test_loopy.py

    knl = lp.tag_inames(knl, dict(i="l.0", j="ilp"))

    knl = lp.preprocess_kernel(knl, ctx.devices[0])
    for k in lp.generate_loop_schedules(knl):
        assert k.temporary_variables["a"].shape == (16, 17)


def test_ilp_write_race_avoidance_private(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[j]: 0<=j<16 }",
            [
                "<> a = 5+j",
                ],
            [])

    knl = lp.tag_inames(knl, dict(j="ilp"))

    knl = lp.preprocess_kernel(knl, ctx.devices[0])
    for k in lp.generate_loop_schedules(knl):
        assert k.temporary_variables["a"].shape == (16,)

# }}}


def test_write_parameter(ctx_factory):
    dtype = np.float32
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n }",
            """
                a = sum((i,j), i*j)
                b = sum(i, sum(j, i*j))
                n = 15
                """,
            [
                lp.GlobalArg("a", dtype, shape=()),
                lp.GlobalArg("b", dtype, shape=()),
                lp.ValueArg("n", np.int32, approximately=1000),
                ],
            assumptions="n>=1")

    import pytest
    with pytest.raises(RuntimeError):
        lp.CompiledKernel(ctx, knl).get_code()


# {{{ arg guessing

def test_arg_shape_guessing(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n }",
            """
                a = 1.5 + sum((i,j), i*j)
                b[i, j] = i*j
                c[i+j, j] = b[j,i]
                """,
            [
                lp.GlobalArg("a", shape=lp.auto),
                lp.GlobalArg("b", shape=lp.auto),
                lp.GlobalArg("c", shape=lp.auto),
                lp.ValueArg("n"),
                ],
            assumptions="n>=1")

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code())


def test_arg_guessing(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n }",
            """
                a = 1.5 + sum((i,j), i*j)
                b[i, j] = i*j
                c[i+j, j] = b[j,i]
                """,
            assumptions="n>=1")

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code())


def test_arg_guessing_with_reduction(ctx_factory):
    #logging.basicConfig(level=logging.DEBUG)
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n }",
            """
                a = 1.5 + sum((i,j), i*j)
                d = 1.5 + sum((i,j), b[i,j])
                b[i, j] = i*j
                c[i+j, j] = b[j,i]
                """,
            assumptions="n>=1")

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code())

# }}}


def test_nonlinear_index(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n }",
            """
                a[i*i] = 17
                """,
            [
                lp.GlobalArg("a", shape="n"),
                lp.ValueArg("n"),
                ],
            assumptions="n>=1")

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code())


def test_triangle_domain(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n and i <= j}",
            "a[i,j] = 17",
            assumptions="n>=1")

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code())


def test_offsets_and_slicing(ctx_factory):
    ctx = ctx_factory()
    queue = cl.CommandQueue(ctx)

    n = 20

    knl = lp.make_kernel(
            "{[i,j]: 0<=i<n and 0<=j<m }",
            """
                b[i,j] = 2*a[i,j]
                """,
            assumptions="n>=1 and m>=1",
            default_offset=lp.auto)

    knl = lp.tag_data_axes(knl, "a,b", "stride:auto,stride:1")

    cknl = lp.CompiledKernel(ctx, knl)

    a_full = cl.clrandom.rand(queue, (n, n), np.float64)
    a_full_h = a_full.get()
    b_full = cl.clrandom.rand(queue, (n, n), np.float64)
    b_full_h = b_full.get()

    a_sub = (slice(3, 10), slice(5, 10))
    a = a_full[a_sub]

    b_sub = (slice(3+3, 10+3), slice(5+4, 10+4))
    b = b_full[b_sub]

    b_full_h[b_sub] = 2*a_full_h[a_sub]

    print(cknl.get_highlighted_code({"a": a.dtype}))
    cknl(queue, a=a, b=b)

