def createSHAILA(synstates, enable_ps):
m = ila.Abstraction("sha")
m.enable_parameterized_synthesis = enable_ps
# I/O interface: this is where commands come from.
cmd = m.inp('cmd', 2)
cmdaddr = m.inp('cmdaddr', 16)
cmddata = m.inp('cmddata', 8)
# response.
dataout = m.reg('dataout', 8)
# internal arch state.
state = m.reg('sha_state', 3)
rdaddr = m.reg('sha_rdaddr', 16)
wraddr = m.reg('sha_wraddr', 16)
oplen = m.reg('sha_len', 16)
# for the uinst.
bytes_read = m.reg('sha_bytes_read', 16)
rd_data = m.reg('sha_rd_data', 512)
hs_data = m.reg('sha_hs_data', 160)
xram = m.mem('XRAM', 16, 8)
sha = m.fun('sha', 160, [512])
# fetch is just looking at the input command.
m.fetch_expr = ila.concat([state, cmd, cmdaddr, cmddata])
m.fetch_valid = (cmd == 1) | (cmd == 2)
# decode
rdcmds = [(state == i) & (cmd == 1) & (cmdaddr == addr)
for addr in xrange(0xfe00, 0xfe10) for i in [0, 1, 2, 3, 4]]
wrcmds = [(state == 0) & (cmd == 2) & (cmdaddr == addr)
for addr in xrange(0xfe00, 0xfe10)]
nopcmds = [(state == i) & (cmd != 1) & (cmdaddr == addr)
for addr in xrange(0xfe00, 0xfe10) for i in [1, 2, 3, 4]]
m.decode_exprs = rdcmds + wrcmds + nopcmds
# read commands.
statebyte = ila.zero_extend(state, 8)
rdaddrbyte = ila.readchunk('rd_addr', rdaddr, 8)
wraddrbyte = ila.readchunk('wr_addr', wraddr, 8)
oplenbyte = ila.readchunk('op_len', oplen, 8)
dataoutnext = ila.choice(
'dataout',
[statebyte, rdaddrbyte, wraddrbyte, oplenbyte,
m.const(0, 8)])
m.set_next('dataout', dataoutnext)
# write commands.
def mb_reg_wr(name, reg):
# multibyte register write.
reg_wr = ila.writechunk('wr_' + name, reg, cmddata)
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
mb_reg_wr('sha_rdaddr', rdaddr)
mb_reg_wr('sha_wraddr', wraddr)
mb_reg_wr('sha_len', oplen)
# state
state_next = ila.choice('state_next', [
m.const(0, 3),
m.const(1, 3),
m.const(2, 3),
m.const(3, 3),
m.const(4, 3),
ila.ite(cmddata == 1, m.const(1, 3), state),
ila.ite(bytes_read < oplen, m.const(1, 3), m.const(4, 3))
])
m.set_next('sha_state', state_next)
# these are for the uinst
# bytes_read
#bytes_read_inc = ila.ite(bytes_read+64 <= oplen, bytes_read+64, oplen)
bytes_read_inc = bytes_read + 64
bytes_read_rst = ila.ite(cmddata == 1, m.const(0, 16), bytes_read)
bytes_read_nxt = ila.choice(
'bytes_read_nxt',
[m.const(0, 16), bytes_read_inc, bytes_read_rst, bytes_read])
m.set_next('sha_bytes_read', bytes_read_nxt)
# rd_data
rdblock_little = ila.loadblk(xram, rdaddr + bytes_read, 64)
rdblock_big = ila.loadblk_big(xram, rdaddr + bytes_read, 64)
rd_data_nxt = ila.choice('rd_data_nxt', rdblock_big, rdblock_little,
rd_data)
m.set_next('sha_rd_data', rd_data_nxt)
# hs_data
sha_hs_data = ila.appfun(sha, [rd_data])
hs_data_nxt = ila.choice('hs_data_nxt', sha_hs_data, hs_data)
m.set_next('sha_hs_data', hs_data_nxt)
# xram write
xram_w_sha_little = ila.storeblk(xram, wraddr, hs_data)
xram_w_sha_big = ila.storeblk_big(xram, wraddr, hs_data)
xram_nxt = ila.choice('xram_nxt', xram, xram_w_sha_little, xram_w_sha_big)
m.set_next('XRAM', xram_nxt)
suffix = 'en' if enable_ps else 'dis'
timefile = open('sha-times-%s.txt' % suffix, 'wt')
