scheduling_bug.py

import numpy as np
import loopy as lp
import pyopencl as cl
import pyopencl.array
import pyopencl.clrandom as cl_random
from time import time

ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)

n = 2**8

a_mat_dev = cl_random.rand(queue, (n, n), dtype=np.float32)
b_mat_dev = cl_random.rand(queue, (n, n), dtype=np.float32)
c_mat_dev = cl.array.zeros(queue, (n, n), dtype=np.float32)

queue.finish()

knl = lp.make_kernel(
        "{ [i, j, k, k1]: 0<=i, j, k, k1<256 }",
        """
        temp_cnst[k] = 2.0 {id=insn_1}
        temp_cnst_2[k1] = 2*temp_cnst[k1] {id=insn_4}
        c[i, j] = reduce(sum, k, a[i,k]*b[k,j]) {id=insn_2}
        c[i, j] = reduce(sum, k1, temp_cnst_2[k1]*a[i,k1]*b[k1,j]) {id=insn_3}
        """,
        [lp.TemporaryVariable("temp_cnst",
                              dtype=np.float32,
                              shape=lp.auto,
                              base_indices=lp.auto,
                              scope=lp.temp_var_scope.PRIVATE),
                              lp.TemporaryVariable("temp_cnst_2",
                              dtype=np.float32,
                              shape=lp.auto,
                              base_indices=lp.auto,
                              scope=lp.temp_var_scope.PRIVATE),
                              '...']
        )
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32,
                                        b=np.float32,
                                        c=np.float32))
knl = lp.set_options(knl, "write_cl")
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0", inner_tag="l.0")
processed_knl = lp.preprocess.preprocess_kernel(knl)
print(processed_knl)
print(lp.has_schedulable_iname_nesting(processed_knl))
kernel_args = {}
kernel_args['a'] = a_mat_dev
kernel_args['b'] = b_mat_dev
kernel_args['c'] = c_mat_dev

evt, (out,) = knl(queue, **kernel_args)

a = a_mat_dev.get()
b = b_mat_dev.get()
c = out.get()

print('Error =', np.linalg.norm(4*a.dot(b)-c)/np.linalg.norm(4*a.dot(b)))