def popc(self, count, x):
"""
Add the number of 1 bits in each word in X to the value in count.
"""
temp = spu.get_active_code().acquire_register()
spu.cntb(temp, x)
spu.sumb(temp, temp, 0)
spu.a(count, count, temp)
spu.get_active_code().release_register(temp)
return
def popc(self, count, x):
"""
Add the number of 1 bits in each word in X to the value in count.
"""
temp = spu.get_active_code().acquire_register()
spu.cntb(temp, x)
spu.sumb(temp, temp, 0)
spu.a(count, count, temp)
spu.get_active_code().release_register(temp)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
# Reserve two variable registers
count = code.acquire_register()
result = code.acquire_register()
# 'Load' the input vector x from register 5
x = code.acquire_register()
spu.ai(x, 5, 0)
# Zero count and result
spu.xor(count, count, count)
spu.xor(result, result, result)
# Inline the popc and reduce operations
self.popc(count, x)
self.reduce_word(result, count)
# Send the result to the caller
spu.wrch(result, dma.SPU_WrOutMbox)
code.release_register(x)
spu.set_active_code(old_code)
return
def init_mm_buffer(self, addr, size, offset=0):
code = spu.get_active_code()
util.set_slot_value(code, self.mm_buffer, 0, addr)
util.set_slot_value(code, self.mm_buffer, 1, size)
util.set_slot_value(code, self.mm_buffer, 2, offset)
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 synthesize(self, code):
"""
Render a vector with 4 pixels.
"""
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.x_offset is None: raise Exception('Please call setup')
if self.result is None: raise Exception('Please set result')
if self.one is None: raise Exception('Please set one')
# Make the part of the result positive and subtract 1
# to transform (-1,-oo) into (0,oo)
self.result.v = spu.fs.ex(0, self.result)
self.result.v = spu.fs.ex(self.result, self.one)
# Convert the result to an unsigned int, scaling by 2^4 to put
# values between 0 and 16 in the gradient. Values outside [0,16]
# are 0 or FF
self.result.v = spu.cfltu.ex(self.result, 169) # 173 - 169 == 4
# self.result.v = spu.sfi.ex(self.result, 255) # 173 - 169 == 4
# Extract the first two bytes from the result into the RGB positions
# and set alpha to 0xFF
self.result.v = spu.shufb.ex(self.result, self.ff, self.uint2rgba)
# Save the result and increment the offset
spu.stqd(self.result, self.x_offset, self.lsa >> 4)
spu.ai(self.x_offset, self.x_offset, 16)
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
# Reserve two variable registers
count = code.acquire_register()
result = code.acquire_register()
# 'Load' the input vector x from register 5
x = code.acquire_register()
spu.ai(x, 5, 0)
# Zero count and result
spu.xor(count, count, count)
spu.xor(result, result, result)
# Inline the popc and reduce operations
self.popc(count, x)
self.reduce_word(result, count)
# Send the result to the caller
spu.wrch(result, dma.SPU_WrOutMbox)
code.release_register(x)
spu.set_active_code(old_code)
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 synthesize(self, code):
"""
Render a vector with 4 pixels.
"""
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.x_offset is None: raise Exception('Please call setup')
if self.result is None: raise Exception('Please set result')
if self.one is None: raise Exception('Please set one')
# Make the part of the result positive and subtract 1
# to transform (-1,-oo) into (0,oo)
self.result.v = spu.fs.ex(0, self.result)
self.result.v = spu.fs.ex(self.result, self.one)
# Convert the result to an unsigned int, scaling by 2^4 to put
# values between 0 and 16 in the gradient. Values outside [0,16]
# are 0 or FF
self.result.v = spu.cfltu.ex(self.result, 169) # 173 - 169 == 4
# self.result.v = spu.sfi.ex(self.result, 255) # 173 - 169 == 4
# Extract the first two bytes from the result into the RGB positions
# and set alpha to 0xFF
self.result.v = spu.shufb.ex(self.result, self.ff, self.uint2rgba)