    import numpy.linalg as la
    assert la.norm(b_full.get() - b_full_h) < 1e-13


def test_vector_ilp_with_prefetch(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{ [i]: 0<=i<n }",
            "out[i] = 2*a[i]",
            [
                # Tests that comma'd arguments interoperate with
                # argument guessing.
                lp.GlobalArg("out,a", np.float32, shape=lp.auto),
                "..."
                ])

    knl = lp.split_iname(knl, "i", 128, inner_tag="l.0")
    knl = lp.split_iname(knl, "i_outer", 4, outer_tag="g.0", inner_tag="ilp")
    knl = lp.add_prefetch(knl, "a", ["i_inner", "i_outer_inner"])

    cknl = lp.CompiledKernel(ctx, knl)
    cknl.cl_kernel_info()

    import re
    code = cknl.get_code()
    assert len(list(re.finditer("barrier", code))) == 1


# {{{ convolutions

def test_convolution(ctx_factory):
    ctx = ctx_factory()

    dtype = np.float32

    knl = lp.make_kernel(
        "{ [iimg, ifeat, icolor, im_x, im_y, f_x, f_y]: \
                -f_w <= f_x,f_y <= f_w \
                and 0 <= im_x < im_w and 0 <= im_y < im_h \
                and 0<=iimg<=nimgs and 0<=ifeat<nfeats and 0<=icolor<ncolors \
                }",
        """
        out[iimg, ifeat, im_x, im_y] = sum((f_x, f_y, icolor), \
            img[iimg, f_w+im_x-f_x, f_w+im_y-f_y, icolor] \
            * f[ifeat, f_w+f_x, f_w+f_y, icolor])
        """,
        [
            lp.GlobalArg("f", dtype, shape=lp.auto),
            lp.GlobalArg("img", dtype, shape=lp.auto),
            lp.GlobalArg("out", dtype, shape=lp.auto),
            "..."
            ],
        assumptions="f_w>=1 and im_w, im_h >= 2*f_w+1 and nfeats>=1 and nimgs>=0",
        flags="annotate_inames",
        defines=dict(ncolors=3))

    f_w = 3

    knl = lp.fix_parameters(knl, f_w=f_w)

    ref_knl = knl

    def variant_0(knl):
        #knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
        knl = lp.set_loop_priority(knl, "iimg,im_x,im_y,ifeat,f_x,f_y")
        return knl

    def variant_1(knl):
        knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
        knl = lp.set_loop_priority(knl, "iimg,im_x_outer,im_y,ifeat,f_x,f_y")
        return knl

    def variant_2(knl):
        knl = lp.split_iname(knl, "im_x", 16, outer_tag="g.0", inner_tag="l.0")
        knl = lp.split_iname(knl, "im_y", 16, outer_tag="g.1", inner_tag="l.1")
        knl = lp.tag_inames(knl, dict(ifeat="g.2"))
        knl = lp.add_prefetch(knl, "f[ifeat,:,:,:]")
        knl = lp.add_prefetch(knl, "img", "im_x_inner, im_y_inner, f_x, f_y")
        return knl

    for variant in [
            variant_0,
            variant_1,
            variant_2
            ]:
        lp.auto_test_vs_ref(ref_knl, ctx, variant(knl),
                parameters=dict(
                    im_w=128, im_h=128, f_w=f_w,
                    nfeats=3, nimgs=3
                    ))


def test_convolution_with_nonzero_base(ctx_factory):
    # This is kept alive as a test for domains that don't start at zero.
    # These are a bad idea for split_iname, which places its origin at zero
    # and therefore produces a first block that is odd-sized.
    #
    # Therefore, for real tests, check test_convolution further up.

