[SYCL] Add group algorithms for MUL/OR/XOR/AND operations

clang/lib/Sema/SPIRVBuiltins.td

@@ -917,11 +917,13 @@ foreach name = ["GroupBroadcast"] in {
   }
 }
 
-foreach name = ["GroupIAdd"] in {
+foreach name = ["GroupIAdd", "GroupNonUniformIMul", "GroupNonUniformBitwiseOr",
+                "GroupNonUniformBitwiseXor", "GroupNonUniformBitwiseAnd"] in {
   def : SPVBuiltin<name, [AIGenTypeN, UInt, UInt, AIGenTypeN], Attr.Convergent>;
 }
 
-foreach name = ["GroupFAdd", "GroupFMin", "GroupFMax"] in {
+foreach name = ["GroupFAdd", "GroupFMin", "GroupFMax",
+                "GroupNonUniformFMul"] in {
   def : SPVBuiltin<name, [FGenTypeN, UInt, UInt, FGenTypeN], Attr.Convergent>;
 }
 

sycl/include/CL/sycl/ONEAPI/functional.hpp

@@ -53,6 +53,7 @@ template <> struct maximum<void> {
 #endif
 
 template <typename T = void> using plus = std::plus<T>;
+template <typename T = void> using multiplies = std::multiplies<T>;
 template <typename T = void> using bit_or = std::bit_or<T>;
 template <typename T = void> using bit_xor = std::bit_xor<T>;
 template <typename T = void> using bit_and = std::bit_and<T>;
@@ -103,6 +104,16 @@ __SYCL_CALC_OVERLOAD(GroupOpISigned, IAdd, ONEAPI::plus<T>)
 __SYCL_CALC_OVERLOAD(GroupOpIUnsigned, IAdd, ONEAPI::plus<T>)
 __SYCL_CALC_OVERLOAD(GroupOpFP, FAdd, ONEAPI::plus<T>)
 
+__SYCL_CALC_OVERLOAD(GroupOpISigned, NonUniformIMul, ONEAPI::multiplies<T>)
+__SYCL_CALC_OVERLOAD(GroupOpIUnsigned, NonUniformIMul, ONEAPI::multiplies<T>)
+__SYCL_CALC_OVERLOAD(GroupOpFP, NonUniformFMul, ONEAPI::multiplies<T>)
+__SYCL_CALC_OVERLOAD(GroupOpISigned, NonUniformBitwiseOr, ONEAPI::bit_or<T>)
+__SYCL_CALC_OVERLOAD(GroupOpIUnsigned, NonUniformBitwiseOr, ONEAPI::bit_or<T>)
+__SYCL_CALC_OVERLOAD(GroupOpISigned, NonUniformBitwiseXor, ONEAPI::bit_xor<T>)
+__SYCL_CALC_OVERLOAD(GroupOpIUnsigned, NonUniformBitwiseXor, ONEAPI::bit_xor<T>)
+__SYCL_CALC_OVERLOAD(GroupOpISigned, NonUniformBitwiseAnd, ONEAPI::bit_and<T>)
+__SYCL_CALC_OVERLOAD(GroupOpIUnsigned, NonUniformBitwiseAnd, ONEAPI::bit_and<T>)
+
 #undef __SYCL_CALC_OVERLOAD
 
 template <typename T, __spv::GroupOperation O, __spv::Scope::Flag S,

sycl/include/CL/sycl/ONEAPI/group_algorithm.hpp

@@ -86,17 +86,39 @@ template <typename T, typename V> struct identity<T, ONEAPI::plus<V>> {
 };
 
 template <typename T, typename V> struct identity<T, ONEAPI::minimum<V>> {
-  static constexpr T value = (std::numeric_limits<T>::max)();
+  static constexpr T value = std::numeric_limits<T>::has_infinity
+                                 ? std::numeric_limits<T>::infinity()
+                                 : (std::numeric_limits<T>::max)();
 };
 
 template <typename T, typename V> struct identity<T, ONEAPI::maximum<V>> {
-  static constexpr T value = std::numeric_limits<T>::lowest();
+  static constexpr T value =
+      std::numeric_limits<T>::has_infinity
+          ? static_cast<T>(-std::numeric_limits<T>::infinity())
+          : std::numeric_limits<T>::lowest();
+};
+
+template <typename T, typename V> struct identity<T, ONEAPI::multiplies<V>> {
+  static constexpr T value = static_cast<T>(1);
+};
+
+template <typename T, typename V> struct identity<T, ONEAPI::bit_or<V>> {
+  static constexpr T value = 0;
+};
+
+template <typename T, typename V> struct identity<T, ONEAPI::bit_xor<V>> {
+  static constexpr T value = 0;
+};
+
+template <typename T, typename V> struct identity<T, ONEAPI::bit_and<V>> {
+  static constexpr T value = ~static_cast<T>(0);
 };
 
