diff options
Diffstat (limited to 'src/cpu/kernels/CpuMulKernel.cpp')
| -rw-r--r-- | src/cpu/kernels/CpuMulKernel.cpp | 46 |
1 files changed, 46 insertions, 0 deletions
diff --git a/src/cpu/kernels/CpuMulKernel.cpp b/src/cpu/kernels/CpuMulKernel.cpp index 82e5445321..81bb85c3dd 100644 --- a/src/cpu/kernels/CpuMulKernel.cpp +++ b/src/cpu/kernels/CpuMulKernel.cpp @@ -34,6 +34,11 @@ #include <arm_neon.h> +namespace +{ + static constexpr size_t default_mws_N1_fp32_neon = 22447; + static constexpr size_t default_mws_V1_fp32_neon = 38982; +} namespace arm_compute { namespace cpu @@ -1909,6 +1914,47 @@ void CpuMulKernel::configure(ITensorInfo *src1, ITensorInfo *src2, ITensorInfo * ICpuKernel::configure(win); } +size_t CpuMulKernel::get_mws(const CPUInfo &platform, size_t thread_count) const +{ + ARM_COMPUTE_UNUSED(thread_count); + +#if defined(ENABLE_FP32_KERNELS) + if(this->_func_float == &mul_F32_F32_F32) + { + size_t mws = ICPPKernel::default_mws; + if(platform.get_cpu_model() == CPUModel::N1) + { + mws = default_mws_N1_fp32_neon; + } + else if(platform.get_cpu_model() == CPUModel::V1) + { + mws = default_mws_V1_fp32_neon; + } + else + { + return ICPPKernel::default_mws; + } + + // tensor is 1D or was re-interpreted as 1D + if(this->window().shape().num_dimensions() == 1) + { + return mws; + } + else + { + // scale mws down by the number of elements along all the dimensions (x, z, w, etc) except the one + // that we parallelize along (the y dimension). This allows for parallelization when the Y_SIZE is small + // but the other sizes are large, which boosts performance. + mws = static_cast<size_t>(mws / (this->window().num_iterations_total() / this->window().num_iterations(1))); + return std::max(static_cast<size_t>(1), mws); + } + } +#else /* ENABLE_FP32_KERNELS */ + ARM_COMPUTE_UNUSED(platform); +#endif /* ENABLE_FP32_KERNELS */ + return ICPPKernel::default_mws; +} + Status CpuMulKernel::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy) { |