From 37d1d87c3c2842bdd4dd97050fce13a1fc2042e9 Mon Sep 17 00:00:00 2001
From: Meifeng Lin <meifeng@caa.columbia.edu>
Date: Tue, 19 Dec 2023 08:03:28 -0600
Subject: [PATCH] bug fix for Intel GPUs

---
 Grid/threads/Accelerator.cc | 11 +++--------
 benchmarks/Benchmark_su3.cc |  4 ++--
 2 files changed, 5 insertions(+), 10 deletions(-)

diff --git a/Grid/threads/Accelerator.cc b/Grid/threads/Accelerator.cc
index 698020a7..126f9bf3 100644
--- a/Grid/threads/Accelerator.cc
+++ b/Grid/threads/Accelerator.cc
@@ -14,15 +14,10 @@ void     acceleratorThreads(uint32_t t) {accelerator_threads = t;};
 #define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
 #define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
 
-#ifdef __CUDA_ARCH__
-#warning "ifdef cuda arch"
-#endif
 
 // fold omptarget into device specific acceleratorInit()
+#if defined(GRID_CUDA) || (defined(GRID_OMPTARGET) && defined(__CUDA_ARCH__))
 #include <cuda_runtime_api.h>
-//#if defined(GRID_CUDA) || (defined(GRID_OMPTARGET) && defined(__CUDA_ARCH__))
-#if defined(GRID_OMPTARGET)
-#warning "using cuda from opmtarget"
 cudaDeviceProp *gpu_props;
 cudaStream_t copyStream;
 cudaStream_t computeStream;
@@ -206,7 +201,7 @@ void acceleratorInit(void)
 #endif
 
 
-#if defined(GRID_SYCL) || (defined(GRID_OMPTARGET) && defined(__SYCL_DEVICE_ONLY__))
+#if defined(GRID_SYCL) //|| (defined(GRID_OMPTARGET) && defined(__SYCL_DEVICE_ONLY__))
 
 cl::sycl::queue *theGridAccelerator;
 cl::sycl::queue *theCopyAccelerator;
@@ -278,7 +273,7 @@ void acceleratorInit(void)
 }
 #endif
 
-#if (!defined(GRID_CUDA)) && (!defined(GRID_SYCL))&& (!defined(GRID_HIP)) && (!defined(GRID_OMPTARGET))
+#if (!defined(GRID_CUDA)) && (!defined(GRID_SYCL))&& (!defined(GRID_HIP))// && (!defined(GRID_OMPTARGET))
 void acceleratorInit(void){}
 #endif
 
diff --git a/benchmarks/Benchmark_su3.cc b/benchmarks/Benchmark_su3.cc
index da59893a..46a170cd 100644
--- a/benchmarks/Benchmark_su3.cc
+++ b/benchmarks/Benchmark_su3.cc
@@ -36,8 +36,8 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-#define LMAX (31)
-#define LMIN (31)
+#define LMAX (8)
+#define LMIN (8)
 #define LADD (8)
 
   int64_t Nwarm=500;