numba-cuda 0.21.1__cp313-cp313-win_amd64.whl

numba_cuda/numba/cuda/tests/cudapy/test_intrinsics.py

@@ -0,0 +1,1219 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: BSD-2-Clause
+
+import itertools
+import numpy as np
+import operator
+import re
+from numba import cuda
+from numba.cuda import int64
+from numba.cuda import HAS_NUMBA
+
+if HAS_NUMBA:
+    from numba.core.errors import TypingError
+else:
+    from numba.cuda.core.errors import TypingError
+from numba.cuda.types import f2
+from numba.cuda.testing import (
+    unittest,
+    CUDATestCase,
+    skip_on_cudasim,
+    skip_unless_cc_53,
+    skip_if_nvjitlink_missing,
+)
+
+
+def simple_threadidx(ary):
+    i = cuda.threadIdx.x
+    ary[0] = i
+
+
+def fill_threadidx(ary):
+    i = cuda.threadIdx.x
+    ary[i] = i
+
+
+def fill3d_threadidx(ary):
+    i = cuda.threadIdx.x
+    j = cuda.threadIdx.y
+    k = cuda.threadIdx.z
+
+    ary[i, j, k] = (i + 1) * (j + 1) * (k + 1)
+
+
+def simple_grid1d(ary):
+    i = cuda.grid(1)
+    ary[i] = i
+
+
+def simple_grid2d(ary):
+    i, j = cuda.grid(2)
+    ary[i, j] = i + j
+
+
+def simple_gridsize1d(ary):
+    i = cuda.grid(1)
+    x = cuda.gridsize(1)
+    if i == 0:
+        ary[0] = x
+
+
+def simple_gridsize2d(ary):
+    i, j = cuda.grid(2)
+    x, y = cuda.gridsize(2)
+    if i == 0 and j == 0:
+        ary[0] = x
+        ary[1] = y
+
+
+def intrinsic_forloop_step(c):
+    startX, startY = cuda.grid(2)
+    gridX = cuda.gridDim.x * cuda.blockDim.x
+    gridY = cuda.gridDim.y * cuda.blockDim.y
+    height, width = c.shape
+
+    for x in range(startX, width, gridX):
+        for y in range(startY, height, gridY):
+            c[y, x] = x + y
+
+
+def simple_popc(ary, c):
+    ary[0] = cuda.popc(c)
+
+
+def simple_bit_count(ary, c):
+    ary[0] = c.bit_count()
+
+
+def simple_fma(ary, a, b, c):
+    ary[0] = cuda.fma(a, b, c)
+
+
+def simple_hadd(ary, a, b):
+    ary[0] = cuda.fp16.hadd(a[0], b[0])
+
+
+def simple_hadd_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hadd(a, b)
+
+
+def simple_hfma(ary, a, b, c):
+    ary[0] = cuda.fp16.hfma(a[0], b[0], c[0])
+
+
+def simple_hfma_scalar(ary, a, b, c):
+    ary[0] = cuda.fp16.hfma(a, b, c)
+
+
+def simple_hsub(ary, a, b):
+    ary[0] = cuda.fp16.hsub(a[0], b[0])
+
+
+def simple_hsub_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hsub(a, b)
+
+
+def simple_hmul(ary, a, b):
+    ary[0] = cuda.fp16.hmul(a[0], b[0])
+
+
+def simple_hmul_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hmul(a, b)
+
+
+def simple_hdiv_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hdiv(a, b)
+
+
+def simple_hdiv_kernel(ary, array_a, array_b):
+    i = cuda.grid(1)
+    if i < ary.size:
+        a = array_a[i]
+        b = array_b[i]
+        ary[i] = cuda.fp16.hdiv(a, b)
+
+
+def simple_hneg(ary, a):
+    ary[0] = cuda.fp16.hneg(a[0])
+
+
+def simple_hneg_scalar(ary, a):
+    ary[0] = cuda.fp16.hneg(a)
+
+
+def simple_habs(ary, a):
+    ary[0] = cuda.fp16.habs(a[0])
+
+
+def simple_habs_scalar(ary, a):
+    ary[0] = cuda.fp16.habs(a)
+
+
+def simple_heq_scalar(ary, a, b):
+    ary[0] = cuda.fp16.heq(a, b)
+
+
+def simple_hne_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hne(a, b)
+
+
+def simple_hge_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hge(a, b)
+
+
+def simple_hgt_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hgt(a, b)
+
+
+def simple_hle_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hle(a, b)
+
+
+def simple_hlt_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hlt(a, b)
+
+
+@cuda.jit(device=True)
+def hlt_func_1(x, y):
+    return cuda.fp16.hlt(x, y)
+
+
+@cuda.jit(device=True)
+def hlt_func_2(x, y):
+    return cuda.fp16.hlt(x, y)
+
+
+def multiple_hcmp_1(r, a, b, c):
+    # float16 predicates used in two separate functions
+    r[0] = hlt_func_1(a, b) and hlt_func_2(b, c)
+
+
+def multiple_hcmp_2(r, a, b, c):
+    # The same float16 predicate used in the caller and callee
+    r[0] = hlt_func_1(a, b) and cuda.fp16.hlt(b, c)
+
+
+def multiple_hcmp_3(r, a, b, c):
+    # Different float16 predicates used in the caller and callee
+    r[0] = hlt_func_1(a, b) and cuda.fp16.hge(c, b)
+
+
+def multiple_hcmp_4(r, a, b, c):
+    # The same float16 predicates used twice in a function
+    r[0] = cuda.fp16.hlt(a, b) and cuda.fp16.hlt(b, c)
+
+
+def multiple_hcmp_5(r, a, b, c):
+    # Different float16 predicates used in a function
+    r[0] = cuda.fp16.hlt(a, b) and cuda.fp16.hge(c, b)
+
+
+def simple_hmax_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hmax(a, b)
+
+
+def simple_hmin_scalar(ary, a, b):
+    ary[0] = cuda.fp16.hmin(a, b)
+
+
+def simple_hsin(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hsin(x[i])
+
+
+def simple_hcos(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hcos(x[i])
+
+
+def simple_hlog(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hlog(x[i])
