[OpenMP] amdgpu bad choice of max_flat_workgroup_size

[ye-luo]

Currently clang sets max_flat_workgroup_size always to 1024 and causes register spill

    .max_flat_workgroup_size: 1024
    .name:           __omp_offloading_32_7a3077cd__ZN11qmcplusplus17einspline_spo_ompIfE18multi_evaluate_vghERKSt6vectorIPNS_6SPOSetESaIS4_EERKS2_IPNS_11ParticleSetESaISA_EEib_l413
    .private_segment_fixed_size: 264
    .sgpr_count:     60
    .sgpr_spill_count: 0
    .symbol:         __omp_offloading_32_7a3077cd__ZN11qmcplusplus17einspline_spo_ompIfE18multi_evaluate_vghERKSt6vectorIPNS_6SPOSetESaIS4_EERKS2_IPNS_11ParticleSetESaISA_EEib_l413.kd
    .vgpr_count:     128
    .vgpr_spill_count: 66
    .wavefront_size: 64

I tested overriding the default using ompx_attribute(__attribute__((amdgpu_flat_work_group_size(128, 256))))
and got 2x kernel speed-up.

    .max_flat_workgroup_size: 256
    .name:           __omp_offloading_32_7a4aee56__ZN11qmcplusplus17einspline_spo_ompIfE18multi_evaluate_vghERKSt6vectorIPNS_6SPOSetESaIS4_EERKS2_IPNS_11ParticleSetESaISA_EEib_l413
    .private_segment_fixed_size: 0
    .sgpr_count:     58
    .sgpr_spill_count: 0
    .symbol:         __omp_offloading_32_7a4aee56__ZN11qmcplusplus17einspline_spo_ompIfE18multi_evaluate_vghERKSt6vectorIPNS_6SPOSetESaIS4_EERKS2_IPNS_11ParticleSetESaISA_EEib_l413.kd
    .vgpr_count:     166
    .vgpr_spill_count: 0
    .wavefront_size: 64

The default 1024 is clearly very bad in this case. When I code cuda, even 1024 is supported, I really use 1024 but mostly 128 or 256.

1. Can max_flat_workgroup_size be chosen at linking when the needed vgpr got figured out?
2. When I specify thread_limit(192) clause, can the compiler take advantage of it?

[arsenm]

1024 is the conservative default is always executable. CUDA, unlike OpenCL, doesn't require checking the supported workgroup size before executing, so for compatibility the backend has to support the maximum workgroup size by default.

The clang-chosen default for OpenCL is maximum 256. If OpenMP only executes kernels under controlled situations the default could be decreased.

[llvmbot]

@llvm/issue-subscribers-openmp

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

[jhuber6]

The problem with OpenMP is that it sets a maximum of 1024 and allows environment variables to change the number of thread. Meaning, even if the number of threads is stated statically in the program we need to assume it can change at runtime.

[ye-luo]

The problem with OpenMP is that it sets a maximum of 1024 and allows environment variables to change the number of thread. Meaning, even if the number of threads is stated statically in the program we need to assume it can change at runtime.

Which environment variables were you referring to?

[ye-luo]

Given the kernel is compiled with max_flat_workgroup_size, environment variable can only reduce but not increase the workgroup size at run, right?

I saw MaxFlatWorkgroupSize in the plugin but it doesn't seem being used. On AMD GPU, I noticed 256 as the default workgroup size at run (shown by rocprof).

[ye-luo]

A related question asked to AOMP ROCm/aomp#614

[jhuber6]

It's just in general there are OMP_NUM_THREADS environment variables that can be set by the user at any time. Because of that we can't just assume the thread count even when constant. I believe AOMP has optimizations that specifically turn it off to spite the standard.

[ye-luo]

OMP_NUM_THREADS never change GPU threads. Every implementation should respect that. I guess you meant OMP_THREAD_LIMIT.

[jdoerfert]

There are (supposed to be) env vars for devices too. That said, in your case it's somewhat easy since you already provide a constant thread_limit. We'll look into this.

[ye-luo]

I don't always set thread_limit. Since the runtime usually go with 256 at run, would prefer a consistent default and users may override via command line option or attributes in the source code.