def save_register(self, reg): # , branch_to_save = False):
code = spu.get_active_code()
offset = code.acquire_register()
size = code.acquire_register()
test = code.acquire_register()
regs = [offset, size, test]
spu.rotqbyi(offset, self.ls_buffer, 4)
spu.rotqbyi(size, self.ls_buffer, 8)
spu.stqx(reg, self.ls_buffer, offset)
spu.ai(offset, offset, 16)
spu.ceq(test, offset, size)
spu.wrch(size, dma.SPU_WrOutMbox)
spu.wrch(offset, dma.SPU_WrOutMbox)
spu.wrch(test, dma.SPU_WrOutMbox)
# !!! STOPPED HERE !!! THESE VALUES ARE WRONG !!!
lbl_ls_full = code.size()
spu.stop(0xB)
self.save_ls_buffer(ls_size = size)
spu.nop(0)
code[lbl_ls_full] = spu.brz(test, (code.size() - lbl_ls_full), ignore_active = True)
code.release_registers(regs)
return
def save_register(self, reg): # , branch_to_save = False):
code = spu.get_active_code()
offset = code.acquire_register()
size = code.acquire_register()
test = code.acquire_register()
regs = [offset, size, test]
spu.rotqbyi(offset, self.ls_buffer, 4)
spu.rotqbyi(size, self.ls_buffer, 8)
spu.stqx(reg, self.ls_buffer, offset)
spu.ai(offset, offset, 16)
spu.ceq(test, offset, size)
spu.wrch(size, dma.SPU_WrOutMbox)
spu.wrch(offset, dma.SPU_WrOutMbox)
spu.wrch(test, dma.SPU_WrOutMbox)
# !!! STOPPED HERE !!! THESE VALUES ARE WRONG !!!
lbl_ls_full = code.size()
spu.stop(0xB)
self.save_ls_buffer(ls_size=size)
spu.nop(0)
code[lbl_ls_full] = spu.brz(test, (code.size() - lbl_ls_full),
ignore_active=True)
code.release_registers(regs)
return
def TestParams():
# Run this with a stop instruction and examine the registers
prgm = Program()
code = prgm.get_stream()
proc = Processor()
# r_sum = code.acquire_register(reg = 1)
r_sum = prgm.gp_return
r_current = prgm.acquire_register()
# Zero the sum
code.add(spu.xor(r_sum, r_sum, r_sum))
for param in [
spu_param_1,
spu_param_2,
spu_param_3,
spu_param_4,
spu_param_5,
spu_param_6,
spu_param_7,
spu_param_8,
spu_param_9,
spu_param_10,
]:
copy_param(code, r_current, param)
code.add(spu.a(r_sum, r_sum, r_current))
code.add(spu.ceqi(r_current, r_sum, 55))
# code.add(spu.ori(code.gp_return, r_current, 0))
code.add(spu.brz(r_current, 2))
code.add(spu.stop(0x200A))
code.add(spu.stop(0x200B))
params = spu_exec.ExecParams()
params.p1 = 1
params.p2 = 2
params.p3 = 3
params.p4 = 4
params.p5 = 5
params.p6 = 6
params.p7 = 7
params.p8 = 8
params.p9 = 9
params.p10 = 10
prgm += code
r = proc.execute(prgm, params=params, stop=True)
assert r[0] == 55
assert r[1] == 0x200A
# print 'int result:', r
return
def SimpleSPU(): """ A very simple SPU that computes 11 + 31 and returns 0xA on success. """ prgm = env.Program() code = prgm.get_stream() proc = env.Processor() spu.set_active_code(code) # Acquire two registers #x = code.acquire_register() x = code.gp_return test = prgm.acquire_register(reg_name = 55) spu.xor(x, x, x) # zero x spu.ai(x, x, 11) # x = x + 11 spu.ai(x, x, 31) # x = x + 31 spu.ceqi(test, x, 42) # test = (x == 42) # If test is false (all 0s), skip the stop(0x100A) instruction spu.brz(test, 2) spu.stop(0x100A) spu.stop(0x100B) prgm.add(code) prgm.print_code(hex = True) r = proc.execute(prgm, mode = 'int', stop = True, debug = True) assert(r[0] == 42) assert(r[1] == 0x100A) prgm = env.Program() code = prgm.get_stream() spu.set_active_code(code) util.load_float(code, code.fp_return, 3.14) prgm.add(code) prgm.print_code(hex = True) r = proc.execute(prgm, mode = 'fp') print r return
def SimpleSPU():
"""
A very simple SPU that computes 11 + 31 and returns 0xA on success.
