diff options
author | Tor Lillqvist <tml@collabora.com> | 2016-06-27 12:36:22 +0300 |
---|---|---|
committer | Tor Lillqvist <tml@collabora.com> | 2016-06-27 13:19:49 +0300 |
commit | a48aefcae40663b63005d49a04fc7b89a473d613 (patch) | |
tree | 0bf0e09f3b89b5f66b2e1ddc0282f4dadbf32fe6 | |
parent | 08fc0da4033b8ea2b3ae67aa06175e839771396b (diff) |
tdf#98515: Bail out early if kernel would have ridiculously many parameters
The Right Thing to do would be to compare the accumulated kernel
parameter size against the CL_DEVICE_MAX_PARAMETER_SIZE of the device,
but let's just do this sanity check for now. Bail out if the kernel
would have more than 50 parameters.
Calculating the accumulated kernel parameter size would be more
complicated and I don't want to touch this code more than
necessary. The kernel compilation will fail anyway if the size of
parameters exceeds the limit and this sanity check is just to make us
bail out a bit earlier.
In tdf#98515 the slowness seems to indeed be caused by the OpenCL code
generation , and there the number of parameters was 999, so this
sanity check does fix the slow loading issue (I checked).
Change-Id: Iead6dfb94ec8e7b2968ffed9423f0f9522a10ce9
-rw-r--r-- | sc/source/core/opencl/formulagroupcl.cxx | 13 |
1 files changed, 13 insertions, 0 deletions
diff --git a/sc/source/core/opencl/formulagroupcl.cxx b/sc/source/core/opencl/formulagroupcl.cxx index 3427d9fea34d..1a9cf1323ea1 100644 --- a/sc/source/core/opencl/formulagroupcl.cxx +++ b/sc/source/core/opencl/formulagroupcl.cxx @@ -2628,6 +2628,19 @@ DynamicKernelSoPArguments::DynamicKernelSoPArguments(const ScCalcConfig& config, const formula::DoubleVectorRefToken* pDVR = static_cast<const formula::DoubleVectorRefToken*>(pChild); + // FIXME: The Right Thing to do would be to compare the accumulated kernel + // parameter size against the CL_DEVICE_MAX_PARAMETER_SIZE of the device, but + // let's just do this sanity check for now. The kernel compilation will + // hopefully fail anyway if the size of parameters exceeds the limit and this + // sanity check is just to make us bail out a bit earlier. + + // The number 50 comes from the fact that the minimum size of + // CL_DEVICE_MAX_PARAMETER_SIZE is 256, which for 32-bit code probably means 64 + // of them. Round down a bit. + + if (pDVR->GetArrays().size() > 50) + throw UnhandledToken(("Kernel would have ridiculously many parameters (" + std::to_string(2 + pDVR->GetArrays().size()) + ")").c_str(), __FILE__, __LINE__); + for (size_t j = 0; j < pDVR->GetArrays().size(); ++j) { SAL_INFO("sc.opencl", "i=" << i << " j=" << j << |