From 7ebc3325348bb297a539306e4a8772661d51cd04 Mon Sep 17 00:00:00 2001
From: Yury Plyakhin <yury.plyakhin@intel.com>
Date: Mon, 20 May 2024 15:05:58 -0700
Subject: [PATCH 1/6] add vec load 1d testing for element_wise_all_ops

---
 sycl/test-e2e/Matrix/element_wise_all_ops.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops.cpp b/sycl/test-e2e/Matrix/element_wise_all_ops.cpp
index 4d6493c23e18e..37eb9e0190066 100644
--- a/sycl/test-e2e/Matrix/element_wise_all_ops.cpp
+++ b/sycl/test-e2e/Matrix/element_wise_all_ops.cpp
@@ -9,6 +9,7 @@
 
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out %}
 
 #include "common.hpp"
 #include "element_wise_all_ops_impl.hpp"

From dfeb73431f8268a73fccfabf52cf27d3a858c0ad Mon Sep 17 00:00:00 2001
From: Yury Plyakhin <yury.plyakhin@intel.com>
Date: Tue, 21 May 2024 18:04:41 -0700
Subject: [PATCH 2/6] split test to 2 to reduce test time

---
 sycl/test-e2e/Matrix/element_wise_all_ops.cpp      |  1 -
 .../Matrix/element_wise_all_ops_1d_cont.cpp        | 14 ++++++++++++++
 2 files changed, 14 insertions(+), 1 deletion(-)
 create mode 100644 sycl/test-e2e/Matrix/element_wise_all_ops_1d_cont.cpp

diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops.cpp b/sycl/test-e2e/Matrix/element_wise_all_ops.cpp
index 37eb9e0190066..4d6493c23e18e 100644
--- a/sycl/test-e2e/Matrix/element_wise_all_ops.cpp
+++ b/sycl/test-e2e/Matrix/element_wise_all_ops.cpp
@@ -9,7 +9,6 @@
 
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
-// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out %}
 
 #include "common.hpp"
 #include "element_wise_all_ops_impl.hpp"
diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops_1d_cont.cpp b/sycl/test-e2e/Matrix/element_wise_all_ops_1d_cont.cpp
new file mode 100644
index 0000000000000..1529e60c4165c
--- /dev/null
+++ b/sycl/test-e2e/Matrix/element_wise_all_ops_1d_cont.cpp
@@ -0,0 +1,14 @@
+//==-------- element_wise_all_ops_1d_cont.cpp - DPC++ joint_matrix ---------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// REQUIRES: aspect-ext_intel_matrix, gpu
+
+// RUN: %{build} -o %t.out
+// RUN: env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out
+
+#include "common.hpp"
+#include "element_wise_all_ops_impl.hpp"

From 887ea6e0c5559ea869037e1c2eaf88658a0242ba Mon Sep 17 00:00:00 2001
From: Yury Plyakhin <yury.plyakhin@intel.com>
Date: Tue, 21 May 2024 18:39:39 -0700
Subject: [PATCH 3/6] [SYCL][Joint Matrix][E2E] Added tests for 1d load/store

---
 sycl/test-e2e/Matrix/element_wise_all_ops_1d.cpp | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
 create mode 100644 sycl/test-e2e/Matrix/element_wise_all_ops_1d.cpp

diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops_1d.cpp b/sycl/test-e2e/Matrix/element_wise_all_ops_1d.cpp
new file mode 100644
index 0000000000000..d3d0396593bb0
--- /dev/null
+++ b/sycl/test-e2e/Matrix/element_wise_all_ops_1d.cpp
@@ -0,0 +1,14 @@
+//==-------- element_wise_all_ops_1d.cpp - DPC++ joint_matrix ---------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// REQUIRES: aspect-ext_intel_matrix, gpu
+
+// RUN: %{build} -o %t.out
+// RUN: env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out
+
+#include "common.hpp"
+#include "element_wise_all_ops_impl.hpp"

From 38ca05fa6666de76cefa8084a5a933f56c41d4c8 Mon Sep 17 00:00:00 2001
From: Yury Plyakhin <yury.plyakhin@intel.com>
Date: Tue, 21 May 2024 18:40:54 -0700
Subject: [PATCH 4/6] more tests

---
 sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp    | 1 +
 sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp | 1 +
 2 files changed, 2 insertions(+)

diff --git a/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp b/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp
index cb1beda1f95d6..d3bf6b5744941 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp
@@ -11,6 +11,7 @@
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out %}
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out %}
 
 #include "common.hpp"
 #include "joint_matrix_bfloat16_packedB_impl.hpp"
diff --git a/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp b/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp
index c537112345be5..c3ba99bfbabde 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp
@@ -12,6 +12,7 @@
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out %}
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out %}
 