    ctx = ctx_factory()

    dtype = np.float32

    knl = lp.make_kernel(
        "{ [iimg, ifeat, icolor, im_x, im_y, f_x, f_y]: \
                -f_w <= f_x,f_y <= f_w \
                and f_w <= im_x < im_w-f_w and f_w <= im_y < im_h-f_w \
                and 0<=iimg<=nimgs and 0<=ifeat<nfeats and 0<=icolor<ncolors \
                }",
        """
        out[iimg, ifeat, im_x-f_w, im_y-f_w] = sum((f_x, f_y, icolor), \
            img[iimg, im_x-f_x, im_y-f_y, icolor] \
            * f[ifeat, f_w+f_x, f_w+f_y, icolor])
        """,
        [
            lp.GlobalArg("f", dtype, shape=lp.auto),
            lp.GlobalArg("img", dtype, shape=lp.auto),
            lp.GlobalArg("out", dtype, shape=lp.auto),
            "..."
            ],
        assumptions="f_w>=1 and im_w, im_h >= 2*f_w+1 and nfeats>=1 and nimgs>=0",
        flags="annotate_inames",
        defines=dict(ncolors=3))

    ref_knl = knl

    f_w = 3

    def variant_0(knl):
        #knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
        knl = lp.set_loop_priority(knl, "iimg,im_x,im_y,ifeat,f_x,f_y")
        return knl

    def variant_1(knl):
        knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
        knl = lp.set_loop_priority(knl, "iimg,im_x_outer,im_y,ifeat,f_x,f_y")
        return knl

    for variant in [
            variant_0,
            variant_1,
            ]:
        lp.auto_test_vs_ref(ref_knl, ctx, variant(knl),
                parameters=dict(
                    im_w=128, im_h=128, f_w=f_w,
                    nfeats=12, nimgs=17
                    ))

# }}}


def test_c_instruction(ctx_factory):
    #logging.basicConfig(level=logging.DEBUG)
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n }",
            [
                lp.CInstruction("i", """
                    x = sin((float) i);
                    """, assignees="x"),
                "a[i*i] = x",
                ],
            [
                lp.GlobalArg("a", shape="n"),
                lp.ValueArg("n"),
                lp.TemporaryVariable("x", np.float32),
                ],
            assumptions="n>=1")

    knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code())


def test_dependent_domain_insn_iname_finding(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel([
            "{[isrc_box]: 0<=isrc_box<nsrc_boxes}",
            "{[isrc,idim]: isrc_start<=isrc<isrc_end and 0<=idim<dim}",
            ],
            """
                <> src_ibox = source_boxes[isrc_box]
                <> isrc_start = box_source_starts[src_ibox]
                <> isrc_end = isrc_start+box_source_counts_nonchild[src_ibox]
                <> strength = strengths[isrc] {id=set_strength}
                """,
            [
                lp.GlobalArg("box_source_starts,box_source_counts_nonchild",
                    None, shape=None),
                lp.GlobalArg("strengths",
                    None, shape="nsources"),
                "..."])

    print(knl)
    assert "isrc_box" in knl.insn_inames("set_strength")

    print(lp.CompiledKernel(ctx, knl).get_highlighted_code(
            dict(
                source_boxes=np.int32,
                box_source_starts=np.int32,
                box_source_counts_nonchild=np.int32,
                strengths=np.float64,
                nsources=np.int32,
                )))


def test_inames_deps_from_write_subscript(ctx_factory):
    knl = lp.make_kernel(
            "{[i,j]: 0<=i,j<n}",
            """
                <> src_ibox = source_boxes[i]
                <int32> something = 5
                a[src_ibox] = sum(j, something) {id=myred}
                """,
            [
                lp.GlobalArg("box_source_starts,box_source_counts_nonchild,a",
                    None, shape=None),
                "..."])

    print(knl)
    assert "i" in knl.insn_inames("myred")


def test_split_reduction(ctx_factory):
    knl = lp.make_kernel(
            "{[i,j,k]: 0<=i,j,k<n}",
            """
                b = sum((i,j,k), a[i,j,k])
                """,
            [
                lp.GlobalArg("box_source_starts,box_source_counts_nonchild,a",
                    None, shape=None),
                "..."])