+
+
+def simple_hlog2(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hlog2(x[i])
+
+
+def simple_hlog10(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hlog10(x[i])
+
+
+def simple_hexp(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hexp(x[i])
+
+
+def simple_hexp2(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hexp2(x[i])
+
+
+def simple_hsqrt(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hsqrt(x[i])
+
+
+def simple_hrsqrt(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hrsqrt(x[i])
+
+
+def numpy_hrsqrt(x, dtype):
+    return x**-0.5
+
+
+def simple_hceil(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hceil(x[i])
+
+
+def simple_hfloor(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hfloor(x[i])
+
+
+def simple_hrcp(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hrcp(x[i])
+
+
+def simple_htrunc(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.htrunc(x[i])
+
+
+def simple_hrint(r, x):
+    i = cuda.grid(1)
+
+    if i < len(r):
+        r[i] = cuda.fp16.hrint(x[i])
+
+
+def simple_cbrt(ary, a):
+    ary[0] = cuda.cbrt(a)
+
+
+def simple_brev(ary, c):
+    ary[0] = cuda.brev(c)
+
+
+def simple_clz(ary, c):
+    ary[0] = cuda.clz(c)
+
+
+def simple_ffs(ary, c):
+    ary[0] = cuda.ffs(c)
+
+
+def simple_round(ary, c):
+    ary[0] = round(c)
+
+
+def simple_round_to(ary, c, ndigits):
+    ary[0] = round(c, ndigits)
+
+
+def branching_with_ifs(a, b, c):
+    i = cuda.grid(1)
+
+    if a[i] > 4:
+        if b % 2 == 0:
+            a[i] = c[i]
+        else:
+            a[i] = 13
+    else:
+        a[i] = 3
+
+
+def branching_with_selps(a, b, c):
+    i = cuda.grid(1)
+
+    inner = cuda.selp(b % 2 == 0, c[i], 13)
+    a[i] = cuda.selp(a[i] > 4, inner, 3)
+
+
+def simple_laneid(ary):
+    i = cuda.grid(1)
+    ary[i] = cuda.laneid
+
+
+def simple_warpsize(ary):
+    ary[0] = cuda.warpsize
+
+
+def nonliteral_grid(x):
+    cuda.grid(x)
+
+
+def nonliteral_gridsize(x):
+    cuda.gridsize(x)
+
+
+class TestCudaIntrinsic(CUDATestCase):
+    def setUp(self):
+        super().setUp()
+        np.random.seed(0)
+
+    def test_simple_threadidx(self):
+        compiled = cuda.jit("void(int32[:])")(simple_threadidx)
+        ary = np.ones(1, dtype=np.int32)
+        compiled[1, 1](ary)
+        self.assertTrue(ary[0] == 0)
+
+    def test_fill_threadidx(self):
+        compiled = cuda.jit("void(int32[:])")(fill_threadidx)
+        N = 10
+        ary = np.ones(N, dtype=np.int32)
+        exp = np.arange(N, dtype=np.int32)
+        compiled[1, N](ary)
+        self.assertTrue(np.all(ary == exp))
+
+    def test_fill3d_threadidx(self):
+        X, Y, Z = 4, 5, 6
+
+        def c_contigous():
+            compiled = cuda.jit("void(int32[:,:,::1])")(fill3d_threadidx)
+            ary = np.zeros((X, Y, Z), dtype=np.int32)
+            compiled[1, (X, Y, Z)](ary)
+            return ary
+
+        def f_contigous():
+            compiled = cuda.jit("void(int32[::1,:,:])")(fill3d_threadidx)
+            ary = np.asfortranarray(np.zeros((X, Y, Z), dtype=np.int32))
+            compiled[1, (X, Y, Z)](ary)
+            return ary
+
+        c_res = c_contigous()
+        f_res = f_contigous()
+        self.assertTrue(np.all(c_res == f_res))
+
+    @skip_on_cudasim("Cudasim does not check types")
+    def test_nonliteral_grid_error(self):
+        with self.assertRaisesRegex(TypingError, "RequireLiteralValue"):
+            cuda.jit("void(int32)")(nonliteral_grid)
+
+    @skip_on_cudasim("Cudasim does not check types")
+    def test_nonliteral_gridsize_error(self):
+        with self.assertRaisesRegex(TypingError, "RequireLiteralValue"):
+            cuda.jit("void(int32)")(nonliteral_gridsize)
+
+    def test_simple_grid1d(self):
+        compiled = cuda.jit("void(int32[::1])")(simple_grid1d)
+        ntid, nctaid = 3, 7
+        nelem = ntid * nctaid
+        ary = np.empty(nelem, dtype=np.int32)
+        compiled[nctaid, ntid](ary)
+        self.assertTrue(np.all(ary == np.arange(nelem)))
+
+    def test_simple_grid2d(self):
+        compiled = cuda.jit("void(int32[:,::1])")(simple_grid2d)
+        ntid = (4, 3)
+        nctaid = (5, 6)
+        shape = (ntid[0] * nctaid[0], ntid[1] * nctaid[1])
+        ary = np.empty(shape, dtype=np.int32)
+        exp = ary.copy()
+        compiled[nctaid, ntid](ary)
+
+        for i in range(ary.shape[0]):
+            for j in range(ary.shape[1]):
+                exp[i, j] = i + j
+
+        self.assertTrue(np.all(ary == exp))
+
+    def test_simple_gridsize1d(self):
+        compiled = cuda.jit("void(int32[::1])")(simple_gridsize1d)
+        ntid, nctaid = 3, 7
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[nctaid, ntid](ary)
+        self.assertEqual(ary[0], nctaid * ntid)
+
+    @skip_on_cudasim("Requires too many threads")
+    def test_issue_9229(self):