# Save the result and increment the offset
spu.stqd(self.result, self.x_offset, self.lsa >> 4)
spu.ai(self.x_offset, self.x_offset, 16)
spu.set_active_code(old_code)
return
def init_mm_buffer(self, addr, size, offset = 0):
code = spu.get_active_code()
util.set_slot_value(code, self.mm_buffer, 0, addr)
util.set_slot_value(code, self.mm_buffer, 1, size)
util.set_slot_value(code, self.mm_buffer, 2, offset)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
stream = spuiter.stream_buffer(code, self.stream_addr, self.stream_size * 4,
self.buffer_size, self.lsa)
ls_data = spuiter.memory_desc('I', self.lsa, self.buffer_size / 4)
popc = syn_popc_var()
x = var.Word(0)
count = var.Word(0)
total = var.Word(0)
for buffer in stream:
for x in spuiter.spu_vec_iter(code, ls_data, addr_reg = buffer):
popc.popc(count, x)
popc.reduce_word(total, count)
# Send the result to the caller
spu.wrch(total, dma.SPU_WrOutMbox)
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
stream = spuiter.stream_buffer(code, self.stream_addr,
self.stream_size * 4, self.buffer_size,
self.lsa)
ls_data = spuiter.memory_desc('I', self.lsa, self.buffer_size / 4)
popc = syn_popc_var()
x = var.Word(0)
count = var.Word(0)
total = var.Word(0)
for buffer in stream:
for x in spuiter.spu_vec_iter(code, ls_data, addr_reg=buffer):
popc.popc(count, x)
popc.reduce_word(total, count)
# Send the result to the caller
spu.wrch(total, dma.SPU_WrOutMbox)
spu.set_active_code(old_code)
return
def test(self, cmp, count_var): code = spu.get_active_code() self._branch_idx = len(code) spu.stop(0xB) # spu.nop(0) self._cmp = cmp self._count = count_var return
def setup(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
self.consts = {}
for const in constants.keys():
self.consts[const] = var.Word(constants[const])
spu.set_active_code(old_code)
return
def setup(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
self.consts = {}
for const in constants.keys():
self.consts[const] = var.Word(constants[const])
spu.set_active_code(old_code)
return
def test(self, cmp, score, x_off, y_off): code = spu.get_active_code() self._branch_idx = len(code) spu.stop(0xB) # spu.nop(0) self._cmp = cmp self._score = score self._x_off = x_off self._y_off = y_off return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.result is None: raise Exception('Please set result')
spu.wrch(self.result, dma.SPU_WrOutMbox)
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.result is None: raise Exception('Please set result')
spu.wrch(self.result, dma.SPU_WrOutMbox)
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
self._load_parameters(code)
log = spu_log.SPULog()
log.setup(code)
if self.renderer is not None:
self.renderer.setup(code)
self.renderer.set_one(log.consts['ONE'])
r1_inc = var.SingleFloat()
r2_inc = var.SingleFloat()
r1 = var.SingleFloat()
r2 = var.SingleFloat()
result = var.SingleFloat()
pattern = var.Word(0)
self.ly_point.set_pattern_reg(pattern)
self.ly_point.set_result_reg(result)
self.ly_point.set_r_regs(r1, r2)
self.ly_point.set_log(log)
self.ly_point.setup(code)
spu.lqa(r1, 0)
spu.lqa(r2, 4)
spu.lqa(r1_inc, 8)
spu.lqa(r2_inc, 12)
spu.lqa(pattern, 16)
for y in spuiter.syn_iter(code, self.h):
spu.lqa(r1, 0)
for x in spuiter.syn_iter(code, self.w / 4):
self.ly_point.synthesize(code)
r1.v = spu.fa.ex(r1, r1_inc)
if self.renderer is not None:
# result.v = spu.fm.ex(r1, r2)
self.renderer.set_result_reg(result)
self.renderer.synthesize(code)
if self.renderer is not None:
self.renderer.row_complete(code)
r2.v = spu.fa.ex(r2, r2_inc)
# return Numeric.where(Numeric.less(results, 0), results, 0)
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
self._load_parameters(code)
log = spu_log.SPULog()
log.setup(code)
if self.renderer is not None:
self.renderer.setup(code)
self.renderer.set_one(log.consts['ONE'])
r1_inc = var.SingleFloat()
r2_inc = var.SingleFloat()
r1 = var.SingleFloat()
r2 = var.SingleFloat()
