[SYCL][FPGA] Implement task_sequence header, properties, and add new fpga_cluster kernel property (intel#12453)

Implement task_sequence header, task_sequence properties, and
fpga_cluster kernel property according to the spec updates
intel#6348 (almost ready to be merged)
Header
- Refactor invocation and response capacity to be on the Create
intrinsic, so Async and Get do not take in respective argument
- Remove max_outstanding argument of Create intrinsic
- Add pipelined and fpga_cluster arguments to Create intrinsic
- Async, Get, and Release intrinsic now accept the TargetExtType
target("spirv.TaskSequenceINTEL") returned from the Create intrinsic,
and do not take in object pointer or task function pointer arguments
- Task Sequence accepts properties

Task Sequence Properties
- Convert invocation and response capacity from template parameters to
properties
- Add balanced property to remove Get intrinsic loop in destructor

FPGA Kernel properties
- Add fpga_cluster property
- Support fpga_cluster and pipelined property for task sequence

Author: bowenxue-intel

clang/lib/CodeGen/CodeGenTypes.cpp

@@ -654,6 +654,11 @@ llvm::Type *CodeGenTypes::ConvertType(QualType T) {
                       "__spv::__spirv_JointMatrixINTEL") {
           ResultType = ConvertSYCLJointMatrixINTELType(RD);
           break;
+        } else if (RD && RD->getQualifiedNameAsString() ==
+                             "__spv::__spirv_TaskSequenceINTEL") {
+          ResultType = llvm::TargetExtType::get(getLLVMContext(),
+                                                "spirv.TaskSequenceINTEL");
+          break;
         }
       }
     }

clang/lib/Driver/ToolChains/Clang.cpp

@@ -10235,7 +10235,8 @@ void SPIRVTranslator::ConstructJob(Compilation &C, const JobAction &JA,
         ",+SPV_INTEL_fpga_buffer_location"
         ",+SPV_INTEL_fpga_argument_interfaces"
         ",+SPV_INTEL_fpga_invocation_pipelining_attributes"
-        ",+SPV_INTEL_fpga_latency_control";
+        ",+SPV_INTEL_fpga_latency_control"
+        ",+SPV_INTEL_task_sequence";
     ExtArg = ExtArg + DefaultExtArg + INTELExtArg;
     if (C.getDriver().IsFPGAHWMode())
       // Enable several extensions on FPGA H/W exclusively

clang/test/CodeGenSYCL/intel-task-sequence.cpp

@@ -0,0 +1,11 @@
+// RUN: %clang_cc1 -triple spir64-unknown-unknown -disable-llvm-passes -emit-llvm %s -o - | FileCheck %s
+
+// Test that SPIR-V codegen generates the expected LLVM struct name for the
+// TaskSequenceINTEL type
+
+namespace __spv {
+  struct __spirv_TaskSequenceINTEL;
+} // namespace __spv
+
+// CHECK: @_Z4func{{.*}}(target("spirv.TaskSequenceINTEL")
+void func(__spv::__spirv_TaskSequenceINTEL *task_seq) {}

clang/test/Driver/sycl-spirv-ext.c

@@ -47,6 +47,7 @@
 // CHECK-DEFAULT-SAME:,+SPV_INTEL_fpga_argument_interfaces
 // CHECK-DEFAULT-SAME:,+SPV_INTEL_fpga_invocation_pipelining_attributes
 // CHECK-DEFAULT-SAME:,+SPV_INTEL_fpga_latency_control
+// CHECK-DEFAULT-SAME:,+SPV_INTEL_task_sequence
 // CHECK-DEFAULT-SAME:,+SPV_INTEL_token_type
 // CHECK-DEFAULT-SAME:,+SPV_INTEL_bfloat16_conversion
 // CHECK-DEFAULT-SAME:,+SPV_INTEL_joint_matrix
@@ -77,6 +78,7 @@
 // CHECK-FPGA-HW-SAME:,+SPV_INTEL_fpga_buffer_location
 // CHECK-FPGA-HW-SAME:,+SPV_INTEL_fpga_argument_interfaces
 // CHECK-FPGA-HW-SAME:,+SPV_INTEL_fpga_latency_control
+// CHECK-FPGA-HW-SAME:,+SPV_INTEL_task_sequence
 // CHECK-FPGA-HW-SAME:,+SPV_INTEL_usm_storage_classes
 // CHECK-FPGA-HW-SAME:,+SPV_INTEL_runtime_aligned
 // CHECK-FPGA-HW-SAME:,+SPV_INTEL_fpga_cluster_attributes,+SPV_INTEL_loop_fuse
@@ -108,6 +110,7 @@
 // CHECK-CPU-SAME:,+SPV_INTEL_fpga_argument_interfaces
 // CHECK-CPU-SAME:,+SPV_INTEL_fpga_invocation_pipelining_attributes
 // CHECK-CPU-SAME:,+SPV_INTEL_fpga_latency_control
+// CHECK-CPU-SAME:,+SPV_INTEL_task_sequence
 // CHECK-CPU-SAME:,+SPV_INTEL_token_type
 // CHECK-CPU-SAME:,+SPV_INTEL_bfloat16_conversion
 // CHECK-CPU-SAME:,+SPV_INTEL_joint_matrix

