From 8b02d61c819f9217fc3ee8ee9b63351f2b1e7b1d Mon Sep 17 00:00:00 2001
From: Matthias Diener <mdiener@illinois.edu>
Date: Mon, 26 Sep 2022 00:20:39 -0500
Subject: [PATCH] PytatoPyOpenCLArrayContext: don't trust the arg limit
reported by the GPU (#198)
* PytatoPyOpenCLArrayContext: don't trust the arg limit reported by the
GPU
Apparently, CUDA doesn't like it when argument sizes get close
to the reported limit.
Avoids PTX/JIT errors of the type
CUDA_ERROR_INVALID_IMAGE: device kernel image is invalid
CUDA_ERROR_FILE_NOT_FOUND: file not found
* restructure and clarify
* add warning, clarify comment
* add comment about device naming
---
arraycontext/impl/pytato/__init__.py | 22 +++++++++++++++++++++-
1 file changed, 21 insertions(+), 1 deletion(-)
diff --git a/arraycontext/impl/pytato/__init__.py b/arraycontext/impl/pytato/__init__.py
index 98a310d..6b9ac6b 100644
--- a/arraycontext/impl/pytato/__init__.py
+++ b/arraycontext/impl/pytato/__init__.py
@@ -387,7 +387,27 @@ class PytatoPyOpenCLArrayContext(_BasePytatoArrayContext):
self.using_svm and dev.type & cl.device_type.GPU
and cl_char.has_coarse_grain_buffer_svm(dev))):
- limit = dev.max_parameter_size
+ if dev.max_parameter_size == 4352:
+ # Nvidia devices and PTXAS declare a limit of 4352 bytes,
+ # which is incorrect. The CUDA documentation at
+ # https://docs.nvidia.com/cuda/cuda-c-programming-guide/#function-parameters
+ # mentions a limit of 4KB, which is also incorrect.
+ # As far as I can tell, the actual limit is around 4080
+ # bytes, at least on a K40. Reducing the limit further
+ # in order to be on the safe side.
+
+ # Note that the naming convention isn't super consistent
+ # for Nvidia GPUs, so that we only use the maximum
+ # parameter size to determine if it is an Nvidia GPU.
+
+ limit = 4096-200
+
+ from warnings import warn
+ warn("Running on an Nvidia GPU, reducing the argument "
+ f"size limit from 4352 to {limit}.")
+ else:
+ limit = dev.max_parameter_size
+
if self._force_svm_arg_limit is not None:
limit = self._force_svm_arg_limit
--
GitLab