[SYCL] Add Experimental Range Rounding (#12690)

hdelan · web-flow · commit 319cf0ee1697 · 2024-04-01T11:58:09.000+01:00
This commit adds new clang command line option
-fsycl-exp-range-rounding. Experimental range rounding maps all 1, 2 or
3 dim range kernels to range rounded kernels which can be rounded in all
dims.

`SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS=x:y:z` can be used for exp
range rounding, where the `y` param will be a factor of the rounded
range in all dims
diff --git a/clang/include/clang/Basic/LangOptions.def b/clang/include/clang/Basic/LangOptions.def
@@ -303,6 +303,7 @@ LANGOPT(
 ENUM_LANGOPT(SYCLRangeRounding, SYCLRangeRoundingPreference, 2,
     SYCLRangeRoundingPreference::On,
     "Preference for SYCL parallel_for range rounding")
+LANGOPT(SYCLExperimentalRangeRounding, 1, 0, "Use experimental parallel for range rounding")
 LANGOPT(SYCLEnableIntHeaderDiags, 1, 0, "Enable diagnostics that require the "
         "SYCL integration header")
 LANGOPT(SYCLAllowVirtualFunctions, 1, 0,
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
@@ -4012,6 +4012,10 @@ def fsycl_disable_range_rounding : Flag<["-"], "fsycl-disable-range-rounding">,
   Alias<fsycl_range_rounding_EQ>, AliasArgs<["disable"]>,
   HelpText<"Deprecated: please use -fsycl-range-rounding=disable instead.">,
   Flags<[Deprecated]>;
+def fsycl_exp_range_rounding : Flag<["-"], "fsycl-exp-range-rounding">,
+  Visibility<[ClangOption, CLOption, DXCOption, CC1Option]>,
+  HelpText<"Use experimental range rounding.">,
+  MarshallingInfoFlag<LangOpts<"SYCLExperimentalRangeRounding">>;
 def fno_sycl_use_footer : Flag<["-"], "fno-sycl-use-footer">, Visibility<[ClangOption, CLOption, DXCOption]>,
   HelpText<"Disable usage of the integration footer during SYCL enabled "
    "compilations.">;
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -5430,6 +5430,9 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
     if (Arg *A = Args.getLastArg(options::OPT_fsycl_range_rounding_EQ))
       A->render(Args, CmdArgs);
 
+    if (Arg *A = Args.getLastArg(options::OPT_fsycl_exp_range_rounding))
+      A->render(Args, CmdArgs);
+
     // Add the Unique ID prefix
     StringRef UniqueID = D.getSYCLUniqueID(Input.getBaseInput());
     if (!UniqueID.empty())
diff --git a/clang/lib/Frontend/InitPreprocessor.cpp b/clang/lib/Frontend/InitPreprocessor.cpp
@@ -589,6 +589,12 @@ static void InitializeStandardPredefinedMacros(const TargetInfo &TI,
     default:
       break;
     }
+
+    // Set __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__ macro for
+    // both host and device compilations if -fsycl-exp-range-rounding
+    // flag is used.
+    if (LangOpts.SYCLExperimentalRangeRounding)
+      Builder.defineMacro("__SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__");
   }
 
   if (LangOpts.DeclareSPIRVBuiltins) {
diff --git a/clang/lib/Sema/SemaSYCL.cpp b/clang/lib/Sema/SemaSYCL.cpp
@@ -5187,6 +5187,12 @@ void SYCLIntegrationHeader::emit(raw_ostream &O) {
     break;
   }
 
+  if (S.getLangOpts().SYCLExperimentalRangeRounding) {
+    O << "#ifndef __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__ \n";
+    O << "#define __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__ 1\n";
+    O << "#endif //__SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__\n\n";
+  }
+
   if (SpecConsts.size() > 0) {
     O << "// Forward declarations of templated spec constant types:\n";
     for (const auto &SC : SpecConsts)
diff --git a/clang/test/Preprocessor/predefined-macros.c b/clang/test/Preprocessor/predefined-macros.c
@@ -339,6 +339,34 @@
 // CHECK-FORCE-RANGE: #define __SYCL_FORCE_PARALLEL_FOR_RANGE_ROUNDING__ 1
 // CHECK-FORCE-NO-RANGE-NOT: #define __SYCL_FORCE_PARALLEL_FOR_RANGE_ROUNDING__ 1
 
