[SYCL] Allow native OpenCL and L0 executable binaries in persistent cache (#6256)

These changes expand the persistent device code cache to allow
executable binaries produced by IGC for both OpenCL and L0 to be cached.

Signed-off-by: Larsen, Steffen <[email protected]>

Author: steffenlarsen

sycl/source/detail/persistent_device_code_cache.cpp

@@ -47,11 +47,17 @@ LockCacheItem::~LockCacheItem() {
                                      FileName);
 }
 
+// Returns true if the specified format is either SPIRV or a native binary.
+static bool IsSupportedImageFormat(RT::PiDeviceBinaryType Format) {
+  return Format == PI_DEVICE_BINARY_TYPE_SPIRV ||
+         Format == PI_DEVICE_BINARY_TYPE_NATIVE;
+}
+
 /* Returns true if specified image should be cached on disk. It checks if
- * cache is enabled, image has SPIRV type and matches thresholds. */
+ * cache is enabled, image has supported format and matches thresholds. */
 bool PersistentDeviceCodeCache::isImageCached(const RTDeviceBinaryImage &Img) {
-  // Cache shoould be enabled and image type should be SPIR-V
-  if (!isEnabled() || Img.getFormat() != PI_DEVICE_BINARY_TYPE_SPIRV)
+  // Cache should be enabled and image type is one of the supported formats.
+  if (!isEnabled() || !IsSupportedImageFormat(Img.getFormat()))
     return false;
 
   // Disable cache for ITT-profiled images.

sycl/source/detail/pi.cpp

@@ -684,22 +684,54 @@ DeviceBinaryImage::getProperty(const char *PropName) const {
   return *It;
 }
 
+// Returns the e_type field from an ELF image.
+static uint16_t getELFHeaderType(const unsigned char *ImgData, size_t ImgSize) {
+  bool IsBigEndian = ImgData[5] == 2;
+  if (IsBigEndian)
+    return (static_cast<uint16_t>(ImgData[16]) << 8) |
+           static_cast<uint16_t>(ImgData[17]);
+  uint16_t HdrType = 0;
+  std::copy(ImgData + 16, ImgData + 16 + sizeof(HdrType),
+            reinterpret_cast<char *>(&HdrType));
+  return HdrType;
+}
+
 RT::PiDeviceBinaryType getBinaryImageFormat(const unsigned char *ImgData,
                                             size_t ImgSize) {
+  // Top-level magic numbers for the recognized binary image formats.
   struct {
     RT::PiDeviceBinaryType Fmt;
     const uint32_t Magic;
   } Fmts[] = {{PI_DEVICE_BINARY_TYPE_SPIRV, 0x07230203},
-              {PI_DEVICE_BINARY_TYPE_LLVMIR_BITCODE, 0xDEC04342}};
+              {PI_DEVICE_BINARY_TYPE_LLVMIR_BITCODE, 0xDEC04342},
+              // 'I', 'N', 'T', 'C' ; Intel native
+              {PI_DEVICE_BINARY_TYPE_NATIVE, 0x43544E49}};
 
   if (ImgSize >= sizeof(Fmts[0].Magic)) {
     detail::remove_const_t<decltype(Fmts[0].Magic)> Hdr = 0;
     std::copy(ImgData, ImgData + sizeof(Hdr), reinterpret_cast<char *>(&Hdr));
 
+    // Check headers for direct formats.
     for (const auto &Fmt : Fmts) {
       if (Hdr == Fmt.Magic)
         return Fmt.Fmt;
     }
+
+    // ELF e_type for recognized binary image formats.
+    struct {
+      RT::PiDeviceBinaryType Fmt;
+      const uint16_t Magic;
+    } ELFFmts[] = {{PI_DEVICE_BINARY_TYPE_NATIVE, 0xFF04}}; // OpenCL executable
+
+    // ELF files need to be parsed separately. The header type ends after 18
+    // bytes.
+    if (Hdr == 0x464c457F && ImgSize >= 18) {
+      uint16_t HdrType = getELFHeaderType(ImgData, ImgSize);
+      for (const auto &ELFFmt : ELFFmts) {
+        if (HdrType == ELFFmt.Magic)
+          return ELFFmt.Fmt;
+      }
+    }
   }
   return PI_DEVICE_BINARY_TYPE_NONE;
 }

sycl/unittests/kernel-and-program/PersistentDeviceCodeCache.cpp

@@ -1,4 +1,4 @@
-//==----- PersistenDeviceCodeCache.cpp --- Persistent cache tests ----------==//
+//==----- PersistentDeviceCodeCache.cpp --- Persistent cache tests ---------==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
@@ -78,7 +78,8 @@ static pi_result redefinedProgramGetInfo(pi_program program,
   return PI_SUCCESS;
 }
 
-class PersistenDeviceCodeCache : public ::testing::Test {
+class PersistentDeviceCodeCache
+    : public ::testing::TestWithParam<pi_device_binary_type> {
 public:
 #ifdef _WIN32
   int setenv(const char *name, const char *value, int overwrite) {
@@ -123,6 +124,8 @@ class PersistenDeviceCodeCache : public ::testing::Test {
     EXPECT_NE(getenv("SYCL_CACHE_DIR"), nullptr)
         << "Please set SYCL_CACHE_DIR environment variable pointing to cache "
            "location.";
+    // Set binary format from parameter.
+    BinStruct.Format = GetParam();
   }
 
   virtual void TearDown() {
@@ -133,7 +136,7 @@ class PersistenDeviceCodeCache : public ::testing::Test {
                                     : nullptr);
   }
 