result = var.SingleFloat()
pattern = var.Word(0)
self.ly_point.set_pattern_reg(pattern)
self.ly_point.set_result_reg(result)
self.ly_point.set_r_regs(r1, r2)
self.ly_point.set_log(log)
self.ly_point.setup(code)
spu.lqa(r1, 0)
spu.lqa(r2, 4)
spu.lqa(r1_inc, 8)
spu.lqa(r2_inc, 12)
spu.lqa(pattern, 16)
for y in spuiter.syn_iter(code, self.h):
spu.lqa(r1, 0)
for x in spuiter.syn_iter(code, self.w / 4):
self.ly_point.synthesize(code)
r1.v = spu.fa.ex(r1, r1_inc)
if self.renderer is not None:
# result.v = spu.fm.ex(r1, r2)
self.renderer.set_result_reg(result)
self.renderer.synthesize(code)
if self.renderer is not None:
self.renderer.row_complete(code)
r2.v = spu.fa.ex(r2, r2_inc)
# return Numeric.where(Numeric.less(results, 0), results, 0)
spu.set_active_code(old_code)
return
def synthesize_constants(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
self._one = code.acquire_register()
spu.xor(self._one, self._one, self._one)
spu.ai(self._one, self._one, 1)
spu.cuflt(self._one, self._one, 155)
if old_code is not None:
spu.set_active_code(old_code)
return
def _compute_ratio(self, ab, c, result):
# Convert ab and c to float
spu.cuflt(ab, ab, 155)
spu.cuflt(c, c, 155)
# Compute ab = ab + c
spu.fa(ab, ab, c)
# Compute c / (ab + c)
fdiv(spu.get_active_code(), result, c, ab, self._one)
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 block(self):
code = spu.get_active_code()
self._block_idx = len(code)
# --> add the branch instruction
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)
# FILL IN HERE
# Return to the loop
idx = len(code)
spu.br(- (idx - self._branch_idx - 1))
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 synthesize(self, code):
if self._x_regs is None: raise Exception("Please set x_regs")
if self._y_regs is None: raise Exception("Please set y_regs")
if self._result is None: raise Exception("Please set result register")
old_code = spu.get_active_code()
spu.set_active_code(code)
regs = []
if self._one is None:
self.synthesize_constants(code)
regs.append(self._one)
ab = code.acquire_register()
c = code.acquire_register()
ab_temp = code.acquire_register()
c_temp = code.acquire_register()
result = code.acquire_register()
regs = regs + [ab, c, ab_temp, c_temp]
nregs = self._n_bits / 128
for i in range(nregs):
# self._ab(self._x_regs[i], self._y_regs[i], ab, ab_temp)
# self._c( self._x_regs[i], self._y_regs[i], c, c_temp)
self._ab_c(self._x_regs[i], self._y_regs[i], ab, c, ab_temp,
c_temp)
self._reduce_word(ab, ab_temp)
self._reduce_word(c, c_temp)
self._compute_ratio(ab_temp, c_temp, result)
print '%d registers,' % (len(regs) + len(self._x_regs) +
len(self._y_regs)),
code.release_registers(regs)
if old_code is not None:
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
# Create and initialize the variables
count = var.Word(0)
result = var.Word(0)
x = var.Word(0)
# 'Load' the input vector x from register 5
x.v = spu.ai.ex(5, 0)
# Inline the popc and reduce operations
self.popc(count, x)
self.reduce_word(result, count)
# Send the result to the caller
spu.wrch(result, dma.SPU_WrOutMbox)
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
# Create and initialize the variables
count = var.Word(0)
result = var.Word(0)
x = var.Word(0)
# 'Load' the input vector x from register 5
x.v = spu.ai.ex(5, 0)
# Inline the popc and reduce operations
self.popc(count, x)
self.reduce_word(result, count)
# Send the result to the caller
spu.wrch(result, dma.SPU_WrOutMbox)
spu.set_active_code(old_code)
return
def synthesize(self, code):
if self._x_regs is None: raise Exception("Please set x_regs")
if self._y_regs is None: raise Exception("Please set y_regs")
if self._result is None: raise Exception("Please set result register")
old_code = spu.get_active_code()
spu.set_active_code(code)
regs = []
if self._one is None:
self.synthesize_constants(code)
regs.append(self._one)
ab = code.acquire_register()
