[ESIMD] Extend tests to cover sycl::half. (#640)

* [ESIMD] Add infra for C++ ops testing, add heavy 'half' tests, extend existing.

Signed-off-by: Konstantin S Bobrovsky <[email protected]>

Author: kbobrovs

SYCL/ESIMD/api/bin_and_cmp_ops_heavy.cpp

@@ -0,0 +1,300 @@
+//==--------------- bin_un_cmp_ops_heavy.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
+//
+//===----------------------------------------------------------------------===//
+// REQUIRES: gpu
+// UNSUPPORTED: cuda || hip
+// RUN: %clangxx -fsycl %s -o %t.out
+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+
+// Tests various binary operations applied to simd objects.
+
+// TODO
+// Arithmetic operations behaviour depends on Gen's control regiter's rounding
+// mode, which is RTNE by default:
+//    cr0.5:4 is 00b = Round to Nearest or Even (RTNE)
+// For half this leads to divergence between Gen and host (emulated) results
+// larger than certain threshold. Might need to tune the cr0 once this feature
+// is available in ESIMD.
+//
+
+#include "../esimd_test_utils.hpp"
+
+#include <CL/sycl.hpp>
+#include <iostream>
+#include <sycl/ext/intel/experimental/esimd.hpp>
+
+using namespace cl::sycl;
+using namespace sycl::ext::intel::experimental::esimd;
+
+template <class T1, class T2, int VL, class OpClass, class Ops> class TestID;
+
+// Result type of a scalar binary Op
+template <class T1, class T2, class OpClass>
+using scalar_comp_t =
+    std::conditional_t<std::is_same_v<OpClass, esimd_test::CmpOp>,
+                       typename simd_mask<8>::element_type,
+                       __SEIEED::computation_type_t<T1, T2>>;
+
+// Result type of a vector binary Op
+template <class T1, class T2, class OpClass, int N = 0>
+using comp_t = std::conditional_t<
+    N == 0, scalar_comp_t<T1, T2, OpClass>,
+    std::conditional_t<std::is_same_v<OpClass, esimd_test::CmpOp>, simd_mask<N>,
+                       simd<__SEIEED::computation_type_t<T1, T2>, N>>>;
+
+// Helpers for printing
+template <class T> auto cast(T val) { return val; }
+template <> auto cast<char>(char val) { return (int)val; }
+template <> auto cast<unsigned char>(unsigned char val) {
+  return (unsigned int)val;
+}
+#ifdef __SYCL_DEVICE_ONLY__
+template <> auto cast<_Float16>(_Float16 val) { return (float)val; }
+#endif
+
+// Main test function.
+// T1, T2 - operand types,
+// VL - vector length,
+// OpClass - binary or comparison operations,
+// VerifyF and InitF - verification and initialization function types
+//   (instantiated within the test function),
+// Ops - a compile-time sequence of operations to test.
+//
+template <class T1, class T2, int VL, class OpClass,
+          template <class, class, class> class VerifyF,
+          template <class, class, class> class InitF, class Ops>
+bool test(Ops ops, queue &q, comp_t<T1, T2, OpClass> epsilon = 0) {
+  // Log test case info
+  std::cout << "Testing T1=" << typeid(T1).name() << " T2=" << typeid(T2).name()
+            << ", VL=" << VL << " ...\n";
+  std::cout << "Operations:";
+  esimd_test::iterate_ops(ops, [=](OpClass op) {
+    std::cout << " '" << esimd_test::Op2Str(op) << "'";
+  });
+  std::cout << "\n";
+
+  // initialize test data
+  constexpr int Size = 1024 * 7;
+  T1 *A = sycl::malloc_shared<T1>(Size, q);
+  T2 *B = sycl::malloc_shared<T2>(Size, q);
+  constexpr int NumOps = (int)Ops::size;
+  int CSize = NumOps * Size;
+  using T = comp_t<T1, T2, OpClass>;
+  // Result array. For each pair of A[i] and B[i] elements it reserves NumOps
+  // elements to store result of all operations under test applied to the A[i]
+  // and B[i]
+  T *C = sycl::malloc_shared<T>(CSize, q);
+  InitF<T1, T2, OpClass> init;
+
+  for (int i = 0; i < Size; ++i) {
+    init(A, B, C, i);
+  }
+
+  // submit the kernel
+  try {
+    auto e = q.submit([&](handler &cgh) {
