From af6a9eedd118b0e3d7f1a2098c10ff90b27b5d0e Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Sat, 11 May 2019 20:39:00 +0200
Subject: [PATCH] Added a function to set the OpenCL kernel standard, either
 1.1 or 1.2

---
 src/clpp11.hpp              |  1 -
 src/routine.cpp             |  1 +
 src/utilities/compile.cpp   |  1 +
 src/utilities/utilities.cpp | 13 +++++++++++++
 src/utilities/utilities.hpp |  4 ++++
 5 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/src/clpp11.hpp b/src/clpp11.hpp
index 786f7dbb..aeb14989 100644
--- a/src/clpp11.hpp
+++ b/src/clpp11.hpp
@@ -485,7 +485,6 @@ class Program {
 
   // Compiles the device program and checks whether or not there are any warnings/errors
   void Build(const Device &device, std::vector<std::string> &options) {
-    options.push_back("-cl-std=CL1.1");
     auto options_string = std::accumulate(options.begin(), options.end(), std::string{" "});
     const cl_device_id dev = device();
     CheckError(clBuildProgram(program_, 1, &dev, options_string.c_str(), nullptr, nullptr));
diff --git a/src/routine.cpp b/src/routine.cpp
index 2df6d563..ed1b58ee 100644
--- a/src/routine.cpp
+++ b/src/routine.cpp
@@ -97,6 +97,7 @@ void Routine::InitProgram(std::initializer_list<const char *> source) {
                                             &has_binary);
   if (has_binary) {
     program_ = std::make_shared<Program>(device_, context_, binary);
+    SetOpenCLKernelStandard(device_, options);
     program_->Build(device_, options);
     ProgramCache::Instance().Store(ProgramKey{ context_(), device_(), precision_, routine_info },
                                     std::shared_ptr<Program>{program_});
diff --git a/src/utilities/compile.cpp b/src/utilities/compile.cpp
index 00cb90cb..aeb7a3e2 100644
--- a/src/utilities/compile.cpp
+++ b/src/utilities/compile.cpp
@@ -111,6 +111,7 @@ std::shared_ptr<Program> CompileFromSource(
   // Compiles the kernel
   auto program = std::make_shared<Program>(context, kernel_string);
   try {
+    SetOpenCLKernelStandard(device, options);
     program->Build(device, options);
   } catch (const CLCudaAPIBuildError &e) {
     if (program->StatusIsCompilationWarningOrError(e.status()) && !silent) {
diff --git a/src/utilities/utilities.cpp b/src/utilities/utilities.cpp
index a0e89c98..29161e74 100644
--- a/src/utilities/utilities.cpp
+++ b/src/utilities/utilities.cpp
@@ -498,6 +498,19 @@ std::string GetDeviceName(const Device& device) {
 
 // =================================================================================================
 
+void SetOpenCLKernelStandard(const Device &device, std::vector<std::string> &options) {
+  // Inclusion of one of the following extensions needs OpenCL 1.2 kernels
+  if (device.HasExtension(kKhronosIntelSubgroups)) {
+    options.push_back("-cl-std=CL1.2");
+  }
+    // Otherwise we fall-back to the default CLBlast OpenCL 1.1
+  else {
+    options.push_back("-cl-std=CL1.1");
+  }
+}
+
+// =================================================================================================
+
 // Solve Bezout's identity
 // a * p + b * q = r = GCD(a, b)
 void EuclidGCD(int a, int b, int &p, int &q, int &r) {
diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp
index 23486d35..b66df118 100644
--- a/src/utilities/utilities.hpp
+++ b/src/utilities/utilities.hpp
@@ -375,6 +375,10 @@ std::string GetDeviceName(const Device& device);
 
 // =================================================================================================
 
+void SetOpenCLKernelStandard(const Device &device, std::vector<std::string> &options);
+
+// =================================================================================================
+
 // Solve Bezout's identity
 // a * p + b * q = r = GCD(a, b)
 void EuclidGCD(int a, int b, int &p, int &q, int &r);