def fb_draw():
code0 = synspu.InstructionStream()
code1 = synspu.InstructionStream()
proc = synspu.Processor()
fb = cell_fb.framebuffer()
cell_fb.fb_open(fb)
draw0 = FBDraw()
draw0.set_buffers(cell_fb.fb_addr(fb, 0), cell_fb.fb_addr(fb, 1))
draw0.set_stride(fb.stride)
draw0.synthesize(code0)
draw1 = FBDraw()
draw1.set_buffers(cell_fb.fb_addr(fb, 1), cell_fb.fb_addr(fb, 0))cell_fb.fb_addr(fb, 0))
draw1.set_stride(fb.stride)
draw1.synthesize(code1)
while True:
# cell_fb.fb_clear(fb, 0)
proc.execute(code0)
cell_fb.fb_wait_vsync(fb)
cell_fb.fb_flip(fb, 0)
# cell_fb.fb_clear(fb, 1)
proc.execute(code1)
cell_fb.fb_wait_vsync(fb)
cell_fb.fb_flip(fb, 1)
cell_fb.fb_close(fb)
return
def TestAll(): import corepy.arch.spu.platform as env prgm = env.Program() code = prgm.get_stream() spu.set_active_code(code) a = code.prgm.acquire_register() b = code.prgm.acquire_register() c = code.prgm.acquire_register() shr(c, a, b) cneq(c, a, b) cge(c, a, b) cgei(c, a, 10) lt(c, a, b) lti(c, a, 10) a_immediate(c, a, 10) a_immediate(c, a, 10000) sf_immediate(c, a, 10000) prgm.add(code) prgm.print_code() proc = env.Processor() proc.execute(prgm) return
def generate(self, results, pattern, r1_range, r2_range, max_init, max_n,
size):
# Setup the range parameter array
r1_inc = (r1_range[1] - r1_range[0]) / size[0]
r2_inc = (r2_range[1] - r2_range[0]) / size[1]
ranges = extarray.extarray('f', [0.0] * 16)
for i in range(4):
ranges[i] = r1_range[0]
ranges[4 + i] = r2_range[0]
ranges[8 + i] = r1_inc
ranges[12 + i] = r2_inc
# Setup the pattern vector
bits = _pattern2vector(pattern)
# Copy the paramters to aligned buffers
#a_ranges = synspu.aligned_memory(len(ranges), typecode='I')
#a_ranges.copy_to(ranges.buffer_info()[0], len(ranges))
#a_pattern = synspu.aligned_memory(len(bits), typecode='I')
#a_pattern.copy_to(bits.buffer_info()[0], len(bits))
renderer = MailboxRenderer()
ly_block = LyapunovBlock()
ly_block.set_size(size[0], size[1])
#ly_block.set_range(a_ranges)
#ly_block.set_pattern(a_pattern)
ly_block.set_range(ranges)
ly_block.set_pattern(bits)
ly_block.set_max_init(max_init)
ly_block.set_max_n(max_n)
ly_block.set_renderer(renderer)
code = synspu.InstructionStream()
ly_block.synthesize(code)
proc = synspu.Processor()
spe_id = proc.execute(code, async=True)
for i in range(size[0] * size[1]):
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'ly said: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
proc.join(spe_id)
# for x in range(size[0]):
# r2 = r2_range[0] + r2_inc
# print 'col:', x, r1, r2
# for y in range(size[1]):
# results[y, x] = lyapunov_point(pattern, r1, r2, max_init, max_n)
# r2 += r2_inc
# r1 += r1_inc
return
def TestSetSlotValue():
import corepy.arch.spu.platform as synspu
import corepy.arch.spu.types.spu_types as var
import corepy.arch.spu.lib.dma as dma
prgm = synspu.Program()
code = prgm.get_stream()
proc = synspu.Processor()
spu.set_active_code(code)
a = var.SignedWord(0x11)
b = var.SignedWord(0x13)
r = var.SignedWord(0xFFFFFFFF)
set_slot_value(code, r, 0, 0x10)
set_slot_value(code, r, 1, a)
set_slot_value(code, r, 2, 0x12)
set_slot_value(code, r, 3, b)
for i in range(4):
spu.wrch(r, dma.SPU_WrOutMbox)
spu.rotqbyi(r, r, 4)
prgm.add(code)
spe_id = proc.execute(prgm, async=True)
for i in range(4):
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
result = synspu.spu_exec.read_out_mbox(spe_id)
assert (result == (i + 0x10))
proc.join(spe_id)