c = code.acquire_register()
ab_temp = code.acquire_register()
c_temp = code.acquire_register()
result = code.acquire_register()
regs = regs + [ab, c, ab_temp, c_temp]
nregs = self._n_bits / 128
for i in range(nregs):
# self._ab(self._x_regs[i], self._y_regs[i], ab, ab_temp)
# self._c( self._x_regs[i], self._y_regs[i], c, c_temp)
self._ab_c(self._x_regs[i], self._y_regs[i], ab, c, ab_temp, c_temp)
self._reduce_word(ab, ab_temp)
self._reduce_word( c, c_temp)
self._compute_ratio(ab_temp, c_temp, result)
print '%d registers,' % (len(regs) + len(self._x_regs) + len(self._y_regs)),
code.release_registers(regs)
if old_code is not None:
spu.set_active_code(old_code)
return
def save_ls_buffer(self, ls_size=None, branch=False):
code = spu.get_active_code()
regs = []
if ls_size is None:
ls_size = code.acquire_register()
regs.append(ls_size)
# Set the main memory address
mm_offset = code.acquire_register()
regs.append(mm_offset)
spu.rotqbyi(mm_offset, self.mm_buffer, 4)
spu.a(mm_offset, mm_offset, self.mm_buffer)
# Tranfer the buffer
md = spuiter.memory_desc('b')
md.set_size_reg(ls_size)
md.set_addr_reg(mm_offset)
md.put(code, self.ls_buffer)
# Increment the main memory offset
mm_size = code.acquire_register()
regs.append(mm_size)
spu.rotqbyi(mm_size, self.mm_buffer, 8)
spu.rotqbyi(mm_offset, self.mm_buffer, 4)
spu.a(mm_offset, mm_offset, mm_size)
util.set_slot_value(code, self.mm_buffer, 2, mm_offset)
# Reset the ls offset
util.set_slot_value(code, self.ls_buffer, 2, 0)
code.release_registers(regs)
return
def save_ls_buffer(self, ls_size = None, branch = False):
code = spu.get_active_code()
regs = []
if ls_size is None:
ls_size = code.acquire_register()
regs.append(ls_size)
# Set the main memory address
mm_offset = code.acquire_register()
regs.append(mm_offset)
spu.rotqbyi(mm_offset, self.mm_buffer, 4)
spu.a(mm_offset, mm_offset, self.mm_buffer)
# Tranfer the buffer
md = spuiter.memory_desc('b')
md.set_size_reg(ls_size)
md.set_addr_reg(mm_offset)
md.put(code, self.ls_buffer)
# Increment the main memory offset
mm_size = code.acquire_register()
regs.append(mm_size)
spu.rotqbyi(mm_size, self.mm_buffer, 8)
spu.rotqbyi(mm_offset, self.mm_buffer, 4)
spu.a(mm_offset, mm_offset, mm_size)
util.set_slot_value(code, self.mm_buffer, 2, mm_offset)
# Reset the ls offset
util.set_slot_value(code, self.ls_buffer, 2, 0)
code.release_registers(regs)
return
def _transfer_data(self, code, kernel, lsa, tag):
"""
Load the data into the SPU memory
"""
# Check the types
if not isinstance(code, spe.InstructionStream):
raise Exception('Code must be an InstructionStream')
if not (isinstance(lsa, int) or issubclass(type(lsa), (spe.Register, spe.Variable))):
raise Exception('lsa must be an integer, Register, or Variable')
old_code = spu.get_active_code()
spu.set_active_code(code)
# Acquire registers for address and size, if they were not supplied by the user
if self.r_addr is None: r_ea_data = code.prgm.acquire_register()
else: r_ea_data = self.r_addr
if self.r_size is None: r_size = code.prgm.acquire_register()
else: r_size = self.r_size
# Create variables
ea_addr = var.SignedWord(reg = r_ea_data)
aligned_size = var.SignedWord(0)
mod_16 = var.SignedWord(0xF)
# Initialize the lsa_addr variable.
if isinstance(lsa, int):
# From a constant
ls_addr = var.SignedWord(lsa)
elif issubclass(type(lsa), (spe.Register, spe.Variable)):
# From a variable
ls_addr = var.SignedWord()
ls_addr.v = lsa
tag_var = var.SignedWord(tag)
cmp = var.SignedWord(0)
# Load the effective address
if self.r_addr is None:
if self.addr % 16 != 0:
print '[get_memory] Misaligned data'
util.load_word(code, ea_addr, self.addr)
# Load the size, rounding up as required to be 16-byte aligned
if self.r_size is None:
rnd_size = self.size * var.INT_SIZES[self.typecode]
if rnd_size < 16:
rnd_size = 16
elif (rnd_size % 16) != 0:
rnd_size += (16 - (rnd_size % 16))
util.load_word(code, aligned_size, rnd_size)
else:
# TODO: !!! UNIT TEST THIS !!!
# Same as above, but using SPU arithemtic to round
size = var.SignedWord(reg = r_size)
sixteen = var.SignedWord(16)
cmp.v = ((size & mod_16) == size)
aligned_size.v = size + (sixteen - (size & mod_16))
spu.selb(aligned_size.reg, size.reg, aligned_size.reg, cmp.reg, order = _mi(spu.selb))
code.release_register(sixteen.reg)
# Use an auxillary register for the moving ea value if the
# caller supplied the address register
if self.r_addr is not None:
ea_load = var.SignedWord(0)
ea_load.v = ea_addr
else:
ea_load = ea_addr # note that this is reference, not .v assignment
# Transfer parameters
buffer_size = var.SignedWord(16384)
remaining = var.SignedWord(0)
transfer_size = var.SignedWord(0)
remaining.v = aligned_size
# Set up the iterators to transfer at most 16k at a time
xfer_iter = syn_iter(code, 0, 16384)
xfer_iter.set_stop_reg(aligned_size.reg)
for offset in xfer_iter:
cmp.v = buffer_size > remaining
spu.selb(transfer_size, buffer_size, remaining, cmp)
# Transfer the data
kernel(code, ls_addr, ea_load, transfer_size, tag_var)
ls_addr.v = ls_addr + buffer_size
ea_load.v = ea_load + buffer_size
remaining.v = remaining - buffer_size
# Set the tag bit to tag
dma.mfc_write_tag_mask(code, 1<<tag);
# Wait for the transfer to complete
dma.mfc_read_tag_status_all(code);
# Release the registers
code.release_register(buffer_size.reg)
code.release_register(remaining.reg)
code.release_register(aligned_size.reg)
code.release_register(transfer_size.reg)
code.release_register(cmp.reg)
code.release_register(ls_addr.reg)
code.release_register(tag_var.reg)
code.release_register(ea_load.reg)
if old_code is not None:
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.x is None: raise Exception("Please set x")
if self.result is None: raise Exception("Please set result")
# exponent
e = var.Word()
# Working values
x = var.Word()
y = var.Word()
z = var.Word()
cmp = var.Bits()
tmp = var.Word()
spu.xor(cmp, cmp, cmp)
spu.xor(tmp, tmp, tmp)
# Set the working x
x.v = self.x
# Extract the exponent
# int e = (((*(unsigned int *) &x) >> 23) & 0xff) - 0x7e;
e.v = x >> self.consts['_23']
e.v = spu.andi.ex(e, 0xff)
e.v = spu.ai.ex(e, 0x382) # 0x382 == (- 0x7E) using 10 bits
# 0b 111 1110
# Extract the mantissa
x.v = x & self.consts['M1'] # *(unsigned int*)&x &= 0x807fffff;
x.v = x | self.consts['M2'] # *(unsigned int*)&x |= 0x3f000000;
# Normalize
x1, x2, e1 = y, z, tmp
# if (x < SQRTHF)
cmp.v = spu.fcgt.ex(self.consts['SQRTHF'], x)
# (True) { ... }
e1.v = spu.ai.ex(e, -1) # e -= 1;
x1.v = spu.fa.ex(x, x) # x = x + x - 1.0;
x1.v = spu.fs.ex(x1, self.consts['ONE']) # "" ""
# (False) { ... }
x2.v = spu.fs.ex(x, self.consts['ONE']) # x = x - 1.0;
# Select the True/False values based on cmp
e.v = spu.selb.ex(e, e1, cmp)
x.v = spu.selb.ex(x2, x1, cmp)
# Compute polynomial
z.v = spu.fm.ex(x, x) # z = x * x;
y.v = spu.fms.ex(
self.consts['C1'],
x, # y = (((((((( 7.0376836292E-2 * x
self.consts['C2']) # - 1.1514610310E-1) * x
y.v = spu.fma.ex(y, x,
self.consts['C3']) # + 1.1676998740E-1) * x
y.v = spu.fms.ex(y, x, self.consts['C4']) # - 1.2420140846E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C5']) # + 1.4249322787E-1) * x
y.v = spu.fms.ex(y, x, self.consts['C6']) # - 1.6668057665E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C7']) # + 2.0000714765E-1) * x
y.v = spu.fms.ex(y, x, self.consts['C8']) # - 2.4999993993E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C9']) # + 3.3333331174E-1)
y.v = spu.fm.ex(y, x) # * x
y.v = spu.fm.ex(y, z) # * z;
y.v = spu.fma.ex(self.consts['C10'], z, y) # y += -0.5 * z;
# Convert to log base 2
z.v = spu.fm.ex(y, self.consts['LOG2EA']) # z = y * LOG2EA;
z.v = spu.fma.ex(x, self.consts['LOG2EA'], z) # z += x * LOG2EA;
z.v = spu.fa.ex(z, y) # z += y;
z.v = spu.fa.ex(z, x) # z += x;
e.v = spu.csflt.ex(e, 155) # z += (float) e;
z.v = spu.fa.ex(z, e) # "" ""
spu.ai(self.result, z, 0) # return z
spu.set_active_code(old_code)
return
def isched(scode):
old_active_code = spu.get_active_code()
spu.set_active_code(None)
# Generate the instruction dependence DAG(s)
blocks = isched_gen_blocks(scode)
# For each instruction, compute the max cycles to the end of the code
g_critpath = critpath_block(blocks)
# Apply heuristics to build an optimized InstructionStream
fcode = scode.prgm.get_stream()
inst_cycle = {} # For each inst, the cycle number it has in the code
lastpos = -1 # Index of last instruction in the stream (excludes labels!)