llvm/include/llvm/SYCLLowerIR/DeviceConfigFile.td

@@ -70,6 +70,7 @@ def AspectExt_intel_matrix : Aspect<"ext_intel_matrix">;
 def AspectExt_oneapi_is_composite : Aspect<"ext_oneapi_is_composite">;
 def AspectExt_oneapi_is_component : Aspect<"ext_oneapi_is_component">;
 def AspectExt_oneapi_graph : Aspect<"ext_oneapi_graph">;
+def AspectExt_intel_fpga_task_sequence : Aspect<"ext_intel_fpga_task_sequence">;
 // Deprecated aspects
 def AspectInt64_base_atomics : Aspect<"int64_base_atomics">;
 def AspectInt64_extended_atomics : Aspect<"int64_extended_atomics">;
@@ -120,7 +121,8 @@ def : TargetInfo<"__TestAspectList",
     AspectExt_oneapi_interop_semaphore_import, AspectExt_oneapi_interop_semaphore_export,
     AspectExt_oneapi_mipmap, AspectExt_oneapi_mipmap_anisotropy, AspectExt_oneapi_mipmap_level_reference, AspectExt_intel_esimd,
     AspectExt_oneapi_ballot_group, AspectExt_oneapi_fixed_size_group, AspectExt_oneapi_opportunistic_group,
-    AspectExt_oneapi_tangle_group, AspectExt_intel_matrix, AspectExt_oneapi_is_composite, AspectExt_oneapi_is_component, AspectExt_oneapi_graph],
+    AspectExt_oneapi_tangle_group, AspectExt_intel_matrix, AspectExt_oneapi_is_composite, AspectExt_oneapi_is_component,
+    AspectExt_oneapi_graph, AspectExt_intel_fpga_task_sequence],
     []>;
 // This definition serves the only purpose of testing whether the deprecated aspect list defined in here and in SYCL RT
 // match.

llvm/lib/SYCLLowerIR/CompileTimePropertiesPass.cpp

@@ -416,6 +416,7 @@ attributeToExecModeMetadata(const Attribute &Attr, Function &F) {
 
   if (AttrKindStr == "sycl-streaming-interface") {
     // generate either:
+    //   !ip_interface !N
     //   !N = !{!"streaming"} or
     //   !N = !{!"streaming", !"stall_free_return"}
     SmallVector<Metadata *, 2> MD;
@@ -428,6 +429,7 @@ attributeToExecModeMetadata(const Attribute &Attr, Function &F) {
 
   if (AttrKindStr == "sycl-register-map-interface") {
     // generate either:
+    //   !ip_interface !N
     //   !N = !{!"csr"} or
     //   !N = !{!"csr", !"wait_for_done_write"}
     SmallVector<Metadata *, 2> MD;
@@ -438,6 +440,20 @@ attributeToExecModeMetadata(const Attribute &Attr, Function &F) {
                                             MDNode::get(Ctx, MD));
   }
 
+  if (AttrKindStr == "sycl-fpga-cluster") {
+    // generate either:
+    //   !stall_free !N
+    //   !N = !{i32 1} or
+    //   !stall_enable !N
+    //   !N = !{i32 1}
+    std::string ClusterType =
+        getAttributeAsInteger<uint32_t>(Attr) ? "stall_enable" : "stall_free";
+    Metadata *ClusterMDArgs[] = {
+        ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Ctx), 1))};
+    return std::pair<std::string, MDNode *>(ClusterType,
+                                            MDNode::get(Ctx, ClusterMDArgs));
+  }
+
   if ((AttrKindStr == SYCL_REGISTER_ALLOC_MODE_ATTR ||
        AttrKindStr == SYCL_GRF_SIZE_ATTR) &&
       !llvm::esimd::isESIMD(F)) {

llvm/test/SYCLLowerIR/CompileTimePropertiesPass/kernel-attributes/fpga-cluster.ll