    knl = lp.split_reduction_outward(knl, "j,k")
    # FIXME: finish test


def test_modulo_indexing(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{[i,j]: 0<=i<n and 0<=j<5}",
            """
                b[i] = sum(j, a[(i+j)%n])
                """,
            [
                lp.GlobalArg("a", None, shape="n"),
                "..."
                ]
            )

    print(knl)
    print(lp.CompiledKernel(ctx, knl).get_highlighted_code(
            dict(
                a=np.float32,
                )))


def test_rob_stroud_bernstein(ctx_factory):
    ctx = ctx_factory()

    # NOTE: tmp would have to be zero-filled beforehand

    knl = lp.make_kernel(
            "{[el, i2, alpha1,alpha2]: \
                    0 <= el < nels and \
                    0 <= i2 < nqp1d and \
                    0 <= alpha1 <= deg and 0 <= alpha2 <= deg-alpha1 }",
            """
                <> xi = qpts[1, i2] {inames=+el}
                <> s = 1-xi
                <> r = xi/s
                <> aind = 0 {id=aind_init,inames=+i2:el}

                <> w = s**(deg-alpha1) {id=init_w}

                tmp[el,alpha1,i2] = tmp[el,alpha1,i2] + w * coeffs[aind] \
                        {id=write_tmp,inames=+alpha2}
                w = w * r * ( deg - alpha1 - alpha2 ) / (1 + alpha2) \
                        {id=update_w,dep=init_w:write_tmp}
                aind = aind + 1 \
                        {id=aind_incr,\
                        dep=aind_init:write_tmp:update_w, \
                        inames=+el:i2:alpha1:alpha2}
                """,
            [
                # Must declare coeffs to have "no" shape, to keep loopy
                # from trying to figure it out the shape automatically.

                lp.GlobalArg("coeffs", None, shape=None),
                "..."
                ],
            assumptions="deg>=0 and nels>=1"
            )

    knl = lp.fix_parameters(knl, nqp1d=7, deg=4)
    knl = lp.split_iname(knl, "el", 16, inner_tag="l.0")
    knl = lp.split_iname(knl, "el_outer", 2, outer_tag="g.0", inner_tag="ilp",
            slabs=(0, 1))
    knl = lp.tag_inames(knl, dict(i2="l.1", alpha1="unr", alpha2="unr"))

    print(lp.CompiledKernel(ctx, knl).get_highlighted_code(
            dict(
                qpts=np.float32,
                coeffs=np.float32,
                tmp=np.float32,
                )))


def test_rob_stroud_bernstein_full(ctx_factory):
    #logging.basicConfig(level=logging.DEBUG)
    ctx = ctx_factory()

    # NOTE: result would have to be zero-filled beforehand

    knl = lp.make_kernel(
            "{[el, i2, alpha1,alpha2, i1_2, alpha1_2, i2_2]: \
                    0 <= el < nels and \
                    0 <= i2 < nqp1d and \
                    0 <= alpha1 <= deg and 0 <= alpha2 <= deg-alpha1 and\
                    \
                    0 <= i1_2 < nqp1d and \
                    0 <= alpha1_2 <= deg and \
                    0 <= i2_2 < nqp1d \
                    }",
            """
                <> xi = qpts[1, i2] {inames=+el}
                <> s = 1-xi
                <> r = xi/s
                <> aind = 0 {id=aind_init,inames=+i2:el}

                <> w = s**(deg-alpha1) {id=init_w}

                <> tmp[alpha1,i2] = tmp[alpha1,i2] + w * coeffs[aind] \
                        {id=write_tmp,inames=+alpha2}
                w = w * r * ( deg - alpha1 - alpha2 ) / (1 + alpha2) \
                        {id=update_w,dep=init_w:write_tmp}
                aind = aind + 1 \
                        {id=aind_incr,\
                        dep=aind_init:write_tmp:update_w, \
                        inames=+el:i2:alpha1:alpha2}

