[SYCL] Add Windows support for device_info

Fixed few warnings caused by inconsistent usage of class and struct keywords. Added a macro to expose global symbols and generate sycl.lib on Windows. Signed-off-by: Vyacheslav N Klochkov <[email protected]>
intel · May 21, 2019 · 3b0defe · 3b0defe
1 parent ed668e0
commit 3b0defe
Show file tree

Hide file tree

Showing 13 changed files with 232 additions and 115 deletions.
diff --git a/sycl/CMakeLists.txt b/sycl/CMakeLists.txt
@@ -5,6 +5,7 @@ project(sycl-solution)
 set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
 
 if(MSVC)
   set_property(GLOBAL PROPERTY USE_FOLDERS ON)
@@ -135,6 +136,7 @@ add_library("${SYCLLibrary}" SHARED
   "${sourceRootPath}/detail/program_manager/program_manager.cpp"
   "${sourceRootPath}/detail/queue_impl.cpp"
   "${sourceRootPath}/detail/os_util.cpp"
+  "${sourceRootPath}/detail/platform_util.cpp"
   "${sourceRootPath}/detail/sampler_impl.cpp"
   "${sourceRootPath}/detail/scheduler/commands.cpp"
   "${sourceRootPath}/detail/scheduler/commands2.cpp"

diff --git a/sycl/include/CL/sycl/detail/buffer_impl.hpp b/sycl/include/CL/sycl/detail/buffer_impl.hpp
@@ -42,7 +42,7 @@ class accessor;
 template <typename T, int dimensions, typename AllocatorT> class buffer;
 class handler;
 class queue;
-template <int dimentions> class id;
+template <int dimentions> struct id;
 template <int dimentions> class range;
 using buffer_allocator = aligned_allocator<char, /*alignment*/ 64>;
 namespace detail {

diff --git a/sycl/include/CL/sycl/detail/helpers.hpp b/sycl/include/CL/sycl/detail/helpers.hpp
@@ -19,10 +19,10 @@ namespace cl {
 namespace sycl {
 class context;
 class event;
-template <int dimensions, bool with_offset> class item;
+template <int dimensions, bool with_offset> struct item;
 template <int dimensions> class group;
 template <int dimensions> class range;
-template <int dimensions> class id;
+template <int dimensions> struct id;
 template <int dimensions> class nd_item;
 namespace detail {
 class context_impl;

diff --git a/sycl/include/CL/sycl/detail/os_util.hpp b/sycl/include/CL/sycl/detail/os_util.hpp
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <stdlib.h>
+
 #ifdef _WIN32
 #define SYCL_RT_OS_WINDOWS
 // Windows platform
@@ -48,6 +50,9 @@ class OSUtil {
   /// Module handle for the executable module - it is assumed there is always
   /// single one at most.
   static const OSModuleHandle ExeModuleHandle;
+
+  /// Returns the amount of RAM available for the operating system.
+  static size_t getOSMemSize();
 };
 
 } // namespace detail

diff --git a/sycl/include/CL/sycl/detail/platform_util.hpp b/sycl/include/CL/sycl/detail/platform_util.hpp
@@ -0,0 +1,40 @@
+//===-- platform_util.hpp - platform utilities ----------------*- C++ -*--===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include <cstdint>
+
+namespace cl {
+namespace sycl {
+namespace detail {
+
+struct PlatformUtil {
+  enum class TypeIndex : unsigned int {
+    Char = 0,
+    Short = 1,
+    Int = 2,
+    Long = 3,
+    Float = 4,
+    Double = 5,
+    Half = 6
+  };
+
+  /// Returns the maximum vector width counted in elements of the given type.
+  static uint32_t getNativeVectorWidth(TypeIndex Index);
+
+  static uint32_t getMaxClockFrequency();
+
+  static uint32_t getMemCacheLineSize();
+
+  static uint64_t getMemCacheSize();
+};
+
+} // namespace detail
+} // namespace sycl
+} // namespace cl
diff --git a/sycl/include/CL/sycl/group.hpp b/sycl/include/CL/sycl/group.hpp
@@ -19,7 +19,7 @@
 namespace cl {
 namespace sycl {
 namespace detail {
-class Builder;
+struct Builder;
 } // namespace detail
 
 template <int dimensions = 1> class group {

diff --git a/sycl/include/CL/sycl/id.hpp b/sycl/include/CL/sycl/id.hpp
@@ -16,8 +16,9 @@ namespace cl {
 namespace sycl {
 template <int dimensions> class range;
 template <int dimensions = 1> struct id : public detail::array<dimensions> {
-public:
+private:
   using base = detail::array<dimensions>;
+public:
   id() = default;
 
