ROCm / AMDMIGraphX

AMD's graph optimization engine.
https://rocm.docs.amd.com/projects/AMDMIGraphX/en/latest/
MIT License
185 stars 86 forks source link

benchmark: No configs to tune #2765

Open attila-dusnoki-htec opened 9 months ago

attila-dusnoki-htec commented 9 months ago

Model(s) to reproduce:

Fails at: /code/AMDMIGraphX/src/targets/gpu/compile_ops.cpp:157: benchmark: No configs to tune

gyulaz-htec commented 8 months ago

I've checked a few models from the list randomly and all of them were failing with the following:

/tmp/comgr-a921d5/input/main.cpp:26:60: error: unused parameter 'x0' [-Werror,-Wunused-parameter]
__device__ __attribute__((const)) auto fused_reduce_op(Tx0 x0,Tx1 x1,Tx2 x2,Tx3 x3,Tx4 x4,Tr r,Tout_idx out_idx) {
                                                           ^
/tmp/comgr-a921d5/input/main.cpp:26:67: error: unused parameter 'x1' [-Werror,-Wunused-parameter]
__device__ __attribute__((const)) auto fused_reduce_op(Tx0 x0,Tx1 x1,Tx2 x2,Tx3 x3,Tx4 x4,Tr r,Tout_idx out_idx) {
                                                                  ^
2 errors generated when compiling for gfx90a.

/code/AMDMIGraphX/src/targets/gpu/compile_hip.cpp:172: compile: hiprtc: HIPRTC_ERROR_COMPILATION: Compilation failed.
/tmp/comgr-faf9ae/input/main.cpp:26:60: error: unused parameter 'x0' [-Werror,-Wunused-parameter]
__device__ __attribute__((const)) auto fused_reduce_op(Tx0 x0,Tx1 x1,Tx2 x2,Tx3 x3,Tx4 x4,Tr r,Tout_idx out_idx) {
                                                           ^
/tmp/comgr-faf9ae/input/main.cpp:26:67: error: unused parameter 'x1' [-Werror,-Wunused-parameter]
__device__ __attribute__((const)) auto fused_reduce_op(Tx0 x0,Tx1 x1,Tx2 x2,Tx3 x3,Tx4 x4,Tr r,Tout_idx out_idx) {
                                                                  ^
2 errors generated when compiling for gfx90a.

what seems to be coming form HIPRTC Running them with MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS=1 doesn't solve the error.