t_elapsed = 0
# synthesis.
sim = lambda s: SHA().simulate(s)
for s in synstates:
st = time.clock()
m.synthesize(s, sim)
dt = time.clock() - st
print >> timefile, '%s %.2f' % (s, dt)
t_elapsed += dt
ast = m.get_next(s)
m.exportOne(ast, 'asts/%s_%s' % (s, suffix))
print 'time: %.2f' % t_elapsed
#m.generateSim('tmp/shasim.hpp')
m.generateSimToDir('sim')
def createAESILA(enable_ps):
m = ila.Abstraction("aes")
m.enable_parameterized_synthesis = enable_ps
# I/O interface: this is where the commands come from.
cmd = m.inp('cmd', 2)
cmdaddr = m.inp('cmdaddr', 16)
cmddata = m.inp('cmddata', 8)
# internal arch state.
state = m.reg('aes_state', 2)
opaddr = m.reg('aes_addr', 16)
oplen = m.reg('aes_len', 16)
ctr = m.reg('aes_ctr', 128)
key0 = m.reg('aes_key0', 128)
# for the uinst.
xram = m.mem('XRAM', 16, 8)
aes = m.fun('aes', 128, [128, 128, 128])
# fetch is just looking at the input command.
m.fetch_expr = ila.concat([cmd, cmdaddr, cmddata
]) # actually, the equivelant instruction
m.fetch_valid = (cmd == 2) # when write to some addresses
# decode
wrcmds = [(cmd == 2) & (cmdaddr == addr)
for addr in xrange(0xff00, 0xff40)]
m.decode_exprs = wrcmds
m.add_assumption((state == 0) | (oplen > 1))
um = m.add_microabstraction('aes_compute', (state != 0))
# write commands.
def mb_reg_wr(name, reg):
# multibyte register write.
reg_wr = ila.writechunk('wr_' + name, reg, cmddata)
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
mb_reg_wr('aes_addr', opaddr)
mb_reg_wr('aes_len', oplen)
mb_reg_wr('aes_ctr', ctr)
mb_reg_wr('aes_key0', key0)
# state
state_next = ila.choice(
'state_next',
[state,
m.const(0, 2),
ila.ite((cmddata == 1), m.const(1, 2), state)])
m.set_next('aes_state', state_next)
# xram
m.set_next('XRAM', xram)
################################
# Micro-ILA
################################
# read data
rd_data = um.reg('rd_data', 128)
enc_data = um.reg('enc_data', 128)
blk_cnt = um.reg('blk_cnt', 16)
uaes_ctr = um.reg('uaes_ctr', 128)
um.set_init('blk_cnt', um.const(0, 16))
um.set_init('uaes_ctr', um.getreg('aes_ctr'))
uxram = m.getmem('XRAM')
um.fetch_expr = state
um.decode_exprs = [(state == i) for i in [1, 2, 3]] # READ/OPERATE/WRITE
# blk_cnt
blk_cnt_inc = blk_cnt + ila.inrange('blkcntrange', um.const(1, 16),
um.const(32, 16))
more_blocks = ila.choice('cond1', (blk_cnt_inc != oplen),
(oplen >= blk_cnt_inc), (oplen > blk_cnt_inc))
blk_cnt_nxt = ila.choice('blk_cnt_nxt', [
m.const(0, 16), blk_cnt, blk_cnt_inc,
ila.ite(more_blocks, blk_cnt_inc, blk_cnt)
])
um.set_next('blk_cnt', blk_cnt_nxt)
# ustate
ustate = um.getreg('aes_state')
ustate_nxt = ila.choice('ustate_next', [