                <> xi2 = qpts[0, i1_2] {dep=aind_incr,inames=+el}
                <> s2 = 1-xi2
                <> r2 = xi2/s2
                <> w2 = s2**deg

                result[el, i1_2, i2_2] = result[el, i1_2, i2_2] + \
                        w2 * tmp[alpha1_2, i2_2] \
                        {inames=el:alpha1_2:i1_2:i2_2}

                w2 = w2 * r2 * (deg-alpha1_2) / (1+alpha1_2)
                """,
            [
                # Must declare coeffs to have "no" shape, to keep loopy
                # from trying to figure it out the shape automatically.

                lp.GlobalArg("coeffs", None, shape=None),
                "..."
                ],
            assumptions="deg>=0 and nels>=1"
            )

    knl = lp.fix_parameters(knl, nqp1d=7, deg=4)

    if 0:
        knl = lp.split_iname(knl, "el", 16, inner_tag="l.0")
        knl = lp.split_iname(knl, "el_outer", 2, outer_tag="g.0", inner_tag="ilp",
                slabs=(0, 1))
        knl = lp.tag_inames(knl, dict(i2="l.1", alpha1="unr", alpha2="unr"))

    from pickle import dumps, loads
    knl = loads(dumps(knl))

    knl = lp.CompiledKernel(ctx, knl).get_highlighted_code(
            dict(
                qpts=np.float32,
                tmp=np.float32,
                coeffs=np.float32,
                result=np.float32,
                ))
    print(knl)


@pytest.mark.parametrize("vec_len", [2, 3, 4, 8, 16])
def test_vector_types(ctx_factory, vec_len):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{ [i,j]: 0<=i<n and 0<=j<vec_len }",
            "out[i,j] = 2*a[i,j]",
            [
                lp.GlobalArg("a", np.float32, shape=lp.auto),
                lp.GlobalArg("out", np.float32, shape=lp.auto),
                "..."
                ])

    knl = lp.fix_parameters(knl, vec_len=vec_len)

    ref_knl = knl

    knl = lp.tag_data_axes(knl, "out", "c,vec")
    knl = lp.tag_inames(knl, dict(j="unr"))

    knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")

    lp.auto_test_vs_ref(ref_knl, ctx, knl,
            parameters=dict(
                n=20000
                ))


def test_tag_data_axes(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{ [i,j,k]: 0<=i,j,k<n }",
            "out[i,j,k] = 15")

    ref_knl = knl

    with pytest.raises(lp.LoopyError):
        lp.tag_data_axes(knl, "out", "N1,N0,N5")

    with pytest.raises(lp.LoopyError):
        lp.tag_data_axes(knl, "out", "N1,N0,c")

    knl = lp.tag_data_axes(knl, "out", "N1,N0,N2")
    knl = lp.tag_inames(knl, dict(j="g.0", i="g.1"))

    lp.auto_test_vs_ref(ref_knl, ctx, knl,
            parameters=dict(n=20))


def test_conditional(ctx_factory):
    #logging.basicConfig(level=logging.DEBUG)
    ctx = ctx_factory()

    knl = lp.make_kernel(
            "{ [i,j]: 0<=i,j<n }",
            """
                <> my_a = a[i,j] {id=read_a}
                <> a_less_than_zero = my_a < 0 {dep=read_a,inames=i:j}
                my_a = 2*my_a {id=twice_a,dep=read_a,if=a_less_than_zero}
                my_a = my_a+1 {id=aplus,dep=twice_a,if=a_less_than_zero}
                out[i,j] = 2*my_a {dep=aplus}
                """,
            [
                lp.GlobalArg("a", np.float32, shape=lp.auto),
                lp.GlobalArg("out", np.float32, shape=lp.auto),
                "..."
                ])

    ref_knl = knl

    lp.auto_test_vs_ref(ref_knl, ctx, knl,
            parameters=dict(
                n=200
                ))


def test_ilp_loop_bound(ctx_factory):
    # The salient bit of this test is that a joint bound on (outer, inner)
    # from a split occurs in a setting where the inner loop has been ilp'ed.
    # In 'normal' parallel loops, the inner index is available for conditionals
    # throughout. In ILP'd loops, not so much.