"""
prgm = env.Program()
code = prgm.get_stream()
proc = env.Processor()
spu.set_active_code(code)
# Acquire two registers
#x = code.acquire_register()
x = code.gp_return
test = prgm.acquire_register(reg_name=55)
spu.xor(x, x, x) # zero x
spu.ai(x, x, 11) # x = x + 11
spu.ai(x, x, 31) # x = x + 31
spu.ceqi(test, x, 42) # test = (x == 42)
# If test is false (all 0s), skip the stop(0x100A) instruction
spu.brz(test, 2)
spu.stop(0x100A)
spu.stop(0x100B)
prgm.add(code)
prgm.print_code(hex=True)
r = proc.execute(prgm, mode='int', stop=True, debug=True)
assert (r[0] == 42)
assert (r[1] == 0x100A)
prgm = env.Program()
code = prgm.get_stream()
spu.set_active_code(code)
util.load_float(code, code.fp_return, 3.14)
prgm.add(code)
prgm.print_code(hex=True)
r = proc.execute(prgm, mode='fp')
print r
return
def TestParams():
# Run this with a stop instruction and examine the registers
prgm = Program()
code = prgm.get_stream()
proc = Processor()
#r_sum = code.acquire_register(reg = 1)
r_sum = prgm.gp_return
r_current = prgm.acquire_register()
# Zero the sum
code.add(spu.xor(r_sum, r_sum, r_sum))
for param in [
spu_param_1, spu_param_2, spu_param_3, spu_param_4, spu_param_5,
spu_param_6, spu_param_7, spu_param_8, spu_param_9, spu_param_10
]:
copy_param(code, r_current, param)
code.add(spu.a(r_sum, r_sum, r_current))
code.add(spu.ceqi(r_current, r_sum, 55))
#code.add(spu.ori(code.gp_return, r_current, 0))
code.add(spu.brz(r_current, 2))
code.add(spu.stop(0x200A))
code.add(spu.stop(0x200B))
params = spu_exec.ExecParams()
params.p1 = 1
params.p2 = 2
params.p3 = 3
params.p4 = 4
params.p5 = 5
params.p6 = 6
params.p7 = 7
params.p8 = 8
params.p9 = 9
params.p10 = 10
prgm += code
r = proc.execute(prgm, params=params, stop=True)
assert (r[0] == 55)
assert (r[1] == 0x200A)
# print 'int result:', r
return
def TestContinueLabel(n_spus=1):
n = 1024
a = extarray.extarray('I', range(n))
buffer_size = 16
if n_spus > 1: code = env.ParallelInstructionStream()
else: code = env.InstructionStream()
current = var.SignedWord(0, code)
test = var.SignedWord(0, code)
four = var.SignedWord(4, code)
stream = stream_buffer(code,
a.buffer_info()[0],
n * 4,
buffer_size,
0,
save=True)
if n_spus > 1: stream = parallel(stream)
md = memory_desc('i', 0, buffer_size)
lsa_iter = spu_vec_iter(code, md)
for buffer in stream:
for current in lsa_iter:
current.v = current + current
test.v = (current == four)
code.add(spu.gbb(test, test))
#lbl_continue = code.add(spu.stop(0xC)) - 1 # Place holder for the continue
#lsa_iter.add_continue(code, 0, lambda lbl, reg = test.reg: spu.brz(reg, lbl))
code.add(spu.brz(test.reg, lsa_iter.continue_label))
current.v = current + current
#lsa_iter.add_continue(code, lbl_continue, lambda next, reg = test.reg: spu.brz(reg, next))
proc = env.Processor()
r = proc.execute(code, n_spus=n_spus)
for i in range(0, n):
if i >= 4:
assert (a[i] == i + i)
else:
#print a[i]
assert (a[i] == i * 4)
return
def TestParams():
# Run this with a stop instruction and examine the registers
code = InstructionStream()
proc = Processor()
r_sum = code.acquire_register()
r_current = code.acquire_register()
# Zero the sum
code.add(spu.xor(r_sum, r_sum, r_sum))
for param in [
spu_param_1, spu_param_2, spu_param_3, spu_param_4, spu_param_5,
spu_param_6, spu_param_7, spu_param_8, spu_param_9, spu_param_10
]:
copy_param(code, r_current, param)
code.add(spu.a(r_sum, r_sum, r_current))
code.add(spu.ceqi(r_current, r_sum, 55))
code.add(spu.brz(r_current, 2))
code.add(spu.stop(0x200A))
code.add(spu.stop(0x200B))
params = spu_exec.ExecParams()
params.p1 = 1
params.p2 = 2
params.p3 = 3
params.p4 = 4
params.p5 = 5
params.p6 = 6
params.p7 = 7
params.p8 = 8
params.p9 = 9
params.p10 = 10
r = proc.execute(code, params=params)
assert (r == 0xA)
# print 'int result:', r
# while True:
# pass
return
def TestContinueLabel(n_spus = 1):
n = 1024
a = extarray.extarray('I', range(n))