-  PersistenDeviceCodeCache() : Plt{default_selector()} {
+  PersistentDeviceCodeCache() : Plt{default_selector()} {
 
     if (Plt.is_host() || Plt.get_backend() != backend::opencl) {
       std::clog << "This test is only supported on OpenCL devices\n";
@@ -199,9 +202,10 @@ class PersistenDeviceCodeCache : public ::testing::Test {
   detail::OSModuleHandle ModuleHandle = detail::OSUtil::ExeModuleHandle;
   platform Plt;
   device Dev;
+  // NOTE: Format is a parameter of the test so use none and set in SetUp.
   pi_device_binary_struct BinStruct{/*Version*/ 1,
                                     /*Kind*/ 4,
-                                    /*Format*/ PI_DEVICE_BINARY_TYPE_SPIRV,
+                                    /*Format*/ PI_DEVICE_BINARY_TYPE_NONE,
                                     /*DeviceTargetSpec*/ nullptr,
                                     /*CompileOptions*/ nullptr,
                                     /*LinkOptions*/ nullptr,
@@ -221,7 +225,7 @@ class PersistenDeviceCodeCache : public ::testing::Test {
 
 /* Checks that key values with \0 symbols are processed correctly
  */
-TEST_F(PersistenDeviceCodeCache, KeysWithNullTermSymbol) {
+TEST_P(PersistentDeviceCodeCache, KeysWithNullTermSymbol) {
   if (Plt.is_host() || Plt.get_backend() != backend::opencl) {
     return;
   }
@@ -254,21 +258,21 @@ TEST_F(PersistenDeviceCodeCache, KeysWithNullTermSymbol) {
 /* Do read/write for the same cache item to/from 300 threads for small device
  * code size. Make sure that there is no data corruption or crashes.
  */
-TEST_F(PersistenDeviceCodeCache, ConcurentReadWriteSmallItem) {
+TEST_P(PersistentDeviceCodeCache, ConcurentReadWriteSmallItem) {
   ConcurentReadWriteCache(0, 300);
 }
 
 /* Do read/write for the same cache item to/from 100 threads for medium device
  * code size. Make sure that there is no data corruption or crashes.
  */
-TEST_F(PersistenDeviceCodeCache, ConcurentReadWriteCacheMediumItem) {
+TEST_P(PersistentDeviceCodeCache, ConcurentReadWriteCacheMediumItem) {
   ConcurentReadWriteCache(1, 100);
 }
 
 /* Do read/write for the same cache item to/from 20 threads from big device
  * code size. Make sure that there is no data corruption or crashes.
  */
-TEST_F(PersistenDeviceCodeCache, ConcurentReadWriteCacheBigItem) {
+TEST_P(PersistentDeviceCodeCache, ConcurentReadWriteCacheBigItem) {
   ConcurentReadWriteCache(2, 20);
 }
 
@@ -279,7 +283,7 @@ TEST_F(PersistenDeviceCodeCache, ConcurentReadWriteCacheBigItem) {
  *  - source file is corrupted;
  *  - binary file is corrupted.
  */
-TEST_F(PersistenDeviceCodeCache, CorruptedCacheFiles) {
+TEST_P(PersistentDeviceCodeCache, CorruptedCacheFiles) {
   if (Plt.is_host() || Plt.get_backend() != backend::opencl) {
     return;
   }
@@ -349,7 +353,7 @@ TEST_F(PersistenDeviceCodeCache, CorruptedCacheFiles) {
  *  - new cache item is created if existing one is locked on write operation;
  *  - cache miss happens on read operation.
  */
-TEST_F(PersistenDeviceCodeCache, LockFile) {
+TEST_P(PersistentDeviceCodeCache, LockFile) {
   if (Plt.is_host() || Plt.get_backend() != backend::opencl) {
     return;
   }
@@ -405,7 +409,7 @@ TEST_F(PersistenDeviceCodeCache, LockFile) {
 // llvm::sys::fs::setPermissions does not make effect on Windows
 /* Checks cache behavior when filesystem read/write operations fail
  */
-TEST_F(PersistenDeviceCodeCache, AccessDeniedForCacheDir) {
+TEST_P(PersistentDeviceCodeCache, AccessDeniedForCacheDir) {
   if (Plt.is_host() || Plt.get_backend() != backend::opencl) {
     return;
   }
@@ -452,4 +456,9 @@ TEST_F(PersistenDeviceCodeCache, AccessDeniedForCacheDir) {
   ASSERT_NO_ERROR(llvm::sys::fs::remove_directories(ItemDir));
 }
 #endif //_WIN32
+
+INSTANTIATE_TEST_SUITE_P(PersistentDeviceCodeCacheImpl,
+                         PersistentDeviceCodeCache,
+                         ::testing::Values(PI_DEVICE_BINARY_TYPE_SPIRV,
+                                           PI_DEVICE_BINARY_TYPE_NATIVE));
 } // namespace