return
def TestFloatArray():
from corepy.arch.spu.platform import InstructionStream, Processor
import corepy.arch.spu.lib.dma as dma
import corepy.arch.spu.platform as env
prgm = env.Program()
code = prgm.get_stream()
spu.set_active_code(code)
x = SingleFloat([1.0, 2.0, 3.0, 4.0])
y = SingleFloat([0.5, 1.5, 2.5, 3.5])
sum = SingleFloat(0.0)
sum.v = spu.fa.ex(x, y)
r = SingleFloat([0.0, 0.0, 0.0, 0.0], reg=code.fp_return)
for i in range(4):
r.v = spu.fa.ex(sum, r)
spu.rotqbyi(sum, sum, 4)
prgm.add(code)
proc = env.Processor()
result = proc.execute(prgm, mode='fp')
x_test = array.array('f', [1.0, 2.0, 3.0, 4.0])
y_test = array.array('f', [0.5, 1.5, 2.5, 3.5])
r_test = 0.0
for i in range(4):
r_test += x_test[i] + y_test[i]
assert (result == r_test)
return
def TestTanimoto(): code = synspu.InstructionStream() proc = synspu.Processor() code.set_debug(True) x_regs = code.acquire_registers(2) y_regs = code.acquire_registers(2) result = code.acquire_register() tan = Tanimoto() tan.set_n_bits(256) tan.set_x_regs(x_regs) tan.set_y_regs(y_regs) tan.set_result_reg(result) tan.synthesize(code) code.print_code() proc.execute(code) # TODO: Do a real test, not just a synthesis test return
def TestDecrementer():
code = synspu.InstructionStream()
spu_write_decr(code, 0x7FFFFFFFl)
spu_start_decr(code)
# Get a message from the PPU
spu_read_in_mbox(code)
reg = spu_read_decr(code)
spu_write_out_mbox(code, reg)
spu_stop_decr(code)
proc = synspu.Processor()
spe_id = proc.execute(code, async=True)
print 'test is sleeping for 1 second'
time.sleep(1)
synspu.spu_exec.write_in_mbox(spe_id, 0x44CAFE)
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'spu said: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
proc.join(spe_id)
return
def TestVecIter(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)
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)
for buffer in stream:
for current in spu_vec_iter(code, md):
current.v = current + current
prgm.add(code)
proc = env.Processor()
r = proc.execute(prgm, n_spus = n_spus)
for i in range(0, n):
assert(a[i] == i + i)
return
def TestMbox():
code = synspu.InstructionStream()
# Send a message to the PPU
spu_write_out_mbox(code, 0xDEADBEEFl)
# Get a message from the PPU
reg = spu_read_in_mbox(code)
# And send it back
code.add(spu.wrch(reg, SPU_WrOutMbox))
proc = synspu.Processor()
spe_id = proc.execute(code, async=True)
synspu.spu_exec.write_in_mbox(spe_id, 0x88CAFE)
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'spe said: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'spe said: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
proc.join(spe_id)
return
def TestStreamBufferDouble(n_spus = 1):
n = 2048
a = extarray.extarray('I', range(n))
buffer_size = 32
if n_spus > 1: prgm = env.ParallelProgram()
else: prgm = env.Program()
code = prgm.get_stream()
current = var.SignedWord(0, code)
addr = a.buffer_info()[0]
n_bytes = n * 4
#print 'addr 0x%(addr)x %(addr)d' % {'addr':a.buffer_info()[0]}, n_bytes, buffer_size
stream = stream_buffer(code, addr, n_bytes, buffer_size, 0, buffer_mode='double', save = True)
if n_spus > 1: stream = parallel(stream)
for buffer in stream:
for lsa in syn_iter(code, buffer_size, 16):
code.add(spu.lqx(current, lsa, buffer))
current.v = current + current
code.add(spu.stqx(current, lsa, buffer))
prgm.add(code)
proc = env.Processor()
r = proc.execute(prgm, n_spus = n_spus)