pipe = 0 # Current pipeline (0 = even, 1 = odd)
cycle = 0 # Current cycle number
for (ind, block) in enumerate(blocks):
if block.label is not None:
fcode.add(block.label)
start = block.start
g_in = block.g_in
g_incnt = block.g_incnt
g_out = block.g_out
while len(start) > 0:
# Apply heuristics to find the best instruction in the queue.
# For each inst in start, compute the minimum stall time
# TODO - cache this instead of computing each time?
# Do this by computing the stall time when an inst is added to start.
# Each time the cycle number is moved forward, reduce the stall time
# by that number of cycles for each inst in start.
# TODO - idea from I think Muchnick -- keep a start Q of no-stall nodes,
# and a Q of nodes that would stall. Then just pull from no-stall nodes
# unless empty, in which case fall back to the stall Q
# would make it easy(er) to do cached stall counts
best = (None, 999)
for s in start:
# Find the stall time of s, or maximum delay for all its deps
maxstall = 0
for d in g_in[s]:
if d[0] == None:
continue
# Compute stall time for this dep
stall = d[1] - (cycle - inst_cycle[d[0]])
if stall > maxstall:
maxstall = stall
best = heurcompare_block(best, (s, maxstall), pipe, g_critpath, blocks, ind)
inst = best[0]
start.remove(inst)
cycle += best[1] + 1
block.inst_cnt -= 1
fcode.add(inst)
# Dual issue? if so, adjust the cycle back one.
# Careful, lastpos starts out as -1. However the pipe also starts out
# as 0, so the first part of the conditional will fail before lastpos
# is used.
# Ah - if a label occurs first in the stream, followed by say an ai,
# this will fail
previnst = fcode[lastpos]
if (pipe == inst.cycles[0] == 1 and
previnst.cycles[0] == 0 and
inst_cycle[previnst] == cycle - 1):
cycle -= 1
inst_cycle[inst] = cycle
lastpos = len(fcode) - 1
pipe = (pipe + 1) & 1
# Evaluate all the instructions that depend on this inst.
# Can any be added to start?
for d in g_out[inst]:
# Skip this d if it's not in the current block
if d not in g_incnt:
continue
g_incnt[d] -= 1
if g_incnt[d] == 0:
start.append(d)
# Does d have depend on any insts in start?
# If so, move those to the front of start
# Why does this still matter? It affects ties in the heuristic..
else:
