def TestSynIterInc():
SIZE = 64
# build and run the kernel
prgm = env.Program()
code = prgm.get_stream()
code.add(cal.dcl_output(reg.o0, USAGE=cal.usage.pos))
ones = prgm.acquire_register((1, 1, 1, 1))
counter = prgm.acquire_register()
code.add(cal.mov(counter, ones))
for i in syn_iter(code, 4, step=1, mode=INC):
code.add(cal.iadd(counter, counter, ones))
code.add(cal.mov(reg.o0, counter.x))
domain = (0, 0, SIZE, SIZE)
proc = env.Processor(0)
ext_output=proc.alloc_remote('i', 1, SIZE)
prgm.set_binding(reg.o0, ext_output)
prgm.add(code)
proc.execute(prgm, domain)
passed = True
for i in xrange(0, SIZE):
if ext_output[i] != 5:
passed = False
print "Passed == ", passed
proc.free(ext_output)
return
def TestRelativeAddressing():
import corepy.arch.cal.platform as env
import corepy.arch.cal.isa as cal
proc = env.Processor(0)
input_mem = proc.alloc_remote('I', 4, 16, 1)
output_mem = proc.alloc_remote('I', 4, 1, 1)
for i in range(16 * 1 * 4):
for j in range(4):
input_mem[i * 4 + j] = i
prgm = env.Program()
code = prgm.get_stream()
cal.set_active_code(code)
cal.dcl_output(o0, USAGE=cal.usage.generic)
cal.dcl_literal(l0, 1, 1, 1, 1)
cal.dcl_literal(l1, 16, 16, 16, 16)
cal.mov(r0, r0('0000'))
cal.mov(r1, r1('0000'))
cal.whileloop()
cal.iadd(r1, r1, g[r0.x])
cal.iadd(r0, r0, l0)
cal.breakc(cal.relop.ge, r0, l1)
cal.endloop()
cal.mov(o0, r1)
prgm.set_binding('g[]', input_mem)
prgm.set_binding('o0', output_mem)
prgm.add(code)
domain = (0, 0, 128, 128)
prgm.print_code()
proc.execute(prgm, domain)
# code.cache_code()
# print code.render_string
if output_mem[0] == 120:
print "Passed relative addressing test"
else:
print "Failed relative addressing test"
proc.free(input_mem)
proc.free(output_mem)
def test_4comp():
proc = env.Processor(0)
prgm = env.Program()
code = prgm.get_stream()
inp = proc.alloc_remote('i', 1, 4, 1)
out = proc.alloc_remote('i', 4, 1, 1)
for i in xrange(0, 4):
inp[i] = i + 1
out[i] = 0
print "inp", inp[0:4]
print "out", out[0:4]
cal.set_active_code(code)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_resource(0, cal.pixtex_type.oned, cal.fmt.float,
UNNORM=True) # positions
r_cnt = prgm.acquire_register()
r = prgm.acquire_registers(4)
cal.mov(r_cnt, r_cnt('0000'))
for i in xrange(0, 4):
cal.sample(0, 0, r[i].x000, r_cnt.x)
cal.add(r_cnt, r_cnt, r_cnt('1111'))
cal.iadd(r[0], r[0], r[1]('0x00'))
cal.iadd(r[0], r[0], r[2]('00x0'))
cal.iadd(r[0], r[0], r[3]('000x'))
cal.iadd(r[0], r[0], r[0])
cal.mov(reg.o0, r[0])
prgm.set_binding(reg.i0, inp)
prgm.set_binding(reg.o0, out)
prgm.add(code)
prgm.print_code()
proc.execute(prgm, (0, 0, 1, 1))
print "inp", inp[0:4]
print "out", out[0:4]
for i in xrange(0, 4):
assert (out[i] == (i + 1) * 2)
return
def TestSynIterIncFloatExtStopExtStart():
SIZE = 64
# build and run the kernel
prgm = env.Program()
code = prgm.get_stream()
code.add(cal.dcl_output(reg.o0, USAGE=cal.usage.pos))
ones = prgm.acquire_register((1, 1, 1, 1))
counter = prgm.acquire_register()
code.add(cal.mov(counter, ones))
stop = prgm.acquire_register((4.0, 4.0, 4.0, 4.0))
start = prgm.acquire_register((2.0, 2.0, 2.0, 2.0))
step = prgm.acquire_register((1.0, 1.0, 1.0, 1.0))
fiter = syn_iter_float(code, stop, step=step, mode=INC)
fiter.set_start_reg(start)
for i in fiter:
code.add(cal.iadd(counter, counter, ones))
code.add(cal.mov(reg.o0, counter.x))
domain = (0, 0, SIZE, SIZE)
proc = env.Processor(0)
ext_output=proc.alloc_remote('i', 1, SIZE, 1)
prgm.set_binding(reg.o0, ext_output)
prgm.add(code)
proc.execute(prgm, domain)
passed = True
for i in xrange(0, SIZE):
if ext_output[i] != 3:
passed = False
print "Passed == ", passed
proc.free(ext_output)
return
def test_foo():
proc = env.Processor(0)
prgm = env.Program()
code = prgm.get_stream()
cal.set_active_code(code)
cb = proc.alloc_remote('i', 1, 4, 1)
out = proc.alloc_remote('i', 4, 1, 1)
gb = proc.alloc_remote('i', 1, 4, 1, True)
for i in xrange(0, 4):
cb[i] = i + 1
out[i] = 42
gb[i] = 67
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_cb('cb0[4]')
cal.mov('r0', 'cb0[0]')
cal.mov('r1', 'cb0[1]')
#cal.mov('r2', 'cb0[2]')
#cal.mov('r3', 'cb0[3]')
cal.mov('o0', 'r0')
cal.mov('g[0]', 'r0')
prgm.set_binding('cb0', cb)
prgm.set_binding('o0', out)
prgm.set_binding('g[]', gb)
prgm.add(code)
prgm.print_code()
proc.execute(prgm, (0, 0, 1, 1))
print "cb ", cb[0:4]
print "out", out[0:4]
print "gb ", gb[0:4]
return
cal.breakc(cal.relop.ge, r_count.y, r_limit)
cal.add(r_sum, r_sum, r_sum('1111'))
cal.add(r_count, r_count, r_count('0100'))
cal.endloop()
cal.add(r_count, r_count, r_count('1000'))
cal.endloop()
cal.mov(reg.o0, r_sum)
return code
if __name__ == '__main__':
SIZE = 64
prgm = env.Program()
prgm.add(generate(prgm))
proc = env.Processor(0)
out = proc.alloc_remote('f', 4, SIZE, SIZE)
prgm.set_binding(reg.o0, out)
proc.execute(prgm, (0, 0, SIZE, SIZE))
print out
prgm.print_code()