@@ -0,0 +1,33 @@
+; Check conversion of sycl-fpga-cluster attribute
+; RUN: opt -passes="compile-time-properties" %s -S -o - | FileCheck %s --check-prefix CHECK-IR
+
+; CHECK-IR-DAG: @stallFree() #0 {{.*}}!stall_free [[MD_TRUE:![0-9]+]] {
+; Function Attrs: convergent norecurse
+define weak_odr dso_local spir_kernel void @stallFree() #0 {
+entry:
+  ret void
+}
+
+; CHECK-IR-DAG: @stallEnable() #1 {{.*}}!stall_enable [[MD_TRUE:![0-9]+]] {
+; Function Attrs: convergent norecurse
+define weak_odr dso_local spir_kernel void @stallEnable() #1 {
+entry:
+  ret void
+}
+
+attributes #0 = { convergent norecurse "frame-pointer"="all" "sycl-fpga-cluster"="0" }
+attributes #1 = { convergent norecurse "frame-pointer"="all" "sycl-fpga-cluster"="1" }
+
+!opencl.spir.version = !{!0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0}
+!spirv.Source = !{!1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1}
+!llvm.ident = !{!2, !2, !2, !2, !2, !2, !2, !2, !2, !2, !2}
+!llvm.module.flags = !{!3, !4}
+
+!0 = !{i32 1, i32 2}
+!1 = !{i32 4, i32 100000}
+!2 = !{!"clang version 13.0.0 (https://github.com/intel/llvm)"}
+!3 = !{i32 1, !"wchar_size", i32 4}
+!4 = !{i32 7, !"frame-pointer", i32 2}
+
+; Confirm the decorations for the functions
+; CHECK-IR-DAG: [[MD_TRUE]] = !{i32 1}

sycl/include/CL/__spirv/spirv_ops.hpp

@@ -1267,6 +1267,25 @@ extern __DPCPP_SYCL_EXTERNAL
     std::enable_if_t<std::is_integral_v<to> && std::is_unsigned_v<to>, to>
     __spirv_ConvertPtrToU(from val) noexcept;
 
+template <typename RetT, typename... ArgsT>
+extern __DPCPP_SYCL_EXTERNAL __spv::__spirv_TaskSequenceINTEL *
+__spirv_TaskSequenceCreateINTEL(RetT (*f)(ArgsT...), int Pipelined = -1,
+                                int ClusterMode = -1,
+                                unsigned int ResponseCapacity = 0,
+                                unsigned int InvocationCapacity = 0) noexcept;
+
+template <typename... ArgsT>
+extern __DPCPP_SYCL_EXTERNAL void
+__spirv_TaskSequenceAsyncINTEL(__spv::__spirv_TaskSequenceINTEL *TaskSequence,
+                               ArgsT... Args) noexcept;
+
+template <typename RetT>
+extern __DPCPP_SYCL_EXTERNAL RetT __spirv_TaskSequenceGetINTEL(
+    __spv::__spirv_TaskSequenceINTEL *TaskSequence) noexcept;
+
+extern __DPCPP_SYCL_EXTERNAL void __spirv_TaskSequenceReleaseINTEL(
+    __spv::__spirv_TaskSequenceINTEL *TaskSequence) noexcept;
+
 #else  // if !__SYCL_DEVICE_ONLY__
 
 template <typename dataT>

sycl/include/CL/__spirv/spirv_types.hpp

@@ -124,6 +124,8 @@ template <typename T, std::size_t R, std::size_t C, MatrixLayout L,
           MatrixUse U = MatrixUse::MatrixA>
 struct __spirv_JointMatrixINTEL;
 
+struct __spirv_TaskSequenceINTEL;
+
 } // namespace __spv
 
 #ifdef __SYCL_DEVICE_ONLY__

sycl/include/sycl/device_aspect_macros.hpp

@@ -318,6 +318,11 @@
 #define __SYCL_ALL_DEVICES_HAVE_ext_oneapi_graph__ 0
 #endif
 
+#ifndef __SYCL_ALL_DEVICES_HAVE_ext_intel_fpga_task_sequence__
+// __SYCL_ASPECT(ext_intel_fpga_task_sequence, 62)
+#define __SYCL_ALL_DEVICES_HAVE_ext_intel_fpga_task_sequence__ 0
+#endif
+
 #ifndef __SYCL_ANY_DEVICE_HAS_host__
 // __SYCL_ASPECT(host, 0)
 #define __SYCL_ANY_DEVICE_HAS_host__ 0
@@ -627,3 +632,8 @@
 // __SYCL_ASPECT(ext_oneapi_graph, 61)
 #define __SYCL_ANY_DEVICE_HAS_ext_oneapi_graph__ 0
 #endif
+
+#ifndef __SYCL_ANY_DEVICE_HAS_ext_intel_fpga_task_sequence__
+// __SYCL_ASPECT(ext_intel_fpga_task_sequence__, 62)
+#define __SYCL_ANY_DEVICE_HAS_ext_intel_fpga_task_sequence__ 0
+#endif