# d depends on inst, and at least 1 other inst not in start.
# look at the insts d depends on. if any are in start, move
# them to the front of the start set.
# how does this help? helps insts closer to getting into start,
# get in sooner. get a larger start set for choosing best inst
for e in g_in[d]:
if e[0] in start:
start.remove(e[0])
start.insert(0, e[0])
# end while len(start) > 0
if block.branch is not None:
fcode.add(block.branch)
# end for block in blocks
spu.set_active_code(old_active_code)
return fcode
def synthesize(self, code):
self._check_inputs()
old_code = spu.get_active_code()
spu.set_active_code(code)
zero = var.Word(reg = code.r_zero)
one = self.log.consts['ONE']
two = self.consts['TWO']
x = var.Word(self.x0)
r = var.Word(0)
cmp = var.Word(0)
x_neg = var.Word(0)
fmax = var.Word(self.max_init)
temp = var.SingleFloat()
fmax.v = spu.cuflt.ex(fmax, 155)
# Init
for i in spuiter.syn_iter(code, self.max_init):
# x = r[i % r_max] * x * (1.0 - x)
self._next_r(r)
temp.v = spu.fs.ex(one, x)
x.v = spu.fm.ex(x, temp)
x.v = spu.fm.ex(r, x)
# if x == float('-infinity'):
# return -10.0
# Derive Exponent
total = var.Word(0)
logx = var.SingleFloat()
for i in spuiter.syn_iter(code, self.max_n):
# x = ri * x * (1.0 - x)
self._next_r(r)
temp.v = spu.fs.ex(one, x)
x.v = spu.fm.ex(x, temp)
x.v = spu.fm.ex(r, x)
# logx = ri - 2.0 * ri * x
logx.v = spu.fm.ex(two, x)
logx.v = spu.fm.ex(r, logx)
logx.v = spu.fs.ex(r, logx)
# abs(logx)
x_neg.v = spu.fs.ex(zero, logx)
cmp.v = spu.fcgt.ex(logx, zero)
logx.v = spu.selb.ex(x_neg, logx, cmp)
# logx.v = spu.selb.ex(logx, x_neg, cmp)
# log(logx)
self.log.set_result(logx)
self.log.set_x(logx)
self.log.synthesize(code)
# total = total + x
total.v = spu.fa.ex(total, logx)
# return total / float(max_n)
fdiv(code, self.result, total, fmax, one)
spu.set_active_code(code)
return
def _get_active_code(self):
return spu.get_active_code()
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
# Sanity checks
if self._x_addr is None: raise Exception("Please set x_addr")
if self._y_addr is None: raise Exception("Please set y_addr")
if self._n_bits is None: raise Exception("Please set n_bits")
if self._m is None: raise Exception("Please set m")
if self._n is None: raise Exception("Please set n")
# Acquire a registers for the bit vectors and result
n_vecs = self._n_bits / 128
x_regs = [code.acquire_register() for i in range(n_vecs)]
y_regs = [code.acquire_register() for i in range(n_vecs)]
result = code.acquire_register()
x_addr = var.Word()
y_addr = var.Word()
if self._save_op is not None:
if self._threshold is not None:
threshold = var.SingleFloat(self._threshold)
else:
threshold = var.SingleFloat(0.0)
bcmp = var.Word(0)
# Setup the Tanimito kernel
tan = Tanimoto()
tan.set_n_bits(self._n_bits)
tan.set_x_regs(x_regs)
tan.set_y_regs(y_regs)
tan.set_result(result)
tan.synthesize_constants(code)
# Setup the save op
save_op = self._save_op
if save_op is not None:
save_op.setup()
# Create the iterators
xiter = spuiter.syn_iter(code, self._m)
yiter = spuiter.syn_iter(code, self._n)
# Synthesize the block comparison loops
x_addr.v = self._x_addr
for x_off in xiter:
x_addr.v = x_addr + 16 * n_vecs
y_addr.v = self._y_addr
self._load_bit_vector(x_addr, x_regs)
for y_off in yiter:
y_addr.v = y_addr + 16 * n_vecs
self._load_bit_vector(y_addr, y_regs)
tan.synthesize(code)
if save_op is not None:
spu.fcgt(bcmp, result, threshold)
save_op.test(bcmp, result, x_off, y_off)
# /x_off
if old_code is not None:
spu.set_active_code(old_code)
return
def synthesize(self, code):
old_code = spu.get_active_code()
spu.set_active_code(code)
if self.x is None: raise Exception("Please set x")
if self.result is None: raise Exception("Please set result")
# exponent
e = var.Word()
# Working values
x = var.Word()
y = var.Word()
z = var.Word()
cmp = var.Bits()
tmp = var.Word()
spu.xor(cmp, cmp, cmp)
spu.xor(tmp, tmp, tmp)
# Set the working x
x.v = self.x
# Extract the exponent
# int e = (((*(unsigned int *) &x) >> 23) & 0xff) - 0x7e;
e.v = x >> self.consts['_23']
e.v = spu.andi.ex(e, 0xff)
e.v = spu.ai.ex(e, 0x382) # 0x382 == (- 0x7E) using 10 bits
# 0b 111 1110
# Extract the mantissa
x.v = x & self.consts['M1'] # *(unsigned int*)&x &= 0x807fffff;
x.v = x | self.consts['M2'] # *(unsigned int*)&x |= 0x3f000000;
