From 05a24078ec9dc03c10bccb5a18a9109556f4f73b Mon Sep 17 00:00:00 2001
From: Adarsh Yoga <adarsh.yoga@intel.com>
Date: Thu, 16 May 2024 16:53:04 +0000
Subject: [PATCH] l2_norm code difference fix

---
 dpbench/benchmarks/default/l2_norm/l2_norm_numba_dpex_p.py       | 1 +
 dpbench/benchmarks/default/l2_norm/l2_norm_numba_mlir_p.py       | 1 +
 dpbench/benchmarks/default/l2_norm/l2_norm_numba_npr.py          | 1 +
 .../l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp     | 1 +
 4 files changed, 4 insertions(+)

diff --git a/dpbench/benchmarks/default/l2_norm/l2_norm_numba_dpex_p.py b/dpbench/benchmarks/default/l2_norm/l2_norm_numba_dpex_p.py
index 8b5df2de..5342a5c8 100644
--- a/dpbench/benchmarks/default/l2_norm/l2_norm_numba_dpex_p.py
+++ b/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])
diff --git a/dpbench/benchmarks/default/l2_norm/l2_norm_numba_mlir_p.py b/dpbench/benchmarks/default/l2_norm/l2_norm_numba_mlir_p.py
index eedd437c..5f2bdc1b 100644
--- a/dpbench/benchmarks/default/l2_norm/l2_norm_numba_mlir_p.py
+++ b/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])
diff --git a/dpbench/benchmarks/default/l2_norm/l2_norm_numba_npr.py b/dpbench/benchmarks/default/l2_norm/l2_norm_numba_npr.py
index db050ae6..48c42eb6 100644
--- a/dpbench/benchmarks/default/l2_norm/l2_norm_numba_npr.py
+++ b/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])
diff --git a/dpbench/benchmarks/default/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp b/dpbench/benchmarks/default/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp
index 8f13d972..6acfa123 100644
--- a/dpbench/benchmarks/default/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp
+++ b/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];
                 }