def do_it(o):
#meat = P.cutoff_comments(g_mul0)
#meat = P.cutoff_comments(g_mul0_only)
#meat = P.cutoff_comments(g_mul0_only_immediate_writes)
#meat = P.cutoff_comments(g_mul0_only_adcx)
#meat = P.cutoff_comments(g_mul0_only_adox_imm_writes)
if 0:
p = list(range(12))
meat += mul1_code(3, P.cutoff_comments(g_muladd_3), p)
m4 = P.cutoff_comments(g_muladd_4)
meat += mul1_code(4, m4, p)
m5 = P.swap_adox_adcx(m4)
q = [int(x, 16) for x in g_perm.split(' ')]
p = P.composition(p, q)
meat += mul1_code(5, m5, p)
p = P.composition(p, q)
meat += mul1_code(6, m4, p)
p = P.composition(p, q)
meat += mul1_code(7, m5, p)
else:
# benchmark only mul0 and tail
#meat += P.cutoff_comments(g_tail_after_mul0_delayed_writes)
#meat += P.cutoff_comments(g_tail_after_mul0_immediate_writes)
pass
code = P.cutoff_comments(g_mul01) + P.cutoff_comments(g_mul01_tail)
code = '\n'.join(code)
code = P.replace_symbolic_names_wr(code, g_var_map)
code = code.split('\n')
code = [replace_ymm_by_xmm(x) for x in code]
cook_asm(o, code, 'mul8x2_zen')
def do_it(o):
meat = mul0_code(P.cutoff_comments(g_mul_012))
p = list(range(12))
meat += mul1_code(3, P.cutoff_comments(g_muladd_3), p)
m4 = P.cutoff_comments(g_muladd_4)
meat += mul1_code(4, m4, p)
m5 = P.swap_adox_adcx(m4)
q = [int(x, 16) for x in g_perm.split(' ')]
p = P.composition(p, q)
meat += mul1_code(5, m5, p)
p = P.composition(p, q)
meat += mul1_code(6, m4, p)
p = P.composition(p, q)
meat += mul1_code(7, m5, p)
cook_asm(o, meat)
def alignment_code(shift):
p = list(range(12))
m2 = P.cutoff_comments(E.g_muladd_2)
m3 = P.swap_adox_adcx(m2)
code = mul1_code(2, m2, p, shift)
q = [int(x, 16) for x in E.g_perm.split(' ')]
for i in range(3, 6):
p = P.composition(p, q)
if i & 1:
code += mul1_code(i, m3, p, shift)
else:
code += mul1_code(i, m2, p, shift)
tail = E.cook_tail(m2)
p = P.composition(p, q)
code += mul1_code(6, tail, p, shift)
return code
def do_it(o):
meat = P.cutoff_comments(g_mul_01)
p = list(range(12))
m2 = P.cutoff_comments(g_muladd_2)
m3 = P.swap_adox_adcx(m2)
meat += mul1_code(2, m2, p)
q = [int(x, 16) for x in g_perm.split(' ')]
for i in range(3, 6):
p = P.composition(p, q)
if i & 1:
meat += mul1_code(i, m3, p)
else:
meat += mul1_code(i, m2, p)
tail = cook_tail(m2)
p = P.composition(p, q)
meat += mul1_code(6, tail, p)
P.cook_asm(o, meat, g_var_map, True)