+        # Ensure that grid and grid size are correct - #9229 showed that they
+        # overflowed an int32.
+        @cuda.jit
+        def f(grid_error, gridsize_error):
+            i1 = cuda.grid(1)
+            i2 = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.x
+            gs1 = cuda.gridsize(1)
+            gs2 = cuda.blockDim.x * cuda.gridDim.x
+            if i1 != i2:
+                grid_error[0] = 1
+            if gs1 != gs2:
+                gridsize_error[0] = 1
+
+        grid_error = np.zeros(1, dtype=np.uint64)
+        gridsize_error = np.zeros(1, dtype=np.uint64)
+
+        # A large enough grid for thread IDs to overflow an int32
+        # (22121216 * 256 = 5663031296, which is greater than 2 ** 32)
+        f[22121216, 256](grid_error, gridsize_error)
+
+        self.assertEqual(grid_error[0], 0)
+        self.assertEqual(gridsize_error[0], 0)
+
+    @skip_on_cudasim("Tests PTX emission")
+    def test_selp(self):
+        sig = (int64[:], int64, int64[:])
+        cu_branching_with_ifs = cuda.jit(sig)(branching_with_ifs)
+        cu_branching_with_selps = cuda.jit(sig)(branching_with_selps)
+
+        n = 32
+        b = 6
+        c = np.full(shape=32, fill_value=17, dtype=np.int64)
+
+        expected = c.copy()
+        expected[:5] = 3
+
+        a = np.arange(n, dtype=np.int64)
+        cu_branching_with_ifs[n, 1](a, b, c)
+        ptx = cu_branching_with_ifs.inspect_asm(sig)
+        self.assertEqual(2, len(re.findall(r"\s+bra\s+", ptx)))
+        np.testing.assert_array_equal(a, expected, err_msg="branching")
+
+        a = np.arange(n, dtype=np.int64)
+        cu_branching_with_selps[n, 1](a, b, c)
+        ptx = cu_branching_with_selps.inspect_asm(sig)
+        self.assertEqual(0, len(re.findall(r"\s+bra\s+", ptx)))
+        np.testing.assert_array_equal(a, expected, err_msg="selp")
+
+    def test_simple_gridsize2d(self):
+        compiled = cuda.jit("void(int32[::1])")(simple_gridsize2d)
+        ntid = (4, 3)
+        nctaid = (5, 6)
+        ary = np.zeros(2, dtype=np.int32)
+        compiled[nctaid, ntid](ary)
+
+        self.assertEqual(ary[0], nctaid[0] * ntid[0])
+        self.assertEqual(ary[1], nctaid[1] * ntid[1])
+
+    def test_intrinsic_forloop_step(self):
+        compiled = cuda.jit("void(int32[:,::1])")(intrinsic_forloop_step)
+        ntid = (4, 3)
+        nctaid = (5, 6)
+        shape = (ntid[0] * nctaid[0], ntid[1] * nctaid[1])
+        ary = np.empty(shape, dtype=np.int32)
+
+        compiled[nctaid, ntid](ary)
+
+        gridX, gridY = shape
+        height, width = ary.shape
+        for i, j in zip(range(ntid[0]), range(ntid[1])):
+            startX, startY = gridX + i, gridY + j
+            for x in range(startX, width, gridX):
+                for y in range(startY, height, gridY):
+                    self.assertTrue(ary[y, x] == x + y, (ary[y, x], x + y))
+
+    def test_3dgrid(self):
+        @cuda.jit
+        def foo(out):
+            x, y, z = cuda.grid(3)
+            a, b, c = cuda.gridsize(3)
+            out[x, y, z] = a * b * c
+
+        arr = np.zeros(9**3, dtype=np.int32).reshape(9, 9, 9)
+        foo[(3, 3, 3), (3, 3, 3)](arr)
+
+        np.testing.assert_equal(arr, 9**3)
+
+    def test_3dgrid_2(self):
+        @cuda.jit
+        def foo(out):
+            x, y, z = cuda.grid(3)
+            a, b, c = cuda.gridsize(3)
+            grid_is_right = (
+                x == cuda.threadIdx.x + cuda.blockIdx.x * cuda.blockDim.x
+                and y == cuda.threadIdx.y + cuda.blockIdx.y * cuda.blockDim.y
+                and z == cuda.threadIdx.z + cuda.blockIdx.z * cuda.blockDim.z
+            )
+            gridsize_is_right = (
+                a == cuda.blockDim.x * cuda.gridDim.x
+                and b == cuda.blockDim.y * cuda.gridDim.y
+                and c == cuda.blockDim.z * cuda.gridDim.z
+            )
+            out[x, y, z] = grid_is_right and gridsize_is_right
+
+        x, y, z = (4 * 3, 3 * 2, 2 * 4)
+        arr = np.zeros((x * y * z), dtype=np.bool_).reshape(x, y, z)
+        foo[(4, 3, 2), (3, 2, 4)](arr)
+
+        self.assertTrue(np.all(arr))
+
+    def test_popc_u1(self):
+        compiled = cuda.jit("void(int32[:], uint8)")(simple_popc)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint8(0xFF))
+        self.assertEqual(ary[0], 8)
+
+    def test_popc_u2(self):
+        compiled = cuda.jit("void(int32[:], uint16)")(simple_popc)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint16(0xFFFF))
+        self.assertEqual(ary[0], 16)
+
+    def test_popc_u4(self):
+        compiled = cuda.jit("void(int32[:], uint32)")(simple_popc)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint32(0xFFFFFFFF))
+        self.assertEqual(ary[0], 32)
+
+    def test_popc_u8(self):
+        compiled = cuda.jit("void(int32[:], uint64)")(simple_popc)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint64(0xFFFFFFFFFFFFFFFF))