    ctx = ctx_factory()
    knl = lp.make_kernel(
            "{ [i,j,k]: 0<=i,j,k<n }",
            """
            out[i,k] = sum(j, a[i,j]*b[j,k])
            """,
            [
                lp.GlobalArg("a,b", np.float32, shape=lp.auto),
                "...",
                ],
            assumptions="n>=1")

    ref_knl = knl

    knl = lp.set_loop_priority(knl, "j,i,k")
    knl = lp.split_iname(knl,  "k", 4, inner_tag="ilp")

    lp.auto_test_vs_ref(ref_knl, ctx, knl,
            parameters=dict(
                n=200
                ))


def test_arg_shape_uses_assumptions(ctx_factory):
    # If arg shape determination does not use assumptions, then it won't find a
    # static shape for out, which is at least 1 x 1 in size, but otherwise of
    # size n x n.

    lp.make_kernel(
            "{ [i,j]: 0<=i,j<n }",
            """
            out[i,j] = 2*a[i,j]
            out[0,0] = 13.0
            """, assumptions="n>=1")


def test_slab_decomposition_does_not_double_execute(ctx_factory):
    ctx = ctx_factory()
    queue = cl.CommandQueue(ctx)

    knl = lp.make_kernel(
        "{ [i]: 0<=i<n }",
        "a[i] = 2*a[i]",
        assumptions="n>=1")

    ref_knl = knl

    for outer_tag in ["for", "g.0"]:
        knl = ref_knl
        knl = lp.split_iname(knl, "i", 4, slabs=(0, 1), inner_tag="unr",
                outer_tag=outer_tag)
        knl = lp.set_loop_priority(knl, "i_outer")

        a = cl.clrandom.rand(queue, 20, np.float32)
        a_ref = a.copy()
        a_knl = a.copy()

        knl = lp.set_options(knl, "write_cl")
        knl(queue, a=a_knl)
        ref_knl(queue, a=a_ref)

        queue.finish()

        assert (a_ref == a_knl).get().all()


def test_multiple_writes_to_local_temporary():
    # Loopy would previously only handle barrier insertion correctly if exactly
    # one instruction wrote to each local temporary. This tests that multiple
    # writes are OK.

    knl = lp.make_kernel(
        "{[i,e]: 0<=i<5 and 0<=e<nelements}",
        """
        <> temp[i, 0] = 17
        temp[i, 1] = 15
        """)
    knl = lp.tag_inames(knl, dict(i="l.0"))

    knl = lp.preprocess_kernel(knl)
    for k in lp.generate_loop_schedules(knl):
        code, _ = lp.generate_code(k)
        print(code)


def test_fd_demo():
    knl = lp.make_kernel(
        "{[i,j]: 0<=i,j<n}",
        "result[i,j] = u[i, j]**2 + -1 + (-4)*u[i + 1, j + 1] \
                + u[i + 1 + 1, j + 1] + u[i + 1 + -1, j + 1] \
                + u[i + 1, j + 1 + 1] + u[i + 1, j + 1 + -1]")
    #assumptions="n mod 16=0")
    knl = lp.split_iname(knl,
            "i", 16, outer_tag="g.1", inner_tag="l.1")
    knl = lp.split_iname(knl,
            "j", 16, outer_tag="g.0", inner_tag="l.0")
    knl = lp.add_prefetch(knl, "u",
            ["i_inner", "j_inner"],
            fetch_bounding_box=True)

