Forums - Why is Elementwise Mul so slow on DSP

1 post / 0 new
or Register
Why is Elementwise Mul so slow on DSP
malila
Join Date: 9 Jun 21
Posts: 3
Posted: Thu, 2021-08-26 23:59
Top


Layers GPU_ub_float_timing(2 runs)       GPU_FP16_ub_float_timing(2 runs)       DSP_ub_float_timing(2 runs)       CPU_ub_float_timing(2 runs)     
layer_012 (Name:Mul_43 Type:elementwise_binary_op) 647 649 645 GPU 653 654 653 GPU 15836 16361 15311 DSP 78752 79322 78182 CPU
layer_015 (Name:Mul_46 Type:elementwise_binary_op) 331 332 330 GPU 326 327 325 GPU 6462 6616 6308 DSP 9130 10907 7353 CPU
layer_018 (Name:Mul_49 Type:elementwise_binary_op) 160 160 160 GPU 161 163 160 GPU 3256 3641 2872 DSP 4684 5105 4264 CPU
layer_021 (Name:Mul_52 Type:elementwise_binary_op) 160 160 160 GPU 165 166 164 GPU 2771 2824 2718 DSP 3612 3636 3589 CPU
layer_024 (Name:Mul_55 Type:elementwise_binary_op) 159 160 159 GPU 163 164 163 GPU 2804 2848 2760 DSP 3619 3650 3588 CPU
layer_025 (Name:Add_56 Type:elementwise_binary_op) 165 166 165 GPU 169 170 168 GPU 271 277 266 DSP 5335 5439 5232 CPU
layer_028 (Name:Mul_59 Type:elementwise_binary_op) 160 161 159 GPU 161 162 161 GPU 3265 3293 3238 DSP 5309 5347 5272 CPU
layer_032 (Name:Mul_63 Type:elementwise_binary_op) 324 326 323 GPU 328 329 328 GPU 6754 6895 6613 DSP 9000 9697 8303 CPU
layer_035 (Name:Mul_66 Type:elementwise_binary_op) 159 160 159 GPU 160 160 160 GPU 3360 3673 3048 DSP 18483 33559 3408 CPU
layer_038 (Name:Mul_69 Type:elementwise_binary_op) 75 75 75 GPU 77 78 76 GPU 1460 1568 1353 DSP 5231 8670 1793 CPU
layer_041 (Name:Mul_72 Type:elementwise_binary_op) 77 78 76 GPU 76 77 76 GPU 1369 1418 1320 DSP 2726 3223 2229 CPU
layer_044 (Name:Mul_75 Type:elementwise_binary_op) 74 75 74 GPU 73 73 73 GPU 1452 1452 1452 DSP 1735 1741 1729 CPU
layer_045 (Name:Add_76 Type:elementwise_binary_op) 85 86 85 GPU 88 90 87 GPU 143 144 143 DSP 1717 1750 1685 CPU
layer_048 (Name:Mul_79 Type:elementwise_binary_op) 76 77 75 GPU 75 76 75 GPU 1510 1546 1474 DSP 1619 1690 1549 CPU
layer_051 (Name:Mul_82 Type:elementwise_binary_op) 74 76 72 GPU 74 75 73 GPU 1570 1627 1513 DSP 1490 1491 1490 CPU
layer_052 (Name:Add_83 Type:elementwise_binary_op) 86 87 85 GPU 90 92 88 GPU 146 147 145 DSP 1456 1469 1443 CPU
 
 
We can get that the time cost on DSP is about 20 times on GPU.  Why is Elementwise Mul so slow on DSP ?
 
  • Up0
  • Down0
Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.