+      cgh.parallel_for<TestID<T1, T2, VL, OpClass, Ops>>(
+          Size / VL, [=](id<1> i) SYCL_ESIMD_KERNEL {
+            unsigned off = i * VL;
+            simd<T1, VL> va(A + off, vector_aligned_tag{});
+            simd<T2, VL> vb(B + off, vector_aligned_tag{});
+
+            // applies each of the input operations to the va and vb vectors,
+            // then invokes the lambda below, passing the result of the
+            // operation, its ID and sequential number within the input sequence
+            esimd_test::apply_ops(
+                ops, va, vb,
+                [=](comp_t<T1, T2, OpClass, VL> res, OpClass op,
+                    unsigned op_num) {
+                  unsigned res_off = off * NumOps + op_num * VL;
+                  res.copy_to(C + res_off, vector_aligned_tag{});
+                });
+          });
+    });
+    e.wait();
+  } catch (sycl::exception const &e) {
+    std::cout << "SYCL exception caught: " << e.what() << '\n';
+    sycl::free(A, q);
+    sycl::free(B, q);
+    sycl::free(C, q);
+    return false;
+  }
+
+  int err_cnt = 0;
+
+  // now verify the results using provided verification function type
+  for (unsigned i = 0; i < Size / VL; ++i) {
+    unsigned off = i * VL;
+
+    for (int j = 0; j < VL; ++j) {
+      T1 a = A[off + j];
+      T2 b = B[off + j];
+
+      esimd_test::apply_ops(
+          ops, a, b, [&](T Gold, OpClass op, unsigned op_num) {
+            unsigned res_off = off * NumOps + op_num * VL;
+            T Res = C[res_off + j];
+            using Tint = esimd_test::int_type_t<sizeof(T)>;
+            Tint ResBits = *(Tint *)&Res;
+            Tint GoldBits = *(Tint *)&Gold;
+            VerifyF<T1, T2, OpClass> verify_f(epsilon);
+
+            if (!verify_f(Gold, Res, op)) {
+              if (++err_cnt < 10) {
+                std::cout << "  failed at index " << (res_off + j) << ", op "
+                          << esimd_test::Op2Str(op) << ": " << cast(Res)
+                          << "(0x" << std::hex << ResBits << ")"
+                          << " != " << std::dec << cast(Gold) << "(0x"
+                          << std::hex << GoldBits << ") [" << std::dec
+                          << cast(a) << " " << esimd_test::Op2Str(op) << " "
+                          << cast(b) << "]\n";
+              }
+            }
+          });
+    }
+  }
+  if (err_cnt > 0) {
+    auto Size1 = NumOps * Size;
+    std::cout << "  pass rate: "
+              << ((float)(Size1 - err_cnt) / (float)Size1) * 100.0f << "% ("
+              << (Size1 - err_cnt) << "/" << Size1 << ")\n";
+  }
+
+  free(A, q);
+  free(B, q);
+  free(C, q);
+  std::cout << (err_cnt > 0 ? "  FAILED\n" : "  Passed\n");
+  return err_cnt == 0;
+}
+
+// Flavours of verification function types.
+
+template <class T1, class T2, class OpClass> struct verify_strict {
+  using T = comp_t<T1, T2, OpClass>;
+
+  verify_strict(T) {}
+
+  bool operator()(T res, T gold, OpClass op) { return res == gold; }
+};
+
+#define EQ(x, y, epsilon)                                                      \
+  ((x) > (y) ? (x) - (y) <= epsilon : (y) - (x) <= epsilon)
+
+template <class T1, class T2, class OpClass> struct verify_epsilon {
+  using T = comp_t<T1, T2, OpClass>;
+  T epsilon;
+  verify_epsilon(T epsilon) : epsilon(epsilon) {}
+
+  bool operator()(T res, T gold, OpClass op) {
+    if constexpr (std::is_same_v<OpClass, esimd_test::BinaryOp>) {
+      if (op == esimd_test::BinaryOp::div) {
+        return EQ(res, gold, epsilon);
+      }
+    }
+    return res == gold;
+  }
+};
+
+template <class T1, class T2, class OpClass> struct verify_n {
+  using T = comp_t<T1, T2, OpClass>;
+  int n;
+  verify_n(int n) : n(n) {}
+
+  bool operator()(T res, T gold, OpClass op) {
+    using Tint = esimd_test::int_type_t<sizeof(T)>;
