def test_peephole_twoconstant_sources(): src = ''' @kernel def main(): a = 1 b = 12 c = a + b d = b - a e = a | b a = -2 b = 3 f = b & a g = a * b ''' compiler = Compiler() main_object = compiler.compile(src)[0] main_object = compiler.opt_peephole(main_object) code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm a@0,1 imm b@0,12 imm c@0,13 imm d@0,11 imm e@0,13 imm a@1,-2 imm b@1,3 imm f@0,2 imm g@0,-6 '''.strip() assert match_code(code, pattern)
def test_builtin_transfer(): ''' Test the transfer builtin functions. ''' compiler = Compiler() src = ''' @kernel def main(): a = 3 b = transferFromNorth(a) c = transferFromEast(b) d = transferFromSouth(c) e = transferFromWest(d) ''' kernel_objects= compiler.compile(src) main_object = kernel_objects[0] code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm a@0, 3 mov out, a@0 mov b@0, north mov out, b@0 mov c@0, east mov out, c@0 mov d@0, south mov out, d@0 mov d@0, west '''.strip() assert match_code(code, pattern)
def test_simple_loop_emit():
compiler = Compiler()
src = '''
@kernel
def main():
b = 5
for i in range(2):
b = i
'''
kernel_objects= compiler.compile(src)
main_object = kernel_objects[0]
code = [InstrAdapter(x) for x in main_object.code]
pattern = '''
imm b@0, 5
ximm tmp_0@0, 2
xlabel loop_intro_1
ximm i@0, 0
ximm inc_1@0, 1
xlabel for_2
ximm tmp_3@0, 0
xemit tmp_5@0, i@1
imm tmp_6@0, 0
xcmp i@1, tmp_3@0
xphi {EQ} tmp_8@0, i@0, i@1
xmov b@1, tmp_8@0
xadd i@1, tmp_8@0, inc_1@0
xcmp i@1, tmp_0@0
xjmp {LT} for_2
'''.strip()
assert match_code(code, pattern)
def test_peephole_addorsub_zero(): ''' Test if an addition/subtraction with a zero operand is converted into a mov instr. ''' src = ''' @kernel def main(a): b = a | 0 c = a + 0 d = a - 0 e = 0 - a # unoptimised case f = 4 g = 3 h = f + g i = f - g ''' compiler = Compiler() main_object = compiler.compile(src)[0] main_object = compiler.opt_peephole(main_object) code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm tmp_0@0,0 mov b@0,a@0 imm tmp_1@0,0 mov c@0,a@0 imm tmp_2@0,0 mov d@0,a@0 imm tmp_3@0,0 sub e@0,tmp_3@0,a@0 imm f@0,4 imm g@0,3 imm h@0,7 imm i@0,1 '''.strip() assert match_code(code, pattern)
def test_get2D_copy_propagation():
''' Test the get2D builtin, this code fetches a value from a 2D buffer with boundary handling.
The difference with the previous test is that it also runs the copy propagation optimiser.
