========================== CUDA UR Reference Document ========================== This document gives general guidelines of how to use UR to load and build programs, and execute kernels on a CUDA device. Device code =========== A CUDA device image may be made of PTX and/or SASS, two different kinds of device code for NVIDIA GPUs. CUDA device images can be generated by a CUDA-capable compiler toolchain. Most CUDA compiler toolchains are capable of generating PTX, SASS and/or bundles of PTX and SASS. When generating device code to be launched using Unified Runtime, it is recommended to use a programming model with explicit kernel parameters, such as OpenCL or CUDA. This is because kernels generated by a programming model with implicit kernel parameters, such as SYCL, cannot guarantee any specific number or ordering of kernel parameters. It has been observed that kernel signatures for the same SYCL kernel may vary significantly when compiled for different architectures. PTX --- PTX is a high level NVIDIA ISA which can be JIT compiled at runtime by the CUDA driver. In UR, this JIT compilation happens at :ref:`urProgramBuild`\, where PTX is assembled into device specific SASS which then can run on device. PTX is forward compatible, so PTX generated for ``.target sm_52`` will be JIT compiled without issue for devices with a greater compute capability than ``sm_52``. Whereas PTX generated for ``sm_80`` cannot be JIT compiled for an ``sm_60`` device. An advantage of using PTX over SASS is that one code can run on multiple devices. However, PTX generated for an older arch may not give access to newer hardware instructions, such as new atomic operations, or tensor core instructions. JIT compilation has some overhead at :ref:`urProgramBuild`\, especially if the program that is being loaded contains multiple kernels. The ``ptxjitcompiler`` keeps a JIT cache, however, so this overhead is only paid the first time that a program is built. JIT caching may be turned off by setting the environment variable ``CUDA_CACHE_DISABLE=1``. SASS ---- SASS is a device specific binary which may be produced by ``ptxas`` or some other tool. SASS is specific to an individual arch and is not portable across arches. A SASS file may be stored as a ``.cubin`` file by NVIDIA tools. UR Programs =========== A ur_program_handle_t has a one to one mapping with the CUDA driver object `CUModule `_. In UR for CUDA, a ur_program_handle_t can be created using :ref:`urProgramCreateWithBinary` with: * A single PTX module, stored as a null terminated ``uint8_t`` buffer. * A single SASS module, stored as an opaque ``uint8_t`` buffer. * A mixed PTX/SASS module, where the SASS module is the assembled PTX module. A ur_program_handle_t is valid only for a single architecture. If a CUDA compatible binary contains device code for multiple NVIDIA architectures, it is the user's responsibility to split these separate device images so that :ref:`urProgramCreateWithBinary` is only called with a device binary for a single device arch. If a program is large and contains many kernels, loading and/or JIT compiling the program may have a high overhead. This can be mitigated by splitting a program into multiple smaller programs (corresponding to PTX/SASS files). In this way, an application will only pay the overhead of loading/compiling kernels that it will likely use. Using PTX Modules in UR ----------------------- A PTX module will be loaded and JIT compiled for the necessary architecture at :ref:`urProgramBuild`\. If the PTX module has been generated for a compute capability greater than the compute capability of the device, then :ref:`urProgramBuild` will fail with the error ``CUDA_ERROR_NO_BINARY_FOR_GPU``. A PTX module passed to :ref:`urProgramBuild` must contain only one PTX file. Separate PTX files are to be handled separately. Arguments may be passed to the ``ptxjitcompiler`` via :ref:`urProgramBuild`\. Currently ``maxrregcount`` is the only supported argument. .. parsed-literal:: :ref:`urProgramBuild`\(ctx, program, "maxrregcount=128"); Using SASS Modules in UR ------------------------ A SASS module will be loaded and checked for compatibility at :ref:`urProgramBuild`\. If the SASS module is incompatible with the device arch then :ref:`urProgramBuild` will fail with the error ``CUDA_ERROR_NO_BINARY_FOR_GPU``. Using Mixed PTX/SASS Bundles in UR ---------------------------------- Mixed PTX/SASS modules can be used to make a program with :ref:`urProgramCreateWithBinary`\. At :ref:`urProgramBuild` the CUDA driver will check whether the bundled SASS is compatible with the active device. If the SASS is compatible then the ur_program_handle_t will be built from the SASS, and if not then the PTX will be used as a fallback and JIT compiled by the CUDA driver. If both PTX and SASS are incompatible with the active device then :ref:`urProgramBuild` will fail with the error ``CUDA_ERROR_NO_BINARY_FOR_GPU``. UR Kernels ========== Once :ref:`urProgramCreateWithBinary` and :ref:`urProgramBuild` have succeeded, kernels can be fetched from programs with :ref:`urKernelCreate`\. :ref:`urKernelCreate` must be called with the exact name of the kernel in the PTX/SASS module. This name will depend on the mangling used when compiling the kernel, so it is recommended to examine the symbols in the PTX/SASS module before trying to extract kernels in UR. .. code-block:: console $ cuobjdump --dump-elf-symbols hello.cubin | grep mykernel _Z13mykernelv At present it is not possible to query the names of the kernels in a UR program for CUDA, so it is necessary to know the (mangled or otherwise) names of kernels in advance or by some other means. UR kernels can be dispatched with :ref:`urEnqueueKernelLaunch`\. The argument ``pGlobalWorkOffset`` can only be used if the kernels have been instrumented to take the extra global offset argument. Use of the global offset is not recommended for non SYCL compiler toolchains. This parameter can be ignored if the user does not wish to use the global offset. Other Notes =========== - The environment variable ``SYCL_PI_CUDA_MAX_LOCAL_MEM_SIZE`` can be set in order to exceed the default max dynamic local memory size. More information can be found `here `_. - The size of primitive datatypes may differ in host and device code. For instance, NVCC treats ``long double`` as 8 bytes for device and 16 bytes for host. - In kernel ``printf`` for NVPTX targets does not support the ``%z`` modifier. Contributors ------------ * Hugh Delaney `hugh.delaney@codeplay.com `_