 template <typename T>
 using native_op_list =
     type_list<ONEAPI::plus<T>, ONEAPI::bit_or<T>, ONEAPI::bit_xor<T>,
-              ONEAPI::bit_and<T>, ONEAPI::maximum<T>, ONEAPI::minimum<T>>;
+              ONEAPI::bit_and<T>, ONEAPI::maximum<T>, ONEAPI::minimum<T>,
+              ONEAPI::multiplies<T>>;
 
 template <typename T, typename BinaryOperation> struct is_native_op {
   static constexpr bool value =

sycl/test/group-algorithm/exclusive_scan.cpp

@@ -7,6 +7,14 @@
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
 
+// TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
+// That requires either adding a switch to clang (-spirv-max-version=1.3) or
+// raising the spirv version from 1.1. to 1.3 for spirv translator
+// unconditionally. Using operators specific for spirv 1.3 and higher with
+// -spirv-max-version=1.1 being set by default causes assert/check fails
+// in spirv translator.
+// RUNx: %clangxx -fsycl -fsycl-targets=%sycl_triple -DSPIRV_1_3 %s -o %t13.out
+
 #include <CL/sycl.hpp>
 #include <algorithm>
 #include <cassert>
@@ -120,10 +128,27 @@ void test(queue q, InputContainer input, OutputContainer output,
   assert(std::equal(output.begin(), output.begin() + N, expected.begin()));
 }
 
+bool isSupportedDevice(device D) {
+  std::string PlatformName = D.get_platform().get_info<info::platform::name>();
+  if (PlatformName.find("Level-Zero") != std::string::npos)
+    return true;
+
+  if (PlatformName.find("OpenCL") != std::string::npos) {
+    std::string Version = D.get_info<info::device::version>();
+    size_t Offset = Version.find("OpenCL");
+    if (Offset == std::string::npos)
+      return false;
+    Version = Version.substr(Offset + 7, 3);
+    if (Version >= std::string("2.0"))
+      return true;
+  }
+
+  return false;
+}
+
 int main() {
   queue q;
-  std::string version = q.get_device().get_info<info::device::version>();
-  if (version < std::string("2.0")) {
+  if (!isSupportedDevice(q.get_device())) {
     std::cout << "Skipping test\n";
     return 0;
   }
@@ -134,14 +159,20 @@ int main() {
   std::iota(input.begin(), input.end(), 0);
   std::fill(output.begin(), output.end(), 0);
 
-#if __cplusplus >= 201402L
   test(q, input, output, plus<>(), 0);
   test(q, input, output, minimum<>(), std::numeric_limits<int>::max());
   test(q, input, output, maximum<>(), std::numeric_limits<int>::lowest());
-#endif
+
   test(q, input, output, plus<int>(), 0);
   test(q, input, output, minimum<int>(), std::numeric_limits<int>::max());
   test(q, input, output, maximum<int>(), std::numeric_limits<int>::lowest());
 
+#ifdef SPIRV_1_3
+  test(q, input, output, multiplies<int>(), 1);
+  test(q, input, output, bit_or<int>(), 0);
+  test(q, input, output, bit_xor<int>(), 0);
+  test(q, input, output, bit_and<int>(), ~0);
+#endif // SPIRV_1_3
+
   std::cout << "Test passed." << std::endl;
 }

sycl/test/group-algorithm/inclusive_scan.cpp

@@ -7,6 +7,14 @@
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
 
+// TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
+// That requires either adding a switch to clang (-spirv-max-version=1.3) or
+// raising the spirv version from 1.1. to 1.3 for spirv translator
+// unconditionally. Using operators specific for spirv 1.3 and higher with
+// -spirv-max-version=1.1 being set by default causes assert/check fails
+// in spirv translator.
+// RUNx: %clangxx -fsycl -fsycl-targets=%sycl_triple -DSPIRV_1_3 %s -o %t13.out
+
 #include <CL/sycl.hpp>
 #include <algorithm>
 #include <cassert>
@@ -120,10 +128,27 @@ void test(queue q, InputContainer input, OutputContainer output,
   assert(std::equal(output.begin(), output.begin() + N, expected.begin()));
 }
 
