On 10/22/2015 10:12 AM, Jakub Jelinek wrote:
@@ -2129,6 +3242,19 @@ nvptx_file_end (void) FOR_EACH_HASH_TABLE_ELEMENT (*needed_fndecls_htab, decl, tree, iter) nvptx_record_fndecl (decl, true); fputs (func_decls.str().c_str(), asm_out_file); + + if (worker_bcast_hwm) + { + /* Define the broadcast buffer. */ + + worker_bcast_hwm = (worker_bcast_hwm + worker_bcast_align - 1) + & ~(worker_bcast_align - 1); + + fprintf (asm_out_file, "// BEGIN VAR DEF: %s\n", worker_bcast_name); + fprintf (asm_out_file, ".shared .align %d .u8 %s[%d];\n", + worker_bcast_align, + worker_bcast_name, worker_bcast_hwm); + }So, is the worker broadcast buffer effectively a file scope .shared variable? My worry is that as .shared is quite limited resource, if you compile many TUs and each allocates its own broadcast buffer you run out of shared memory. Is there any way how to share the broadcast buffers in between different TUs (other than LTO)?
I think LTO is the mechanism, nvptx-lto1 only ever produces one assembly file. So I'm not really concerned about this.
One other thing about this occurred to me yesterday - I was worried about thread-safety with a single static buffer - couldn't code execute multiple kernels at the same time? I googled a bit, and could not actually find a definitive answer as to whether all shared memory is allocated at kernel launch, or just the dynamic portion?
Bernd