m.const(0, 2),
m.const(1, 2),
m.const(2, 2),
m.const(3, 2), ustate,
ila.ite(more_blocks, m.const(1, 2), m.const(0, 2))
]) # change 4
um.set_next('aes_state', ustate_nxt)
# rd_data
rdblock = ila.loadblk(uxram, opaddr + blk_cnt, 16)
rd_data_nxt = ila.choice('rd_data_nxt', rdblock, rd_data)
um.set_next('rd_data', rd_data_nxt)
# enc_data
aes_key = key0
aes_enc_data = ila.appfun(aes, [uaes_ctr, aes_key, rd_data])
enc_data_nxt = ila.ite(state == 2, aes_enc_data, enc_data)
um.set_next('enc_data', enc_data_nxt)
#print um.get_next('enc_data')
uaes_ctr_nxt = ila.choice(
'uaes_ctr_nxt', uaes_ctr, uaes_ctr +
ila.inrange('uaes_ctr_nxt_range', m.const(1, 128), m.const(128, 128)))
um.set_next('uaes_ctr', uaes_ctr_nxt)
# xram write
xram_w_addr = opaddr + blk_cnt
xram_w_aes = ila.storeblk(uxram, xram_w_addr, enc_data)
xram_nxt = ila.choice('xram_nxt', uxram, xram_w_aes)
um.set_next('XRAM', xram_nxt)
return m, um
def createAESILA(synstates, enable_ps):
m = ila.Abstraction("aes")
m.enable_parameterized_synthesis = enable_ps
# I/O interface: this is where the commands come from.
cmd = m.inp('cmd', 2)
cmdaddr = m.inp('cmdaddr', 16)
cmddata = m.inp('cmddata', 8)
# response.
dataout = m.reg('dataout', 8)
# internal arch state.
state = m.reg('aes_state', 2)
opaddr = m.reg('aes_addr', 16)
oplen = m.reg('aes_len', 16)
keysel = m.reg('aes_keysel', 1)
ctr = m.reg('aes_ctr', 128)
key0 = m.reg('aes_key0', 128)
key1 = m.reg('aes_key1', 128)
# for the uinst.
byte_cnt = m.reg('byte_cnt', 16)
rd_data = m.reg('rd_data', 128)
enc_data = m.reg('enc_data', 128)
xram = m.mem('XRAM', 16, 8)
aes = m.fun('aes', 128, [128, 128, 128])
# fetch is just looking at the input command.
m.fetch_expr = ila.concat([state, cmd, cmdaddr, cmddata])
m.fetch_valid = (cmd == 1) | (cmd == 2)
# decode
rdcmds = [(state == i) & (cmd == 1) & (cmdaddr == addr)
for addr in xrange(0xff00, 0xff40) for i in [0, 1, 2, 3]]
wrcmds = [(state == 0) & (cmd == 2) & (cmdaddr == addr)
for addr in xrange(0xff00, 0xff40)]
nopcmds = [(state == i) & (cmd != 1) & (cmdaddr == addr)
for addr in xrange(0xff00, 0xff40) for i in [1, 2, 3]]
m.decode_exprs = rdcmds + wrcmds + nopcmds
# read commands
statebyte = ila.zero_extend(state, 8)
opaddrbyte = ila.readchunk('rd_addr', opaddr, 8)
oplenbyte = ila.readchunk('rd_len', oplen, 8)
keyselbyte = ila.zero_extend(keysel, 8)
ctrbyte = ila.readchunk('rd_ctr', ctr, 8)
key0byte = ila.readchunk('rd_key0', key0, 8)
key1byte = ila.readchunk('rd_key1', key1, 8)
dataoutnext = ila.choice('dataout', [
statebyte, opaddrbyte, oplenbyte, keyselbyte, ctrbyte, key0byte,
key1byte,
m.const(0, 8)
])
m.set_next('dataout', dataoutnext)
