clean code

composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_fp16_bfp16.hpp

@@ -413,7 +413,7 @@ struct GridwiseBatchGemmXdlops_gkmkpack_gknkpack_gmn_v2
         constexpr auto a_g_k_m_kpack_block_desc = make_native_tensor_descriptor_aligned(
             Sequence<1, KPerBlock, MPerBlock, KPack>{}, Number<max_align>{});
 
-        auto a_blockwise_copy = BlockwiseGenericTensorSliceCopy_v4<
+        auto a_blockwise_copy = BlockwiseGenericTensorSliceCopy_v5<
             BlockSize,
             decltype(a_g_k_m_kpack_global_desc),
             decltype(a_g_k_m_kpack_block_desc),
@@ -509,14 +509,14 @@ struct GridwiseBatchGemmXdlops_gkmkpack_gknkpack_gmn_v2
         for(index_t k_block_data_begin = 0; k_block_data_begin < K - KPerBlock;
             k_block_data_begin += KPerBlock)
         {
-            ABFloat p_a_thread_buffer[a_blockwise_copy.GetThreadBufferSize()];
+            // ABFloat p_a_thread_buffer[a_blockwise_copy.GetThreadBufferSize()];
 
             // load next data from device mem
             a_blockwise_copy.MoveSrcSliceWindow(blockwise_a_copy_src_step, True);
             b_blockwise_copy.MoveSrcSliceWindow(blockwise_b_copy_src_step, True);
 
-            // a_blockwise_copy.RunLoadThreadBuffer(p_a_global);
-            a_blockwise_copy.RunLoadThreadBuffer(p_a_global, p_a_thread_buffer);
+            a_blockwise_copy.RunLoadThreadBuffer(p_a_global);
+            // a_blockwise_copy.RunLoadThreadBuffer(p_a_global, p_a_thread_buffer);
 
             b_blockwise_copy.RunLoadThreadBuffer(p_b_global);
 
@@ -535,8 +535,8 @@ struct GridwiseBatchGemmXdlops_gkmkpack_gknkpack_gmn_v2
             block_sync_lds();
 
             // store next data to LDS
-            // a_blockwise_copy.RunStoreThreadBuffer(p_a_block);
-            a_blockwise_copy.RunStoreThreadBuffer(p_a_thread_buffer, p_a_block);
+            a_blockwise_copy.RunStoreThreadBuffer(p_a_block);
+            // a_blockwise_copy.RunStoreThreadBuffer(p_a_thread_buffer, p_a_block);
 
             b_blockwise_copy.RunStoreThreadBuffer(p_b_block);
         }

composable_kernel/include/tensor_operation/xdlops_gemm.hpp

@@ -727,9 +727,6 @@ struct XdlopsGemm_t
 #else
         const index_t laneId = get_thread_local_1d_id() % mfma_type.wave_size;
 
-        FloatA a[K * MRepeats];
-        FloatB b[K * NRepeats];
-
         constexpr index_t data_size       = sizeof(FloatA) / sizeof(data_type);
         constexpr index_t a_reg_buff_size = K * MRepeats * data_size;
         constexpr index_t b_reg_buff_size = K * NRepeats * data_size;
@@ -748,13 +745,11 @@ struct XdlopsGemm_t
 
         constexpr index_t KRepeats = sizeof(FloatA) / (sizeof(data_type) * mfma_type.k_base);
 
-        auto pa = reinterpret_cast<const data_type*>(&a);
-        auto pb = reinterpret_cast<const data_type*>(&b);
-
         constexpr index_t AStride = K * KRepeats;
         constexpr index_t BStride = K * KRepeats;
 
         static_if<!IsKReduction>{}([&](auto) {
+#if 0
             for(index_t m_i = 0; m_i < MRepeats; ++m_i)
                 for(index_t k_i = 0; k_i < K; ++k_i)
                     a[k_i + m_i * K] = p_a_wave[k_i * M + laneId + MPerXdlops * m_i];
@@ -774,6 +769,7 @@ struct XdlopsGemm_t
                                                     BStride>(
                     &pa[k_i * mfma_type.k_base], &pb[k_i * mfma_type.k_base], p_c_thread);
             }
+#endif
         })
             .Else([&](auto) {
                 const index_t blk_id = laneId / mfma_type.num_threads_blk;

composable_kernel/include/utility/float_type.amd.hpp.in

@@ -36,7 +36,6 @@ union float_vec4_t
     StaticallyIndexedArray<float, 4> s1;
     StaticallyIndexedArray<float2_t, 2> s2;
     StaticallyIndexedArray<float4_t, 1> s4;
-    float n[4];
     __host__ __device__ constexpr float_vec4_t() { s4(Number<0>{}) = 0; }
 
     template <index_t vs>