x64: matmul: bugfix copy B padding calculation

(cherry picked from commit c4abad0)
oneapi-src · Jan 30, 2025 · 9b3abcd · 9b3abcd
1 parent d00e508
commit 9b3abcd
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 0 deletions.
diff --git a/src/cpu/x64/matmul/brgemm_matmul.cpp b/src/cpu/x64/matmul/brgemm_matmul.cpp
@@ -1754,6 +1754,7 @@ struct brgemm_matmul_t<isa>::brg_matmul_exec_ctx_t {
     bool packed_sparse_weights() const { return bgmmc_.packed_sparse_weights; }
 
     int get_current_K_pad(int current_K_iters) const {
+        if (current_K_iters % bgmmc_.wei_k_blk == 0) return 0;
         return bgmmc_.extendable_k ? bgmmc_.wei_k_blk
                         - rnd_up(
                                 current_K_iters % bgmmc_.wei_k_blk, vnni_factor)

diff --git a/tests/benchdnn/inputs/matmul/harness_matmul_regression_bf16 b/tests/benchdnn/inputs/matmul/harness_matmul_regression_bf16
@@ -17,3 +17,11 @@
 --reset
 --dt=bf16:bf16:bf16
 2x1280:1280x65_n"parallel_reduction"
+
+# Shape to better test k tail.
+--reset
+--dt=bf16:bf16:bf16
+--stag=ab
+--wtag=ab
+--dtag=ab
+8x2664:2664x256_n"k tails"