change ways but still could not use immediate data as ds_read

cmake/EnableCompilerWarnings.cmake

@@ -66,6 +66,7 @@ else()
             -Wunreachable-code
             -Wunused
             -Wno-reserved-identifier
+            -v --save-temps -Wno-gnu-line-marker 
 	    # -Werror
             -Wno-option-ignored
             -Wsign-compare

example/ck_tile/03_gemm/gemm_basic.hpp

@@ -82,7 +82,8 @@ auto create_args(int argc, char* argv[])
         .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
         .insert("warmup", "50", "number of iterations before benchmark the kernel")
         .insert("repeat", "100", "number of iterations to benchmark the kernel")
-        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
+        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
+        .insert("init", "0", "0:random, 1:linear, 2:constant(1)");
 
     bool result = arg_parser.parse(argc, argv);
     return std::make_tuple(result, arg_parser);

example/ck_tile/03_gemm/run_gemm_example.inc

@@ -69,6 +69,7 @@ int run_gemm_example_with_layouts(int argc,
     ck_tile::index_t batch_size = arg_parser.get_int("b");
     int n_warmup                = arg_parser.get_int("warmup");
     int n_repeat                = arg_parser.get_int("repeat");
+    ck_tile::index_t init_method = arg_parser.get_int("init");
 
     using namespace ck_tile::literals;
 
@@ -114,14 +115,16 @@ int run_gemm_example_with_layouts(int argc,
         f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
 
     // TODO: add different init types
-
+    if (init_method == 0) {
         ck_tile::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
         ck_tile::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
-    // ck_tile::FillMonotonicSeq<ADataType>{}(a_m_k);
-    // ck_tile::FillMonotonicSeq<BDataType>{}(b_k_n);
-    // ck_tile::FillConstant<ADataType>{1.f}(a_m_k);
-    // ck_tile::FillConstant<BDataType>{1.f}(b_k_n);
-
+    } else if (init_method == 1) {
+        ck_tile::FillMonotonicSeq<ADataType>{}(a_m_k);
+        ck_tile::FillMonotonicSeq<BDataType>{}(b_k_n);
+    } else {
+        ck_tile::FillConstant<ADataType>{1.f}(a_m_k);
+        ck_tile::FillConstant<BDataType>{1.f}(b_k_n);
+    }
     ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
     ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
     ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp

@@ -374,29 +374,29 @@ struct BlockwiseGemmXdlops_pipeline_v4
     {
         // schedule
         constexpr auto num_ds_read_inst =
-            HotLoopInstList::A_LDS_Read_Inst_Num + HotLoopInstList::B_LDS_Read_Inst_Num;
+            HotLoopInstList::A_LDS_Read_Inst_Num + HotLoopInstList::B_LDS_Read_Inst_Num; //16
         constexpr auto num_ds_write_inst =
-            HotLoopInstList::A_LDS_Write_Inst_Num + HotLoopInstList::B_LDS_Write_Inst_Num;
+            HotLoopInstList::A_LDS_Write_Inst_Num + HotLoopInstList::B_LDS_Write_Inst_Num; //8
         ;
         constexpr auto num_buffer_load_inst =
-            HotLoopInstList::A_Buffer_Load_Inst_Num + HotLoopInstList::B_Buffer_Load_Inst_Num;
+            HotLoopInstList::A_Buffer_Load_Inst_Num + HotLoopInstList::B_Buffer_Load_Inst_Num; //8
         ;
-        constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
+        constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num; //64
 
-        constexpr auto num_issue = num_buffer_load_inst;
+        constexpr auto num_issue = num_buffer_load_inst; // 8
 
         static_for<0, num_issue, 1>{}([&](auto i) {
             ignore = i;
-            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA : 1
             __builtin_amdgcn_sched_group_barrier(
-                0x100, num_ds_read_inst / num_buffer_load_inst, 0); // DS read
-            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0);      // MFMA
+                0x100, num_ds_read_inst / num_buffer_load_inst, 0); // DS read : 2
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0);      // MFMA: 1
             __builtin_amdgcn_sched_group_barrier(
-                0x200, num_ds_write_inst / num_buffer_load_inst, 0); // DS write
-            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0);       // MFMA
-            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);       // VMEM read
+                0x200, num_ds_write_inst / num_buffer_load_inst, 0); // DS write : 1
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0);       // MFMA : 1
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);       // VMEM read :1
             __builtin_amdgcn_sched_group_barrier(
-                0x008, num_mfma_inst / num_buffer_load_inst - 3, 0); // MFMA
+                0x008, num_mfma_inst / num_buffer_load_inst - 3, 0); // MFMA : 5
         });
     }
 

include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v2.hpp

@@ -184,7 +184,6 @@ struct BlockGemmARegBRegCRegV2
             a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
         constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
         return a_block_dstr;
-        // return make_static_distributed_tensor<ADataType>(a_block_dstr);
     }
     
     CK_TILE_DEVICE static constexpr auto MakeBBlockDistribution()
@@ -208,8 +207,11 @@ struct BlockGemmARegBRegCRegV2
     template <typename BlockWindow, typename BlockTensor>
     CK_TILE_DEVICE static auto PrefetchLds(const BlockWindow& block_window, BlockTensor& block_tensor)
     {
-        auto tileDist = BlockTensor::get_tile_distribution();//.get_static_tile_distribution_encoding()
+        auto tileDist = BlockTensor::get_tile_distribution();
         return load_tile(block_tensor, make_tile_window(block_window, tileDist));
+        
+        // load_tile_raw(block_tensor, make_tile_window_linear_raw(block_window, tileDist));
+        // return;
     }
     
     // C = A * B