Prevent AMD OpenCL on OS X prior to 10.10.4

OS X release 10.9.x has been confirmed to produce invalid OpenCL
kernels for AMD (but correct for NVIDIA), and we have tested that
it is fixed in 10.10.4. The fix might actually have appeared
already in 10.10.3, but it's not worth tracking down the exact
point since 10.10.5 is already out.
This change will issue a CMake warning when compiling GROMACS
on earlier OS X releases, and at runtime we check the version
again and mark AMD GPUs as incompatible.

Also fixed docs about the length of time JIT compilation takes
on AMD.

Refs #1783

Change-Id: I0202faea60c39daae6621d2bb9ba828aab5532a0

cmake/gmxManageOpenCL.cmake

@@ -57,6 +57,11 @@ if (UNIX)
     set(OPENCL_DEFINITIONS ${OPENCL_DEFINITIONS} " -Wno-comment")
 endif()
 
+# Yes Virginia, Darwin kernel version 14.4 corresponds to OS X 10.4.
+if(APPLE AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "14.4")
+        message(WARNING "OS X prior to version 10.10.4 produces incorrect AMD OpenCL code at runtime. You will not be able to use AMD GPUs on this host unless you upgrade your operating system.");
+endif()
+
 add_definitions(${OPENCL_DEFINITIONS})
 
 include_directories(${OPENCL_INCLUDE_DIRS})

docs/install-guide/index.rst

@@ -630,6 +630,8 @@ do so anyway, because NVIDIA OpenCL support is part of the CUDA
 package, a C++ compiler supported by your CUDA installation is
 required.
 
+On Mac OS, an AMD GPU can be used only with OS version 10.10.4 and
+higher; earlier OS versions are known to run incorrectly.
 
 Static linking
 ^^^^^^^^^^^^^^

docs/user-guide/mdrun-performance.rst

@@ -537,12 +537,9 @@ The following devices are known to work correctly:
    - NVIDIA: GeForce GTX 660M, GeForce GTX 660Ti, GeForce GTX 750Ti,
      GeForce GTX 780, GTX Titan
 
-Building an OpenCL program can take a significant amount of
-time. NVIDIA implements a mechanism to cache the result of the
-build. As a consequence, only the first run will take longer (because
-of the kernel builds), and the following runs will be very fast. AMD
-drivers, on the other hand, implement no caching and the initial phase
-of running an OpenCL program can be very slow. This is not normally a
+Building the OpenCL program can take a few seconds when :ref:`gmx
+mdrun` starts up, because the kernels that run on the
+GPU can only be compiled at run time. This is not normally a
 problem for long production MD, but you might prefer to do some kinds
 of work on just the CPU (e.g. see ``-nb`` above).
 

src/gromacs/gmxlib/gpu_utils/gpu_utils_ocl.cpp

@@ -46,6 +46,9 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#if __APPLE__
+#    include <sys/sysctl.h>
+#endif
 
 #include <memory.h>
 
@@ -81,6 +84,36 @@ static bool is_compatible_gpu(int stat)
     return (stat == egpuCompatible);
 }
 
+/*! \brief Return true if executing on OS X earlier than 10.10.4
+ *
+ * Uses the BSD sysctl() interfaces to extract the kernel version.
+ *
+ * \return true if version is 14.4 or later (= OS X version 10.10.4),
+ *         otherwise false.
+ */
+static bool
+runningOnWorkingOSXVersionForAmd()
+{
+#ifdef __APPLE__
+    int    mib[2];
+    char   kernelVersion[256];
+    size_t len = sizeof(kernelVersion);
+
+    mib[0] = CTL_KERN;
+    mib[1] = KERN_OSRELEASE;
+
+    sysctl(mib, sizeof(mib)/sizeof(mib[0]), kernelVersion, &len, NULL, 0);
+
+    int major = strtod(kernelVersion, NULL);
+    int minor = strtod(strchr(kernelVersion, '.')+1, NULL);
+
+    // Kernel 14.4 corresponds to OS X 10.10.4
+    return (major > 14 || (major == 14 && minor >= 4));
+#else
+    return false;
+#endif
+}
+
 /*! \brief Returns true if the gpu characterized by the device properties is
  *  supported by the native gpu acceleration.
  * \returns             true if the GPU properties passed indicate a compatible
@@ -89,15 +122,18 @@ static bool is_compatible_gpu(int stat)
 static int is_gmx_supported_gpu_id(struct gmx_device_info_t *ocl_gpu_device)
 {
     /* Only AMD and NVIDIA GPUs are supported for now */
-    if ((OCL_VENDOR_NVIDIA == ocl_gpu_device->vendor_e) ||
-        (OCL_VENDOR_AMD == ocl_gpu_device->vendor_e))
+    switch (ocl_gpu_device->vendor_e)
     {
-        return egpuCompatible;
+        case OCL_VENDOR_NVIDIA:
+            return egpuCompatible;
+        case OCL_VENDOR_AMD:
+            return runningOnWorkingOSXVersionForAmd() ? egpuCompatible : egpuIncompatible;
+        default:
+            return egpuIncompatible;
     }
-
-    return egpuIncompatible;
 }
 
+
 /*! \brief Returns an ocl_vendor_id_t value corresponding to the input OpenCL vendor name.
  *
  *  \param[in] vendor_name String with OpenCL vendor name.