for i in range(0, len(a)):
assert(a[i] == i + i)
return
def test_stream_popc():
code = synspu.InstructionStream()
proc = synspu.Processor()
bits = array.array('I', range(1024))
for i in range(0, 1024, 4):
bits[i] = 0x01010101 # 4 bits
bits[i + 1] = 0xFFFFFFFF # 32 bits
bits[i + 2] = 0x10101010 # 4 bits
bits[i + 3] = 0xFF0FF0F0 # 20 bits = 60 bits total
# bits[i] = 1
# bits[i+1] = 2
# bits[i+2] = 3
# bits[i+3] = 4
#abits = synspu.aligned_memory(len(bits), typecode = 'I')
#abits.copy_to(bits.buffer_info()[0], len(bits))
popc = syn_popc_stream()
popc.set_stream_addr(bits.buffer_info()[0])
popc.set_stream_size(len(bits))
popc.synthesize(code)
count = proc.execute(code, mode='mbox')
print '-->', count
assert (count == 60 * 1024 / 4)
return
def TestAll():
import corepy.arch.spu.platform as env
code = env.InstructionStream()
spu.set_active_code(code)
a = code.acquire_register()
b = code.acquire_register()
c = code.acquire_register()
shr(c, a, b)
cneq(c, a, b)
cge(c, a, b)
cgei(c, a, 10)
lt(c, a, b)
lti(c, a, 10)
a_immediate(c, a, 10)
a_immediate(c, a, 10000)
sf_immediate(c, a, 10000)
code.print_code()
proc = env.Processor()
proc.execute(code)
return
def __init__(self):
self.spus = spu_bank()
self.n_spus = 0
self.proc = synspu.Processor()
self.prog = None
return
def DoubleBufferExample(n_spus=6):
"""
stream_buffer is an iterator that streams data from main memory to
SPU local store in blocked buffers. The buffers can be managed
using single or double buffering semantics. The induction variable
returned by the buffer returns the address of the current buffer.
Note: stream_buffer was designed before memory descriptors and has
not been updated to support them yet. The interface will
change slightly when the memory classes are finalized.
"""
n = 30000
buffer_size = 16
# Create an array and align the data
a = extarray.extarray('I', range(n))
addr = a.buffer_info()[0]
n_bytes = n * 4
if n_spus > 1: code = env.ParallelInstructionStream()
else: code = env.InstructionStream()
current = SignedWord(0, code)
two = SignedWord(2, code)
# Create the stream buffer, parallelizing it if using more than 1 SPU
stream = stream_buffer(code,
addr,
n_bytes,
buffer_size,
0,
buffer_mode='double',
save=True)
if n_spus > 1: stream = parallel(stream)
# Loop over the buffers
for buffer in stream:
# Create an iterators that computes the address offsets within the
# buffer. Note: this will be supported by var/vec iters soon.
for lsa in syn_iter(code, buffer_size, 16):
code.add(spu.lqx(current, lsa, buffer))
current.v = current - two
code.add(spu.stqx(current, lsa, buffer))
# Run the synthetic program and copy the results back to the array
proc = env.Processor()
r = proc.execute(code, n_spus=n_spus)
for i in range(2, len(a)):
try:
assert (a[i] == i - 2)
except:
print 'DoubleBuffer error:', a[i], i - 2
return
def TestFloats():
import math
code = synspu.InstructionStream()
proc = synspu.Processor()
spu.set_active_code(code)
code.set_debug(True)
# Create a simple SPU program that computes log for all values bettween
# .01 and 10.0 with .01 increments
start = .65
stop = .75
inc = .01
sp_step = 0x3C23D70A
# r_current = var.Word(0x3C23D70A) # .01 in single precision
r_current = var.Word(0x3F266666)
r_step = var.Word(sp_step) # .01 in single precision
result = var.Word(0)
log = SPULog()
log.setup(code)
log.set_result(result)
log.set_x(r_current)
log_iter = syn_iter(code, int((stop - start) / inc))