+// RUN: %clang_cc1 %s -E -dM -fsycl-is-device \
+// RUN:   -triple spir64-unknown-unknown -fsycl-exp-range-rounding -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-RANGE
+
+// RUN: %clang_cc1 %s -E -dM -fsycl-is-device \
+// RUN:   -triple spir64_fpga-unknown-unknown -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-NO-RANGE
+
+// RUN: %clang_cc1 %s -E -dM -fsycl-is-device -fsycl-exp-range-rounding \
+// RUN:   -triple spir64_fpga-unknown-unknown -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-RANGE
+
+// RUN: %clang_cc1 %s -E -dM -fsycl-is-device -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-NO-RANGE
+
+// RUN: %clang_cc1 %s -E -dM -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-NO-RANGE
+
+// RUN: %clang_cc1 %s -E -dM -fsycl-is-host \
+// RUN: -triple x86_64-unknown-linux-gnu -fsycl-exp-range-rounding -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-RANGE
+
+// RUN: %clang_cc1 %s -E -dM -fsycl-is-host -o - \
+// RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-EXP-NO-RANGE
+
+// CHECK-EXP-RANGE: #define __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__ 1
+// CHECK-EXP-NO-RANGE-NOT: #define __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__ 1
+
 // RUN: %clang_cc1 %s -E -dM -o - -x hip -triple x86_64-unknown-linux-gnu \
 // RUN:   -fgpu-default-stream=per-thread \
 // RUN:   | FileCheck -match-full-lines %s --check-prefix=CHECK-PTH
diff --git a/sycl/doc/design/ParallelForRangeRounding.md b/sycl/doc/design/ParallelForRangeRounding.md
@@ -47,3 +47,35 @@ the range rounded kernel, `-fsycl-range-rounding=force`. The user can also tell
 the SYCL implementation to only produce the unrounded kernel using the flag
 `-fsycl-range-rounding=disable`. By default both kernels will be generated,
 which is equivalent to `-fsycl-range-rounding=on`.
+
+## Experimental Range Rounding
+
+Experimental range rounding will perform rounding in all dimensions.
+Experimental range rounding can be set using the `-fsycl-exp-range-rounding`
+flag.
+
+Some oddly shaped ranges and how they might round:
+
+```
+{43} -> {64}
+{43, 79} -> {64, 96}
+{43, 79, 7} -> {64, 96, 8}
+```
+
+The user can specify the factor that they want the rounded range to be a
+multiple of in all dimensions using the
+`SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS` environment variable. When
+experimental range rounding is used, only the middle value in
+`SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS` is used.
+If `SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS` is set to `1:256:1`, the rounded
+range will divide `256` in all dimensions
+
+```
+{43} -> {256}
+{43, 257} -> {256, 512}
+{43, 257, 7} -> {256, 512, 256}
+```
+
+`-fsycl-range-rounding=disable` will override `-fsycl-exp-range-rounding`. If
+both are used in conjunction then no range rounding will happen.
+
diff --git a/sycl/include/sycl/handler.hpp b/sycl/include/sycl/handler.hpp
@@ -1241,6 +1241,29 @@ class __SYCL_EXPORT handler {
       DidAdjust = true;
     };
 
