From e2140e26214e1f5285fbd56535323b54fe943870 Mon Sep 17 00:00:00 2001
From: Noel Chalmers <noel.chalmers@gmail.com>
Date: Thu, 9 Nov 2023 10:10:07 -0600
Subject: [PATCH] Resize the argument buffer to be able to fit large arguments
 passed by value (#714)

---
 src/occa/internal/modes/hip/kernel.cpp | 22 ++++++++++++++++------
 src/occa/internal/modes/hip/kernel.hpp |  2 +-
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/src/occa/internal/modes/hip/kernel.cpp b/src/occa/internal/modes/hip/kernel.cpp
index bbf514087..e721a24e0 100644
--- a/src/occa/internal/modes/hip/kernel.cpp
+++ b/src/occa/internal/modes/hip/kernel.cpp
@@ -72,15 +72,25 @@ namespace occa {
 
     void kernel::deviceRun() const {
       const int args = (int) arguments.size();
-      if (!args) {
-        vArgs.resize(1);
-      } else if ((int) vArgs.size() < args) {
-        vArgs.resize(args);
+
+      int offset = 0;
+      for (int i = 0; i < args; ++i) {
+        const kernelArgData &arg = arguments[i];
+        const udim_t argSize = arg.size();
+
+        offset += offset % std::min(static_cast<size_t>(argSize),
+                                    sizeof(void*)); //align
+        offset += argSize;
+      }
+
+      size_t argBufferSize = std::max(offset,1);
+      if (vArgs.size() < argBufferSize) {
+        vArgs.resize(argBufferSize);
       }
 
       // HIP expects kernel arguments to be aligned
-      char *dataPtr = (char*) &(vArgs[0]);
-      int offset = 0;
+      std::byte *dataPtr = vArgs.data();
+      offset = 0;
       for (int i = 0; i < args; ++i) {
         const kernelArgData &arg = arguments[i];
         const udim_t argSize = arg.size();
diff --git a/src/occa/internal/modes/hip/kernel.hpp b/src/occa/internal/modes/hip/kernel.hpp
index 95eb49e4c..3a98b9a89 100644
--- a/src/occa/internal/modes/hip/kernel.hpp
+++ b/src/occa/internal/modes/hip/kernel.hpp
@@ -15,7 +15,7 @@ namespace occa {
       hipModule_t hipModule;
       hipFunction_t hipFunction;
 
-      mutable std::vector<void*> vArgs;
+      mutable std::vector<std::byte> vArgs;
 
     public:
       kernel(modeDevice_t *modeDevice_,