Change the config of the XeTLA kernel for 4k*2k to better performance. (

#1151) Change it to a better config from manually tuned result.
intel · May 20, 2024 · e84f153 · e84f153
1 parent 1bd2c8e
commit e84f153
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/benchmarks/xetla_kernel/softmax/softmax_config.hpp b/benchmarks/xetla_kernel/softmax/softmax_config.hpp
@@ -48,9 +48,9 @@ class mat1_4096x2048_bf16_cfg0 {
 public:
   static constexpr size_t mat_n = 2048;
   static constexpr size_t mat_m = 4096;
-  static constexpr size_t wg_n = mat_n;
+  static constexpr size_t wg_n = mat_n / 2;
   static constexpr size_t wg_m = 4; // 1 4 8 16
-  static constexpr size_t sg_n = mat_n;
+  static constexpr size_t sg_n = mat_n / 2;
   static constexpr size_t sg_m = 1;
   using data_type_in = sycl::ext::oneapi::bfloat16;
   using data_type_out = sycl::ext::oneapi::bfloat16;