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 synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.buffers is None: raise Exception('Please set buffers')
if self.stride is None: raise Exception('Please set stride')
# Draw a square
color = var.SignedWord(0x0F0F0FFF)
fb0 = var.Word(self.buffers[0])
fb1 = var.Word(self.buffers[1])
stride = var.Word(self.stride)
addr = var.Word(0)
# Draw one line
line_pixels = 256
for i in spuiter.syn_iter(code, line_pixels*4, step = 16):
spu.stqx(color, addr, i)
# Transfer the line to the frame buffer
md_fb = spuiter.memory_desc('I', size = line_pixels)
md_fb.set_addr_reg(addr.reg)
addr.v = fb0
for i in spuiter.syn_iter(code, 128):
md_fb.put(code, 0)
addr.v = addr + stride
spu.set_active_code(old_code)
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 block(self):
code = spu.get_active_code()
self._block_idx = len(code)
# --> add the branch instruction (use brz (?) to always branch, nop to never branch)
code[self._branch_idx] = spu.nop(0, ignore_active=True)
# code[self._branch_idx] = spu.brnz(self._cmp, self._block_idx - self._branch_idx, ignore_active = True)
# code[self._branch_idx] = spu.brz(self._cmp, self._block_idx - self._branch_idx, ignore_active = True)
# Pack result into vector
# [x][y][score][--]
# Zero the save value
spu.xor(self._save_value, self._save_value, self._save_value)
# Copy the score
spu.selb(self._save_value, self._save_value, self._score,
self._word_mask)
spu.rotqbyi(self._save_value, self._save_value, 12)
# Copy the y value
spu.selb(self._save_value, self._save_value, self._y_off,
self._word_mask)
spu.rotqbyi(self._save_value, self._save_value, 12)
# Copy the x value
spu.selb(self._save_value, self._save_value, self._x_off,
self._word_mask)
# Save value to local store
spu.stqx(self._save_value, self._count, self._md_results.r_addr)
self._count.v = self._count.v + 16
# --> MemorySave test
cmp = self._save_value # reuse the save register
spu.ceq.ex(cmp, self._count, self._md_results.r_size)
if self._save_op is not None:
self._save_op.test(cmp, self._count)
# Just reset for now
spu.selb(self._count, self._count, 0, cmp)
# Return to the loop
idx = len(code)
spu.br(-(idx - self._branch_idx - 1))
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 block(self):
code = spu.get_active_code()
self._block_idx = len(code)
# --> add the branch instruction (use brz (?) to always branch, nop to never branch)
code[self._branch_idx] = spu.nop(0, ignore_active = True)
# code[self._branch_idx] = spu.brnz(self._cmp, self._block_idx - self._branch_idx, ignore_active = True)
# code[self._branch_idx] = spu.brz(self._cmp, self._block_idx - self._branch_idx, ignore_active = True)
# Pack result into vector
# [x][y][score][--]
# Zero the save value
spu.xor(self._save_value, self._save_value, self._save_value)
# Copy the score
spu.selb(self._save_value, self._save_value, self._score, self._word_mask)
spu.rotqbyi(self._save_value, self._save_value, 12)
# Copy the y value
spu.selb(self._save_value, self._save_value, self._y_off, self._word_mask)
spu.rotqbyi(self._save_value, self._save_value, 12)
# Copy the x value
spu.selb(self._save_value, self._save_value, self._x_off, self._word_mask)
# Save value to local store
spu.stqx(self._save_value, self._count, self._md_results.r_addr)
self._count.v = self._count.v + 16
# --> MemorySave test
cmp = self._save_value # reuse the save register
spu.ceq.ex(cmp, self._count, self._md_results.r_size)
if self._save_op is not None:
self._save_op.test(cmp, self._count)
# Just reset for now
spu.selb(self._count, self._count, 0, cmp)
# Return to the loop
idx = len(code)
spu.br(- (idx - self._branch_idx - 1))
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 = array.array('I', range(n))
addr = a.buffer_info()[0]
n_bytes = n * 4
if n_spus > 1: code = ParallelInstructionStream()
else: code = 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 = 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 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 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
spu.il(r_cnt, 0)
spu.il(r_sum, 16 * 4)
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):
spu.il(r_cnt, 0)
spu.il(r_sum, 16 * 4)
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)
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 store_current(self): return self.code.add(spu.stqx(self.r_current, self.r_addr, self.r_count))
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
