Userptr also require size cache alignment, otherwise, the remained memory may be allocated in CPU side, when gpu flush the last cacheline to memory, will override the value changed by CPU.
Signed-off-by: Yang Rong <[email protected]> --- src/cl_mem.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/src/cl_mem.c b/src/cl_mem.c index 229bc0a..9e796ef 100644 --- a/src/cl_mem.c +++ b/src/cl_mem.c @@ -295,7 +295,8 @@ cl_mem_allocate(enum cl_mem_type type, assert(host_ptr != NULL); /* userptr not support tiling */ if (!is_tiled) { - if (ALIGN((unsigned long)host_ptr, cacheline_size) == (unsigned long)host_ptr) { + if ((ALIGN((unsigned long)host_ptr, cacheline_size) == (unsigned long)host_ptr) && + (ALIGN((unsigned long)sz, cacheline_size) == (unsigned long)sz)) { void* aligned_host_ptr = (void*)(((unsigned long)host_ptr) & (~(page_size - 1))); mem->offset = host_ptr - aligned_host_ptr; mem->is_userptr = 1; @@ -851,6 +852,7 @@ _cl_mem_new_image(cl_context ctx, cl_get_device_info(ctx->device, CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE, sizeof(cacheline_size), &cacheline_size, NULL); if (ALIGN((unsigned long)data, cacheline_size) == (unsigned long)data && ALIGN(h, cl_buffer_get_tiling_align(ctx, CL_NO_TILE, 1)) == h && + ALIGN(h * pitch * depth, cacheline_size) == h * pitch * depth && //h and pitch should same as aligned_h and aligned_pitch if enable userptr ((image_type != CL_MEM_OBJECT_IMAGE3D && image_type != CL_MEM_OBJECT_IMAGE1D_ARRAY && image_type != CL_MEM_OBJECT_IMAGE2D_ARRAY) || pitch * h == slice_pitch)) { tiling = CL_NO_TILE; enableUserptr = 1; -- 2.1.4 _______________________________________________ Beignet mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/beignet