for i in log_iter:
log.synthesize(code)
spu.fa(r_current, r_current, r_step)
spu.wrch(result, dma.SPU_WrOutMbox)
# code.print_code()
spe_id = proc.execute(code, mode='async')
x = start
for i in range(int((stop - start) / inc)):
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
slog = synspu.spu_exec.read_out_mbox(spe_id)
print '%.3f 0x%08X %.08f %.08f ' % (x, slog, _sp_to_float(slog),
math.log(x, 2))
x += inc
proc.join(spe_id)
return
def RunTest(test):
import corepy.arch.spu.platform as env
#from corepy.arch.spu.platform import InstructionStream, Processor
prgm = env.Program()
code = prgm.get_stream()
spu.set_active_code(code)
test()
prgm.add(code)
prgm.print_code()
proc = env.Processor()
proc.execute(prgm)
return
def test_c_popc():
code = synspu.NativeInstructionStream("spu_popc")
proc = synspu.Processor()
params = synspu.spu_exec.ExecParams()
params.p7 = 0x01010101 # 4 bits
params.p8 = 0xFFFFFFFF # 32 bits
params.p9 = 0x10101010 # 4 bits
params.p10 = 0xFF0FF0F0 # 20 bits = 60 bits total
count = proc.execute(code, mode='mbox', params=params)
assert (count == 60)
print 'test_syn_c passed'
return
def TestStreamBufferSingle(n_spus = 1):
n = 1024
a = extarray.extarray('I', range(n))
buffer_size = 128
if n_spus > 1: prgm = env.ParallelProgram()
else: prgm = env.Program()
code = prgm.get_stream()
current = var.SignedWord(0, code)
addr = a.buffer_info()[0]
stream = stream_buffer(code, addr, n * 4, buffer_size, 0, save = True)
if n_spus > 1: stream = parallel(stream)
#r_bufsize = code.acquire_register()
#r_lsa = code.acquire_register()
#r_current = code.acquire_register()
for buffer in stream:
#util.load_word(code, r_bufsize, buffer_size)
#code.add(spu.il(r_lsa, 0))
#loop = code.size()
#code.add(spu.lqx(r_current, buffer, r_lsa))
#code.add(spu.a(r_current, r_current, r_current))
#code.add(spu.stqx(r_current, buffer, r_lsa))
#code.add(spu.ai(r_bufsize, r_bufsize, -16))
#code.add(spu.ai(r_lsa, r_lsa, 16))
#code.add(spu.brnz(r_bufsize, loop - code.size()))
for lsa in syn_iter(code, buffer_size, 16):
code.add(spu.lqx(current, lsa, buffer))
current.v = current + current
#current.v = 5
code.add(spu.stqx(current, lsa, buffer))
prgm.add(code)
proc = env.Processor()
r = proc.execute(prgm, n_spus = n_spus)
for i in range(0, n):
assert(a[i] == i + i)
return
def TestSaveBuffer1():
import array
code = synspu.InstructionStream()
proc = synspu.Processor()
code.set_debug(True)
spu.set_active_code(code)
n = 2**14
data = array.array('I', range(n))
#data = synspu.aligned_memory(n, typecode = 'I')
#data.copy_to(data_array.buffer_info()[0], len(data_array))
save_buffer = SaveBuffer()
save_buffer.setup()
save_buffer.init_ls_buffer(0, 128)
save_buffer.init_mm_buffer(data.buffer_info()[0], n)
value = var.SignedWord(0xCAFEBABE)
for i in spuiter.syn_iter(code, n / 4):
save_buffer.save_register(value)
code.print_code()
spe_id = proc.execute(code, mode='async')
for i in range(n / 4):
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'size: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'offset: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'test: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
proc.join(spe_id)
#data.copy_from(data_array.buffer_info()[0], len(data_array))
print data[:10]
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 test_syn(kernel):
code = synspu.InstructionStream()
proc = synspu.Processor()
popc = kernel()
