Hi Julian! On 2023-09-06T02:34:30-0700, Julian Brown <jul...@codesourcery.com> wrote: > This patch works around behaviour of the 2D and 3D memcpy operations in > the CUDA driver runtime. Particularly in Fortran, the "base pointer" > of an array (used for either source or destination of a host/device copy) > may lie outside of data that is actually stored on the device. The fix > is to make sure that we use the first element of data to be transferred > instead, and adjust parameters accordingly.
Do you (a) have a stand-alone test case for this (that is, not depending on your other pending patches, so that this could go in directly -- together with the before-FAIL test case). Do you (b) know if is this a bug in our use of the CUDA Driver API or rather in CUDA itself? If the latter, have you reported this to Nvidia? (I didn't quickly understand <https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__MEM.html#group__CUDA__MEM_1g27f885b30c34cc20a663a671dbf6fc27> "cuMemcpy2D" etc.) Grüße Thomas > 2023-09-05 Julian Brown <jul...@codesourcery.com> > > libgomp/ > * plugin/plugin-nvptx.c (GOMP_OFFLOAD_memcpy2d): Adjust parameters to > avoid out-of-bounds array checks in CUDA runtime. > (GOMP_OFFLOAD_memcpy3d): Likewise. > --- > libgomp/plugin/plugin-nvptx.c | 67 +++++++++++++++++++++++++++++++++++ > 1 file changed, 67 insertions(+) > > diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c > index 00d4241ae02b..cefe288a8aab 100644 > --- a/libgomp/plugin/plugin-nvptx.c > +++ b/libgomp/plugin/plugin-nvptx.c > @@ -1827,6 +1827,35 @@ GOMP_OFFLOAD_memcpy2d (int dst_ord, int src_ord, > size_t dim1_size, > data.srcXInBytes = src_offset1_size; > data.srcY = src_offset0_len; > > + if (data.srcXInBytes != 0 || data.srcY != 0) > + { > + /* Adjust origin to the actual array data, else the CUDA 2D memory > + copy API calls below may fail to validate source/dest pointers > + correctly (especially for Fortran where the "virtual origin" of an > + array is often outside the stored data). */ > + if (src_ord == -1) > + data.srcHost = (const void *) ((const char *) data.srcHost > + + data.srcY * data.srcPitch > + + data.srcXInBytes); > + else > + data.srcDevice += data.srcY * data.srcPitch + data.srcXInBytes; > + data.srcXInBytes = 0; > + data.srcY = 0; > + } > + > + if (data.dstXInBytes != 0 || data.dstY != 0) > + { > + /* As above. */ > + if (dst_ord == -1) > + data.dstHost = (void *) ((char *) data.dstHost > + + data.dstY * data.dstPitch > + + data.dstXInBytes); > + else > + data.dstDevice += data.dstY * data.dstPitch + data.dstXInBytes; > + data.dstXInBytes = 0; > + data.dstY = 0; > + } > + > CUresult res = CUDA_CALL_NOCHECK (cuMemcpy2D, &data); > if (res == CUDA_ERROR_INVALID_VALUE) > /* If pitch > CU_DEVICE_ATTRIBUTE_MAX_PITCH or for device-to-device > @@ -1895,6 +1924,44 @@ GOMP_OFFLOAD_memcpy3d (int dst_ord, int src_ord, > size_t dim2_size, > data.srcY = src_offset1_len; > data.srcZ = src_offset0_len; > > + if (data.srcXInBytes != 0 || data.srcY != 0 || data.srcZ != 0) > + { > + /* Adjust origin to the actual array data, else the CUDA 3D memory > + copy API call below may fail to validate source/dest pointers > + correctly (especially for Fortran where the "virtual origin" of an > + array is often outside the stored data). */ > + if (src_ord == -1) > + data.srcHost > + = (const void *) ((const char *) data.srcHost > + + (data.srcZ * data.srcHeight + data.srcY) > + * data.srcPitch > + + data.srcXInBytes); > + else > + data.srcDevice > + += (data.srcZ * data.srcHeight + data.srcY) * data.srcPitch > + + data.srcXInBytes; > + data.srcXInBytes = 0; > + data.srcY = 0; > + data.srcZ = 0; > + } > + > + if (data.dstXInBytes != 0 || data.dstY != 0 || data.dstZ != 0) > + { > + /* As above. */ > + if (dst_ord == -1) > + data.dstHost = (void *) ((char *) data.dstHost > + + (data.dstZ * data.dstHeight + data.dstY) > + * data.dstPitch > + + data.dstXInBytes); > + else > + data.dstDevice > + += (data.dstZ * data.dstHeight + data.dstY) * data.dstPitch > + + data.dstXInBytes; > + data.dstXInBytes = 0; > + data.dstY = 0; > + data.dstZ = 0; > + } > + > CUDA_CALL (cuMemcpy3D, &data); > return true; > } > -- > 2.41.0 ----------------- Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955