buffer_size = 16
if n_spus > 1: prgm = env.ParallelProgram()
else: prgm = env.Program()
code = prgm.get_stream()
current = var.SignedWord(0, code)
test = var.SignedWord(0, code)
four = var.SignedWord(4, code)
stream = stream_buffer(code, a.buffer_info()[0], n * 4, buffer_size, 0, save = True)
if n_spus > 1: stream = parallel(stream)
md = memory_desc('i', 0, buffer_size)
lsa_iter = spu_vec_iter(code, md)
for buffer in stream:
for current in lsa_iter:
current.v = current + current
test.v = (current == four)
code.add(spu.gbb(test, test))
#lbl_continue = code.add(spu.stop(0xC)) - 1 # Place holder for the continue
#lsa_iter.add_continue(code, 0, lambda lbl, reg = test.reg: spu.brz(reg, lbl))
code.add(spu.brz(test.reg, lsa_iter.continue_label))
current.v = current + current
#lsa_iter.add_continue(code, lbl_continue, lambda next, reg = test.reg: spu.brz(reg, next))
prgm.add(code)
proc = env.Processor()
r = proc.execute(prgm, n_spus = n_spus)
for i in range(0, n):
if i >= 4:
assert(a[i] == i + i)
else:
#print a[i]
assert(a[i] == i * 4)
return
r_data = prgm.acquire_register()
r_cmp = prgm.acquire_register()
r_lsa = prgm.acquire_register()
spu.il(r_lsa, 0x1000)
lbl_incloop = prgm.get_label("incloop")
code.add(lbl_incloop)
spu.lqx(r_data, r_cnt, r_lsa)
spu.ai(r_data, r_data, 2)
spu.stqx(r_data, r_cnt, r_lsa)
spu.ai(r_cnt, r_cnt, 16)
spu.ceq(r_cmp, r_cnt, r_sum)
spu.brz(r_cmp, lbl_incloop)
dma.spu_write_out_mbox(code, code.r_zero)
prgm += code
t3 = time.time()
id = proc.execute(prgm, async = True, mode = 'int')
t1 = time.time()
for i in xrange(0, ITERS):
#env.spu_exec.write_in_mbox(id, 1)
#env.spu_exec.write_in_mbox(id, 1)
env.spu_exec.write_in_mbox(id, i)
#cnt = env.spu_exec.stat_in_mbox(id)
r_data = prgm.acquire_register()
r_cmp = prgm.acquire_register()
r_lsa = prgm.acquire_register()
spu.il(r_lsa, 0x1000)
lbl_incloop = prgm.get_label("incloop")
code.add(lbl_incloop)
spu.lqx(r_data, r_cnt, r_lsa)
spu.ai(r_data, r_data, 2)
spu.stqx(r_data, r_cnt, r_lsa)
spu.ai(r_cnt, r_cnt, 16)
spu.ceq(r_cmp, r_cnt, r_sum)
spu.brz(r_cmp, lbl_incloop)
dma.spu_write_out_mbox(code, code.r_zero)
prgm += code
t3 = time.time()
id = proc.execute(prgm, async=True, mode='int')
t1 = time.time()
for i in xrange(0, ITERS):
#env.spu_exec.write_in_mbox(id, 1)
#env.spu_exec.write_in_mbox(id, 1)
env.spu_exec.write_in_mbox(id, i)
#cnt = env.spu_exec.stat_in_mbox(id)
#print "cnt %x" % cnt
def SimpleSPU():
"""
A very simple SPU that computes 11 + 31 and returns 0xA on success.
"""
code = InstructionStream()
proc = Processor()
spu.set_active_code(code)
# Acquire two registers
#x = code.acquire_register()
x = code.gp_return
test = code.acquire_register()
lbl_brz = code.get_label("BRZ")
lbl_skip = code.get_label("SKIP")
spu.hbrr(lbl_brz, lbl_skip)
spu.xor(x, x, x) # zero x
spu.ai(x, x, 11) # x = x + 11
spu.ai(x, x, 31) # x = x + 31
spu.ceqi(test, x, 42) # test = (x == 42)
# If test is false (all 0s), skip the stop(0x100A) instruction
code.add(lbl_brz)
spu.brz(test, lbl_skip)
spu.stop(0x100A)
code.add(lbl_skip)
spu.stop(0x100B)
code.print_code(hex=True, pro=True, epi=True)
r = proc.execute(code, mode='int', stop=True)
print "ret", r
assert (r[0] == 42)
assert (r[1] == 0x100A)
code = InstructionStream()
spu.set_active_code(code)
lbl_loop = code.get_label("LOOP")
lbl_break = code.get_label("BREAK")
r_cnt = code.acquire_register()
r_stop = code.acquire_register()
r_cmp = code.acquire_register()
r_foo = code.gp_return
spu.ori(r_foo, code.r_zero, 0)
spu.ori(r_cnt, code.r_zero, 0)
util.load_word(code, r_stop, 10)
code.add(lbl_loop)
spu.ceq(r_cmp, r_cnt, r_stop)
spu.brnz(r_cmp, lbl_break)
spu.ai(r_cnt, r_cnt, 1)
spu.a(r_foo, r_foo, r_cnt)
spu.br(lbl_loop)
code.add(lbl_break)
code.print_code()
r = proc.execute(code, mode='int', stop=True)
print "ret", r
assert (r[0] == 55)
return
def SimpleSPU():
"""
A very simple SPU that computes 11 + 31 and returns 0xA on success.