'''
compiler = Compiler()
main_object = compiler.compile(get2D_simple_src)[0]
print '\n'.join(str(InstrAdapter(x)) for x in main_object.code)
main_object = compiler.opt_copy_propagation(main_object)
code = [InstrAdapter(x) for x in main_object.code]
pattern = '''
imm in_ptr@0, 0
imm bwidth@0, 8
imm bheight@0, 8
imm tmp_0@0, 0
lid y@0, tmp_0@0
imm tmp_1@0, 1
lid x@0, tmp_1@0
imm cg_tmp_1@0, 0
imm cg_tmp_2@0, 0
add cg_tmp_3@0, x@0, bwidth@0
cmp x@0, cg_tmp_1@0
phi {LT} cg_tmp_6@0, cg_tmp_3@0, x@0
sub cg_tmp_7@0, cg_tmp_6@0, bwidth@0
cmp x@0, bwidth@0
phi {GE} cg_tmp_8@0, cg_tmp_7@0, cg_tmp_6@0
add cg_tmp_9@0, y@0, bheight@0
cmp y@0, cg_tmp_2@0
phi {LT} cg_tmp_12@0, cg_tmp_9@0, y@0
sub cg_tmp_13@0, cg_tmp_12@0, bheight@0
cmp y@0, bheight@0
phi {GE} cg_tmp_14@0, cg_tmp_13@0, cg_tmp_12@0
mul cg_tmp_15@0, cg_tmp_14@0, bwidth@0
add cg_tmp_16@0, cg_tmp_15@0, cg_tmp_8@0
add cg_tmp_17@0, in_ptr@0, cg_tmp_16@0
memr cg_tmp_19@0, cg_tmp_17@0
mov out, cg_tmp_19@0
mov cg_tmp_20@0, west
cmp x@0, cg_tmp_1@0
phi {LT} cg_tmp_21@0, cg_tmp_20@0, cg_tmp_19@0
mov out, cg_tmp_21@0
mov cg_tmp_22@0, east
cmp x@0, bwidth@0
phi {GE} cg_tmp_23@0, cg_tmp_22@0, cg_tmp_21@0
mov out, cg_tmp_23@0
mov cg_tmp_24@0, north
cmp y@0, cg_tmp_2@0
phi {LT} cg_tmp_25@0, cg_tmp_24@0, cg_tmp_23@0
mov out, cg_tmp_25@0
mov cg_tmp_26@0, south
cmp y@0, bheight@0
phi {GE} cg_tmp_27@0, cg_tmp_26@0, cg_tmp_25@0
'''
assert match_code(code, pattern)
def test_undefined_variable_detection(): ''' Test if an undefined variable gives the correct error. ''' compiler = Compiler() src = ''' @kernel def main(a): b = 3 a += 1 c = a + d ''' try: main_object = compiler.compile(src)[0] except UndefinedVariableException, e: print str(e) return # correct execution
def test_builtin_loadwest(): compiler = Compiler() src = ''' @kernel def main(): b = loadWest() ''' kernel_objects= compiler.compile(src) main_object = kernel_objects[0] main_object = compiler.replace_phi_nodes(main_object) code = [InstrAdapter(x) for x in main_object.code] pattern = ''' mov b@0, west '''.strip() assert match_code(code, pattern)
def test_compile_add(): compiler = Compiler() src = ''' @kernel def main(): b = 4 a = b + 2 ''' kernel_objects= compiler.compile(src) main_object = kernel_objects[0] code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm b@0,4 imm tmp_0@0,2 add a@0,b@0,tmp_0@0 '''.strip() assert match_code(code, pattern)
def test_inplace_operator(): compiler = Compiler() src = ''' @kernel def main(): b = 5 b += 1 ''' kernel_objects= compiler.compile(src) main_object = kernel_objects[0] code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm b@0,5 imm tmp_0@0,1 add b@1,b@0,tmp_0@0 '''.strip() assert match_code(code, pattern)
def test_builtin_sendout(): compiler = Compiler() src = ''' @kernel def main(): a = 3 sendOut(a) ''' kernel_objects= compiler.compile(src) main_object = kernel_objects[0] main_object = compiler.replace_phi_nodes(main_object) code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm a@0, 3 mov out, a@0 '''.strip() assert match_code(code, pattern)