+        self.assertEqual(ary[0], 64)
+
+    def test_bit_count_u1(self):
+        compiled = cuda.jit("void(int32[:], uint8)")(simple_bit_count)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint8(0xFF))
+        self.assertEqual(ary[0], 8)
+
+    def test_bit_count_u2(self):
+        compiled = cuda.jit("void(int32[:], uint16)")(simple_bit_count)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint16(0xFFFF))
+        self.assertEqual(ary[0], 16)
+
+    def test_bit_count_u4(self):
+        compiled = cuda.jit("void(int32[:], uint32)")(simple_bit_count)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint32(0xFFFFFFFF))
+        self.assertEqual(ary[0], 32)
+
+    def test_bit_count_u8(self):
+        compiled = cuda.jit("void(int32[:], uint64)")(simple_bit_count)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, np.uint64(0xFFFFFFFFFFFFFFFF))
+        self.assertEqual(ary[0], 64)
+
+    def test_fma_f4(self):
+        compiled = cuda.jit("void(f4[:], f4, f4, f4)")(simple_fma)
+        ary = np.zeros(1, dtype=np.float32)
+        compiled[1, 1](ary, 2.0, 3.0, 4.0)
+        np.testing.assert_allclose(ary[0], 2 * 3 + 4)
+
+    def test_fma_f8(self):
+        compiled = cuda.jit("void(f8[:], f8, f8, f8)")(simple_fma)
+        ary = np.zeros(1, dtype=np.float64)
+        compiled[1, 1](ary, 2.0, 3.0, 4.0)
+        np.testing.assert_allclose(ary[0], 2 * 3 + 4)
+
+    @skip_unless_cc_53
+    def test_hadd(self):
+        compiled = cuda.jit("void(f2[:], f2[:], f2[:])")(simple_hadd)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.array([3.0], dtype=np.float16)
+        arg2 = np.array([4.0], dtype=np.float16)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg1 + arg2, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hadd_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)")(simple_hadd_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.1415926)
+        arg2 = np.float16(3.0)
+        compiled[1, 1](ary, arg1, arg2)
+        ref = arg1 + arg2
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_on_cudasim("Compilation unsupported in the simulator")
+    @skip_if_nvjitlink_missing("Numbast generated bindings")
+    def test_hadd_ptx(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)", lto=True)(simple_hadd_scalar)
+        args = (f2[:], f2, f2)
+        ptx = compiled.inspect_lto_ptx(args)
+        self.assertIn("add.f16", ptx)
+
+    @skip_unless_cc_53
+    def test_hfma(self):
+        compiled = cuda.jit("void(f2[:], f2[:], f2[:], f2[:])")(simple_hfma)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.array([2.0], dtype=np.float16)
+        arg2 = np.array([3.0], dtype=np.float16)
+        arg3 = np.array([4.0], dtype=np.float16)
+        compiled[1, 1](ary, arg1, arg2, arg3)
+        np.testing.assert_allclose(
+            ary[0], arg1 * arg2 + arg3, rtol=self.FLOAT16_RTOL
+        )
+
+    @skip_unless_cc_53
+    def test_hfma_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2, f2, f2)")(simple_hfma_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(2.0)
+        arg2 = np.float16(3.0)
+        arg3 = np.float16(4.0)
+        compiled[1, 1](ary, arg1, arg2, arg3)
+        ref = arg1 * arg2 + arg3
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_on_cudasim("Compilation unsupported in the simulator")
+    @skip_if_nvjitlink_missing("Numbast generated bindings")
+    def test_hfma_ptx(self):
+        compiled = cuda.jit("void(f2[:], f2, f2, f2)", lto=True)(
+            simple_hfma_scalar
+        )
+        args = (f2[:], f2, f2, f2)
+        ptx = compiled.inspect_lto_ptx(args)
+        self.assertIn("fma.rn.f16", ptx)
+
+    @skip_unless_cc_53
+    def test_hsub(self):
+        compiled = cuda.jit("void(f2[:], f2[:], f2[:])")(simple_hsub)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.array([3.0], dtype=np.float16)
+        arg2 = np.array([4.0], dtype=np.float16)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg1 - arg2, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hsub_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)")(simple_hsub_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.1415926)
+        arg2 = np.float16(1.57)
+        compiled[1, 1](ary, arg1, arg2)
+        ref = arg1 - arg2
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_on_cudasim("Compilation unsupported in the simulator")
+    @skip_if_nvjitlink_missing("Numbast generated bindings")
+    def test_hsub_ptx(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)", lto=True)(simple_hsub_scalar)
+        args = (f2[:], f2, f2)
+        ptx = compiled.inspect_lto_ptx(args)
+        self.assertIn("sub.f16", ptx)
+
+    @skip_unless_cc_53
+    def test_hmul(self):
+        compiled = cuda.jit(simple_hmul)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.array([3.0], dtype=np.float16)
