fix code execution with Intel CPU driver

silx-kit · Dec 7, 2024 · b23d02a · b23d02a
1 parent af92780
commit b23d02a
Show file tree

Hide file tree

Showing 4 changed files with 13 additions and 9 deletions.
diff --git a/src/pyFAI/opencl/azim_csr.py b/src/pyFAI/opencl/azim_csr.py
@@ -1301,8 +1301,6 @@ def medfilt(self, data, dark=None, dummy=None, delta_dummy=None,
 
         res = Integrate1dtpl(self.bin_centers, avgint, sem, merged[:, 0], merged[:, 2], merged[:, 4], merged[:, 6],
                              std, sem, merged[:, 7])
-        # print(self.get_buffer("merged8")[0])
-        # raise RuntimeError()
         return res
 
 

diff --git a/src/pyFAI/opencl/test/test_ocl_azim_csr.py b/src/pyFAI/opencl/test/test_ocl_azim_csr.py
@@ -122,10 +122,10 @@ def integrate_ng(self, block_size=None, method_called="integrate_ng", extra=None
 
 
             # histogram of normalization
-            print(ref.sum_normalization)
-            print(res.normalization)
+            # print(ref.sum_normalization)
+            # print(res.normalization)
             err = abs((res.normalization - ref.sum_normalization))
-            print(err)
+            # print(err)
             self.assertLess(err.max(), 5e-4, "normalization content is the same: %s<5e-5" % (err.max))
 
             # histogram of signal
@@ -136,7 +136,7 @@ def integrate_ng(self, block_size=None, method_called="integrate_ng", extra=None
 
             # Intensities are not that different:
             delta = ref.intensity - res.intensity
-            print(delta)
+            # print(delta)
             self.assertLessEqual(abs(delta).max(), 1e-5, "intensity is almost the same")
 
 

diff --git a/src/pyFAI/resources/openCL/collective/reduction.cl b/src/pyFAI/resources/openCL/collective/reduction.cl
@@ -17,7 +17,9 @@ int inline sum_int_reduction(local int* shared)
             shared[tid] += shared[tid+stride];
     }
     barrier(CLK_LOCAL_MEM_FENCE);
-    return shared[0];
+    int res = shared[0];
+    barrier(CLK_LOCAL_MEM_FENCE);
+    return res;
 }
 
 /* sum all elements in a shared memory, same size as the workgroup size 0
@@ -35,7 +37,9 @@ int inline sum_int_atomic(local int* shared)
     if (tid)
         atomic_add(shared, value);
     barrier(CLK_LOCAL_MEM_FENCE);
-    return shared[0];
+    int res = shared[0];
+    barrier(CLK_LOCAL_MEM_FENCE);
+    return res;
 }
 
 /*

diff --git a/src/pyFAI/resources/openCL/medfilt.cl b/src/pyFAI/resources/openCL/medfilt.cl
@@ -70,7 +70,9 @@ float2 inline sum_float2_reduction(local float* shared)
 
     }
     barrier(CLK_LOCAL_MEM_FENCE);
-    return (float2)(shared[0], shared[1]);
+    float2 res = (float2)(shared[0], shared[1]);
+    barrier(CLK_LOCAL_MEM_FENCE);
+    return res;
 }