Merge: Fix for unified reduction kernel

pratikvn · web-flow · commit 00226d03360d · 2021-11-24T20:57:50.000+01:00
Small fix for unified kernel reduction, change lambda to capture by value. issue first found on Summit with gcc-7.50.0 and cuda-10.1, 11.3 and 11.4 Related PR: #926
diff --git a/cuda/base/kernel_launch_reduction.cuh b/cuda/base/kernel_launch_reduction.cuh
@@ -458,7 +458,7 @@ void run_kernel_row_reduction(std::shared_ptr<const CudaExecutor> exec,
     } else {
         select_run_generic_kernel_row_reduction(
             subwarp_sizes(),
-            [&](int compiled_subwarp_size) {
+            [cols](int compiled_subwarp_size) {
                 return compiled_subwarp_size >= cols ||
                        compiled_subwarp_size == config::warp_size;
             },
@@ -488,7 +488,7 @@ void run_kernel_col_reduction(std::shared_ptr<const CudaExecutor> exec,
     if (cols <= config::warp_size) {
         select_generic_col_reduction_small(
             subwarp_sizes(),
-            [&](int compiled_subwarp_size) {
+            [cols](int compiled_subwarp_size) {
                 return compiled_subwarp_size >= cols ||
                        compiled_subwarp_size == config::warp_size;
             },
diff --git a/dpcpp/base/kernel_launch_reduction.dp.hpp b/dpcpp/base/kernel_launch_reduction.dp.hpp
@@ -576,7 +576,7 @@ void run_kernel_row_reduction_stage1(std::shared_ptr<const DpcppExecutor> exec,
     } else {
         select_generic_kernel_row_reduction_2d(
             subsubgroup_sizes(),
-            [&](int compiled_ssg_size) {
+            [cols](int compiled_ssg_size) {
                 return compiled_ssg_size >= cols ||
                        compiled_ssg_size == sg_size;
             },
@@ -612,7 +612,7 @@ void run_kernel_col_reduction_stage1(std::shared_ptr<const DpcppExecutor> exec,
     if (cols <= sg_size) {
         select_generic_col_reduction_small(
             subsubgroup_sizes(),
-            [&](int compiled_ssg_size) {
+            [cols](int compiled_ssg_size) {
                 return compiled_ssg_size >= cols ||
                        compiled_ssg_size == sg_size;
             },
diff --git a/hip/base/kernel_launch_reduction.hip.hpp b/hip/base/kernel_launch_reduction.hip.hpp
@@ -466,7 +466,7 @@ void run_kernel_row_reduction(std::shared_ptr<const HipExecutor> exec,
     } else {
         select_run_generic_kernel_row_reduction(
             subwarp_sizes(),
-            [&](int compiled_subwarp_size) {
+            [cols](int compiled_subwarp_size) {
                 return compiled_subwarp_size >= cols ||
                        compiled_subwarp_size == config::warp_size;
             },
@@ -496,7 +496,7 @@ void run_kernel_col_reduction(std::shared_ptr<const HipExecutor> exec,
     if (cols <= config::warp_size) {
         select_generic_col_reduction_small(
             subwarp_sizes(),
-            [&](int compiled_subwarp_size) {
+            [cols](int compiled_subwarp_size) {
                 return compiled_subwarp_size >= cols ||
                        compiled_subwarp_size == config::warp_size;
             },
