feature: set device allocation chunking as default

Device allocation chunking only applies for multi-tile mode for implicit scaling

Related-To: NEO-9051

Signed-off-by: John Falkowski <john.falkowski@intel.com>
Source: f0175b3

Author: johntfalk

shared/source/debug_settings/debug_variables_base.inl

@@ -524,7 +524,7 @@ DECLARE_DEBUG_VARIABLE(bool, PrintBOChunkingLogs, false, "Print some logs on BO
 DECLARE_DEBUG_VARIABLE(bool, EnableBOChunkingPrefetch, false, "Enables prefetching of Shared Memory chunks")
 DECLARE_DEBUG_VARIABLE(bool, EnableBOChunkingDevMemPrefetch, false, "Enables prefetching of Device Memory chunks")
 DECLARE_DEBUG_VARIABLE(bool, EnableBOChunkingPreferredLocationHint, false, "Enables preferred location advise on chunks")
-DECLARE_DEBUG_VARIABLE(int32_t, EnableBOChunking, 0, "Enables use of chunking of BOs in the KMD, mask: 0 = no chunking, 1 = shared allocations only, 2 = device allocations only, 3 = shared and device allocations .")
+DECLARE_DEBUG_VARIABLE(int32_t, EnableBOChunking, -1, "Enables use of chunking of BOs in the KMD, mask: -1 = default, 0 = no chunking, 1 = shared allocations only, 2 = multi-tile device allocations only, 3 = shared and multi-tile device allocations .")
 DECLARE_DEBUG_VARIABLE(int32_t, NumberOfBOChunks, 2, "Number of chunks to use. Must be a power of two)")
 DECLARE_DEBUG_VARIABLE(int32_t, MinimalAllocationSizeForChunking, -1, "2097152: default, >0: size in B. Minimal size an allocation should have to use chunking.")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceAutoGrfCompilationMode, -1, "Adds build option -*-intel-enable-auto-large-GRF-mode to force kernel compilation")

shared/source/os_interface/linux/drm_memory_manager.cpp

@@ -2288,7 +2288,7 @@ GraphicsAllocation *DrmMemoryManager::createSharedUnifiedMemoryAllocation(const
 
     // Dont chunk for sizes less than chunkThreshold or if debugging is enabled
     if (!executionEnvironment.isDebuggingEnabled() &&
-        (drm.getChunkingMode() & 0x01) &&
+        (drm.getChunkingMode() & chunkingModeShared) &&
         !(chunkingSize & (MemoryConstants::chunkThreshold - 1)) &&
         size >= drm.getMinimalSizeForChunking()) {
         numHandles = 1;

shared/source/os_interface/linux/drm_neo.cpp

@@ -1157,7 +1157,11 @@ bool Drm::isChunkingAvailable() {
             int ret = ioctlHelper->isChunkingAvailable();
             if (ret) {
                 chunkingAvailable = true;
-                chunkingMode = DebugManager.flags.EnableBOChunking.get();
+                if (DebugManager.flags.EnableBOChunking.get() == -1) {
+                    chunkingMode = chunkingModeDevice;
+                } else {
+                    chunkingMode = DebugManager.flags.EnableBOChunking.get();
+                }
             }
 
             if (DebugManager.flags.MinimalAllocationSizeForChunking.get() != -1) {
@@ -1167,8 +1171,8 @@ bool Drm::isChunkingAvailable() {
             printDebugString(DebugManager.flags.PrintBOChunkingLogs.get(), stdout,
                              "Chunking available: %d; enabled for: shared allocations %d, device allocations %d; minimalChunkingSize: %zd\n",
                              chunkingAvailable,
-                             (chunkingMode & 0x01),
-                             (chunkingMode & 0x02),
+                             (chunkingMode & chunkingModeShared),
+                             (chunkingMode & chunkingModeDevice),
                              minimalChunkingSize);
         });
     }

shared/source/os_interface/linux/drm_neo.h

@@ -35,6 +35,8 @@ enum EngineType : uint32_t;
 
 namespace NEO {
 constexpr uint32_t contextPrivateParamBoost = 0x80000000;
+constexpr uint32_t chunkingModeShared = 1;
+constexpr uint32_t chunkingModeDevice = 2;
 
 enum class AllocationType;
 enum class CachePolicy : uint32_t;

shared/test/common/test_files/igdrcl.config

@@ -522,7 +522,7 @@ EnableBOChunkingPrefetch = 0
 EnableBOChunkingDevMemPrefetch = 0
 EnableBOChunkingPreferredLocationHint = 0
 NumberOfBOChunks = 2
-EnableBOChunking = 0
+EnableBOChunking = -1
 MinimalAllocationSizeForChunking = -1
 DirectSubmissionControllerMaxTimeout = -1
 ExitOnSubmissionNumber = -1

shared/test/unit_test/os_interface/linux/drm_residency_handler_prelim_tests.cpp

@@ -1483,10 +1483,34 @@ TEST(DrmSetPairTests, whenQueryingForSetPairAvailableAndNoDebugKeyThenFalseIsRet
     EXPECT_EQ(0u, drm.context.setPairQueryCalled);
 }
 
-TEST(DrmChunkingTests, whenQueryingForChunkingAvailableAndNoDebugKeyThenFalseIsReturned) {
+TEST(DrmChunkingTests, whenQueryingForChunkingAvailableAndDefaultDebugVariableThenTrueIsReturned) {
+    DebugManagerStateRestore restorer;
+    DebugManager.flags.UseKmdMigration.set(1);
+
+    auto executionEnvironment = std::make_unique<MockExecutionEnvironment>();
+    DrmQueryMock drm{*executionEnvironment->rootDeviceEnvironments[0]};
+
+    drm.context.chunkingQueryValue = 1;
+    drm.context.chunkingQueryReturn = 0;
+    EXPECT_FALSE(drm.chunkingAvailable);
+
+    EXPECT_EQ(0u, drm.context.chunkingQueryCalled);
+    drm.callBaseIsChunkingAvailable = true;
+    EXPECT_TRUE(drm.isChunkingAvailable());
+    EXPECT_TRUE(drm.getChunkingAvailable());
+    EXPECT_EQ(1u, drm.context.chunkingQueryCalled);
+    EXPECT_EQ(2u, drm.getChunkingMode());
+}
+
+TEST(DrmChunkingTests, whenQueryingForChunkingAvailableAndDisableDebugVariableThenFalseIsReturned) {
+    DebugManagerStateRestore restorer;
+    DebugManager.flags.EnableBOChunking.set(0);
+    DebugManager.flags.UseKmdMigration.set(1);
+
     auto executionEnvironment = std::make_unique<MockExecutionEnvironment>();
     DrmQueryMock drm{*executionEnvironment->rootDeviceEnvironments[0]};
-    drm.context.chunkingQueryValue = 0;
+
+    drm.context.chunkingQueryValue = 1;
     drm.context.chunkingQueryReturn = 0;
     EXPECT_FALSE(drm.chunkingAvailable);