[SYCL][ESIMD] Add tests for accessor-based gather/scatter and scalar I/O

SYCL/ESIMD/accessor_gather_scatter.cpp

@@ -0,0 +1,124 @@
+//==------- accessor_gather_scatter.cpp  - DPC++ ESIMD on-device test ------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// TODO enable on Windows
+// REQUIRES: linux && gpu && opencl
+// RUN: %clangxx-esimd -fsycl %s -o %t.out
+// RUN: %ESIMD_RUN_PLACEHOLDER %t.out
+//
+// The test checks functionality of the gather/scatter accessor-based ESIMD
+// intrinsics.
+
+#include "esimd_test_utils.hpp"
+
+#include <CL/sycl.hpp>
+#include <CL/sycl/INTEL/esimd.hpp>
+#include <iostream>
+
+using namespace cl::sycl;
+
+template <typename T>
+using Acc =
+    accessor<T, 1, access_mode::read_write, access::target::global_buffer>;
+
+#define MASKED_LANE_NUM_REV 1
+
+template <typename T, unsigned VL, unsigned STRIDE> struct Kernel {
+  Acc<T> acc;
+  Kernel(Acc<T> acc) : acc(acc) {}
+
+  void operator()(id<1> i) const SYCL_ESIMD_KERNEL {
+    using namespace sycl::INTEL::gpu;
+    uint32_t ii = static_cast<uint32_t>(i.get(0));
+    // every STRIDE threads (subgroups with sg_size=1) access contiguous block
+    // of STRIDE*VL elements
+    uint32_t global_offset = (ii / STRIDE) * VL * STRIDE + ii % STRIDE;
+    simd<uint32_t, VL> offsets(0, STRIDE);
+    simd<T, VL> v = gather<T, VL>(acc, offsets, global_offset);
+    v += ii;
+    simd<uint16_t, VL> pred = 1;
+    pred.template select<1, 1>(VL - MASKED_LANE_NUM_REV) =
+        0; // mask out the last lane
+    scatter<T, VL>(acc, v, offsets, global_offset, pred);
+  }
+};
+
+template <typename T, unsigned VL, unsigned STRIDE> bool test(queue q) {
+  size_t size = VL * STRIDE * 117;
+
+  std::cout << "Testing T=" << typeid(T).name() << " VL=" << VL
+            << " STRIDE=" << STRIDE << "...\n";
+  T *A = new T[size];
+
+  for (unsigned i = 0; i < size; ++i) {
+    A[i] = (T)i;
+  }
+
+  {
+    buffer<T, 1> buf(A, range<1>(size));
+    range<1> glob_range{size / VL};
+
+    auto e = q.submit([&](handler &cgh) {
+      auto acc = buf.template get_access<access::mode::read_write>(cgh);
+      Kernel<T, VL, STRIDE> kernel(acc);
+      cgh.parallel_for(glob_range, kernel);
+    });
+  }
+
+  int err_cnt = 0;
+
+  for (unsigned i = 0; i < size; ++i) {
+    T gold = (T)i;
+    // the sequential number of sub group block (STRIDE in each) i falls into
+    unsigned sg_block_num = i / (VL * STRIDE);
+    // the start of the i index block this sg block covers
+    unsigned sg_block_start_i = sg_block_num * VL * STRIDE;
+    // the local id (within block) of the sg covering this i
+    unsigned sg_local_id = (i - sg_block_start_i) % STRIDE;
+    // the global id of the sg covering this i
+    unsigned sg_global_id = sg_local_id + sg_block_num * STRIDE;
+
+    unsigned lane_id = ((i % (VL * STRIDE)) - sg_local_id) / STRIDE;
+
+    gold += lane_id == VL - MASKED_LANE_NUM_REV ? 0 : sg_global_id;
+
+    if (A[i] != gold) {
+      if (++err_cnt < 35) {
+        std::cout << "failed at index " << i << ": " << A[i] << " != " << gold
+                  << " (gold)\n";
+      }
+    }
+  }
+  if (err_cnt > 0) {
+    std::cout << "  pass rate: "
+              << ((float)(size - err_cnt) / (float)size) * 100.0f << "% ("
+              << (size - err_cnt) << "/" << size << ")\n";
+  }
+
+  delete[] A;
+
+  std::cout << (err_cnt > 0 ? "  FAILED\n" : "  Passed\n");
+  return err_cnt > 0 ? false : true;
+}
+
+int main(void) {
+  queue q(esimd_test::ESIMDSelector{}, esimd_test::createExceptionHandler());
+
+  auto dev = q.get_device();
+  std::cout << "Running on " << dev.get_info<info::device::name>() << "\n";
+
+  bool passed = true;
+  passed &= test<char, 8, 1>(q);
+  passed &= test<char, 16, 3>(q);
+  passed &= test<short, 8, 8>(q);
+  passed &= test<short, 16, 1>(q);
+  passed &= test<int, 8, 2>(q);
+  passed &= test<int, 16, 1>(q);
+  passed &= test<float, 8, 2>(q);
+  passed &= test<float, 16, 1>(q);
+  return passed ? 0 : 1;
+}