+bool isSupportedDevice(device D) {
+  std::string PlatformName = D.get_platform().get_info<info::platform::name>();
+  if (PlatformName.find("Level-Zero") != std::string::npos)
+    return true;
+
+  if (PlatformName.find("OpenCL") != std::string::npos) {
+    std::string Version = D.get_info<info::device::version>();
+    size_t Offset = Version.find("OpenCL");
+    if (Offset == std::string::npos)
+      return false;
+    Version = Version.substr(Offset + 7, 3);
+    if (Version >= std::string("2.0"))
+      return true;
+  }
+
+  return false;
+}
+
 int main() {
   queue q;
-  std::string version = q.get_device().get_info<info::device::version>();
-  if (version < std::string("2.0")) {
+  if (!isSupportedDevice(q.get_device())) {
     std::cout << "Skipping test\n";
     return 0;
   }
@@ -134,14 +159,20 @@ int main() {
   std::iota(input.begin(), input.end(), 0);
   std::fill(output.begin(), output.end(), 0);
 
-#if __cplusplus >= 201402L
   test(q, input, output, plus<>(), 0);
   test(q, input, output, minimum<>(), std::numeric_limits<int>::max());
   test(q, input, output, maximum<>(), std::numeric_limits<int>::lowest());
-#endif
+
   test(q, input, output, plus<int>(), 0);
   test(q, input, output, minimum<int>(), std::numeric_limits<int>::max());
   test(q, input, output, maximum<int>(), std::numeric_limits<int>::lowest());
 
+#ifdef SPIRV_1_3
+  test(q, input, output, multiplies<int>(), 1);
+  test(q, input, output, bit_or<int>(), 0);
+  test(q, input, output, bit_xor<int>(), 0);
+  test(q, input, output, bit_and<int>(), ~0);
+#endif // SPIRV_1_3
+
   std::cout << "Test passed." << std::endl;
 }

sycl/test/group-algorithm/reduce.cpp

@@ -7,6 +7,14 @@
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
 
+// TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
+// That requires either adding a switch to clang (-spirv-max-version=1.3) or
+// raising the spirv version from 1.1. to 1.3 for spirv translator
+// unconditionally. Using operators specific for spirv 1.3 and higher with
+// -spirv-max-version=1.1 being set by default causes assert/check fails
+// in spirv translator.
+// RUNx: %clangxx -fsycl -fsycl-targets=%sycl_triple -DSPIRV_1_3 %s -o %t13.out
+
 #include <CL/sycl.hpp>
 #include <algorithm>
 #include <cassert>
@@ -58,10 +66,27 @@ void test(queue q, InputContainer input, OutputContainer output,
          std::accumulate(input.begin(), input.end(), init, binary_op));
 }
 
+bool isSupportedDevice(device D) {
+  std::string PlatformName = D.get_platform().get_info<info::platform::name>();
+  if (PlatformName.find("Level-Zero") != std::string::npos)
+    return true;
+
+  if (PlatformName.find("OpenCL") != std::string::npos) {
+    std::string Version = D.get_info<info::device::version>();
+    size_t Offset = Version.find("OpenCL");
+    if (Offset == std::string::npos)
+      return false;
+    Version = Version.substr(Offset + 7, 3);
+    if (Version >= std::string("2.0"))
+      return true;
+  }
+
+  return false;
+}
+
 int main() {
   queue q;
-  std::string version = q.get_device().get_info<info::device::version>();
-  if (version < std::string("2.0")) {
+  if (!isSupportedDevice(q.get_device())) {
     std::cout << "Skipping test\n";
     return 0;
   }
@@ -72,14 +97,20 @@ int main() {
   std::iota(input.begin(), input.end(), 0);
   std::fill(output.begin(), output.end(), 0);
 
-#if __cplusplus >= 201402L
   test(q, input, output, plus<>(), 0);
   test(q, input, output, minimum<>(), std::numeric_limits<int>::max());
   test(q, input, output, maximum<>(), std::numeric_limits<int>::lowest());
-#endif
+
   test(q, input, output, plus<int>(), 0);
   test(q, input, output, minimum<int>(), std::numeric_limits<int>::max());
   test(q, input, output, maximum<int>(), std::numeric_limits<int>::lowest());
 
+#ifdef SPIRV_1_3
+  test(q, input, output, multiplies<int>(), 1);
+  test(q, input, output, bit_or<int>(), 0);
+  test(q, input, output, bit_xor<int>(), 0);
+  test(q, input, output, bit_and<int>(), ~0);
+#endif // SPIRV_1_3
+
   std::cout << "Test passed." << std::endl;
 }