+        arg2 = np.array([4.0], dtype=np.float16)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg1 * arg2, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hmul_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)")(simple_hmul_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.1415926)
+        arg2 = np.float16(1.57)
+        compiled[1, 1](ary, arg1, arg2)
+        ref = arg1 * arg2
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_on_cudasim("Compilation unsupported in the simulator")
+    @skip_if_nvjitlink_missing("Numbast generated bindings")
+    def test_hmul_ptx(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)", lto=True)(simple_hmul_scalar)
+        args = (f2[:], f2, f2)
+        ptx = compiled.inspect_lto_ptx(args)
+        self.assertIn("mul.f16", ptx)
+
+    @skip_unless_cc_53
+    def test_hdiv_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)")(simple_hdiv_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.1415926)
+        arg2 = np.float16(1.57)
+
+        compiled[1, 1](ary, arg1, arg2)
+        ref = arg1 / arg2
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hdiv(self):
+        compiled = cuda.jit("void(f2[:], f2[:], f2[:])")(simple_hdiv_kernel)
+        arry1 = np.random.randint(-65504, 65505, size=500).astype(np.float16)
+        arry2 = np.random.randint(-65504, 65505, size=500).astype(np.float16)
+        ary = np.zeros_like(arry1, dtype=np.float16)
+
+        compiled.forall(ary.size)(ary, arry1, arry2)
+        ref = arry1 / arry2
+        np.testing.assert_allclose(ary, ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hneg(self):
+        compiled = cuda.jit("void(f2[:], f2[:])")(simple_hneg)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.array([3.0], dtype=np.float16)
+        compiled[1, 1](ary, arg1)
+        np.testing.assert_allclose(ary[0], -arg1, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hneg_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2)")(simple_hneg_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.1415926)
+        compiled[1, 1](ary, arg1)
+        ref = -arg1
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_on_cudasim("Compilation unsupported in the simulator")
+    @skip_if_nvjitlink_missing("Numbast generated bindings")
+    def test_hneg_ptx(self):
+        compiled = cuda.jit("void(f2[:], f2)", lto=True)(simple_hneg_scalar)
+        args = (f2[:], f2)
+        ptx = compiled.inspect_lto_ptx(args)
+        self.assertIn("neg.f16", ptx)
+
+    @skip_unless_cc_53
+    def test_habs(self):
+        compiled = cuda.jit(simple_habs)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.array([-3.0], dtype=np.float16)
+        compiled[1, 1](ary, arg1)
+        np.testing.assert_allclose(ary[0], abs(arg1), rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_habs_scalar(self):
+        compiled = cuda.jit("void(f2[:], f2)")(simple_habs_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(-3.1415926)
+        compiled[1, 1](ary, arg1)
+        ref = abs(arg1)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
+
+    @skip_on_cudasim("Compilation unsupported in the simulator")
+    @skip_if_nvjitlink_missing("Numbast generated bindings")
+    def test_habs_ptx(self):
+        compiled = cuda.jit("void(f2[:], f2)", lto=True)(simple_habs_scalar)
+        args = (f2[:], f2)
+        ptx = compiled.inspect_lto_ptx(args)
+        self.assertIn("abs.f16", ptx)
+
+    @skip_unless_cc_53
+    def test_fp16_intrinsics_common(self):
+        kernels = (
+            simple_hsin,
+            simple_hcos,
+            simple_hlog,
+            simple_hlog2,
+            simple_hlog10,
+            simple_hsqrt,
+            simple_hceil,
+            simple_hfloor,
+            simple_hrcp,
+            simple_htrunc,
+            simple_hrint,
+            simple_hrsqrt,
+        )
+        exp_kernels = (simple_hexp, simple_hexp2)
+        expected_functions = (
+            np.sin,
+            np.cos,
+            np.log,
+            np.log2,
+            np.log10,
+            np.sqrt,
+            np.ceil,
+            np.floor,
+            np.reciprocal,
+            np.trunc,
+            np.rint,
+            numpy_hrsqrt,
+        )
+        expected_exp_functions = (np.exp, np.exp2)
+
+        # Generate random data
+        N = 32
+        np.random.seed(1)
+        x = np.random.randint(1, 65505, size=N).astype(np.float16)
+        r = np.zeros_like(x)
+        for kernel, fn in zip(kernels, expected_functions):
+            with self.subTest(fn=fn):
+                kernel = cuda.jit("void(f2[:], f2[:])")(kernel)
+                kernel[1, N](r, x)
+                expected = fn(x, dtype=np.float16)
+                np.testing.assert_allclose(r, expected, rtol=self.FLOAT16_RTOL)
+
+        x2 = np.random.randint(1, 10, size=N).astype(np.float16)