def compile_code(src_code, args, block_size, nr_reg = 8, no_sequencer=True): ''' Compile a kernel and return a codegen function. ''' comp = CompilerDriver(nr_reg, no_sequencer=True) main_object = comp.run(src_code) patched_object = Compiler.patch_arguments_before_run(main_object, [args[argname] for argname in main_object.arguments]) def codegen_func(code, block_size, args): for x in patched_object.code: yield InstrAdapter(x, use_reg_wrapper=True) code = Code() code.set_generator(codegen_func, block_size, args) return code
def test_2D_const_list(): ''' Test code generation with 2D constants list. Note that this looks easy to optimise, but it is not so simple, because q is only know right before execution. ''' from blip.code.BlipCompiler import CompilerDriver, Compiler src = ''' @kernel def main(q): a = [[1, 2, 3], [4, 5, 6]] acc = 0 for i in range(2): for j in range(3): acc += q*a[i][j] return acc ''' comp = CompilerDriver(16, no_sequencer=True) main_object = comp.run(src) patched_object = Compiler.patch_arguments_before_run(main_object, [41]) code = [InstrAdapter(x) for x in patched_object.code] pattern = ''' imm r0,41 imm r1,0 imm r2,0 imm r3,0 imm r4,1 mul r5,r0,r4 add r4,r1,r5 imm r1,1 imm r5,2 mul r6,r0,r5 add r5,r4,r6 imm r4,2 imm r6,3 mul r7,r0,r6 add r6,r5,r7 imm r5,1 imm r7,0 imm r8,4 mul r9,r0,r8 add r8,r6,r9 imm r6,1 imm r9,5 mul r10,r0,r9 add r9,r8,r10 imm r8,2 imm r10,6 mul r11,r0,r10 add r0,r9,r11 mov r9,r0 '''.strip() assert match_code(code, pattern)
def test_replace_phi_nodes():
compiler = Compiler()
src = '''
@kernel
def main(a):
b = 3 if a > 0 else 1
'''
kernel_objects= compiler.compile(src)
main_object = kernel_objects[0]
main_object = compiler.replace_phi_nodes(main_object)
code = [InstrAdapter(x) for x in main_object.code]
pattern = '''
imm tmp_0@0,0
imm tmp_2@0,3
imm tmp_3@0,1
cmp a@0,tmp_0@0
mov b@0,tmp_3@0
mov {GT} b@0,tmp_2@0
'''.strip()
assert match_code(code, pattern)
def test_compile_conditional():
compiler = Compiler()
src = '''
@kernel
def main(a):
b = 3 if a > 0 else 1
return b
'''
kernel_objects= compiler.compile(src)
main_object = kernel_objects[0]
code = [InstrAdapter(x) for x in main_object.code]
pattern = '''
imm tmp_0@0,0
imm tmp_2@0,3
imm tmp_3@0,1
cmp a@0,tmp_0@0
phi {GT} b@0,tmp_2@0,tmp_3@0
mov main___return@0,b@0
'''.strip()
assert match_code(code, pattern)
def test_simple_loop_emit_noseq(): compiler = Compiler(no_sequencer=True) src = ''' @kernel def main(): b = 5 for i in range(2): b = i ''' kernel_objects= compiler.compile(src) main_object = kernel_objects[0] code = [InstrAdapter(x) for x in main_object.code] pattern = ''' imm b@0,5 imm i@0,0 mov b@1,i@0 imm i@1,1 mov b@2,i@1 '''.strip() assert match_code(code, pattern)
def test_patch_reg_arguments():
