[Executorch] Refactor op_add to support op_sub broadcasting #8255

kimishpatel · 2025-02-06T06:39:07Z

Stack from ghstack (oldest at bottom):

Summary:
Refactor op_add to conslidate commong broadcasting related improvements

Test Plan:
Previously added tests

Reviewers:

Subscribers:

Tasks:

Tags:

cc @larryliu0820 @manuelcandales

Differential Revision: D69491817

Summary: Refactoring broadcast handling utils that were added for op_mul. This is in prepartion use these utils to handle broadcast for other ops such as add, sub, div. Plus remove a redundant test Test Plan: optimized_kernels_test in CI Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]

Summary: This brings add op to feature parity, wrt, broadcasting, to mul op in optimized kernels lib Test Plan: tests added Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]

Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]

pytorch-bot · 2025-02-06T06:39:11Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/8255

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (1 Unrelated Failure)

As of commit 17c29b0 with merge base a01571f ():

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

pull / unittest / macos / macos-job (gh) (trunk failure)
exir/tests/test_joint_graph.py::TestJointGraph::test_joint_graph

This comment was automatically generated by Dr. CI and updates every 15 minutes.

…_sub broadcasting" Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]

Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]

swolchok

looks like this stack is going to need some changes in response to comments on this diff and previous, pausing review here

swolchok · 2025-02-06T19:42:58Z

kernels/optimized/cpu/op_add_sub_impl.cpp

+        out,
+        "Failed to resize output tensor.");
+
+    ET_SWITCH_REALB_TYPES(a_type, ctx, "add.out", CTYPE, [&]() {


need to put const char *op_name in the template parameters and fix this

swolchok · 2025-02-06T19:43:14Z

kernels/optimized/cpu/op_add_sub_impl.cpp

+    ET_SWITCH_REALHBBF16_TYPES(a_type, ctx, "add.out", CTYPE_A, [&]() {
+      ET_SWITCH_REALHBBF16_TYPES(b_type, ctx, "add.out", CTYPE_B, [&]() {


ditto op name

kimishpatel · 2025-02-07T00:53:43Z

looks like this stack is going to need some changes in response to comments on this diff and previous, pausing review here

yeah sounds good. Let me address your comments in the previous diffs

…_sub broadcasting" Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

…_sub broadcasting" Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

…_sub broadcasting" Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

…_sub broadcasting" Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales [ghstack-poisoned]

swolchok · 2025-02-11T21:55:46Z

kernels/optimized/cpu/op_add_sub_impl.h

+      if constexpr (is_sub) {
+        if (selected_optimized_path ==
+                ElementwiseOptimizedPath::kBroadcast2dBy1dReverseArguments ||
+            selected_optimized_path ==
+                ElementwiseOptimizedPath::kBroadcastLastDimReverseArguments ||
+            selected_optimized_path ==
+                ElementwiseOptimizedPath::kBroadcastNdByNdReverseArguments) {
+          auto add_lambda = [&alpha_val_vec](auto x, auto y) {
+            return y - alpha_val_vec * x;
+          };
+          return torch::executor::handle_broadcast_elementwise<CTYPE>(
+              ctx, add_lambda, a, b, out, selected_optimized_path, alpha);
+        } else {
+          auto add_lambda = [&alpha_val_vec](auto x, auto y) {
+            return x - alpha_val_vec * y;
+          };
+          return torch::executor::handle_broadcast_elementwise<CTYPE>(
+              ctx, add_lambda, a, b, out, selected_optimized_path, alpha);
+        }
+      } else {
+        if (selected_optimized_path ==
+                ElementwiseOptimizedPath::kBroadcast2dBy1dReverseArguments ||
+            selected_optimized_path ==
+                ElementwiseOptimizedPath::kBroadcastLastDimReverseArguments ||
+            selected_optimized_path ==
+                ElementwiseOptimizedPath::kBroadcastNdByNdReverseArguments) {
+          // Reason we swap out args here is because
+          // handle_broadcast_elementwise handles this selected_optimized_path
+          // option a bit differently. This should really be resolved in
+          // handle_broadcast_elementwise. However, the current blocker is that
+          // handle_broadcast_elementwise tries to be agnostic of op. This
+          // should be fixed, likely by moving lambda creation to
+          // handle_broadcast_elementwise and it be aware of which op is being
+          // executed.
+          auto add_lambda = [&alpha_val_vec](auto x, auto y) {
+            return y + alpha_val_vec * x;
+          };
+          return torch::executor::handle_broadcast_elementwise<CTYPE>(
+              ctx, add_lambda, a, b, out, selected_optimized_path, alpha);
+        } else {
+          auto add_lambda = [&alpha_val_vec](auto x, auto y) {
+            return x + alpha_val_vec * y;
+          };
+          return torch::executor::handle_broadcast_elementwise<CTYPE>(
+              ctx, add_lambda, a, b, out, selected_optimized_path, alpha);
+        }
+      }
+    });


I think you can just select the lambdas based on is_sub rather than duplicating the rest of the code under this if constexpr: https://godbolt.org/z/Esdz1exKj

thats a good suggestion. I tried doing the same in a different way which didnt quite work, but i can try out your suggestion

kimishpatel · 2025-02-11T23:36:58Z

@kimishpatel has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

…_sub broadcasting" Summary: Refactor op_add to conslidate commong broadcasting related improvements Test Plan: Previously added tests Reviewers: Subscribers: Tasks: Tags: cc larryliu0820 manuelcandales Differential Revision: [D69491817](https://our.internmc.facebook.com/intern/diff/D69491817) [ghstack-poisoned]