+        for kernel, fn in zip(exp_kernels, expected_exp_functions):
+            with self.subTest(fn=fn):
+                kernel = cuda.jit("void(f2[:], f2[:])")(kernel)
+                kernel[1, N](r, x2)
+                expected = fn(x2, dtype=np.float16)
+                np.testing.assert_allclose(r, expected, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hexp10(self):
+        @cuda.jit()
+        def hexp10_vectors(r, x):
+            i = cuda.grid(1)
+
+            if i < len(r):
+                r[i] = cuda.fp16.hexp10(x[i])
+
+        # Generate random data
+        N = 32
+        np.random.seed(1)
+        x = np.random.rand(N).astype(np.float16)
+        r = np.zeros_like(x)
+
+        # Run the kernel
+        hexp10_vectors[1, N](r, x)
+        np.testing.assert_allclose(r, 10**x, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_fp16_comparison(self):
+        fns = (
+            simple_heq_scalar,
+            simple_hne_scalar,
+            simple_hge_scalar,
+            simple_hgt_scalar,
+            simple_hle_scalar,
+            simple_hlt_scalar,
+        )
+        ops = (
+            operator.eq,
+            operator.ne,
+            operator.ge,
+            operator.gt,
+            operator.le,
+            operator.lt,
+        )
+
+        for fn, op in zip(fns, ops):
+            with self.subTest(op=op):
+                kernel = cuda.jit("void(b1[:], f2, f2)")(fn)
+
+                expected = np.zeros(1, dtype=np.bool_)
+                got = np.zeros(1, dtype=np.bool_)
+                arg2 = np.float16(2)
+                arg3 = np.float16(3)
+                arg4 = np.float16(4)
+
+                # Check with equal arguments
+                kernel[1, 1](got, arg3, arg3)
+                expected = op(arg3, arg3)
+                self.assertEqual(expected, got[0])
+
+                # Check with LHS < RHS
+                kernel[1, 1](got, arg3, arg4)
+                expected = op(arg3, arg4)
+                self.assertEqual(expected, got[0])
+
+                # Check with LHS > RHS
+                kernel[1, 1](got, arg3, arg2)
+                expected = op(arg3, arg2)
+                self.assertEqual(expected, got[0])
+
+    @skip_unless_cc_53
+    def test_multiple_float16_comparisons(self):
+        functions = (
+            multiple_hcmp_1,
+            multiple_hcmp_2,
+            multiple_hcmp_3,
+            multiple_hcmp_4,
+            multiple_hcmp_5,
+        )
+        for fn in functions:
+            with self.subTest(fn=fn):
+                compiled = cuda.jit("void(b1[:], f2, f2, f2)")(fn)
+                ary = np.zeros(1, dtype=np.bool_)
+                arg1 = np.float16(2.0)
+                arg2 = np.float16(3.0)
+                arg3 = np.float16(4.0)
+                compiled[1, 1](ary, arg1, arg2, arg3)
+                self.assertTrue(ary[0])
+
+    @skip_unless_cc_53
+    def test_hmax(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)")(simple_hmax_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.0)
+        arg2 = np.float16(4.0)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg2, rtol=self.FLOAT16_RTOL)
+        arg1 = np.float16(5.0)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg1, rtol=self.FLOAT16_RTOL)
+
+    @skip_unless_cc_53
+    def test_hmin(self):
+        compiled = cuda.jit("void(f2[:], f2, f2)")(simple_hmin_scalar)
+        ary = np.zeros(1, dtype=np.float16)
+        arg1 = np.float16(3.0)
+        arg2 = np.float16(4.0)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg1, rtol=self.FLOAT16_RTOL)
+        arg1 = np.float16(5.0)
+        compiled[1, 1](ary, arg1, arg2)
+        np.testing.assert_allclose(ary[0], arg2, rtol=self.FLOAT16_RTOL)
+
+    def test_cbrt_f32(self):
+        compiled = cuda.jit("void(float32[:], float32)")(simple_cbrt)
+        ary = np.zeros(1, dtype=np.float32)
+        cbrt_arg = 2.0
+        compiled[1, 1](ary, cbrt_arg)
+        np.testing.assert_allclose(ary[0], cbrt_arg ** (1 / 3))
+
+    def test_cbrt_f64(self):
+        compiled = cuda.jit("void(float64[:], float64)")(simple_cbrt)
+        ary = np.zeros(1, dtype=np.float64)
+        cbrt_arg = 6.0
+        compiled[1, 1](ary, cbrt_arg)
+        np.testing.assert_allclose(ary[0], cbrt_arg ** (1 / 3))
+
+    def test_brev_u4(self):
+        compiled = cuda.jit("void(uint32[:], uint32)")(simple_brev)
+        ary = np.zeros(1, dtype=np.uint32)
+        compiled[1, 1](ary, 0x000030F0)
+        self.assertEqual(ary[0], 0x0F0C0000)
+
+    @skip_on_cudasim('only get given a Python "int", assumes 32 bits')
+    def test_brev_u8(self):
+        compiled = cuda.jit("void(uint64[:], uint64)")(simple_brev)
+        ary = np.zeros(1, dtype=np.uint64)
+        compiled[1, 1](ary, 0x000030F0000030F0)
+        self.assertEqual(ary[0], 0x0F0C00000F0C0000)
