disable OpenCL prefetch to L1 cache

Disables prefetch into L1 for standard builtin prefetch, as out of bound
address will generate page fault (only prefetch to L3 supports out of
bound).

Author: pkwasnie-intel

IGC/BiFModule/Implementation/prefetch.cl

@@ -9,37 +9,40 @@ SPDX-License-Identifier: MIT
 #include "IGCBiF_Intrinsics_Lsc.cl"
 
 extern __constant int __UseLSC;
+extern __constant int __ForceL1Prefetch;
 
 // Mapping from OpenCL prefetch to LSC prefetch. Uniform immediate offset can
 // be used to save on base pointer arithmetics, but offset can't be variable.
-#define LSC_PREFETCH(type, p, num_elements)                                      \
-    if (num_elements == 1)                                                       \
-    {                                                                            \
-        __builtin_IB_lsc_prefetch_global_##type(p, 0, LSC_LDCC_L1C_L3C);         \
-    }                                                                            \
-    else if (num_elements == 2)                                                  \
-    {                                                                            \
-        __builtin_IB_lsc_prefetch_global_##type(p, 0, LSC_LDCC_L1C_L3C);         \
-        __builtin_IB_lsc_prefetch_global_##type(p, 1, LSC_LDCC_L1C_L3C);         \
-    }                                                                            \
-    else if (num_elements == 3)                                                  \
-    {                                                                            \
-        __builtin_IB_lsc_prefetch_global_##type(p, 0, LSC_LDCC_L1C_L3C);         \
-        __builtin_IB_lsc_prefetch_global_##type(p, 1, LSC_LDCC_L1C_L3C);         \
-        __builtin_IB_lsc_prefetch_global_##type(p, 2, LSC_LDCC_L1C_L3C);         \
-    }                                                                            \
-    else if (num_elements == 4)                                                  \
-    {                                                                            \
-        __builtin_IB_lsc_prefetch_global_##type(p, 0, LSC_LDCC_L1C_L3C);         \
-        __builtin_IB_lsc_prefetch_global_##type(p, 1, LSC_LDCC_L1C_L3C);         \
-        __builtin_IB_lsc_prefetch_global_##type(p, 2, LSC_LDCC_L1C_L3C);         \
-        __builtin_IB_lsc_prefetch_global_##type(p, 3, LSC_LDCC_L1C_L3C);         \
-    }                                                                            \
-    else                                                                         \
-    {                                                                            \
-        for (int i = 0; i < num_elements; ++i)                                   \
-            __builtin_IB_lsc_prefetch_global_##type(p + i, 0, LSC_LDCC_L1C_L3C); \
-    }                                                                            \
+#define LSC_PREFETCH(type, p, num_elements)                                            \
+    /* Warning: out of bound L1 prefetch will generate page fault */                   \
+    enum LSC_LDCC cacheOpt = __ForceL1Prefetch ? LSC_LDCC_L1C_L3C : LSC_LDCC_L1UC_L3C; \
+    if (num_elements == 1)                                                             \
+    {                                                                                  \
+        __builtin_IB_lsc_prefetch_global_##type(p, 0, cacheOpt);                       \
+    }                                                                                  \
+    else if (num_elements == 2)                                                        \
+    {                                                                                  \
+        __builtin_IB_lsc_prefetch_global_##type(p, 0, cacheOpt);                       \
+        __builtin_IB_lsc_prefetch_global_##type(p, 1, cacheOpt);                       \
+    }                                                                                  \
+    else if (num_elements == 3)                                                        \
+    {                                                                                  \
+        __builtin_IB_lsc_prefetch_global_##type(p, 0, cacheOpt);                       \
+        __builtin_IB_lsc_prefetch_global_##type(p, 1, cacheOpt);                       \
+        __builtin_IB_lsc_prefetch_global_##type(p, 2, cacheOpt);                       \
+    }                                                                                  \
+    else if (num_elements == 4)                                                        \
+    {                                                                                  \
+        __builtin_IB_lsc_prefetch_global_##type(p, 0, cacheOpt);                       \
+        __builtin_IB_lsc_prefetch_global_##type(p, 1, cacheOpt);                       \
+        __builtin_IB_lsc_prefetch_global_##type(p, 2, cacheOpt);                       \
+        __builtin_IB_lsc_prefetch_global_##type(p, 3, cacheOpt);                       \
+    }                                                                                  \
+    else                                                                               \
+    {                                                                                  \
+        for (int i = 0; i < num_elements; ++i)                                         \
+            __builtin_IB_lsc_prefetch_global_##type(p + i, 0, cacheOpt);               \
+    }                                                                                  \
 
 //Prefetch function
 

IGC/Compiler/Optimizer/BuiltInFuncImport.cpp

@@ -1047,6 +1047,7 @@ void BIImport::InitializeBIFlags(Module& M)
         useHighAccuracyMathFuncs = OCLContext->m_InternalOptions.UseHighAccuracyMathFuncs;
 
         initializeVarWithValue("__UseLSC", pCtx->platform.hasLSC());
+        initializeVarWithValue("__ForceL1Prefetch", IGC_IS_FLAG_ENABLED(ForcePrefetchToL1Cache) ? 1 : 0);
     }
 
     initializeVarWithValue("__EnableSWSrgbWrites", IGC_GET_FLAG_VALUE(cl_khr_srgb_image_writes));

IGC/common/igc_flags.h

@@ -526,6 +526,7 @@ DECLARE_IGC_REGKEY(bool, LateInlineUnmaskedFunc,        false, "Postpone inlinin
 DECLARE_IGC_REGKEY(bool, ForceFormatConversionDG2Plus,  false,
     "Forces SW image format conversion for R10G10B10A2_UNORM, R11G11B10_FLOAT, R10G10B10A2_UINT image formats on DG2+ platforms", true)
 DECLARE_IGC_REGKEY(bool, EnableDivergentBarrierWA, false, "Generate continuation code to handle shaders that places barriers in divergent control flow", false)
+DECLARE_IGC_REGKEY(bool, ForcePrefetchToL1Cache, false, "Forces standard builtin prefetch to use L1 cache", false)
 
 DECLARE_IGC_GROUP("Performance experiments")
 DECLARE_IGC_REGKEY(bool, ForceNonCoherentStatelessBTI,  false, "Enable gneeration of non cache coherent stateless messages", false)