 // This tests support of row major layout for matrix B which does automatic VNNI
 // transform. This is currently only available on AMX and XMX of PVC

From d5dece0feef6744e32f9afd047e3a29b0aee034d Mon Sep 17 00:00:00 2001
From: Yury Plyakhin <yury.plyakhin@intel.com>
Date: Tue, 21 May 2024 20:37:42 -0700
Subject: [PATCH 5/6] [SYCL][Joint Matrix][E2E] Added tests for gather/scatter
 load/store

---
 .../Matrix/element_wise_all_ops_gather_scatter.cpp | 14 ++++++++++++++
 .../test-e2e/Matrix/joint_matrix_annotated_ptr.cpp |  1 +
 .../Matrix/joint_matrix_bfloat16_packedB.cpp       |  1 +
 sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp    |  1 +
 .../Matrix/joint_matrix_rowmajorA_rowmajorB.cpp    |  1 +
 5 files changed, 18 insertions(+)
 create mode 100644 sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp

diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp b/sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp
new file mode 100644
index 0000000000000..b3d4e6b4e779e
--- /dev/null
+++ b/sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp
@@ -0,0 +1,14 @@
+//==---- element_wise_all_ops_gather_scatter.cpp - DPC++ joint_matrix ------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// REQUIRES: aspect-ext_intel_matrix, gpu
+
+// RUN: %{build} -o %t.out
+// RUN: env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out
+
+#include "common.hpp"
+#include "element_wise_all_ops_impl.hpp"
diff --git a/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp b/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp
index 7aad02b2066a1..a3930fe5a8973 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp
@@ -9,6 +9,7 @@
 
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out %}
 
 #include "common.hpp"
 
diff --git a/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp b/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp
index d3bf6b5744941..bf3a65d2a16d7 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_bfloat16_packedB.cpp
@@ -12,6 +12,7 @@
 // RUN: %{run} %t.out
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out %}
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out %}
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out %}
 
 #include "common.hpp"
 #include "joint_matrix_bfloat16_packedB_impl.hpp"
diff --git a/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp b/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp
index f592057ce94d5..f43dcc018ed9d 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp
@@ -9,6 +9,7 @@
 
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out %}
 
 #include "common.hpp"
 
diff --git a/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp b/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp
index c3ba99bfbabde..bfcf624abe97b 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB.cpp
@@ -13,6 +13,7 @@
 // RUN: %{run} %t.out
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=2 %{run} %t.out %}
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out %}
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out %}
 
 // This tests support of row major layout for matrix B which does automatic VNNI
 // transform. This is currently only available on AMX and XMX of PVC

From 20c7bf3bba93b506f331b0af94f21194c225ac07 Mon Sep 17 00:00:00 2001
From: Yury Plyakhin <yury.plyakhin@intel.com>
Date: Tue, 28 May 2024 14:27:24 -0700
Subject: [PATCH 6/6] addressed feedback

---
 ...l_ops_gather_scatter.cpp => element_wise_all_ops_scalar.cpp} | 2 +-
 sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp             | 1 +
 sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp                 | 1 +
 3 files changed, 3 insertions(+), 1 deletion(-)
 rename sycl/test-e2e/Matrix/{element_wise_all_ops_gather_scatter.cpp => element_wise_all_ops_scalar.cpp} (85%)

diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp b/sycl/test-e2e/Matrix/element_wise_all_ops_scalar.cpp
similarity index 85%
rename from sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp
rename to sycl/test-e2e/Matrix/element_wise_all_ops_scalar.cpp
index b3d4e6b4e779e..0604a7f345963 100644
--- a/sycl/test-e2e/Matrix/element_wise_all_ops_gather_scatter.cpp
+++ b/sycl/test-e2e/Matrix/element_wise_all_ops_scalar.cpp
@@ -1,4 +1,4 @@
-//==---- element_wise_all_ops_gather_scatter.cpp - DPC++ joint_matrix ------==//
+//==---- element_wise_all_ops_scalar.cpp - DPC++ joint_matrix ------==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
diff --git a/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp b/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp
index a3930fe5a8973..e21bdaa06a328 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_annotated_ptr.cpp
@@ -10,6 +10,7 @@
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out %}
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out %}
 
 #include "common.hpp"
 
diff --git a/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp b/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp
index f43dcc018ed9d..4ce91037fc668 100644
--- a/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp
+++ b/sycl/test-e2e/Matrix/joint_matrix_int8_vnni.cpp
@@ -10,6 +10,7 @@
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
 // RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=0 %{run} %t.out %}
+// RUN: %if gpu %{ env IGC_JointMatrixLoadStoreOpt=1 %{run} %t.out %}
 
 #include "common.hpp"