+
+    def test_clz_i4(self):
+        compiled = cuda.jit("void(int32[:], int32)")(simple_clz)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x00100000)
+        self.assertEqual(ary[0], 11)
+
+    def test_clz_u4(self):
+        """
+        Although the CUDA Math API
+        (http://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH__INTRINSIC__INT.html)
+        only says int32 & int64 arguments are supported in C code, the LLVM
+        IR input supports i8, i16, i32 & i64 (LLVM doesn't have a concept of
+        unsigned integers, just unsigned operations on integers).
+        http://docs.nvidia.com/cuda/nvvm-ir-spec/index.html#bit-manipulations-intrinics
+        """
+        compiled = cuda.jit("void(int32[:], uint32)")(simple_clz)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x00100000)
+        self.assertEqual(ary[0], 11)
+
+    def test_clz_i4_1s(self):
+        compiled = cuda.jit("void(int32[:], int32)")(simple_clz)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0xFFFFFFFF)
+        self.assertEqual(ary[0], 0)
+
+    def test_clz_i4_0s(self):
+        compiled = cuda.jit("void(int32[:], int32)")(simple_clz)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x0)
+        self.assertEqual(ary[0], 32, "CUDA semantics")
+
+    @skip_on_cudasim('only get given a Python "int", assumes 32 bits')
+    def test_clz_i8(self):
+        compiled = cuda.jit("void(int32[:], int64)")(simple_clz)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x000000000010000)
+        self.assertEqual(ary[0], 47)
+
+    def test_ffs_i4(self):
+        compiled = cuda.jit("void(int32[:], int32)")(simple_ffs)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x00100000)
+        self.assertEqual(ary[0], 21)
+        compiled[1, 1](ary, 0x80000000)
+        self.assertEqual(ary[0], 32)
+
+    def test_ffs_u4(self):
+        compiled = cuda.jit("void(int32[:], uint32)")(simple_ffs)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x00100000)
+        self.assertEqual(ary[0], 21)
+        compiled[1, 1](ary, 0x80000000)
+        self.assertEqual(ary[0], 32)
+
+    def test_ffs_i4_1s(self):
+        compiled = cuda.jit("void(int32[:], int32)")(simple_ffs)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0xFFFFFFFF)
+        self.assertEqual(ary[0], 1)
+
+    def test_ffs_i4_0s(self):
+        compiled = cuda.jit("void(int32[:], int32)")(simple_ffs)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x0)
+        self.assertEqual(ary[0], 0)
+
+    @skip_on_cudasim('only get given a Python "int", assumes 32 bits')
+    def test_ffs_i8(self):
+        compiled = cuda.jit("void(int32[:], int64)")(simple_ffs)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary, 0x000000000010000)
+        self.assertEqual(ary[0], 17)
+        compiled[1, 1](ary, 0x100000000)
+        self.assertEqual(ary[0], 33)
+
+    def test_simple_laneid(self):
+        compiled = cuda.jit("void(int32[:])")(simple_laneid)
+        count = 2
+        ary = np.zeros(count * 32, dtype=np.int32)
+        exp = np.tile(np.arange(32, dtype=np.int32), count)
+        compiled[1, count * 32](ary)
+        self.assertTrue(np.all(ary == exp))
+
+    def test_simple_warpsize(self):
+        compiled = cuda.jit("void(int32[:])")(simple_warpsize)
+        ary = np.zeros(1, dtype=np.int32)
+        compiled[1, 1](ary)
+        self.assertEqual(ary[0], 32, "CUDA semantics")
+
+    def test_round_f4(self):
+        compiled = cuda.jit("void(int64[:], float32)")(simple_round)
+        ary = np.zeros(1, dtype=np.int64)
+
+        for i in [-3.0, -2.5, -2.25, -1.5, 1.5, 2.25, 2.5, 2.75]:
+            compiled[1, 1](ary, i)
+            self.assertEqual(ary[0], round(i))
+
+    def test_round_f8(self):
+        compiled = cuda.jit("void(int64[:], float64)")(simple_round)
+        ary = np.zeros(1, dtype=np.int64)
+
+        for i in [-3.0, -2.5, -2.25, -1.5, 1.5, 2.25, 2.5, 2.75]:
+            compiled[1, 1](ary, i)
+            self.assertEqual(ary[0], round(i))
+
+    def test_round_to_f4(self):
+        compiled = cuda.jit("void(float32[:], float32, int32)")(simple_round_to)
+        ary = np.zeros(1, dtype=np.float32)
+        np.random.seed(123)
+        vals = np.random.random(32).astype(np.float32)
+        np.concatenate((vals, np.array([np.inf, -np.inf, np.nan])))
+        digits = (
+            # Common case branch of round_to_impl
+            -5,
+            -4,
+            -3,
+            -2,
+            -1,
+            0,
+            1,
+            2,
+            3,
+            4,
+            5,
+            # The algorithm currently implemented can only round to 13 digits
+            # with single precision. Note that this doesn't trigger the
+            # "overflow safe" branch of the implementation, which can only be
+            # hit when using double precision.