   /* The following constructor is only available in the id struct

diff --git a/sycl/include/CL/sycl/item.hpp b/sycl/include/CL/sycl/item.hpp
@@ -16,10 +16,10 @@
 namespace cl {
 namespace sycl {
 namespace detail {
-class Builder;
+struct Builder;
 }
 template <int dimensions> struct id;
-template <int dimensions> struct range;
+template <int dimensions> class range;
 template <int dimensions = 1, bool with_offset = true> struct item {
 
   item() = delete;
@@ -86,8 +86,9 @@ template <int dimensions = 1, bool with_offset = true> struct item {
 
 protected:
   // For call constructor inside conversion operator
-  friend class item<dimensions, false>;
-  friend class detail::Builder;
+  friend struct item<dimensions, false>;
+  friend struct item<dimensions, true>;
+  friend struct detail::Builder;
 
   template <size_t W = with_offset>
   item(typename std::enable_if<(W == true), const range<dimensions>>::type &R,

diff --git a/sycl/include/CL/sycl/nd_item.hpp b/sycl/include/CL/sycl/nd_item.hpp
@@ -22,7 +22,7 @@
 namespace cl {
 namespace sycl {
 namespace detail {
-class Builder;
+struct Builder;
 }
 template <int dimensions = 1> struct nd_item {
 

diff --git a/sycl/include/CL/sycl/range.hpp b/sycl/include/CL/sycl/range.hpp
@@ -16,8 +16,8 @@ namespace sycl {
 template <int dimensions> struct id;
 template <int dimensions = 1>
 class range : public detail::array<dimensions> {
-public:
   using base = detail::array<dimensions>;
+public:
   /* The following constructor is only available in the range class
   specialization where: dimensions==1 */
   template <int N = dimensions>

diff --git a/sycl/source/detail/device_info.cpp b/sycl/source/detail/device_info.cpp
@@ -7,9 +7,10 @@
 //===----------------------------------------------------------------------===//
 
 #include <CL/sycl/detail/device_info.hpp>
+#include <CL/sycl/detail/os_util.hpp>
+#include <CL/sycl/detail/platform_util.hpp>
 #include <CL/sycl/device.hpp>
 #include <chrono>
-#include <sys/sysinfo.h>
 #include <thread>
 
 #ifdef __GNUG__
@@ -21,22 +22,6 @@ namespace cl {
 namespace sycl {
 namespace detail {
 
-// Used by methods that duplicate OpenCL behaviour in order to get CPU info
-// TODO add Windows support
-// TODO add support for x86-64 ABI selected using ifdef.
-static void cpuid(unsigned int cpuid_info[], unsigned int type) {
-  unsigned int eax, ebx, ecx, edx;
-  __asm__ __volatile__("mov  %%ebx, %%edi\n\r"
-                       "cpuid\n\r"
-                       "xchg %%edi, %%ebx\n\r"
-                       : "=a"(eax), "=D"(ebx), "=c"(ecx), "=d"(edx)
-                       : "a"(type));
-  cpuid_info[0] = eax;
-  cpuid_info[1] = ebx;
-  cpuid_info[2] = ecx;
-  cpuid_info[3] = edx;
-}
-
 vector_class<info::fp_config> read_fp_bitfield(cl_device_fp_config bits) {
   vector_class<info::fp_config> result;
   if (bits & CL_FP_DENORM)
@@ -156,109 +141,51 @@ cl_uint get_device_info_host<info::device::preferred_vector_width_half>() {
   return 0;
 }
 
