save clear_tile

include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp

@@ -537,7 +537,6 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
         {
             // STAGE 1, Q@K Gemm0
             auto st_acc = SPTBlockTileType{};
-            clear_tile(st_acc);
 
             q_block_tile = load_tile(q_dram_window);
             move_tile_window(q_dram_window, {kM0, 0});
@@ -551,7 +550,8 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             d_block_tile = load_tile(d_dram_window);
             move_tile_window(d_dram_window, {kM0});
 
-            gemm_0(st_acc, q_reg_tensor, k_reg_tensor);
+            st_acc = gemm_0(q_reg_tensor, k_reg_tensor);
+
             auto dot_reg_tensor = load_tile(dot_lds_read_window);
 
             HotLoopScheduler::template GemmStagedScheduler<0>();
@@ -670,9 +670,8 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             __builtin_amdgcn_sched_barrier(0);
             // STAGE 4, OGrad@V Gemm2
             auto dpt_acc = SPGradTBlockTileType{};
-            clear_tile(dpt_acc);
 
-            gemm_2(dpt_acc, do_reg_tensor, v_reg_tensor);
+            dpt_acc = gemm_2(do_reg_tensor, v_reg_tensor);
 
             block_sync_lds();
 
@@ -804,10 +803,9 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
 
         // Tail
         auto st_acc = SPTBlockTileType{};
-        clear_tile(st_acc);
 
         // STAGE 1, Q@K Gemm0
-        gemm_0(st_acc, q_reg_tensor, k_reg_tensor);
+        st_acc = gemm_0(q_reg_tensor, k_reg_tensor);
 
         // STAGE 2, Scale, Add bias, Mask, Softmax, Dropout
         if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
@@ -919,10 +917,10 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
 
         // STAGE 4, OGrad@V Gemm2
         auto dpt_acc = SPGradTBlockTileType{};
-        clear_tile(dpt_acc);
 
         auto qt_reg_tensor = load_tile(qt_lds_read_window);
-        gemm_2(dpt_acc, do_reg_tensor, v_reg_tensor);
+
+        dpt_acc = gemm_2(do_reg_tensor, v_reg_tensor);
 
         HotLoopScheduler::template GemmStagedScheduler<2>();