# write commands.
def mb_reg_wr(name, reg):
# multibyte register write.
reg_wr = ila.writechunk('wr_' + name, reg, cmddata)
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
mb_reg_wr('aes_addr', opaddr)
mb_reg_wr('aes_len', oplen)
mb_reg_wr('aes_ctr', ctr)
mb_reg_wr('aes_key0', key0)
mb_reg_wr('aes_key1', key1)
# bit-level registers
def bit_reg_wr(name, reg, sz):
# bitwise register write
assert reg.type.bitwidth == sz
reg_wr = cmddata[sz - 1:0]
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
bit_reg_wr('aes_keysel', keysel, 1)
# state
state_next = ila.choice('state_next', [
m.const(0, 2),
m.const(1, 2),
m.const(2, 2),
m.const(3, 2),
ila.ite(cmddata == 1, m.const(1, 2), state),
ila.ite(byte_cnt + 16 < oplen, m.const(1, 2), m.const(0, 2))
])
m.set_next('aes_state', state_next)
# these are for the uinst
# byte_cnt
byte_cnt_inc = byte_cnt + 16
byte_cnt_rst = ila.ite(cmddata == 1, m.const(0, 16), byte_cnt)
byte_cnt_nxt = ila.choice(
'byte_cnt_nxt', [m.const(0, 16), byte_cnt_inc, byte_cnt_rst, byte_cnt])
m.set_next('byte_cnt', byte_cnt_nxt)
# rd_data
rdblock = ila.loadblk(xram, opaddr + byte_cnt, 16)
rd_data_nxt = ila.choice('rd_data_nxt', rdblock, rd_data)
m.set_next('rd_data', rd_data_nxt)
# enc_data
aes_key = ila.ite(keysel == 0, key0, key1)
aes_enc_data = ila.appfun(aes, [ctr, aes_key, rd_data])
enc_data_nxt = ila.ite(state == 2, aes_enc_data, enc_data)
m.set_next('enc_data', enc_data_nxt)
# xram write
xram_w_aes = ila.storeblk(xram, opaddr + byte_cnt, enc_data)
xram_nxt = ila.choice('xram_nxt', xram, xram_w_aes)
m.set_next('XRAM', xram_nxt)
# synthesize.
timefile = open('aes-times-%s.txt' % ('en' if enable_ps else 'dis'), 'wt')
sim = lambda s: AES().simulate(s)
for s in synstates:
st = time.clock()
m.synthesize(s, sim)
t_elapsed = time.clock() - st
print >> timefile, s
print >> timefile, '%.2f' % (t_elapsed)
ast = m.get_next(s)
m.exportOne(ast, 'asts/%s_%s' % (s, 'en' if enable_ps else 'dis'))
m.generateSimToDir('sim')
def model(num_regs, reg_size, paramsyn):
reg_field_width = int(math.log(num_regs, 2))
assert (1 << reg_field_width) == num_regs
# create the alu.
alu = alu_sim(reg_field_width, reg_size)
sys = ila.Abstraction("alu")
sys.enable_parameterized_synthesis = paramsyn
# state elements.
rom = sys.mem('rom', alu.ROM_ADDR_WIDTH, alu.OPCODE_WIDTH)
pc = sys.reg('pc', alu.ROM_ADDR_WIDTH)
opcode = rom[pc]
regs = [sys.reg('r%d' % i, alu.REG_SIZE) for i in xrange(alu.NUM_REGS)]
# get the two sources.
rs = ila.choice('rs', regs)
rt = ila.choice('rt', regs)
rs = [rs + rt, rs - rt, rs & rt, rs | rt, ~rs, -rs, ~rt, -rt]
# set next.
sys.set_next('pc', ila.choice('pc', pc + 1, pc + 2))
# set rom next.
sys.set_next('rom', rom)
regs_next = []
for i in xrange(alu.NUM_REGS):
ri_next = ila.choice('result%d' % i, rs + [regs[i]])
sys.set_next('r%d' % i, ri_next)