''' Patching of arguments before run, now with registers as arguments. '''
from blip.code.BlipCompiler import Compiler, NamedValue
src = '''
@kernel
def main(p, q):
b = p - 2
return q + b
'''
compiler = Compiler()
main_object = compiler.compile(src)[0]
patched_object = Compiler.patch_arguments_before_run(main_object, [NamedValue('test_value@0'), 41])
code = [InstrAdapter(x) for x in patched_object.code]
pattern = '''
mov p@0,test_value@0
imm q@0,41
imm tmp_0@0,2
sub b@0,p@0,tmp_0@0
add tmp_1@0,q@0,b@0
mov main___return@0,tmp_1@0
'''.strip()
assert match_code(code, pattern)
def test_patch_arguments(): from blip.code.BlipCompiler import CompilerDriver, Compiler src = ''' @kernel def main(q): return q + 1 ''' comp = CompilerDriver(8) main_object = comp.run(src) patched_object = Compiler.patch_arguments_before_run(main_object, [41]) code = [InstrAdapter(x) for x in patched_object.code] pattern = ''' imm r0,41 imm r1,1 add r2,r0,r1 mov _, r2 '''.strip() assert match_code(code, pattern)
def test_get2D():
''' Test the get2D builtin, this code fetches a value from a 2D buffer with boundary handling. '''
compiler = Compiler()
main_object = compiler.compile(get2D_simple_src)[0]
code = [InstrAdapter(x) for x in main_object.code]
pattern = '''
# initiate parameters
imm in_ptr@0, 0
imm bwidth@0, 8
imm bheight@0, 8
imm tmp_0@0, 0
lid y@0, tmp_0@0
imm tmp_1@0, 1
lid x@0, tmp_1@0
# copy parameters
mov cg_tmp_18@0, in_ptr@0
mov cg_tmp_4@0, x@0
mov cg_tmp_10@0, y@0
mov cg_tmp_5@0, bwidth@0
mov cg_tmp_11@0, bheight@0
# start of get2D
imm cg_tmp_1@0, 0
imm cg_tmp_2@0, 0
add cg_tmp_3@0, cg_tmp_4@0, cg_tmp_5@0 # cg_tmp_3 = x + bwidth
cmp cg_tmp_4@0, cg_tmp_1@0 # comp x, 0
phi {LT} cg_tmp_6@0, cg_tmp_3@0, cg_tmp_4@0 # cg_tmp_6 = phi(x < 0, cg_tmp_3, x)
sub cg_tmp_7@0, cg_tmp_6@0, cg_tmp_5@0 # cg_tmp_7 = cg_tmp_6 - bwidth
cmp cg_tmp_4@0, cg_tmp_5@0 # comp x, bwidth
phi {GE} cg_tmp_8@0, cg_tmp_7@0, cg_tmp_6@0 # cg_tmp_4 = phi(x >= bwidth, cg_tmp_7, cg_tmp_6)
add cg_tmp_9@0, cg_tmp_10@0, cg_tmp_11@0 # cg_tmp_9 = y + bheight
cmp cg_tmp_10@0, cg_tmp_2@0
phi {LT} cg_tmp_12@0, cg_tmp_9@0, cg_tmp_10@0
sub cg_tmp_13@0, cg_tmp_12@0, cg_tmp_11@0
cmp cg_tmp_10@0, cg_tmp_11@0
phi {GE} cg_tmp_14@0, cg_tmp_13@0, cg_tmp_12@0
mul cg_tmp_15@0, cg_tmp_14@0, cg_tmp_5@0
add cg_tmp_16@0, cg_tmp_15@0, cg_tmp_8@0
add cg_tmp_17@0, cg_tmp_18@0, cg_tmp_16@0
memr cg_tmp_19@0, cg_tmp_17@0
mov out, cg_tmp_19@0
mov cg_tmp_20@0, west
cmp cg_tmp_4@0, cg_tmp_1@0
phi {LT} cg_tmp_21@0, cg_tmp_20@0, cg_tmp_19@0
mov out, cg_tmp_21@0
mov cg_tmp_22@0, east
cmp cg_tmp_4@0, cg_tmp_5@0
phi {GE} cg_tmp_23@0, cg_tmp_22@0, cg_tmp_21@0
mov out, cg_tmp_23@0
mov cg_tmp_24@0, north
cmp cg_tmp_10@0, cg_tmp_2@0
phi {LT} cg_tmp_25@0, cg_tmp_24@0, cg_tmp_23@0
mov out, cg_tmp_25@0
mov cg_tmp_26@0, south
cmp cg_tmp_10@0, cg_tmp_11@0
phi {GE} cg_tmp_27@0, cg_tmp_26@0, cg_tmp_25@0
mov cg_tmp_28@0, cg_tmp_27@0
mov v@0, cg_tmp_28@0
'''.strip()
assert match_code(code, pattern)