+#ifdef __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__
+    size_t GoodExpFactor = 1;
+    switch (Dims) {
+    case 1:
+      GoodExpFactor = 32; // Make global range multiple of {32}
+      break;
+    case 2:
+      GoodExpFactor = 16; // Make global range multiple of {16, 16}
+      break;
+    case 3:
+      GoodExpFactor = 8; // Make global range multiple of {8, 8, 8}
+      break;
+    }
+
+    // Check if rounding parameters have been set through environment:
+    // SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS=MinRound:PreferredRound:MinRange
+    this->GetRangeRoundingSettings(MinFactorX, GoodExpFactor, MinRangeX);
+
+    for (auto i = 0; i < Dims; ++i)
+      if (UserRange[i] % GoodExpFactor) {
+        Adjust(i, ((UserRange[i] / GoodExpFactor) + 1) * GoodExpFactor);
+      }
+#else
     // Perform range rounding if there are sufficient work-items to
     // need rounding and the user-specified range is not a multiple of
     // a "good" value.
@@ -1251,6 +1274,7 @@ class __SYCL_EXPORT handler {
       // will yield a rounded-up value for the total range.
       Adjust(0, ((RoundedRange[0] + GoodFactor - 1) / GoodFactor) * GoodFactor);
     }
+#endif // __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__
 #ifdef __SYCL_FORCE_PARALLEL_FOR_RANGE_ROUNDING__
     // If we are forcing range rounding kernels to be used, we always want the
     // rounded range kernel to be generated, even if rounding isn't needed
diff --git a/sycl/test-e2e/Basic/parallel_for_range_roundup.cpp b/sycl/test-e2e/Basic/parallel_for_range_roundup.cpp
@@ -5,6 +5,12 @@
 // RUN: %{build} -fsycl-range-rounding=force -o %t.out
 // RUN: env SYCL_PARALLEL_FOR_RANGE_ROUNDING_TRACE=1 %{run} %t.out | FileCheck %s --check-prefix=CHECK-DEFAULT
 