# set the fetch expressions.
sys.fetch_expr = opcode
# now set the decode expressions.
sys.decode_exprs = [opcode == i for i in xrange(alu.NUM_OPCODES)]
# synthesize pc first.
sys.synthesize('pc', lambda s: alu.alusim(s))
pc_next = sys.get_next('pc')
assert sys.areEqual(pc_next, pc + 1)
# now synthesize.
st = time.clock()
sys.synthesize(lambda s: alu.alusim(s))
et = time.clock()
print 'time for synthesis: %.3f' % (et - st)
regs_next = aluexpr(rom, pc, regs)
for i in xrange(alu.NUM_REGS):
rn1 = sys.get_next('r%d' % i)
rn2 = regs_next[i]
assert sys.areEqual(rn1, rn2)
# addr 16 bit, data 8 bit
xram = sys.mem('xram', 8, 8)
wrrd = sys.reg('wrrd', 8)
data = sys.const(0xfe, 8)
addr = sys.const(0x04, 8)
xram = ila.store(xram, addr, data)
wrrd_next = xram[addr]
sys.set_next('wrrd', wrrd_next)
wrrdblx = sys.reg('wrrdblx', 24)
datablx = sys.const(0x0f00fe, 24)
xram = ila.storeblk(xram, addr, datablx)
wrrdblx_next = ila.loadblk(xram, addr, 3)
sys.set_next('wrrdblx', wrrdblx_next)
#sys.add_assumption(opcode == 0x80)
#print sys.syn_elem("r0", sys.get_next('r0'), alusim)
expFile = "tmp/test_ila_export.txt"
sys.exportAll(expFile)
# now import into a new abstraction and check.
sysp = ila.Abstraction("alu")
sysp.importAll(expFile)
romp = sysp.getmem('rom')
pcp = sysp.getreg('pc')
regsp = [sysp.getreg('r%d' % i) for i in xrange(alu.NUM_REGS)]
regs_next = aluexpr(romp, pcp, regsp)
for i in xrange(alu.NUM_REGS):
rn1 = sysp.get_next('r%d' % i)
rn2 = regs_next[i]
assert sysp.areEqual(rn1, rn2)
#os.unlink(expFile)
#simFile = "tmp/test_ila_sim.hpp"
#sys.generateSim(simFile)
path = 'tmp/dir'
if not os.path.exists(path):
os.makedirs(path)
sys.generateSimToDir(path)
def createShaIla():
m = ila.Abstraction("sha")
m.enable_parameterized_synthesis = 0
# I/O interface
cmd = m.inp('cmd', 2)
cmdaddr = m.inp('cmdaddr', 16)
cmddata = m.inp('cmddata', 8)
# response
dataout = m.reg('dataout', 8)
# arch states
state = m.reg('sha_state', 3)
rdaddr = m.reg('sha_rdaddr', 16)
wraddr = m.reg('sha_wraddr', 16)
oplen = m.reg('sha_len', 16)
xram = m.mem('XRAM', 16, 8)
# child-ILA states
bytes_read = m.reg('sha_bytes_read', 16)
rd_data = m.reg('sha_rd_data', 512)
hs_data = m.reg('sha_hs_data', 160)
sha = m.fun('sha', 160, [512])
# fetch
m.fetch_expr = ila.concat([state, cmd, cmdaddr, cmddata])
m.fetch_valid = (cmd == 1) | (cmd == 2)
# read commands.
statebyte = ila.zero_extend(state, 8)
rdaddrbyte = ila.readchunk('rd_addr', rdaddr, 8)
wraddrbyte = ila.readchunk('wr_addr', wraddr, 8)
oplenbyte = ila.readchunk('op_len', oplen, 8)
dataoutnext = ila.choice(
'dataout',
[statebyte, rdaddrbyte, wraddrbyte, oplenbyte,
m.const(0, 8)])
m.set_next('dataout', dataoutnext)
