Do not clear cthread buffer if needed.

- Add output stream operators for LoopSched and PiplineVer

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp

@@ -50,3 +50,15 @@ constexpr auto GridwiseGemmPipeline_Selector()
 }
 
 } // namespace ck
+
+inline std::ostream& operator<<(std::ostream& os, const ck::PipelineVersion& p)
+{
+    switch(p)
+    {
+    case ck::PipelineVersion::v1: os << "PipelineVersion::v1"; break;
+    case ck::PipelineVersion::v2: os << "PipelineVersion::v2"; break;
+    case ck::PipelineVersion::v4: os << "PipelineVersion::v4"; break;
+    default: os << "";
+    }
+    return os;
+}

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp

@@ -155,7 +155,8 @@ struct GridwiseGemmPipeline_v1<2>
                                const BBlockTransferStep& b_block_copy_step,
                                const BlockwiseGemm& blockwise_gemm,
                                CThreadBuffer& c_thread_buf,
-                               index_t num_loop)
+                               index_t num_loop,
+                               bool clear_c_thread_buf = true)
     {
         // preload data into LDS
         {
@@ -173,7 +174,8 @@ struct GridwiseGemmPipeline_v1<2>
         }
 
         // Initialize C
-        c_thread_buf.Clear();
+        if(clear_c_thread_buf)
+            c_thread_buf.Clear();
 
         // main body
         if constexpr(HasMainLoop)
@@ -298,7 +300,8 @@ struct GridwiseGemmPipelineInterwave_v1<1>
                                const BBlockTransferStep& b_block_copy_step,
                                const BlockwiseGemm& blockwise_gemm,
                                CThreadBuffer& c_thread_buf,
-                               index_t num_loop)
+                               index_t num_loop,
+                               bool clear_c_thread_buf = true)
     {
         // preload data into LDS
         a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
@@ -308,7 +311,8 @@ struct GridwiseGemmPipelineInterwave_v1<1>
         b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
 
         // Initialize C
-        c_thread_buf.Clear();
+        if(clear_c_thread_buf)
+            c_thread_buf.Clear();
 
         a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
         b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp

@@ -49,7 +49,8 @@ struct GridwiseGemmPipeline_v2
                                const BBlockTransferStep& b_block_copy_step,
                                const BlockwiseGemm& blockwise_gemm,
                                CThreadBuffer& c_thread_buf,
-                               index_t num_loop)
+                               index_t num_loop,
+                               bool clear_c_thread_buf = true)
     {
         // global read 0
         a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
@@ -60,7 +61,8 @@ struct GridwiseGemmPipeline_v2
         b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
 
         // Initialize C
-        c_thread_buf.Clear();
+        if(clear_c_thread_buf)
+            c_thread_buf.Clear();
 
         // LDS write 0
         a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp

@@ -68,7 +68,8 @@ struct GridwiseGemmPipeline_v4<1>
                                const BBlockTransferStep& b_block_copy_step,
                                const BlockwiseGemm& blockwise_gemm,
                                CThreadBuffer& c_thread_buf,
-                               index_t num_loop)
+                               index_t num_loop,
+                               bool clear_c_thread_buf = true)
     {
         static_assert(ABlockBuffers::Size() == 1 && BBlockBuffers::Size() == 1);
         auto& a_block_buf = a_block_bufs.At(I0);
@@ -81,7 +82,8 @@ struct GridwiseGemmPipeline_v4<1>
         b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
 
         // Initialize C
-        c_thread_buf.Clear();
+        if(clear_c_thread_buf)
+            c_thread_buf.Clear();
 
         // main body
         if constexpr(HasMainLoop)
@@ -164,7 +166,8 @@ struct GridwiseGemmPipeline_v4<2>
                                const BBlockTransferStep& b_block_copy_step,
                                const BlockwiseGemm& blockwise_gemm,
                                CThreadBuffer& c_thread_buf,
-                               index_t num_loop)
+                               index_t num_loop,
+                               bool clear_c_thread_buf = true)
     {
         static_assert(ABlockBuffers::Size() == 2 && BBlockBuffers::Size() == 2);
         auto& a_block_buf1 = a_block_bufs.At(I0);
@@ -179,7 +182,8 @@ struct GridwiseGemmPipeline_v4<2>
         b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
 
         // Initialize C
-        c_thread_buf.Clear();
+        if(clear_c_thread_buf)
+            c_thread_buf.Clear();
 
         // main body
         if constexpr(HasMainLoop)

include/ck/utility/loop_scheduler.hpp

@@ -3,6 +3,8 @@
 
 #pragma once
 
+#include <ostream>
+
 #include "ck/utility/common_header.hpp"
 #include "ck/tensor_description/tensor_adaptor.hpp"
 
@@ -24,3 +26,14 @@ constexpr LoopScheduler make_default_loop_scheduler()
 }
 
 } // namespace ck
+
+inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
+{
+    switch(s)
+    {
+    case ck::LoopScheduler::Default: os << "Default"; break;
+    case ck::LoopScheduler::Interwave: os << "Interwave"; break;
+    default: os << "";
+    }
+    return os;
+}