    #n = 1000
    #u = cl.clrandom.rand(queue, (n+2, n+2), dtype=np.float32)

    knl = lp.set_options(knl, write_cl=True)
    knl = lp.add_and_infer_dtypes(knl, dict(u=np.float32))
    knl = lp.preprocess_kernel(knl)
    knl = lp.get_one_scheduled_kernel(knl)
    code, inf = lp.generate_code(knl)
    print(code)

    assert "double" not in code


def test_fd_1d(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
        "{[i]: 0<=i<n}",
        "result[i] = u[i+1]-u[i]")

    knl = lp.add_and_infer_dtypes(knl, {"u": np.float32})
    ref_knl = knl

    knl = lp.split_iname(knl, "i", 16)
    knl = lp.extract_subst(knl, "u_acc", "u[j]", parameters="j")
    knl = lp.precompute(knl, "u_acc", "i_inner", default_tag="for")
    knl = lp.assume(knl, "n mod 16 = 0")

    lp.auto_test_vs_ref(
            ref_knl, ctx, knl,
            parameters=dict(n=2048))


def test_make_copy_kernel(ctx_factory):
    ctx = ctx_factory()
    queue = cl.CommandQueue(ctx)

    intermediate_format = "f,f,sep"

    a1 = np.random.randn(1024, 4, 3)

    cknl1 = lp.make_copy_kernel(intermediate_format)

    cknl1 = lp.fix_parameters(cknl1, n2=3)

    cknl1 = lp.set_options(cknl1, write_cl=True)
    evt, a2 = cknl1(queue, input=a1)

    cknl2 = lp.make_copy_kernel("c,c,c", intermediate_format)
    cknl2 = lp.fix_parameters(cknl2, n2=3)

    evt, a3 = cknl2(queue, input=a2)

    assert (a1 == a3).all()


def test_set_arg_order():
    knl = lp.make_kernel(
            "{ [i,j]: 0<=i,j<n }",
            "out[i,j] = a[i]*b[j]")

    knl = lp.set_argument_order(knl, "out,a,n,b")


def test_affine_map_inames():
    knl = lp.make_kernel(
        "{[e, i,j,n]: 0<=e<E and 0<=i,j,n<N}",
        "rhsQ[e, n+i, j] = rhsQ[e, n+i, j] - D[i, n]*x[i,j]")

    knl = lp.affine_map_inames(knl,
            "i", "i0",
            "i0 = n+i")

    print(knl)


def test_precompute_confusing_subst_arguments(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
        "{[i,j]: 0<=i<n and 0<=j<5}",
        """
        D(i):=a[i+1]-a[i]
        b[i,j] = D(j)
        """)

    knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))

    ref_knl = knl

    knl = lp.tag_inames(knl, dict(j="g.1"))
    knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")

    from loopy.symbolic import get_dependencies
    assert "i_inner" not in get_dependencies(knl.substitutions["D"].expression)
    knl = lp.precompute(knl, "D")

    lp.auto_test_vs_ref(
            ref_knl, ctx, knl,
            parameters=dict(n=12345))


def test_precompute_nested_subst(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
        "{[i,j]: 0<=i<n and 0<=j<5}",
        """
        E:=a[i]
        D:=E*E
        b[i] = D
        """)

    knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))

    ref_knl = knl

    knl = lp.tag_inames(knl, dict(j="g.1"))
    knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")

    from loopy.symbolic import get_dependencies
    assert "i_inner" not in get_dependencies(knl.substitutions["D"].expression)
    knl = lp.precompute(knl, "D", "i_inner")

    # There's only one surviving 'E' rule.
    assert len([
        rule_name
        for rule_name in knl.substitutions
        if rule_name.startswith("E")]) == 1

