensor *input, ICLTensor *output);
+ void configure(const ICLTensor *input, ICLTensor *output, int mult_transpose1xW_width = 1);
/** Static function to check if given info will lead to a valid configuration of @ref CLGEMMTranspose1xWKernel
*
- * @param[in] input Input tensor. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
- * @param[in] output Output tensor. Data type supported: same as @p input.
+ * @param[in] input Input tensor. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+ * @param[in] output Output tensor. Data type supported: same as @p input.
+ * @param[in] mult_transpose1xW_width Multiplication factor for the width of the 1xW transposed block
*
* @return a status
*/
- static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output, int mult_transpose1xW_width);
// Inherited methods overridden:
void run(const Window &window, cl::CommandQueue &queue) override;