+// RUN: %{build} -fsycl-exp-range-rounding -o %t.out
+// RUN: env SYCL_PARALLEL_FOR_RANGE_ROUNDING_TRACE=1 %{run} %t.out | FileCheck %s --check-prefix=CHECK-EXP
+
+// RUN: %{build} -fsycl-range-rounding=force -fsycl-exp-range-rounding -o %t.out
+// RUN: env SYCL_PARALLEL_FOR_RANGE_ROUNDING_TRACE=1 %{run} %t.out | FileCheck %s --check-prefix=CHECK-EXP
+//
 // These tests test 3 things:
 //
 // 1. The user range is the same as the in kernel range (using BufRange) as
@@ -243,3 +249,59 @@ int main() {
 // CHECK-DEFAULT-NEXT:  Counter = 540672
 // CHECK-DEFAULT-NEXT:  Correct kernel indexes used
 // CHECK-DEFAULT-NEXT:  Counter = 540672
+
+// CHECK-EXP:       parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  Size seen by user at Dim 0 = 1500
+// CHECK-EXP-NEXT:  Counter = 1500
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  Counter = 1500
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 48
+// CHECK-EXP-NEXT:  Size seen by user at Dim 0 = 1500
+// CHECK-EXP-NEXT:  Size seen by user at Dim 1 = 33
+// CHECK-EXP-NEXT:  Counter = 49500
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 48
+// CHECK-EXP-NEXT:  Counter = 49500
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 40
+// CHECK-EXP-NEXT:  Size seen by user at Dim 0 = 1500
+// CHECK-EXP-NEXT:  Size seen by user at Dim 1 = 33
+// CHECK-EXP-NEXT:  Size seen by user at Dim 2 = 64
+// CHECK-EXP-NEXT:  Counter = 3168000
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 40
+// CHECK-EXP-NEXT:  Counter = 3168000
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 0 from 1500 to 1504
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 40
+// CHECK-EXP-NEXT:  Counter = 3168000
+// CHECK-EXP-NEXT:  Size seen by user at Dim 0 = 256
+// CHECK-EXP-NEXT:  Counter = 256
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  Counter = 256
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 48
+// CHECK-EXP-NEXT:  Size seen by user at Dim 0 = 256
+// CHECK-EXP-NEXT:  Size seen by user at Dim 1 = 33
+// CHECK-EXP-NEXT:  Counter = 8448
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 48
+// CHECK-EXP-NEXT:  Counter = 8448
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 40
+// CHECK-EXP-NEXT:  Size seen by user at Dim 0 = 256
+// CHECK-EXP-NEXT:  Size seen by user at Dim 1 = 33
+// CHECK-EXP-NEXT:  Size seen by user at Dim 2 = 64
+// CHECK-EXP-NEXT:  Counter = 540672
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 40
+// CHECK-EXP-NEXT:  Counter = 540672
+// CHECK-EXP-NEXT:  Correct kernel indexes used
+// CHECK-EXP-NEXT:  parallel_for range adjusted at dim 1 from 33 to 40
+// CHECK-EXP-NEXT:  Counter = 540672
diff --git a/sycl/test-e2e/PerformanceTests/ParallelFor/parallel_for_range_roundup.cpp b/sycl/test-e2e/PerformanceTests/ParallelFor/parallel_for_range_roundup.cpp
@@ -0,0 +1,66 @@
+// RUN: echo "Running parallel_for benchmark without range rounding"
+// RUN: %{build} -fsycl-range-rounding=disable -o %t.out
+// RUN: %{run} %t.out
+
+// RUN: echo "Running parallel_for benchmark with normal range rounding"
+// RUN: %{build} -fsycl-range-rounding=force -o %t.out
+// RUN: %{run} %t.out
+
+// RUN: echo "Running parallel_for benchmark with experimental range rounding"
+// RUN: %{build} -fsycl-exp-range-rounding -fsycl-range-rounding=force -o %t.out
+// RUN: %{run} %t.out
+
+#include <chrono>
+#include <iostream>
+#include <sycl/sycl.hpp>
+
+class FillData;
+class Compute;
+
+int main() {
+  constexpr static size_t width{788};
+  constexpr static size_t height{1888};
+  constexpr static size_t N{width * height};
+  constexpr static size_t iterations{1000};
+
+  sycl::queue q{};
+  float *A{sycl::malloc_device<float>(N, q)};
+  float *B{sycl::malloc_device<float>(N, q)};
+  float *C{sycl::malloc_device<float>(N, q)};
+
+  q.submit([&](sycl::handler &cgh) {
+     cgh.parallel_for<FillData>(
+         sycl::range<2>{height, width}, [=](sycl::id<2> id) {
+           unsigned int row{static_cast<unsigned int>(id[0])};
+           unsigned int col{static_cast<unsigned int>(id[1])};
+           unsigned int ix{row * static_cast<unsigned int>(width) + col};
+           A[ix] = id[0];
+           B[ix] = id[1];
+         });
+   }).wait_and_throw();
+
+  auto start{std::chrono::steady_clock::now()};
+  for (size_t i{0}; i < iterations; ++i) {
+    q.submit([&](sycl::handler &cgh) {
+       cgh.parallel_for<Compute>(
+           sycl::range<2>{height, width}, [=](sycl::id<2> id) {
+             unsigned int row{static_cast<unsigned int>(id[0])};
+             unsigned int col{static_cast<unsigned int>(id[1])};
+             unsigned int ix{row * static_cast<unsigned int>(width) + col};
+             if (ix >= static_cast<unsigned int>(N)) {
+               return;
+             }
+             if (A[ix] > B[ix]) {
+               C[ix] = A[ix];
+             } else {
+               C[ix] = B[ix];
+             }
+           });
+     }).wait_and_throw();
+  }
+  auto end{std::chrono::steady_clock::now()};
+  std::cout << std::chrono::duration_cast<std::chrono::milliseconds>(end -
+                                                                     start)
+                   .count()
+            << " ms" << std::endl;
+}

Original file line number	Diff line number	Diff line change
`@@ -589,6 +589,12 @@ static void InitializeStandardPredefinedMacros(const TargetInfo &TI,`
`589`	`589`	`default:`
`590`	`590`	`break;`
`591`	`591`	`}`
	`592`	`+`
	`593`	`+ // Set __SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__ macro for`
	`594`	`+ // both host and device compilations if -fsycl-exp-range-rounding`
	`595`	`+ // flag is used.`
	`596`	`+ if (LangOpts.SYCLExperimentalRangeRounding)`
	`597`	`+ Builder.defineMacro("__SYCL_EXP_PARALLEL_FOR_RANGE_ROUNDING__");`
`592`	`598`	`}`
`593`	`599`
`594`	`600`	`if (LangOpts.DeclareSPIRVBuiltins) {`