# write commands.
def mb_reg_wr(name, reg):
reg_wr = ila.writechunk('wr_' + name, reg, cmddata)
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
mb_reg_wr('sha_rdaddr', rdaddr)
mb_reg_wr('sha_wraddr', wraddr)
mb_reg_wr('sha_len', oplen)
# state
state_choice = ila.choice('state_choice', [
m.const(0, 3),
m.const(1, 3),
m.const(2, 3),
m.const(3, 3),
m.const(4, 3)
])
rd_nxt = ila.ite(bytes_read < oplen, m.const(1, 3), m.const(4, 3))
state_nxt = ila.choice('state_nxt', [
rd_nxt, state_choice,
ila.ite(cmddata == 1, m.const(1, 3), state), state
])
m.set_next('sha_state', state_nxt)
# bytes_read
bytes_read_inc = bytes_read + 64
bytes_read_rst = ila.ite(cmddata == 1, m.const(0, 16), bytes_read)
bytes_read_nxt = ila.choice(
'bytes_read_nxt',
[m.const(0, 16), bytes_read_inc, bytes_read_rst, bytes_read])
m.set_next('sha_bytes_read', bytes_read_nxt)
# rd_data
rdblock_little = ila.loadblk(xram, rdaddr + bytes_read, 64)
rdblock_big = ila.loadblk_big(xram, rdaddr + bytes_read, 64)
rd_data_nxt = ila.choice('rd_data_nxt',
[rdblock_big, rdblock_little, rd_data])
m.set_next('sha_rd_data', rd_data_nxt)
# hs_data
sha_hs_data = ila.appfun(sha, [rd_data])
hs_data_nxt = ila.choice('sh_data_nxt', sha_hs_data, hs_data)
m.set_next('sha_hs_data', hs_data_nxt)
# xram
xram_w_sha_little = ila.storeblk(xram, wraddr, hs_data)
xram_w_sha_big = ila.storeblk_big(xram, wraddr, hs_data)
xram_nxt = ila.choice('xram_nxt',
[xram_w_sha_little, xram_w_sha_big, xram])
m.set_next('XRAM', xram_nxt)
return m
def main():
sys = ila.Abstraction("test")
iram = sys.mem('iram', 8, 8)
addr = sys.reg('addr', 8)
print iram, iram.type
data = iram[addr]
addrp = sys.reg('addrp', 8)
datap = iram[addrp]
t = sys.bool(True)
f = sys.bool(False)
assert sys.areEqual((addr != addrp) | (data == datap), t)
print data, data.type
datap = data+1
print datap, datap.type
iramp = ila.store(iram, addr, data+1)
print iramp, iramp.type
assert sys.areEqual(iramp[addr], data+1)
assert not sys.areEqual(data, data+1)
m = ila.MemValues(8, 8, 0xff)
print m
for i in xrange(0x80, 0x90):
m[i] = i-0x80
print m
for i in xrange(0x0, 0x100):
if i >= 0x80 and i < 0x90:
assert m[i] == i-0x80
else:
assert m[i] == 0xff
print m
m1 = sys.const(m)
assert m.default == 0xff
m.default = 0x0
print m
assert m[0] == 0
print m.values
m2 = sys.const(m)
# assert not sys.areEqual(m1[addr], m2[addr])
ante = ((addr >= 0x80) & (addr < 0x90))
conseq = (m1[addr] == m2[addr])
assert sys.areEqual(ila.implies(ante, conseq), t)
assert not sys.areEqual(conseq, t)
r1 = iram[addr]+1
r2 = iram[addr]+iram[addr+1]
r = ila.choice('r', r1, r2)