popc.synthesize(code)
params = synspu.spu_exec.ExecParams()
params.p7 = 0x01010101 # 4 bits
params.p8 = 0xFFFFFFFF # 32 bits
params.p9 = 0x10101010 # 4 bits
params.p10 = 0xFF0FF0F0 # 20 bits = 60 bits total
count = proc.execute(code, mode='mbox', params=params)
assert (count == 60)
return
def TestLog():
code = synspu.InstructionStream()
proc = synspu.Processor()
spu.set_active_code(code)
# Create a simple SPU program that computes log for 10 values and
# sends the result back using the mailbox
log = SPULog()
values = []
result = code.acquire_register()
N = 10
x = 1
for i in range(N):
val = var.Word(x)
spu.cuflt(val, val, 155)
values.append(val)
x = x * 10
log.setup(code)
log.set_result(result)
for i in range(N):
log.set_x(values[i])
log.synthesize(code)
spu.wrch(result, dma.SPU_WrOutMbox)
spe_id = proc.execute(code, mode='async')
x = 1
for i in range(N):
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'log said: 0x%08X (%d)' % (
synspu.spu_exec.read_out_mbox(spe_id), x)
x = x * 10
proc.join(spe_id)
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 TestFloatScalar():
from corepy.arch.spu.platform import InstructionStream, Processor
import corepy.arch.spu.lib.dma as dma
import corepy.arch.spu.platform as env
prgm = env.Program()
code = prgm.get_stream()
spu.set_active_code(code)
x = SingleFloat(1.0)
y = SingleFloat(2.0)
r = SingleFloat(0.0, reg=code.fp_return)
r.v = spu.fa.ex(x, y)
prgm.add(code)
proc = env.Processor()
result = proc.execute(prgm, mode='fp')
assert (result == (1.0 + 2.0))
return
def TestSignal():
code = synspu.InstructionStream()
# Get a signal from the PPU
reg = spu_read_signal1(code)
# And send it back
code.add(spu.wrch(reg, SPU_WrOutMbox))
proc = synspu.Processor()
spe_id = proc.execute(code, async=True)
synspu.spu_exec.write_signal(spe_id, 1, 0xCAFEBABEl)
while synspu.spu_exec.stat_out_mbox(spe_id) == 0:
pass
print 'sig said: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
proc.join(spe_id)
return
import corepy.lib.extarray as extarray
import corepy.arch.spu.isa as spu
import corepy.arch.spu.platform as env
import corepy.arch.spu.lib.dma as dma
from corepy.arch.spu.lib.util import load_word
import time
if __name__ == '__main__':
ITERS = 500000
#ITERS = 15
prgm = env.Program()
code = prgm.get_stream()
proc = env.Processor()
spu.set_active_code(code)
psmap = extarray.extarray('I', 131072 / 4)
data = extarray.extarray('I', range(0, 16))
r_sum = prgm.gp_return
r_cnt = prgm.acquire_register()
spu.xor(r_sum, r_sum, r_sum)
load_word(code, r_cnt, ITERS)
lbl_loop = prgm.get_label("loop")
code.add(lbl_loop)
reg = dma.spu_read_in_mbox(code)
def TestSPUIter():
size = 32
data = extarray.extarray('I', range(size))
prgm = env.Program()
code = prgm.get_stream()
r_ea_data = prgm.acquire_register()
r_ls_data = prgm.acquire_register()
r_size = prgm.acquire_register()
r_tag = prgm.acquire_register()
#print 'array ea: %X' % (data.buffer_info()[0])
#print 'r_zero = %s, ea_data = %s, ls_data = %s, r_size = %s, r_tag = %s' % (
# str(code.r_zero), str(r_ea_data), str(r_ls_data), str(r_size), str(r_tag))
# Load the effective address
util.load_word(code, r_ea_data, data.buffer_info()[0])
# Load the size
util.load_word(code, r_size, size * 4)
# Load the tag
code.add(spu.ai(r_tag, code.r_zero, 12))
# Load the lsa
code.add(spu.ai(r_ls_data, code.r_zero, 0))
# Load the data into address 0
dma.mfc_get(code, r_ls_data, r_ea_data, r_size, r_tag)
