Add CUDA 8.6 Compute capability

[Former user]

Please add CUDA 8.6 Compute capability to list of valid architectures (RTX 30* GPUs)

[Cade Brown]

I believe this is already supported, you can check https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#virtual-architecture-feature-list for the list of architectures

We implement this as Ampere, so in your make.inc, add Ampere to the GPU_TARGET variable. You can also specify the architecture version via sm_XY. For example, you could add sm_80 toGPU_TARGET to achieve the same effect

‌

[Mark Gates]

I guess what is meant is that the Makefile doesn’t allow sm_86, it has only sm_80 for Ampere. PR #5 allows any sm for CMake, but it hasn’t been merged in yet.

[Former user]

Yes, I am speaking about ability to pass sm_86 key. I think it will be nice to have it in Makefile as well for consistency.

[Mark Gates]

In SLATE, a not-yet-released change in the Makefile does it this way, which avoids having a list of known sm architectures:

    # Generate flags for which CUDA architectures to build.
    # cuda_arch_ is a local copy to modify.
    cuda_arch_ = $(cuda_arch)
    ifneq ($(findstring kepler, $(cuda_arch_)),)
        cuda_arch_ += sm_30
    endif
    [... and so on for maxwell, pascal, volta, turing, ampere ...]

    # Warn about unrecognized architectures.
    cuda_arch_unknown = $(filter-out sm_% kepler maxwell pascal volta turing ampere, $(cuda_arch))
    ifneq ($(cuda_arch_unknown),)
        $(error ERROR: unknown `$(cuda_arch_unknown)` in cuda_arch)
    endif

    # Extract architectures XX from sm_XX in cuda_arch and sort numerically.
    sms      := $(patsubst sm_%,%,$(filter sm_%, $(cuda_arch_)))
    sms_sort := $(shell printf "%s\n" $(sms) | sort -n)

    # code=sm_XX is binary, code=compute_XX is PTX
    gencode_sm      = -gencode arch=compute_$(sm),code=sm_$(sm)
    gencode_compute = -gencode arch=compute_$(sm),code=compute_$(sm)

    # Get gencode options for all sm_XX in cuda_arch_.
    nv_sm      := $(foreach sm,$(sms_sort),$(gencode_sm))
    nv_compute := $(foreach sm,$(sms_sort),$(gencode_compute))

    ifeq ($(nv_sm),)
        $(error ERROR: unknown `cuda_arch=$(cuda_arch)`. Set cuda_arch to one or more of kepler, maxwell, pascal, volta, turing, ampere, or valid sm_XX from nvcc -h)
    else
        # Get last option (last 2 words) of nv_compute.
        nwords := $(words $(nv_compute))
        nwords_1 := $(shell expr $(nwords) - 1)
        nv_compute_last := $(wordlist $(nwords_1), $(nwords), $(nv_compute))
    endif

    # Use all sm_XX (binary), and the last compute_XX (PTX) for forward compatibility.
    NVCCFLAGS += $(nv_sm) $(nv_compute_last)

‌

[Cade Brown]

I’ve implemented Mark’s suggestion, and tested on my own CUDA machine (with an up-to-date NVCC). It should be available as https://bitbucket.org/icl/magma/commits/6e3a460f9badffb1c391b2d34bb1477ad8eb7367 (which is part of the ‘master’ branch)

‌

You can do it exactly as you say:

‌

GPU_TARGET = Pascal Volta Turing Ampere sm_86

In your make.inc

[Former user]

Thanks for quick help!

Should I close this one?

[Cade Brown]

Yes, go ahead and mark as resolved/close the issue

[Former user]

Resolved in https://bitbucket.org/icl/magma/commits/6e3a460f9badffb1c391b2d34bb1477ad8eb7367