print sys.syn_elem("foo", r, foo)
def bar(d):
print d
ram = d['iram']
ram_ = ila.MemValues(8, 8, ram.default)
print ram
print ram_
for (ad, da) in ram.values:
ram_[ad] = da
addr = d['addr']
print ram_, addr, ram[addr]
if addr != 0:
ram_[addr] = ram_[addr]+1
print ram_
return { "bar": ram_ }
r1 = ila.store(iram, addr, iram[addr]+1)
r2 = ila.store(iram, addr, iram[addr]+2)
r3 = ila.ite(addr != 0, r1, iram)
rp = ila.choice('rp', r1, r2, r3)
expr = sys.syn_elem("bar", rp, bar)
assert sys.areEqual(expr, r3)
ila.setloglevel(3, "")
data = sys.const(0xdeadbeef, 32)
print data
iramp = ila.storeblk(iram, addrp, data)
d0 = iramp[addrp]
d1 = iramp[addrp+1]
d2 = iramp[addrp+2]
d3 = iramp[addrp+3]
datablk = ila.loadblk(iramp, addrp, 4)
assert sys.areEqual(datablk, data)
assert sys.areEqual(datablk, ila.concat([d3, d2, d1, d0]))
assert sys.areEqual(ila.concat(d0, d1), sys.const(0xefbe, 16))
assert sys.areEqual(ila.concat(d2, d3), sys.const(0xadde, 16))
def createAESILA(enable_ps):
m = ila.Abstraction("aes")
m.enable_parameterized_synthesis = enable_ps
# I/O interface: this is where the commands come from.
cmd = m.inp('cmd', 2)
cmdaddr = m.inp('cmdaddr', 16)
cmddata = m.inp('cmddata', 8)
# response.
dataout = m.reg('dataout', 8)
# internal arch state.
state = m.reg('aes_state', 2)
opaddr = m.reg('aes_addr', 16)
oplen = m.reg('aes_len', 16)
keysel = m.reg('aes_keysel', 1)
ctr = m.reg('aes_ctr', 128)
key0 = m.reg('aes_key0', 128)
key1 = m.reg('aes_key1', 128)
# for the uinst.
xram = m.mem('XRAM', 16, 8)
aes = m.fun('aes', 128, [128, 128, 128])
# fetch is just looking at the input command.
m.fetch_expr = ila.concat([state, cmd, cmdaddr, cmddata])
m.fetch_valid = (cmd == 1) | (cmd == 2)
# decode
rdcmds = [(state == i) & (cmd == 1) & (cmdaddr == addr)
for addr in xrange(0xff00, 0xff40) for i in [0, 1, 2, 3]]
wrcmds = [(state == 0) & (cmd == 2) & (cmdaddr == addr)
for addr in xrange(0xff00, 0xff40)]
nopcmds = [
((state != 0) & (cmd != 1)) | ((state == 0) & (cmd != 1) & (cmd != 2))
]
m.decode_exprs = rdcmds + wrcmds + nopcmds
# read commands
statebyte = ila.zero_extend(state, 8)
opaddrbyte = ila.readchunk('rd_addr', opaddr, 8)
oplenbyte = ila.readchunk('rd_len', oplen, 8)
keyselbyte = ila.zero_extend(keysel, 8)
ctrbyte = ila.readchunk('rd_ctr', ctr, 8)
key0byte = ila.readchunk('rd_key0', key0, 8)
key1byte = ila.readchunk('rd_key1', key1, 8)
dataoutnext = ila.choice('dataout', [
statebyte, opaddrbyte, oplenbyte, keyselbyte, ctrbyte, key0byte,
key1byte,
m.const(0, 8)
])
m.set_next('dataout', dataoutnext)
# write commands.
def mb_reg_wr(name, reg):