-// SSE4.2 has 16 byte (XMM) registers
-static const cl_uint NATIVE_VECTOR_WIDTH_SSE42[] = {16, 8, 4, 2, 4, 2, 0};
-// AVX supports 32 byte (YMM) registers only for floats and doubles
-static const cl_uint NATIVE_VECTOR_WIDTH_AVX[] = {16, 8, 4, 2, 8, 4, 0};
-// AVX2 has a full set of 32 byte (YMM) registers
-static const cl_uint NATIVE_VECTOR_WIDTH_AVX2[] = {32, 16, 8, 4, 8, 4, 0};
-// AVX512 has 64 byte (ZMM) registers
-static const cl_uint NATIVE_VECTOR_WIDTH_AVX512[] = {64, 32, 16, 8, 16, 8, 0};
-
-cl_uint get_native_vector_width(size_t idx) {
-#if (__GNUG__ && GCC_VERSION > 40900)
-  if (__builtin_cpu_supports("avx512f")) {
-    return NATIVE_VECTOR_WIDTH_AVX512[idx];
-  }
-#endif
-
-  if (__builtin_cpu_supports("avx2")) {
-    return NATIVE_VECTOR_WIDTH_AVX2[idx];
-  }
-  if (__builtin_cpu_supports("avx")) {
-    return NATIVE_VECTOR_WIDTH_AVX[idx];
-  }
-  return NATIVE_VECTOR_WIDTH_SSE42[idx];
-}
-
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_char>() {
-  return get_native_vector_width(0);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Char);
 }
 
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_short>() {
-  return get_native_vector_width(1);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Short);
 }
 
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_int>() {
-  return get_native_vector_width(2);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Int);
 }
 
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_long>() {
-  return get_native_vector_width(3);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Long);
 }
 
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_float>() {
-  return get_native_vector_width(4);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Float);
 }
 
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_double>() {
-  return get_native_vector_width(5);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Double);
 }
 
 template <>
 cl_uint get_device_info_host<info::device::native_vector_width_half>() {
-  return get_native_vector_width(6);
+  return PlatformUtil::getNativeVectorWidth(PlatformUtil::TypeIndex::Half);
 }
 
 template <> cl_uint get_device_info_host<info::device::max_clock_frequency>() {
-  throw runtime_error(
-      "max_clock_frequency parameter is not supported for host device");
-  unsigned int cpuInfo[4] = {0 - 1u};
-  string_class buff(sizeof(cpuInfo) * 3 + 1, 0);
-  size_t offset = 0;
-
-  for (unsigned int i = 0x80000002; i <= 0x80000004; i++) {
-    cpuid(cpuInfo, i);
-    std::copy(reinterpret_cast<char *>(cpuInfo),
-              reinterpret_cast<char *>(cpuInfo) + sizeof(cpuInfo),
-              buff.begin() + offset);
-    offset += sizeof(cpuInfo);
-  }
-  std::size_t found = buff.rfind("Hz");
-  // Bail out if frequency is not found in CPUID string
-  if (found == std::string::npos)
-    return 0;
-
-  buff = buff.substr(0, found);
-
-  cl_uint freq = 0;
-  switch (buff[buff.size() - 1]) {
-  case 'M':
-    freq = 1;
-    break;
-  case 'G':
-    freq = 1000;
-    break;
-  }
-  buff = buff.substr(buff.rfind(' '), buff.length());
-  freq *= std::stod(buff);
-  return freq;
+  return PlatformUtil::getMaxClockFrequency();
 }
 
 template <> cl_uint get_device_info_host<info::device::address_bits>() {
   return sizeof(void *) * 8;
 }
 
 template <> cl_ulong get_device_info_host<info::device::global_mem_size>() {
-  struct sysinfo meminfo;
-  sysinfo(&meminfo);
-  return meminfo.totalram * meminfo.mem_unit;
+  return static_cast<cl_ulong>(OSUtil::getOSMemSize());
 }
 
 template <> cl_ulong get_device_info_host<info::device::max_mem_alloc_size>() {
@@ -362,16 +289,12 @@ get_device_info_host<info::device::global_mem_cache_type>() {
 
 template <>
 cl_uint get_device_info_host<info::device::global_mem_cache_line_size>() {
-  unsigned int viCPUInfo[4] = {(unsigned int)-1};
-  cpuid(viCPUInfo, 0x80000006);
-  return viCPUInfo[2] & 0xff;
+  return PlatformUtil::getMemCacheLineSize();
 }
 
 template <>
 cl_ulong get_device_info_host<info::device::global_mem_cache_size>() {
-  unsigned int viCPUInfo[4] = {(unsigned int)-1};
-  cpuid(viCPUInfo, 0x80000006);
-  return ((viCPUInfo[2] >> 16) & 0xffff) * 1024;
+  return PlatformUtil::getMemCacheSize();
 }
 
 template <>