    # That rule should use the newly created prefetch inames,
    # not the prior 'i_inner'
    assert "i_inner" not in get_dependencies(knl.substitutions["E"].expression)

    lp.auto_test_vs_ref(
            ref_knl, ctx, knl,
            parameters=dict(n=12345))


def test_poisson(ctx_factory):
    # Stolen from Peter Coogan and Rob Kirby for FEM assembly
    ctx = ctx_factory()

    nbf = 5
    nqp = 5
    sdim = 3

    knl = lp.make_kernel(
            "{ [c,i,j,k,ell,ell2]: \
            0 <= c < nels and \
            0 <= i < nbf and \
            0 <= j < nbf and \
            0 <= k < nqp and \
            0 <= ell < sdim and \
            0 <= ell2 < sdim }",
            """
            dpsi(bf,k0,dir) := sum(ell2, DFinv[c,ell2,dir] * DPsi[bf,k0,ell2] )
            Ael[c,i,j] = J[c] * w[k] * sum(ell, dpsi(i,k,ell) * dpsi(j,k,ell))
            """,
            assumptions="nels>=1 and nbf >= 1 and nels mod 4 = 0")

    knl = lp.fix_parameters(knl, nbf=nbf, sdim=sdim, nqp=nqp)

    ref_knl = knl

    knl = lp.set_loop_priority(knl, ["c", "j", "i", "k"])

    def variant_1(knl):
        knl = lp.precompute(knl, "dpsi", "i,k,ell", default_tag='for')
        knl = lp.set_loop_priority(knl, "c,i,j")
        return knl

    def variant_2(knl):
        knl = lp.precompute(knl, "dpsi", "i,ell", default_tag='for')
        knl = lp.set_loop_priority(knl, "c,i,j")
        return knl

    def add_types(knl):
        return lp.add_and_infer_dtypes(knl, dict(
            w=np.float32,
            J=np.float32,
            DPsi=np.float32,
            DFinv=np.float32,
            ))

    for variant in [
            #variant_1,
            variant_2
            ]:
        knl = variant(knl)

        lp.auto_test_vs_ref(
                add_types(ref_knl), ctx, add_types(knl),
                parameters=dict(n=5, nels=15, nbf=5, sdim=2, nqp=7))


def test_auto_test_can_detect_problems(ctx_factory):
    ctx = ctx_factory()

    knl = lp.make_kernel(
        "{[i,j]: 0<=i,j<n}",
        """
        a[i,j] = 25
        """)

    knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))

    ref_knl = knl

    knl = lp.link_inames(knl, "i,j", "i0")

    from loopy.diagnostic import AutomaticTestFailure
    with pytest.raises(AutomaticTestFailure):
        lp.auto_test_vs_ref(
                ref_knl, ctx, knl,
                parameters=dict(n=123))


def test_generate_c_snippet():
    from loopy.target.c import CTarget

    from pymbolic import var
    I = var("I")  # noqa
    f = var("f")
    df = var("df")
    q_v = var("q_v")
    eN = var("eN")  # noqa
    k = var("k")
    u = var("u")

    from functools import partial
    l_sum = partial(lp.Reduction, "sum")

    Instr = lp.ExpressionInstruction  # noqa

    knl = lp.make_kernel(
        "{[I, k]: 0<=I<nSpace and 0<=k<nQuad}",
        [
            Instr(f[I], l_sum(k, q_v[k, I]*u)),
            Instr(df[I], l_sum(k, q_v[k, I])),
            ],
        [
            lp.GlobalArg("q_v", np.float64, shape="nQuad, nSpace"),
            lp.GlobalArg("f,df", np.float64, shape="nSpace"),
            lp.ValueArg("u", np.float64),
            "...",
            ],
        target=CTarget(),
        assumptions="nQuad>=1")

    if 0:  # enable to play with prefetching
        # (prefetch currently requires constant sizes)
        knl = lp.fix_parameters(knl, nQuad=5, nSpace=3)
        knl = lp.add_prefetch(knl, "q_v", "k,I", default_tag=None)