SYCL/ESIMD/accessor_load_store.cpp

@@ -0,0 +1,116 @@
+//==------- accessor_load_store.cpp  - DPC++ ESIMD on-device test ----------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// TODO enable on Windows
+// REQUIRES: linux && gpu && opencl
+// RUN: %clangxx-esimd -fsycl %s -o %t.out
+// RUN: %ESIMD_RUN_PLACEHOLDER %t.out
+//
+// The test checks functionality of the scalar load/store accessor-based ESIMD
+// intrinsics.
+
+#include "esimd_test_utils.hpp"
+
+#include <CL/sycl.hpp>
+#include <CL/sycl/INTEL/esimd.hpp>
+#include <iostream>
+
+using namespace cl::sycl;
+
+template <typename T>
+using Acc =
+    accessor<T, 1, access_mode::read_write, access::target::global_buffer>;
+
+template <typename T> struct Kernel {
+  Acc<T> acc;
+  Kernel(Acc<T> acc) : acc(acc) {}
+
+  void operator()(id<1> i) const SYCL_ESIMD_KERNEL {
+    using namespace sycl::INTEL::gpu;
+    uint32_t ii = static_cast<uint32_t>(i.get(0));
+    T v = scalar_load<T>(acc, ii);
+    v += ii;
+    scalar_store<T>(acc, ii, v);
+  }
+};
+
+template <typename T> struct char_to_int {
+  using type = typename std::conditional<
+      sizeof(T) == 1,
+      typename std::conditional<std::is_signed<T>::value, int, unsigned>::type,
+      T>::type;
+};
+
+template <typename T> bool test(queue q, size_t size) {
+  std::cout << "Testing T=" << typeid(T).name() << "...\n";
+  T *A = new T[size];
+
+  for (unsigned i = 0; i < size; ++i) {
+    A[i] = (T)i;
+  }
+
+  {
+    buffer<T, 1> buf(A, range<1>(size));
+    range<1> glob_range{size};
+
+    auto e = q.submit([&](handler &cgh) {
+      auto acc = buf.template get_access<access::mode::read_write>(cgh);
+      Kernel<T> kernel(acc);
+      cgh.parallel_for(glob_range, kernel);
+    });
+  }
+
+  int err_cnt = 0;
+
+  for (unsigned i = 0; i < size; ++i) {
+    T gold = (T)i + (T)i;
+
+    if (A[i] != gold) {
+      if (++err_cnt < 10) {
+        using T1 = typename char_to_int<T>::type;
+        std::cout << "failed at index " << i << ": " << (T1)A[i]
+                  << " != " << (T1)gold << " (gold)\n";
+      }
+    }
+  }
+  if (err_cnt > 0) {
+    std::cout << "  pass rate: "
+              << ((float)(size - err_cnt) / (float)size) * 100.0f << "% ("
+              << (size - err_cnt) << "/" << size << ")\n";
+  }
+
+  delete[] A;
+
+  std::cout << (err_cnt > 0 ? "  FAILED\n" : "  Passed\n");
+  return err_cnt > 0 ? false : true;
+}
+
+int main(int argc, char **argv) {
+  // TODO the test fails with 1- and 2-byte types when size is not multiple
+  // of 4. Supposed reason - wrapping the memory buffer into image1d_buffer
+  size_t size = 128; // 117 - fails for char and short
+
+  if (argc > 1) {
+    size = atoi(argv[1]);
+    size = size == 0 ? 128 : size;
+  }
+  std::cout << "Using size=" << size << "\n";
+  queue q(esimd_test::ESIMDSelector{}, esimd_test::createExceptionHandler());
+
+  auto dev = q.get_device();
+  std::cout << "Running on " << dev.get_info<info::device::name>() << "\n";
+
+  bool passed = true;
+  passed &= test<char>(q, size);
+  passed &= test<unsigned char>(q, size);
+  passed &= test<short>(q, size);
+  passed &= test<unsigned short>(q, size);
+  passed &= test<int>(q, size);
+  passed &= test<unsigned int>(q, size);
+  passed &= test<float>(q, size);
+  return passed ? 0 : 1;
+}