Merge pull request #352 from IntelPython/fix/l2_norm_diff

Address difference in l2_norm implementations

Author: adarshyoga

.github/workflows/conda-package.yml

@@ -33,7 +33,7 @@ jobs:
       fail-fast: false
       matrix:
         python: ['3.9', '3.10', '3.11']
-        os: [ubuntu-latest, windows-latest]
+        os: [ubuntu-latest, windows-2019]
 
     runs-on: ${{ matrix.os }}
 

dpbench/benchmarks/default/l2_norm/l2_norm_numba_dpex_p.py

@@ -10,6 +10,7 @@
 @dpjit
 def l2_norm(a, d):
     for i in nb.prange(a.shape[0]):
+        d[i] = 0.0
         for k in range(a.shape[1]):
             d[i] += np.square(a[i, k])
         d[i] = np.sqrt(d[i])

dpbench/benchmarks/default/l2_norm/l2_norm_numba_mlir_p.py

@@ -10,6 +10,7 @@
 @nb.njit(parallel=True, fastmath=True)
 def _l2_norm(a, d):
     for i in numba.prange(a.shape[0]):
+        d[i] = 0.0
         for k in range(a.shape[1]):
             d[i] += np.square(a[i, k])
         d[i] = np.sqrt(d[i])

dpbench/benchmarks/default/l2_norm/l2_norm_numba_npr.py

@@ -9,6 +9,7 @@
 @nb.njit(parallel=True, fastmath=True)
 def l2_norm(a, d):
     for i in nb.prange(a.shape[0]):
+        d[i] = 0.0
         for k in range(a.shape[1]):
             d[i] += np.square(a[i, k])
         d[i] = np.sqrt(d[i])

dpbench/benchmarks/default/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp

@@ -22,6 +22,7 @@ void l2_norm_impl(queue Queue,
         .submit([&](handler &h) {
             h.parallel_for<theKernel<FpTy>>(range<1>{npoints}, [=](id<1> myID) {
                 size_t i = myID[0];
+                d[i] = 0.0;
                 for (size_t k = 0; k < dims; k++) {
                     d[i] += a[i * dims + k] * a[i * dims + k];
                 }