# multibyte register write.
reg_wr = ila.writechunk('wr_' + name, reg, cmddata)
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
mb_reg_wr('aes_addr', opaddr)
mb_reg_wr('aes_len', oplen)
mb_reg_wr('aes_ctr', ctr)
mb_reg_wr('aes_key0', key0)
mb_reg_wr('aes_key1', key1)
# bit-level registers
def bit_reg_wr(name, reg, sz):
# bitwise register write
assert reg.type.bitwidth == sz
reg_wr = cmddata[sz - 1:0]
reg_nxt = ila.choice('nxt_' + name, [reg_wr, reg])
m.set_next(name, reg_nxt)
bit_reg_wr('aes_keysel', keysel, 1)
# these are for the uinst
um = m.add_microabstraction('aes_compute', state != 0)
um.fetch_expr = state
um.decode_exprs = [state == i for i in [1, 2, 3]]
# read data
rd_data = um.reg('rd_data', 128)
enc_data = um.reg('enc_data', 128)
byte_cnt = um.reg('byte_cnt', 16)
um.set_init('byte_cnt', um.const(0, 16))
uxram = m.getmem('XRAM')
usim = lambda s: AES().simMicro(s)
# ustate
ustate = um.getreg('aes_state')
ustate_nxt = ila.choice('ustate_next', [
m.const(0, 2),
m.const(1, 2),
m.const(2, 2),
m.const(3, 2),
ila.ite(byte_cnt + 16 < oplen, m.const(1, 2), m.const(0, 2))
])
um.set_next('aes_state', ustate_nxt)
# byte_cnt
byte_cnt_inc = ila.ite(byte_cnt + 16 < oplen, byte_cnt + 16, byte_cnt)
byte_cnt_nxt = ila.choice('byte_cnt_nxt', [byte_cnt_inc, byte_cnt])
um.set_next('byte_cnt', byte_cnt_nxt)
# rd_data
rdblock = ila.loadblk(uxram, opaddr + byte_cnt, 16)
rd_data_nxt = ila.choice('rd_data_nxt', rdblock, rd_data)
um.set_next('rd_data', rd_data_nxt)
# enc_data
aes_key = ila.ite(keysel == 0, key0, key1)
aes_enc_data = ila.appfun(aes, [ctr, aes_key, rd_data])
enc_data_nxt = ila.ite(state == 2, aes_enc_data, enc_data)
um.set_next('enc_data', enc_data_nxt)
print um.get_next('enc_data')
# xram write
xram_w_aes = ila.storeblk(uxram, opaddr + byte_cnt, enc_data)
xram_nxt = ila.choice('xram_nxt', uxram, xram_w_aes)
um.set_next('XRAM', xram_nxt)
suffix = ('en' if enable_ps else 'dis')
timefile = open('aes-times-uinst-%s.txt' % suffix, 'wt')
# micro-synthesis
t_elapsed = 0
for s in ['aes_state', 'byte_cnt', 'rd_data', 'XRAM']:
st = time.clock()
um.synthesize(s, usim)
dt = time.clock() - st
t_elapsed += dt
ast = um.get_next(s)
print '%s: %s' % (s, str(ast))
m.exportOne(ast, 'uasts/u%s_%s' % (s, suffix))
print >> timefile, 'u_%s' % s
print >> timefile, '%.2f' % dt
sim = lambda s: AES().simMacro(s)
# state
state_next = ila.choice(
'state_next',
[state, ila.ite(cmddata == 1, m.const(1, 2), state)])
m.set_next('aes_state', state_next)
# xram
m.set_next('XRAM', xram)
# synthesize.
for s in [
'aes_state', 'aes_addr', 'aes_len', 'aes_keysel', 'aes_ctr',
'aes_key0', 'aes_key1', 'dataout'
]:
st = time.clock()
m.synthesize(s, sim)
dt = time.clock() - st
t_elapsed += dt
ast = m.get_next(s)
print '%s: %s' % (s, str(ast))
m.exportOne(ast, 'uasts/%s_%s' % (s, suffix))
print >> timefile, '%s' % s
print >> timefile, '%.2f' % dt
# connect to the uinst
m.connect_microabstraction('aes_state', um)
m.connect_microabstraction('XRAM', um)
print 'aes_state: %s' % str(m.get_next('aes_state'))
print 'XRAM: %s' % str(m.get_next('XRAM'))
print 'time: %.2f' % (t_elapsed)
m.generateSimToDir('sim_aes_u')