cpu: x64: binary: Supporting per_w broadcast strategy (#2778)

jstachowintel · web-flow · commit 21636077606a · 2025-03-21T15:42:51.000+01:00
diff --git a/src/cpu/x64/jit_uni_binary.cpp b/src/cpu/x64/jit_uni_binary.cpp
@@ -29,6 +29,7 @@ namespace x64 {
 static bcast_set_t get_supported_postops_bcast_strategies() {
     return {broadcasting_strategy_t::scalar, broadcasting_strategy_t::per_oc,
             broadcasting_strategy_t::per_oc_spatial,
+            broadcasting_strategy_t::per_w,
             broadcasting_strategy_t::no_broadcast};
 }
 
diff --git a/src/cpu/x64/jit_uni_binary_kernel.cpp b/src/cpu/x64/jit_uni_binary_kernel.cpp
@@ -29,6 +29,7 @@ namespace x64 {
 static bcast_set_t get_supported_postops_bcast_strategies() {
     return {broadcasting_strategy_t::scalar, broadcasting_strategy_t::per_oc,
             broadcasting_strategy_t::per_oc_spatial,
+            broadcasting_strategy_t::per_w,
             broadcasting_strategy_t::no_broadcast};
 }
 
diff --git a/tests/benchdnn/inputs/binary/harness_binary_regression b/tests/benchdnn/inputs/binary/harness_binary_regression
@@ -6,3 +6,6 @@
 
 # Mixed src1/post-op src broadcast
 --reset --attr-post-ops=add:f32:2 1x17:1x1
+
+# per_w broadcasting strategy
+--reset --attr-post-ops=mul:f32:4+add:f32:4 --alg=add 1x20x768:1x20x1

Original file line number	Diff line number	Diff line change
`@@ -29,6 +29,7 @@ namespace x64 {`
`29`	`29`	`static bcast_set_t get_supported_postops_bcast_strategies() {`
`30`	`30`	`return {broadcasting_strategy_t::scalar, broadcasting_strategy_t::per_oc,`
`31`	`31`	`broadcasting_strategy_t::per_oc_spatial,`
	`32`	`+ broadcasting_strategy_t::per_w,`
`32`	`33`	`broadcasting_strategy_t::no_broadcast};`
`33`	`34`	`}`
`34`	`35`