MasterJH5574 commented on code in PR #18823:
URL: https://github.com/apache/tvm/pull/18823#discussion_r3033144294
##########
src/target/target_kind.cc:
##########
@@ -427,8 +427,41 @@ TVM_REGISTER_TARGET_KIND("vulkan", kDLVulkan)
// Tags
.set_default_keys({"vulkan", "gpu"});
+/*!
+ * \brief Update WebGPU target attributes for subgroup-enabled lowering.
+ * Runtime routing on the WebLLM side guarantees subgroup size == 32.
+ * Runtime routing on the WebLLM side guarantees
+ * maxComputeInvocationsPerWorkgroup >= 1024.
+ * This is intentionally constrained for the subgroup-enabled WASM variant.
+ * When supports_subgroups is true, canonicalize thread_warp_size to 32 so
+ * TIR lowering can emit subgroup shuffle reductions.
+ */
+ffi::Map<ffi::String, ffi::Any> UpdateWebGPUAttrs(ffi::Map<ffi::String,
ffi::Any> target) {
Review Comment:
Can we move this function to line ~258, since we have a lot of
UpdateXXXAttrs there?
##########
web/emcc/webgpu_runtime.cc:
##########
@@ -40,6 +40,8 @@
#include "../../src/runtime/metadata.h"
#include "../../src/runtime/workspace_pool.h"
#include "../../src/support/bytes_io.h"
+#include "3rdparty/tvm-ffi/src/ffi/extra/json_parser.cc"
+#include "3rdparty/tvm-ffi/src/ffi/extra/json_writer.cc"
Review Comment:
Just curious: are these two lines related to this PR, or they are for some
other purpose?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