# Set the tag bit to 12
dma.mfc_write_tag_mask(code, 1<<12);
# Wait for the transfer to complete
dma.mfc_read_tag_status_all(code);
# Increment the data values by 1 using an unrolled loop (no branches)
# r_current = code.acquire_register()
current = var.SignedWord(0, code)
# Use an SPU iter
for lsa in syn_iter(code, size * 4, 16):
code.add(spu.lqx(current, code.r_zero, lsa))
# code.add(spu.ai(1, r_current, r_current))
current.v = current + current
code.add(spu.stqx(current, code.r_zero, lsa))
# code.prgm.release_register(r_current)
#current.release_register(code)
# Store the values back to main memory
# Load the tag
code.add(spu.ai(r_tag, code.r_zero, 13))
# Load the data into address 0
dma.mfc_put(code, r_ls_data, r_ea_data, r_size, r_tag)
# Set the tag bit to 12
dma.mfc_write_tag_mask(code, 1<<13);
# Wait for the transfer to complete
dma.mfc_read_tag_status_all(code);
# Cleanup
prgm.release_register(r_ea_data)
prgm.release_register(r_ls_data)
prgm.release_register(r_size)
prgm.release_register(r_tag)
# Stop for debugging
# code.add(spu.stop(0xA))
# Execute the code
prgm.add(code)
proc = env.Processor()
r = proc.execute(prgm)
for i in range(0, size):
assert(data[i] == i + i)
return
def TestSPUParallelIter(data, size, n_spus = 6, buffer_size = 16, run_code = True):
import time
# n_spus = 8
# buffer_size = 16 # 16 ints/buffer
# n_buffers = 4 # 4 buffers/spu
# n_buffers = size / buffer_size
# size = buffer_size * n_buffers * n_spus
# data = array.array('I', range(size + 2))
#data = env.aligned_memory(n, typecode = 'I')
#data.copy_to(data_array.buffer_info()[0], len(data_array))
# print 'Data align: 0x%X, %d' % (data.buffer_info()[0], data.buffer_info()[0] % 16)
code = env.ParallelInstructionStream()
# code = env.InstructionStream()
r_zero = code.acquire_register()
r_ea_data = code.acquire_register()
r_ls_data = code.acquire_register()
r_size = code.acquire_register()
r_tag = code.acquire_register()
# Load zero
util.load_word(code, r_zero, 0)
# print 'array ea: 0x%X 0x%X' % (data.buffer_info()[0], long(data.buffer_info()[0]))
# print 'r_zero = %d, ea_data = %d, ls_data = %d, r_size = %d, r_tag = %d' % (
# r_zero, r_ea_data, r_ls_data, r_size, r_tag)
# Load the effective address
if data.buffer_info()[0] % 16 == 0:
util.load_word(code, r_ea_data, data.buffer_info()[0])
else:
util.load_word(code, r_ea_data, data.buffer_info()[0] + 8)
ea_start = data.buffer_info()[0]
# Iterate over each buffer
for ea in parallel(syn_range(code, ea_start, ea_start + size * 4 , buffer_size * 4)):
# ea = var.SignedWord(code = code, reg = r_ea_data)
# print 'n_iters:', size / buffer_size
# for i in syn_range(code, size / buffer_size):
# code.add(spu.stop(0xB))
# Load the size
util.load_word(code, r_size, buffer_size * 4)
# Load the tag
code.add(spu.ai(r_tag, r_zero, 12))
# Load the lsa
code.add(spu.ai(r_ls_data, r_zero, 0))
# Load the data into address 0
dma.mfc_get(code, r_ls_data, ea, r_size, r_tag)
# Set the tag bit to 12
dma.mfc_write_tag_mask(code, 1<<12);
# Wait for the transfer to complete
dma.mfc_read_tag_status_all(code);
# Increment the data values by 1 using an unrolled loop (no branches)
# r_current = code.acquire_register()
current = var.SignedWord(0, code)
count = var.SignedWord(0, code)
# Use an SPU iter
for lsa in syn_iter(code, buffer_size * 4, 16):
code.add(spu.lqx(current, r_zero, lsa))
# code.add(spu.ai(1, r_current, r_current))
current.v = current + current
