ROCm / Tensile

Stretching GPU performance for GEMMs and tensor contractions.
MIT License
210 stars 145 forks source link

kernel.cpp without assembly kernel implement #1832

Closed DoubleClark closed 1 month ago

DoubleClark commented 9 months ago

os: Ubuntu20.04 rocm:5.7.1 gpu: AMD MI210 gfx90a tensile_version: Tensile 4.38.0 for ROCm 5.7.1

Hello, could you please to give an example about how to generate the assembly kernel code in Tensile, the generation of kernel library [*.hsaco] is success, but when i check the kernel.cpp file in ./1_BenchmarkProblems/Cijk_Ailk_Bljk_HHS_BH_00/00_Final/source/Kernels.cpp, there is no GEMM kernel implement, only some kernel like Cijk_H_PostGSU, which not looks like GEMM kernel .

I use the offical config file and command line shown as below for assembly kernel generation, ../Tensile/bin/Tensile ../Tensile/Configs/rocblas_hpa_hgemm_nn_inc1_asm_full.yaml ./

Besides, when i try to generate source code, run command line below, everything seems normal. ../Tensile/bin/Tensile ../Tensile/Configs/rocblas_hpa_hgemm_hip_lite_v1.yaml ./

Q1:Does Tensile support generate assembly kernel? If support, may i get one example config file? i try to use the asm_new.sh to generate the .s file, i suppose this is the asm file i need, but do we have a guide about how to use this file

Thansk for your help and experience.

nakajee commented 9 months ago

You can find many Tensile config yaml files for asm kernels under Tensile/Tests. You can find files with "asm" for asm kernels.

ppanchad-amd commented 1 month ago

@DoubleClark Do you still need assistance with this ticket? If not, please close the ticket. Thanks!