[SYCL][CUDA] Don't infer memcpy src/dst pitches from copied region (intel#3563)

Previously, when passing 0 for row/slice pitches for source and/or
destination to `commonEnqueueMemBufferCopyRect`, the function would
infer the row/slice pitches from the copied region.

This does not work when copying with an offset, as `cuMemcpy3DAsync`
checks whether the pitches are at least as large as the offset + copied
range, otherwise returning CUDA_ERROR_INVALID_VALUE.

For 1D / 2D copies, no / not all pitches are required, and can instead
be specified as 0. The mechanism is thus removed and replaced with
assertions to check that for 2D and 3D copies, non-zero pitches are
provided.

In some situations, a zero row/slice pitch is passed for source and/or
destination to `commonEnqueueMemBufferCopyRect`. In this case, the
pitches are inferred from the copied region. However, if source and/or
destination have offsets in the x or y dimension, those must be taken
into account as well (CUDA checks this, and will otherwise return
`CUDA_ERROR_INVALID_VALUE`).

Author: psalz

sycl/plugins/cuda/pi_cuda.cpp

@@ -3654,12 +3654,18 @@ static pi_result commonEnqueueMemBufferCopyRect(
   assert(src_type == CU_MEMORYTYPE_DEVICE || src_type == CU_MEMORYTYPE_HOST);
   assert(dst_type == CU_MEMORYTYPE_DEVICE || dst_type == CU_MEMORYTYPE_HOST);
 
-  src_row_pitch = (!src_row_pitch) ? region->width_bytes : src_row_pitch;
-  src_slice_pitch = (!src_slice_pitch) ? (region->height_scalar * src_row_pitch)
-                                       : src_slice_pitch;
-  dst_row_pitch = (!dst_row_pitch) ? region->width_bytes : dst_row_pitch;
-  dst_slice_pitch = (!dst_slice_pitch) ? (region->height_scalar * dst_row_pitch)
-                                       : dst_slice_pitch;
+  src_row_pitch = (!src_row_pitch) ? region->width_bytes + src_offset->x_bytes
+                                   : src_row_pitch;
+  src_slice_pitch =
+      (!src_slice_pitch)
+          ? ((region->height_scalar + src_offset->y_scalar) * src_row_pitch)
+          : src_slice_pitch;
+  dst_row_pitch = (!dst_row_pitch) ? region->width_bytes + dst_offset->x_bytes
+                                   : dst_row_pitch;
+  dst_slice_pitch =
+      (!dst_slice_pitch)
+          ? ((region->height_scalar + dst_offset->y_scalar) * dst_row_pitch)
+          : dst_slice_pitch;
 
   CUDA_MEMCPY3D params = {};