code.add(spu.stqx(current, r_zero, lsa))
count.v = count + 1
code.add(spu.stqx(count, r_zero, 0))
# code.release_register(r_current)
current.release_registers(code)
# Store the values back to main memory
# Load the tag
code.add(spu.ai(r_tag, r_zero, 13))
# Load the data into address 0
dma.mfc_put(code, r_ls_data, ea.reg, r_size, r_tag)
# Set the tag bit to 13
dma.mfc_write_tag_mask(code, 1<<13);
# Wait for the transfer to complete
dma.mfc_read_tag_status_all(code);
# code.add(spu.stop(0xB))
# Update ea
# ea.v = ea + (buffer_size * 4)
# /for ea address
# Cleanup
code.release_register(r_zero)
code.release_register(r_ea_data)
code.release_register(r_ls_data)
code.release_register(r_size)
code.release_register(r_tag)
if not run_code:
return code
# Stop for debugging
# code.add(spu.stop(0xA))
# Execute the code
proc = env.Processor()
#data.copy_from(data_array.buffer_info()[0], len(data_array))
def print_blocks():
for i in range(0, size, buffer_size):
# print data[i:(i + buffer_size)]
print data[i + buffer_size],
print ''
# print_blocks()
s = time.time()
r = proc.execute(code, n_spus = n_spus)
# r = proc.execute(code)
t = time.time() - s
# print_blocks()
return t
def TestTanimotoBlock(n_vecs = 4):
code = synspu.InstructionStream()
proc = synspu.Processor()
code.set_debug(True)
spu.set_active_code(code)
tb = TanimotoBlock()
ls_save = LocalSave()
mm_save = MemorySave()
code.set_debug(True)
# Input block parameters
m = 128
n = 64
# n_vecs = 9
n_bits = 128 * n_vecs
# Main memory results buffer
# max_results = 2**16
max_results = 16384
words_per_result = 4
mm_results_data = array.array('I', [12 for i in range(max_results * words_per_result)])
#mm_results_buffer = synspu.aligned_memory(max_results * words_per_result, typecode = 'I')
# mm_results_buffer.copy_to(mm_results_data.buffer_info()[0], len(mm_results_data))
mm_results = spuiter.memory_desc('I')
#mm_results.from_array(mm_results_buffer)
mm_results.from_array(mm_results_data)
mm_save.set_md_save_buffer(mm_results)
# Local Results buffer
buffer_size = var.SignedWord(16384)
buffer_addr = var.SignedWord(m * n * n_vecs * 4)
ls_results = spuiter.memory_desc('B')
ls_results.set_size_reg(buffer_size)
ls_results.set_addr_reg(buffer_addr)
ls_save.set_md_results(ls_results)
ls_save.set_mm_save_op(mm_save)
# Setup the TanimotoBlock class
tb.set_n_bits(n_bits)
tb.set_block_size(m, n)
tb.set_x_addr(0)
tb.set_y_addr(m * n_vecs * 16)
tb.set_save_op(ls_save)
# Main test loop
n_samples = 10000
for samples in spuiter.syn_iter(code, n_samples):
tb.synthesize(code)
spu.wrch(buffer_size, dma.SPU_WrOutMbox)
spu.stop(0x2000)
# "Function" Calls
ls_save.block()
mm_save.block()
# code.print_code()
start = time.time()
spe_id = proc.execute(code, async=True)
while synspu.spu_exec.stat_out_mbox(spe_id) == 0: pass
# print 'tb said: 0x%X' % (synspu.spu_exec.read_out_mbox(spe_id))
stop = time.time()
# mm_results_buffer.copy_from(mm_results_data.buffer_info()[0], len(mm_results_data))
proc.join(spe_id)
total = stop - start
bits_sec = (m * n * n_bits * n_samples) / total / 1e9
ops_per_compare = 48 * 4 + 8 # 48 SIMD instructions, 8 scalar
insts_per_compare = 56
gops = (m * n * n_vecs * n_samples * ops_per_compare ) / total / 1e9
ginsts = (m * n * n_vecs * n_samples * insts_per_compare ) / total / 1e9
print '%.6f sec, %.2f Gbits/sec, %.2f GOps, %.2f GInsts, %d insts' % (
total, bits_sec, gops, ginsts, code.size())