"""
prgm = env.Program()
code = prgm.get_stream()
proc = env.Processor()
spu.set_active_code(code)
# Acquire two registers
#x = code.acquire_register()
x = prgm.gp_return
test = prgm.acquire_register()
lbl_brz = prgm.get_label("BRZ")
lbl_skip = prgm.get_label("SKIP")
spu.hbrr(lbl_brz, lbl_skip)
spu.xor(x, x, x) # zero x
spu.ai(x, x, 11) # x = x + 11
spu.ai(x, x, 31) # x = x + 31
spu.ceqi(test, x, 42) # test = (x == 42)
# If test is false (all 0s), skip the stop(0x100A) instruction
code.add(lbl_brz)
spu.brz(test, lbl_skip)
spu.stop(0x100A)
code.add(lbl_skip)
spu.stop(0x100B)
prgm.add(code)
prgm.print_code()
r = proc.execute(prgm, mode = 'int', stop = True)
print "ret", r
assert(r[0] == 42)
assert(r[1] == 0x100A)
prgm = env.Program()
code = prgm.get_stream()
spu.set_active_code(code)
lbl_loop = prgm.get_label("LOOP")
lbl_break = prgm.get_label("BREAK")
r_cnt = prgm.acquire_register()
r_stop = prgm.acquire_register()
r_cmp = prgm.acquire_register()
r_foo = prgm.gp_return
spu.ori(r_foo, prgm.r_zero, 0)
spu.ori(r_cnt, prgm.r_zero, 0)
util.load_word(code, r_stop, 10)
code.add(lbl_loop)
spu.ceq(r_cmp, r_cnt, r_stop)
spu.brnz(r_cmp, lbl_break)
spu.ai(r_cnt, r_cnt, 1)
spu.a(r_foo, r_foo, r_cnt)
spu.br(lbl_loop)
code.add(lbl_break)
prgm.add(code)
prgm.print_code()
r = proc.execute(prgm, mode = 'int', stop = True)
print "ret", r
assert(r[0] == 55)
return
code = prgm.get_stream()
spu.set_active_code(code)
r_cnt = prgm.acquire_register()
r_cmp = prgm.acquire_register()
r_sum = prgm.acquire_register()
spu.il(r_cnt, 32)
spu.il(r_sum, 0)
lbl_loop = prgm.get_unique_label("LOOP")
code.add(lbl_loop)
spu.ai(r_sum, r_sum, 1)
spu.ceqi(r_cmp, r_cnt, 2)
spu.brz(r_cmp, lbl_loop)
spu.ai(r_sum, r_sum, 10)
#src = prgm.acquire_register()
#tmp = prgm.acquire_registers(3)
#dst = prgm.acquire_registers(2)
#spu.il(tmp[0], 1)
#spu.il(tmp[1], 2)
#spu.il(tmp[2], 3)
#spu.fma(src, tmp[0], tmp[1], tmp[2])
#spu.fa(dst[0], src, src)
#spu.fnms(src, tmp[0], tmp[1], tmp[2])
#spu.fs(dst[1], src, src)
code = prgm.get_stream()
spu.set_active_code(code)
r_cnt = prgm.acquire_register()
r_cmp = prgm.acquire_register()
r_sum = prgm.acquire_register()
spu.il(r_cnt, 32)
spu.il(r_sum, 0)
lbl_loop = prgm.get_unique_label("LOOP")
code.add(lbl_loop)
spu.ai(r_sum, r_sum, 1)
spu.ceqi(r_cmp, r_cnt, 2)
spu.brz(r_cmp, lbl_loop)
spu.ai(r_sum, r_sum, 10)
#src = prgm.acquire_register()
#tmp = prgm.acquire_registers(3)
#dst = prgm.acquire_registers(2)
#spu.il(tmp[0], 1)
#spu.il(tmp[1], 2)
#spu.il(tmp[2], 3)
#spu.fma(src, tmp[0], tmp[1], tmp[2])
#spu.fa(dst[0], src, src)
#spu.fnms(src, tmp[0], tmp[1], tmp[2])
#spu.fs(dst[1], src, src)
