[SYCL][ESIMD] Use intrinsic for named_barrier_signal (#12982)

We don't want to use raw_send here. I manually tested this on PVC with the required driver version. --------- Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>
intel · Mar 15, 2024 · d4a9254 · d4a9254
1 parent 865ed57
commit d4a9254
Show file tree

Hide file tree

Showing 6 changed files with 15 additions and 54 deletions.
diff --git a/llvm/lib/SYCLLowerIR/CMakeLists.txt b/llvm/lib/SYCLLowerIR/CMakeLists.txt
@@ -14,9 +14,9 @@ if (NOT TARGET LLVMGenXIntrinsics)
  if (NOT DEFINED LLVMGenXIntrinsics_SOURCE_DIR)
  set(LLVMGenXIntrinsics_GIT_REPO https://github.com/intel/vc-intrinsics.git)
 
- # Date:  Dec 14, 2023
- # Add support for XeHPCVG platform
- set(LLVMGenXIntrinsics_GIT_TAG da892e1982b6c25b9a133f85b4ac97142d8a3def)
+ # Date: 13 Feb 2024
+ # Add an intrinsic for named barrier arrive/signal operation
+ set(LLVMGenXIntrinsics_GIT_TAG f9c34404d0ea9abad83875a10bd48d88cea90ebd)
 
  message(STATUS "vc-intrinsics repo is missing. Will try to download it from ${LLVMGenXIntrinsics_GIT_REPO}")
  include(FetchContent)

diff --git a/llvm/lib/SYCLLowerIR/ESIMD/LowerESIMD.cpp b/llvm/lib/SYCLLowerIR/ESIMD/LowerESIMD.cpp
@@ -494,8 +494,7 @@ class ESIMDIntrinDescTable {
  {"dpasw", {"dpasw", {a(0), a(1), a(2), t(0)}}},
  {"dpasw_nosrc0", {"dpasw.nosrc0", {a(0), a(1), t(0)}}},
  {"nbarrier", {"nbarrier", {a(0), a(1), a(2)}}},
- {"raw_send_nbarrier_signal",
- {"raw.send.noresult", {a(0), ai1(4), a(1), a(2), a(3)}}},
+ {"nbarrier_arrive", {"nbarrier.arrive", {a(0), a(1), a(2), a(3)}}},
  {"lsc_load_slm",
  {"lsc.load.slm",
  {ai1(0), c8(lsc_subopcode::load), t8(1), t8(2), t16(3), t32(4), t8(5),

diff --git a/llvm/test/SYCLLowerIR/esimd_lower_nbarriers.ll b/llvm/test/SYCLLowerIR/esimd_lower_nbarriers.ll
diff --git a/sycl/include/sycl/ext/intel/experimental/esimd/detail/memory_intrin.hpp b/sycl/include/sycl/ext/intel/experimental/esimd/detail/memory_intrin.hpp
@@ -41,26 +41,18 @@ __ESIMD_INTRIN void __esimd_nbarrier(uint8_t mode, uint8_t id,
 /// @param count - number of named barriers
 __ESIMD_INTRIN void __esimd_nbarrier_init(uint8_t count) __ESIMD_INTRIN_END;
 
-/// Raw send signal to perform signal operation on named barriers
+/// Perform signal operation on named barriers
 /// Available only on PVC
-/// @tparam Ty - message element type
+/// @param id - barrier id
 ///
-/// @tparam N - message length
+/// @param thread_role - thread role
 ///
-/// @param is_sendc - is sendc
+/// @param num_producers - number of producers
 ///
-/// @param extended_descriptor - extended message descriptor
-///
-/// @param descriptor - message descriptor
-///
-/// @param msg_var - source operand of send message
-///
-/// @param pred - predicate for enabled channels
-template <typename Ty, int N>
-__ESIMD_INTRIN void __esimd_raw_send_nbarrier_signal(
- uint32_t is_sendc, uint32_t extended_descriptor, uint32_t descriptor,
- __ESIMD_DNS::vector_type_t<Ty, N> msg_var,
- uint16_t pred = 1) __ESIMD_INTRIN_END;
+/// @param num_consumers - number of consumers
+__ESIMD_INTRIN void
+__esimd_nbarrier_arrive(uint8_t id, uint8_t thread_role, uint8_t num_producers,
+ uint8_t num_consumers) __ESIMD_INTRIN_END;
 
 /// 2D USM pointer block load.
 /// Supported platforms: PVC

diff --git a/sycl/include/sycl/ext/intel/experimental/esimd/memory.hpp b/sycl/include/sycl/ext/intel/experimental/esimd/memory.hpp
@@ -393,19 +393,10 @@ __ESIMD_API void named_barrier_signal(uint8_t barrier_id,
  uint8_t producer_consumer_mode,
  uint32_t num_producers,
  uint32_t num_consumers) {
- constexpr uint32_t gateway = 3;
- constexpr uint32_t barrier = 4;
- constexpr uint32_t descriptor = 1 << 25 | // Message length: 1 register
- 0 << 12 | // Fence Data Ports: No fence
- barrier; // Barrier subfunction
-
- __ESIMD_DNS::vector_type_t<uint32_t, 8> payload = 0;
- payload[2] = (num_consumers & 0xff) << 24 | (num_producers & 0xff) << 16 |
- producer_consumer_mode << 14 | (barrier_id & 0b11111) << 0;
  __esimd_fence(__ESIMD_NS::fence_mask::global_coherent_fence |
  __ESIMD_NS::fence_mask::local_barrier);
- __esimd_raw_send_nbarrier_signal<uint32_t, 8>(
- 0 /*sendc*/, gateway, descriptor, payload, 1 /*pred*/);
+ __esimd_nbarrier_arrive(barrier_id, producer_consumer_mode, num_producers,
+  num_consumers);
 }
 
 /// Create explicit scoreboard dependency to avoid device code motion

diff --git a/sycl/test/esimd/nbarriers.cpp b/sycl/test/esimd/nbarriers.cpp
@@ -16,7 +16,7 @@ void caller(int x) {
  named_barrier_init<7>();
  named_barrier_wait(2);
  // CHECK: call spir_func void @_Z13__esimd_fenceh(i8 noundef zeroext 33)
- // CHECK-NEXT: call spir_func void @_Z32__esimd_raw_send_nbarrier_signal{{.*}}
+ // CHECK-NEXT: call spir_func void @_Z23__esimd_nbarrier_arrive{{.*}}
  named_barrier_signal(0, 0, 4, 4);
  });
 }