+    Tint res_bits = *(Tint *)&res;
+    Tint gold_bits = *(Tint *)&gold;
+    return (abs(gold_bits - res_bits) > n) ? false : true;
+  }
+};
+
+// Flavours of initialization function types.
+
+template <class T1, class T2, class OpClass> struct init_default {
+  using T = comp_t<T1, T2, OpClass>;
+
+  void operator()(T1 *A, T2 *B, T *C, int i) {
+    A[i] = (i % 3) * 90 + 10; /*10, 100, 190, 10, ...*/
+    if constexpr (std::is_unsigned_v<T2>) {
+      B[i] = (i % 3) * 99 + 1 /*1, 100, 199, 1, ...*/;
+    } else {
+      B[i] = (i % 4) * 180 - 170; /*-170, 10, 190, 370, -170,...*/
+    }
+    C[i] = 0;
+  }
+};
+
+template <class T1, class T2, class OpClass> struct init_for_shift {
+  using T = comp_t<T1, T2, OpClass>;
+
+  void operator()(T1 *A, T2 *B, T *C, int i) {
+    if constexpr (std::is_unsigned_v<T1>) {
+      A[i] = (i % 3) + 100; /*100, 101, 102, 100, ...*/
+    } else {
+      A[i] = (i % 4) * 100 - 150; /*-150, -50, 50, 150, -150, ...*/
+    }
+    B[i] = (i % 3);
+    C[i] = 0;
+  }
+};
+
+// shortcuts for less clutter
+template <class T1, class T2, class C> using VSf = verify_strict<T1, T2, C>;
+template <class T1, class T2, class C> using VEf = verify_epsilon<T1, T2, C>;
+template <class T1, class T2, class C> using VNf = verify_n<T1, T2, C>;
+template <class T1, class T2, class C> using IDf = init_default<T1, T2, C>;
+template <class T1, class T2, class C> using ISf = init_for_shift<T1, T2, C>;
+
+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;
+  using BinOp = esimd_test::BinaryOp;
+
+  auto arith_ops = esimd_test::ArithBinaryOps;
+  passed &= test<unsigned char, int, 1, BinOp, VSf, IDf>(arith_ops, q);
+  passed &= test<char, float, 7, BinOp, VEf, IDf>(arith_ops, q, 0.000001f);
+  passed &= test<short, double, 7, BinOp, VSf, IDf>(arith_ops, q);
+  passed &= test<float, float, 32, BinOp, VEf, IDf>(arith_ops, q, 0.000001f);
+  passed &= test<half, char, 1, BinOp, verify_n, IDf>(arith_ops, q, 1);
+  passed &= test<half, unsigned int, 32, BinOp, VSf, IDf>(arith_ops, q, 1);
+  passed &= test<double, half, 7, BinOp, VSf, IDf>(arith_ops, q);
+  passed &= test<short, uint64_t, 7, BinOp, VSf, IDf>(arith_ops, q);
+
+  auto int_ops =
+      esimd_test::IntBinaryOpsNoShift; // different data needed for shift
+  passed &= test<unsigned char, unsigned int, 1, BinOp, VSf, IDf>(int_ops, q);
+  passed &= test<char, uint64_t, 1, BinOp, VSf, IDf>(int_ops, q);
+  passed &= test<uint64_t, char, 32, BinOp, VSf, IDf>(int_ops, q);
+  passed &= test<int, short, 1, BinOp, VSf, IDf>(int_ops, q);
+  passed &= test<short, int, 8, BinOp, VSf, IDf>(int_ops, q);
+  passed &= test<int, int, 7, BinOp, VSf, IDf>(int_ops, q);
+
+  auto sh_ops = esimd_test::BinaryOpSeq<BinOp::shl, BinOp::shr>{};
+  passed &= test<unsigned char, unsigned int, 1, BinOp, VSf, ISf>(sh_ops, q);
+  passed &= test<char, int64_t, 1, BinOp, VSf, ISf>(sh_ops, q);
+  passed &= test<uint64_t, char, 32, BinOp, VSf, ISf>(sh_ops, q);
+  passed &= test<int, short, 1, BinOp, VSf, ISf>(sh_ops, q);
+  passed &= test<short, int, 8, BinOp, VSf, ISf>(sh_ops, q);
+  passed &= test<int, int, 7, BinOp, VSf, ISf>(sh_ops, q);
+
+  using CmpOp = esimd_test::CmpOp;
+  auto cmp_ops = esimd_test::CmpOps;
+  passed &= test<unsigned char, int, 1, CmpOp, VSf, IDf>(cmp_ops, q);
+  passed &= test<char, float, 7, CmpOp, VSf, IDf>(cmp_ops, q);
+  passed &= test<short, double, 7, CmpOp, VSf, IDf>(cmp_ops, q);
+  passed &= test<float, float, 32, CmpOp, VSf, IDf>(cmp_ops, q);
+  passed &= test<half, char, 1, CmpOp, VSf, IDf>(cmp_ops, q, 1);
+  passed &= test<half, unsigned int, 32, CmpOp, VSf, IDf>(cmp_ops, q, 1);
+  passed &= test<double, half, 7, CmpOp, VSf, IDf>(cmp_ops, q);
+  passed &= test<short, uint64_t, 7, CmpOp, VSf, IDf>(cmp_ops, q);
+
+  std::cout << (passed ? "Test PASSED\n" : "Test FAILED\n");
+  return passed ? 0 : 1;
+}