Address PR comments

intel · Feb 8, 2024 · b97483c · b97483c
1 parent 4bb7456
commit b97483c
Show file tree

Hide file tree

Showing 6 changed files with 25 additions and 10 deletions.
diff --git a/llvm/lib/SYCLLowerIR/ESIMD/LowerESIMD.cpp b/llvm/lib/SYCLLowerIR/ESIMD/LowerESIMD.cpp
@@ -1011,9 +1011,9 @@ static void translateScatterStore(CallInst &CI, bool IsSLM) {
                                ESIMDIntrinDesc::GenXArgConversion::TO_I64);
   Align AlignValue(Val.getZExtValue());
 
+  auto ValsOp = CI.getArgOperand(0);
   auto OffsetsOp = CI.getArgOperand(1);
   auto MaskOp = CI.getArgOperand(2);
-  auto ValsOp = CI.getArgOperand(0);
   auto DataType = ValsOp->getType();
 
   // Convert the mask from <N x i16> to <N x i1>.

diff --git a/sycl/test-e2e/ESIMD/unified_memory_api/scatter_usm.cpp b/sycl/test-e2e/ESIMD/unified_memory_api/scatter_usm.cpp
@@ -10,8 +10,8 @@
 
 // The test verifies esimd::scatter() functions accepting USM pointer
 // and optional compile-time esimd::properties.
-// The scatter() calls in this test do not use cache-hint
-// properties to not impose using DG2/PVC features.
+// The scatter() calls in this test do not use cache-hint properties
+// or VS > 1 (number of stores per offset) to not impose using PVC features.
 
 #include "Inputs/scatter.hpp"
 

diff --git a/sycl/test-e2e/ESIMD/unified_memory_api/scatter_usm_legacy.cpp b/sycl/test-e2e/ESIMD/unified_memory_api/scatter_usm_legacy.cpp
@@ -12,7 +12,7 @@
 // The test verifies esimd::scatter() functions accepting USM pointer
 // and optional compile-time esimd::properties.
 // The scatter() calls in this test do not use cache-hint properties
-// or VS > 1 (number of loads per offset) to not impose using PVC features.
+// or VS > 1 (number of stores per offset) to not impose using PVC features.
 //
 // TODO: Remove this test when GPU driver issue with llvm.masked.scatter is
 // resolved and ESIMD starts using llvm.masked.scatter by default.

diff --git a/sycl/test-e2e/ESIMD/unified_memory_api/slm_scatter.cpp b/sycl/test-e2e/ESIMD/unified_memory_api/slm_scatter.cpp
@@ -8,9 +8,9 @@
 // RUN: %{build} -fsycl-device-code-split=per_kernel -D__ESIMD_GATHER_SCATTER_LLVM_IR -o %t.out
 // RUN: %{run} %t.out
 
-// The test verifies esimd::slm_scatter() functions accepting
-// optional compile-time esimd::properties.
-// The scatter() calls in this test do not use DG2/PVC features.
+// The test verifies esimd::slm_scatter() functions accepting  optional
+// compile-time esimd::properties. The slm_scatter() calls in this test do not
+// use VS > 1 (number of stores per offset) to not impose using PVC features.
 
 #include "Inputs/scatter.hpp"
 

diff --git a/sycl/test-e2e/ESIMD/unified_memory_api/slm_scatter_legacy.cpp b/sycl/test-e2e/ESIMD/unified_memory_api/slm_scatter_legacy.cpp
@@ -9,9 +9,9 @@
 // RUN: %{build} -fsycl-device-code-split=per_kernel -o %t.out
 // RUN: %{run} %t.out
 
-// The test verifies esimd::scatter() functions accepting  optional compile-time
-// esimd::properties. The scatter() calls in this test do not use VS > 1 (number
-// of loads per offset) to not impose using PVC features.
+// The test verifies esimd::slm_scatter() functions accepting  optional
+// compile-time esimd::properties. The slm_scatter() calls in this test do not
+// use VS > 1 (number of stores per offset) to not impose using PVC features.
 //
 // TODO: Remove this test when GPU driver issue with llvm.masked.scatter is
 // resolved and ESIMD starts using llvm.masked.scatter by default.

diff --git a/sycl/test/esimd/memory_properties.cpp b/sycl/test/esimd/memory_properties.cpp
@@ -1281,6 +1281,14 @@ test_gather_scatter(AccType &acc, LocalAccType &local_acc, float *ptrf,
   scatter<float, 32, 2>(ptrf, ioffset_n16_view, usm_view, mask_n16);
 
   scatter<float, 32, 2>(ptrf, ioffset_n16_view, usm_view);
+
+  simd<uint32_t, 10> ioffset_n10(byte_offset32, 8);
+  simd<float, 10> usm_n10;
+
+  // Check special case to verify that for cases when N is not power of 2 llvm
+  // intrinsic is used
+  // CHECK-COUNT-1: call void @llvm.masked.scatter.v10f32.v10p4(<10 x float> {{[^)]+}}, <10 x ptr addrspace(4)> {{[^)]+}}, i32 4, <10 x i1> {{[^)]+}})
+  scatter(ptrf, ioffset_n10, usm_n10);
 }
 
 // CHECK-LABEL: define {{.*}} @_Z23test_slm_gather_scatter{{.*}}
@@ -1429,4 +1437,11 @@ test_slm_gather_scatter(int byte_offset32) {
   slm_scatter<float, 32, 2>(ioffset_n16_view, slm, mask_n16, props_align4);
   slm_scatter<float, 32, 2>(ioffset_n16, slm_view, mask_n16, props_align4);
   slm_scatter<float, 32, 2>(ioffset_n16_view, slm_view, mask_n16, props_align4);
+
+  simd<uint32_t, 10> ioffset_n10(byte_offset32, 8);
+  simd<float, 10> usm_n10;
+  // Check special case to verify that for cases when N is not power of 2 llvm
+  // intrinsic is used
+  // CHECK-COUNT-1: call void @llvm.masked.scatter.v10f32.v10p3(<10 x float> {{[^)]+}}, <10 x ptr addrspace(3)> {{[^)]+}}, i32 4, <10 x i1> {{[^)]+}})
+  slm_scatter(ioffset_n10, usm_n10);
 }