fix test failures

Author: kbobrovs

sycl/include/sycl/ext/intel/experimental/esimd/memory.hpp

@@ -439,8 +439,8 @@ template <typename T, typename AccessorTy, CacheHint L1H = CacheHint::None,
           CacheHint L3H = CacheHint::None>
 ESIMD_INLINE ESIMD_NODEBUG void scalar_store(AccessorTy acc, uint32_t offset,
                                              T val) {
-  scatter<T>(acc, simd<T, 1, AccessorTy, L1H, L3H>(val),
-             simd<uint32_t, 1>(offset));
+  scatter<T, 1, AccessorTy, L1H, L3H>(acc, simd<T, 1>(val),
+                                      simd<uint32_t, 1>(offset));
 }
 
 /// Store a scalar value into the Shared Local Memory.

sycl/test/esimd/gather_scatter.cpp

@@ -11,43 +11,43 @@ using namespace cl::sycl;
 
 void kernel(
     accessor<int, 1, access::mode::read_write, access::target::device> &buf)
-    __attribute__((sycl_device)) {
+    SYCL_ESIMD_FUNCTION {
   simd<uint32_t, 32> offsets(0, 1);
   simd<int, 32> v1(0, 1);
 
-  auto v0 = gather<int, 32>(buf.get_pointer(), offsets);
+  auto v0 = gather<int, 32>(buf, offsets);
 
   v0 = v0 + v1;
 
-  scatter<int, 32>(buf.get_pointer(), v0, offsets);
+  scatter<int, 32>(buf, v0, offsets);
 }
 
 void kernel(
     accessor<uint8_t, 1, access::mode::read_write, access::target::device> &buf)
-    __attribute__((sycl_device)) {
+    SYCL_ESIMD_FUNCTION {
   simd<uint32_t, 32> offsets(0, 1);
   simd<uint8_t, 32> v1(0, 1);
 
-  simd<uint8_t, 32> v0 = gather<uint8_t, 32>(buf.get_pointer(), offsets);
+  simd<uint8_t, 32> v0 = gather<uint8_t, 32>(buf, offsets);
 
   // We honor integer promotion rules: uint8_t + uint8_t yields int
   // So we need to convert it back to simd<uint8_t, 32>
   v0 = convert<uint8_t>(v0 + v1);
 
-  scatter<uint8_t, 32>(buf.get_pointer(), v0, offsets);
+  scatter<uint8_t, 32>(buf, v0, offsets);
 }
 
 void kernel(accessor<uint16_t, 1, access::mode::read_write,
                      access::target::device> &buf)
-    __attribute__((sycl_device)) {
+    SYCL_ESIMD_FUNCTION {
   simd<uint32_t, 32> offsets(0, 1);
   simd<uint16_t, 32> v1(0, 1);
 
-  simd<uint16_t, 32> v0 = gather<uint16_t, 32>(buf.get_pointer(), offsets);
+  simd<uint16_t, 32> v0 = gather<uint16_t, 32>(buf, offsets);
 
   // We honor integer promotion rules: uint16_t + uint16_t yields int
   // So we need to convert it back to simd<uint16_t, 32>
   v0 = convert<uint16_t>(v0 + v1);
 
-  scatter<uint16_t, 32>(buf.get_pointer(), v0, offsets);
+  scatter<uint16_t, 32>(buf, v0, offsets);
 }

sycl/test/esimd/intrins_trans.cpp

@@ -129,7 +129,7 @@ SYCL_ESIMD_FUNCTION SYCL_EXTERNAL simd<float, 16> foo() {
     // CHECK: %[[SI3_VAL:[0-9a-zA-Z_.]+]] = ptrtoint i32 addrspace(1)* %{{[0-9a-zA-Z_.]+}} to i32
     // CHECK: store i32 %[[SI3_VAL]], i32 addrspace(4)* %[[SI3_ADDR:[0-9a-zA-Z_.]+]]
     // CHECK: %[[SI3:[0-9a-zA-Z_.]+]] = load i32, i32 addrspace(4)* %[[SI3_ADDR]]
-    // CHECK: %{{[0-9a-zA-Z_.]+}} = call <8 x i32> @llvm.genx.gather.scaled2.v8i32.v8i32(i32 2, i16 0, i32 %[[SI3]], i32 %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}})
+    // CHECK: %{{[0-9a-zA-Z_.]+}} = call <8 x i32> @llvm.genx.gather.masked.scaled2.v8i32.v8i32.v8i1(i32 2, i16 0, i32 %[[SI3]], i32 %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}}, <8 x i1> %{{[0-9a-zA-Z_.]+}})
 
     // 4-byte element scatter
     scatter<int, 8>(acc, v, offsets, 100, pred);
@@ -143,7 +143,7 @@ SYCL_ESIMD_FUNCTION SYCL_EXTERNAL simd<float, 16> foo() {
     // CHECK: %[[SI5_VAL:[0-9a-zA-Z_.]+]] = ptrtoint i32 addrspace(1)* %{{[0-9a-zA-Z_.]+}} to i32
     // CHECK: store i32 %[[SI5_VAL]], i32 addrspace(4)* %[[SI5_ADDR:[0-9a-zA-Z_.]+]]
     // CHECK: %[[SI5:[0-9a-zA-Z_.]+]] = load i32, i32 addrspace(4)* %[[SI5_ADDR]]
-    // CHECK: %{{[0-9a-zA-Z_.]+}} = call <8 x i32> @llvm.genx.gather.scaled2.v8i32.v8i32(i32 0, i16 0, i32 %[[SI5]], i32 %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}})
+    // CHECK: %{{[0-9a-zA-Z_.]+}} = call <8 x i32> @llvm.genx.gather.masked.scaled2.v8i32.v8i32.v8i1(i32 0, i16 0, i32 %[[SI5]], i32 %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}}, <8 x i1> %{{[0-9a-zA-Z_.]+}})
 
     // 1-byte element scatter
     scatter<unsigned char, 8>(acc, v1, offsets, 100, pred);