sycl/include/sycl/ext/intel/experimental/fpga_kernel_properties.hpp

@@ -1,5 +1,5 @@
-//==----- fpga_kernel_properties.hpp - SYCL properties associated with FPGA
-// kernel properties ---==//
+//===--------------------- fpga_kernel_properties.hpp ---------------------===//
+// SYCL properties associated with FPGA kernel properties
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
@@ -12,11 +12,15 @@
 #include <sycl/ext/oneapi/properties/property.hpp>
 #include <sycl/ext/oneapi/properties/property_value.hpp>
 
+#include <cstdint>     // for uint16_t
+#include <type_traits> // for true_type
+
 namespace sycl {
 inline namespace _V1 {
 namespace ext::intel::experimental {
 
 template <typename T, typename PropertyListT> class fpga_kernel_attribute;
+template <auto &f, typename PropertyListT> class task_sequence;
 
 enum class streaming_interface_options_enum : std::uint16_t {
   accept_downstream_stall,
@@ -28,6 +32,11 @@ enum class register_map_interface_options_enum : std::uint16_t {
   wait_for_done_write,
 };
 
+enum class fpga_cluster_options_enum : std::uint16_t {
+  stall_free,
+  stall_enable
+};
+
 struct streaming_interface_key
     : oneapi::experimental::detail::compile_time_property_key<
           oneapi::experimental::detail::PropKind::StreamingInterface> {
@@ -54,6 +63,15 @@ struct pipelined_key : oneapi::experimental::detail::compile_time_property_key<
       std::integral_constant<int, pipeline_directive_or_initiation_interval>>;
 };
 
+struct fpga_cluster_key
+    : oneapi::experimental::detail::compile_time_property_key<
+          oneapi::experimental::detail::PropKind::FPGACluster> {
+  template <fpga_cluster_options_enum option>
+  using value_t = ext::oneapi::experimental::property_value<
+      fpga_cluster_key,
+      std::integral_constant<fpga_cluster_options_enum, option>>;
+};
+
 template <streaming_interface_options_enum option =
               streaming_interface_options_enum::accept_downstream_stall>
 inline constexpr streaming_interface_key::value_t<option> streaming_interface;
@@ -84,6 +102,18 @@ inline constexpr pipelined_key::value_t<
     pipeline_directive_or_initiation_interval>
     pipelined;
 
+template <fpga_cluster_options_enum option =
+              fpga_cluster_options_enum::stall_free>
+inline constexpr fpga_cluster_key::value_t<option> fpga_cluster;
+
+inline constexpr fpga_cluster_key::value_t<
+    fpga_cluster_options_enum::stall_free>
+    stall_free_clusters;
+
+inline constexpr fpga_cluster_key::value_t<
+    fpga_cluster_options_enum::stall_enable>
+    stall_enable_clusters;
+
 } // namespace ext::intel::experimental
 
 namespace ext::oneapi::experimental {
@@ -103,6 +133,22 @@ struct is_property_key_of<
     intel::experimental::fpga_kernel_attribute<T, PropertyListT>>
     : std::true_type {};
 
+template <typename T, typename PropertyListT>
+struct is_property_key_of<
+    intel::experimental::fpga_cluster_key,
+    intel::experimental::fpga_kernel_attribute<T, PropertyListT>>
+    : std::true_type {};
+
+template <auto &f, typename PropertyListT>
+struct is_property_key_of<intel::experimental::pipelined_key,
+                          intel::experimental::task_sequence<f, PropertyListT>>
+    : std::true_type {};
+
+template <auto &f, typename PropertyListT>
+struct is_property_key_of<intel::experimental::fpga_cluster_key,
+                          intel::experimental::task_sequence<f, PropertyListT>>
+    : std::true_type {};
+
 namespace detail {
 template <intel::experimental::streaming_interface_options_enum Stall_Free>
 struct PropertyMetaInfo<
@@ -124,6 +170,14 @@ struct PropertyMetaInfo<intel::experimental::pipelined_key::value_t<Value>> {
   static constexpr int value = Value;
 };
 
+template <intel::experimental::fpga_cluster_options_enum ClusterType>
+struct PropertyMetaInfo<
+    intel::experimental::fpga_cluster_key::value_t<ClusterType>> {
+  static constexpr const char *name = "sycl-fpga-cluster";
+  static constexpr intel::experimental::fpga_cluster_options_enum value =
+      ClusterType;
+};
+
 } // namespace detail
 } // namespace ext::oneapi::experimental
 } // namespace _V1