From acb555c4c7ba5a9314d6a4d562d3dfdb16f3816a Mon Sep 17 00:00:00 2001
From: Maxiwell <maxiwell@linux.ibm.com>
Date: Tue, 19 Apr 2022 19:13:55 -0300
Subject: [PATCH] ppc64le: Optimizing the MlasMaximumPool() to use VSX
 instructions (#11216)

It runs on Power8, Power9, and Power10
---
 onnxruntime/core/mlas/lib/pooling.cpp | 36 +++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)

diff --git a/onnxruntime/core/mlas/lib/pooling.cpp b/onnxruntime/core/mlas/lib/pooling.cpp
index 5655435581..d9ce10481e 100644
--- a/onnxruntime/core/mlas/lib/pooling.cpp
+++ b/onnxruntime/core/mlas/lib/pooling.cpp
@@ -1538,6 +1538,42 @@ Return Value:
             c -= 8;
         }
 
+#elif defined(MLAS_TARGET_POWER)
+
+        while (c >= 32) {
+            auto MaximumVector0 = vec_splats(std::numeric_limits<T8Bits>::lowest());
+            auto MaximumVector1 = vec_splats(std::numeric_limits<T8Bits>::lowest());
+
+            for (size_t k = 0; k < KernelSize; k++) {
+                auto InputVector0 = vec_xl(0,  &Input[k][ChannelOffset]);
+                auto InputVector1 = vec_xl(16, &Input[k][ChannelOffset]);
+
+                MaximumVector0 = vec_max(MaximumVector0, InputVector0);
+                MaximumVector1 = vec_max(MaximumVector1, InputVector1);
+            }
+
+            vec_xst(MaximumVector0, 0, (T8Bits *) Output);
+            vec_xst(MaximumVector1, 16, (T8Bits *) Output);
+
+            Output += 32;
+            ChannelOffset += 32;
+            c -= 32;
+        }
+
+        while (c >= 16) {
+            auto MaximumVector = vec_splats(std::numeric_limits<T8Bits>::lowest());
+
+            for (size_t k = 0; k < KernelSize; k++) {
+                auto InputVector = vec_xl(0,  &Input[k][ChannelOffset]);
+                MaximumVector = vec_max(MaximumVector, InputVector);
+            }
+            vec_xst(MaximumVector, 0, (T8Bits *) Output);
+
+            Output += 16;
+            ChannelOffset += 16;
+            c -= 16;
+        }
+
 #endif
 
         while (c > 0) {