threshold.hpp

Go to the documentation of this file.

//=======================================================================
// Copyright (c) 2014-2023 Baptiste Wicht
// Distributed under the terms of the MIT License.
// (See accompanying file LICENSE or copy at
//  http://opensource.org/licenses/MIT)
//=======================================================================

#pragma once

namespace etl {

#ifdef ETL_DEBUG_THRESHOLDS

constexpr size_t gemm_std_max    = 75 * 75;   
constexpr size_t gemm_cublas_min = 180 * 180; 

constexpr size_t gemm_rr_small_threshold    = 1000; 
constexpr size_t gemm_rr_medium_threshold   = 2000; 
constexpr size_t gemm_nt_rr_small_threshold = 1000; 
constexpr size_t gemm_cc_small_threshold    = 1000; 

constexpr size_t gevm_rm_small_threshold = 1000; 
constexpr size_t gevm_cm_small_threshold = 1000; 

constexpr size_t gemv_rm_small_threshold = 1000; 
constexpr size_t gemv_cm_small_threshold = 1000; 

constexpr size_t parallel_threshold = 2 * 1024; 

constexpr size_t sum_parallel_threshold     = 1024 * 2; 
constexpr size_t vec_sum_parallel_threshold = 1024 * 2; 

constexpr size_t conv1_parallel_threshold_conv   = 100; 
constexpr size_t conv1_parallel_threshold_kernel = 16;  

constexpr size_t fft1_many_threshold_transforms = 16;  
constexpr size_t fft1_many_threshold_n          = 768; 

constexpr size_t fft2_many_threshold_transforms = 16;   
constexpr size_t fft2_many_threshold_n          = 1024; 

constexpr size_t stream_threshold = 1024; 

#else

constexpr size_t gemm_std_max    = 75 * 75;   
constexpr size_t gemm_cublas_min = 180 * 180; 

constexpr size_t gemm_rr_small_threshold    = 100 * 100; 
constexpr size_t gemm_rr_medium_threshold   = 400 * 400; 
constexpr size_t gemm_nt_rr_small_threshold = 500 * 500; 
constexpr size_t gemm_cc_small_threshold    = 40000;     

constexpr size_t gevm_rm_small_threshold = 72000;   
constexpr size_t gevm_cm_small_threshold = 4000000; 

constexpr size_t gemv_rm_small_threshold = 4500000; 
constexpr size_t gemv_cm_small_threshold = 2400000; 

constexpr size_t parallel_threshold = 64 * 1024; 

constexpr size_t sum_parallel_threshold     = 1024 * 32;  
constexpr size_t vec_sum_parallel_threshold = 1024 * 128; 

constexpr size_t conv1_parallel_threshold_conv   = 100; 
constexpr size_t conv1_parallel_threshold_kernel = 16;  

constexpr size_t fft1_many_threshold_transforms = 16;  
constexpr size_t fft1_many_threshold_n          = 768; 

constexpr size_t fft2_many_threshold_transforms = 16;   
constexpr size_t fft2_many_threshold_n          = 1024; 

constexpr size_t stream_threshold = cache_size; 

#endif

} //end of namespace etl