+            13,
+        )
+        for val, ndigits in itertools.product(vals, digits):
+            with self.subTest(val=val, ndigits=ndigits):
+                compiled[1, 1](ary, val, ndigits)
+                self.assertPreciseEqual(
+                    ary[0], round(val, ndigits), prec="single"
+                )
+
+    # CPython on most platforms uses rounding based on dtoa.c, whereas the CUDA
+    # round-to implementation uses CPython's fallback implementation, which has
+    # slightly different behavior at the edges of the domain. Since the CUDA
+    # simulator executes using CPython, we need to skip this test when the
+    # simulator is active.
+    @skip_on_cudasim("Overflow behavior differs on CPython")
+    def test_round_to_f4_overflow(self):
+        # Test that the input value is returned when y in round_ndigits
+        # overflows.
+        compiled = cuda.jit("void(float32[:], float32, int32)")(simple_round_to)
+        ary = np.zeros(1, dtype=np.float32)
+        val = np.finfo(np.float32).max
+        # An unusually large number of digits is required to hit the "y
+        # overflows" branch of the implementation because the typing results in
+        # the computation of y as float64.
+        ndigits = 300
+        compiled[1, 1](ary, val, ndigits)
+        self.assertEqual(ary[0], val)
+
+    def test_round_to_f4_halfway(self):
+        compiled = cuda.jit("void(float32[:], float32, int32)")(simple_round_to)
+        ary = np.zeros(1, dtype=np.float32)
+        # Value chosen to trigger the "round to even" branch of the
+        # implementation
+        val = 0.3425
+        ndigits = 3
+        compiled[1, 1](ary, val, ndigits)
+        self.assertPreciseEqual(ary[0], round(val, ndigits), prec="single")
+
+    def test_round_to_f8(self):
+        compiled = cuda.jit("void(float64[:], float64, int32)")(simple_round_to)
+        ary = np.zeros(1, dtype=np.float64)
+        np.random.seed(123)
+        vals = np.random.random(32)
+        np.concatenate((vals, np.array([np.inf, -np.inf, np.nan])))
+        digits = (-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5)
+
+        for val, ndigits in itertools.product(vals, digits):
+            with self.subTest(val=val, ndigits=ndigits):
+                compiled[1, 1](ary, val, ndigits)
+                self.assertPreciseEqual(
+                    ary[0], round(val, ndigits), prec="exact"
+                )
+
+        # Trigger the "overflow safe" branch of the implementation
+        val = 0.12345678987654321 * 10e-15
+        ndigits = 23
+        with self.subTest(val=val, ndigits=ndigits):
+            compiled[1, 1](ary, val, ndigits)
+            self.assertPreciseEqual(ary[0], round(val, ndigits), prec="double")
+
+    # Skipped on cudasim for the same reasons as test_round_to_f4 above.
+    @skip_on_cudasim("Overflow behavior differs on CPython")
+    def test_round_to_f8_overflow(self):
+        # Test that the input value is returned when y in round_ndigits
+        # overflows.
+        compiled = cuda.jit("void(float64[:], float64, int32)")(simple_round_to)
+        ary = np.zeros(1, dtype=np.float64)
+        val = np.finfo(np.float64).max
+        # Unlike test_round_to_f4_overflow, a reasonable number of digits can
+        # be used for this test to overflow y in round_ndigits.
+        ndigits = 12
+        compiled[1, 1](ary, val, ndigits)
+        self.assertEqual(ary[0], val)
+
+    def test_round_to_f8_halfway(self):
+        compiled = cuda.jit("void(float64[:], float64, int32)")(simple_round_to)
+        ary = np.zeros(1, dtype=np.float64)
+        # Value chosen to trigger the "round to even" branch of the
+        # implementation, with a value that is not exactly representable with a
+        # float32, but only a float64.
+        val = 0.5425
+        ndigits = 3
+        compiled[1, 1](ary, val, ndigits)
+        self.assertPreciseEqual(ary[0], round(val, ndigits), prec="double")
+
+
+if __name__ == "__main__":
+    unittest.main()