return
def generate(self,
results,
patterns,
r1_range,
r2_range,
max_init,
max_n,
size,
n_spus=6):
# Connect to the framebuffer
#fb = cell_fb.framebuffer()
#cell_fb.fb_open(fb)
buffer = extarray.extarray('B', size[0] * size[1] * 4)
buffer.clear()
# Setup the range parameter array
r1_inc = (r1_range[1] - r1_range[0]) / size[0]
r2_inc = (r2_range[1] - r2_range[0]) / size[1]
ranges = [0 for i in range(n_spus)]
#a_ranges = [0 for i in range(n_spus)]
# Slice and dice for parallel execution
spu_slices = [[size[0], size[1] / n_spus] for ispu in range(n_spus)]
spu_slices[-1][1] += size[1] % n_spus
offset = 0.0
for ispu in range(n_spus):
ranges[ispu] = extarray.extarray('f', [0.0] * 16)
for i in range(4):
ranges[ispu][
i] = r1_range[0] + float(i) * r1_inc # horizontal is simd
ranges[ispu][4 + i] = r2_range[0] + offset
ranges[ispu][8 + i] = r1_inc * 4.0
ranges[ispu][12 + i] = r2_inc
# print ranges
# Copy the paramters to aligned buffers
#a_ranges[ispu] = synspu.aligned_memory(len(ranges[ispu]), typecode='I')
#a_ranges[ispu].copy_to(ranges[ispu].buffer_info()[0], len(ranges[ispu]))
offset += r2_inc * spu_slices[ispu][1]
# Setup the pattern vector
for pattern in patterns:
if len(pattern) != len(patterns[0]):
raise Exception('All patterns must be the same length')
bits = [_pattern2vector(pattern) for pattern in patterns]
#a_pattern = synspu.aligned_memory(len(bits[0]), typecode='I')
pattern = extarray.extarray('I', len(bits[0]))
# Create the instruction streams
codes = []
n = len(patterns) * 10
offset = 0
for ispu in range(n_spus):
renderer = FBRenderer()
renderer.set_lsa(0x100)
#renderer.set_addr(cell_fb.fb_addr(fb, 0) + offset)
renderer.set_addr(buffer.buffer_info()[0] + offset)
renderer.set_width(size[0])
#renderer.set_stride(fb.stride)
renderer.set_stride(size[0])
ly_block = LyapunovBlock()
ly_block.set_size(*spu_slices[i])
#ly_block.set_range(a_ranges[ispu])
ly_block.set_range(ranges[ispu])
#ly_block.set_pattern(a_pattern)
ly_block.set_pattern(pattern)
ly_block.set_max_init(max_init)
ly_block.set_max_n(max_n)
ly_block.set_renderer(renderer)
code = synspu.InstructionStream()
# code.set_debug(True)
codes.append(code)
#offset += spu_slices[i][1] * fb.stride * 4
offset += spu_slices[i][1] * size[0] * 4
# for i in spuiter.syn_range(code, n):
ly_block.synthesize(code)
# code.print_code()
proc = synspu.Processor()
#cell_fb.fb_clear(fb, 0)
buffer.clear()
import time
ids = [0 for i in range(n_spus)]
start = time.time()
ipattern = 0
n_patterns = len(patterns)
len_bits = len(bits[0])
pattern_inc = 1
for i in range(n):
#a_pattern.copy_to(bits[ipattern].buffer_info()[0], len_bits)
# TODO - better/faster
for j in xrange(0, len_bits):
pattern[j] = bits[ipattern][j]
for ispu in range(n_spus):
ids[ispu] = proc.execute(codes[ispu], async=True)
for ispu in range(n_spus):
proc.join(ids[ispu])
#cell_fb.fb_wait_vsync(fb)
#cell_fb.fb_flip(fb, 0)
# TODO - write buffer to image file
#im = Image.frombuffer("RGBA", size, buffer.tostring(), "raw", "RGBA", 0, 1)
imgbuf = Image.new("RGBA", size)
arr = [(buffer[i + 3], buffer[i + 2], buffer[i + 1], 0xFF)
for i in xrange(0, len(buffer), 4)]
imgbuf.putdata(arr)
imgbuf.save("lyapunov_%d.png" % ipattern)
ipattern += pattern_inc
if (ipattern == (n_patterns - 1)) or (ipattern == 0):
pattern_inc *= -1
print ipattern
stop = time.time()
print '%.2f fps (%.6f)' % (float(n) / (stop - start), (stop - start))
#cell_fb.fb_close(fb)
return