# Normalize
x1, x2, e1 = y, z, tmp
# if (x < SQRTHF)
cmp.v = spu.fcgt.ex(self.consts['SQRTHF'], x)
# (True) { ... }
e1.v = spu.ai.ex(e, -1) # e -= 1;
x1.v = spu.fa.ex(x, x) # x = x + x - 1.0;
x1.v = spu.fs.ex(x1, self.consts['ONE']) # "" ""
# (False) { ... }
x2.v = spu.fs.ex(x, self.consts['ONE']) # x = x - 1.0;
# Select the True/False values based on cmp
e.v = spu.selb.ex(e, e1, cmp)
x.v = spu.selb.ex(x2, x1, cmp)
# Compute polynomial
z.v = spu.fm.ex(x, x) # z = x * x;
y.v = spu.fms.ex(self.consts['C1'], x, # y = (((((((( 7.0376836292E-2 * x
self.consts['C2']) # - 1.1514610310E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C3']) # + 1.1676998740E-1) * x
y.v = spu.fms.ex(y, x, self.consts['C4']) # - 1.2420140846E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C5']) # + 1.4249322787E-1) * x
y.v = spu.fms.ex(y, x, self.consts['C6']) # - 1.6668057665E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C7']) # + 2.0000714765E-1) * x
y.v = spu.fms.ex(y, x, self.consts['C8']) # - 2.4999993993E-1) * x
y.v = spu.fma.ex(y, x, self.consts['C9']) # + 3.3333331174E-1)
y.v = spu.fm.ex(y, x) # * x
y.v = spu.fm.ex(y, z) # * z;
y.v = spu.fma.ex(self.consts['C10'], z, y) # y += -0.5 * z;
# Convert to log base 2
z.v = spu.fm.ex( y, self.consts['LOG2EA']) # z = y * LOG2EA;
z.v = spu.fma.ex(x, self.consts['LOG2EA'], z) # z += x * LOG2EA;
z.v = spu.fa.ex(z, y) # z += y;
z.v = spu.fa.ex(z, x) # z += x;
e.v = spu.csflt.ex(e, 155) # z += (float) e;
z.v = spu.fa.ex(z, e) # "" ""
spu.ai(self.result, z, 0) # return z
spu.set_active_code(old_code)
return
def isched(scode):
old_active_code = spu.get_active_code()
spu.set_active_code(None)
# Generate the instruction dependence DAG(s)
blocks = isched_gen_blocks(scode)
# For each instruction, compute the max cycles to the end of the code
g_critpath = critpath_block(blocks)
# Apply heuristics to build an optimized InstructionStream
fcode = scode.prgm.get_stream()
inst_cycle = {} # For each inst, the cycle number it has in the code
lastpos = -1 # Index of last instruction in the stream (excludes labels!)
pipe = 0 # Current pipeline (0 = even, 1 = odd)
cycle = 0 # Current cycle number
for (ind, block) in enumerate(blocks):
if block.label is not None:
fcode.add(block.label)
start = block.start
g_in = block.g_in
g_incnt = block.g_incnt
g_out = block.g_out
while len(start) > 0:
# Apply heuristics to find the best instruction in the queue.
# For each inst in start, compute the minimum stall time
# TODO - cache this instead of computing each time?
# Do this by computing the stall time when an inst is added to start.
# Each time the cycle number is moved forward, reduce the stall time
# by that number of cycles for each inst in start.
# TODO - idea from I think Muchnick -- keep a start Q of no-stall nodes,
# and a Q of nodes that would stall. Then just pull from no-stall nodes
# unless empty, in which case fall back to the stall Q
# would make it easy(er) to do cached stall counts
best = (None, 999)
for s in start:
# Find the stall time of s, or maximum delay for all its deps
maxstall = 0
for d in g_in[s]:
if d[0] == None:
continue
# Compute stall time for this dep
stall = d[1] - (cycle - inst_cycle[d[0]])
if stall > maxstall:
maxstall = stall
best = heurcompare_block(best, (s, maxstall), pipe, g_critpath,
blocks, ind)
inst = best[0]
start.remove(inst)
cycle += best[1] + 1
block.inst_cnt -= 1
fcode.add(inst)
# Dual issue? if so, adjust the cycle back one.
# Careful, lastpos starts out as -1. However the pipe also starts out
# as 0, so the first part of the conditional will fail before lastpos
# is used.
# Ah - if a label occurs first in the stream, followed by say an ai,
# this will fail
previnst = fcode[lastpos]
if (pipe == inst.cycles[0] == 1 and previnst.cycles[0] == 0
and inst_cycle[previnst] == cycle - 1):
cycle -= 1
inst_cycle[inst] = cycle
lastpos = len(fcode) - 1
pipe = (pipe + 1) & 1
# Evaluate all the instructions that depend on this inst.
# Can any be added to start?
for d in g_out[inst]:
# Skip this d if it's not in the current block
if d not in g_incnt:
continue
g_incnt[d] -= 1
if g_incnt[d] == 0:
start.append(d)
# Does d have depend on any insts in start?
# If so, move those to the front of start
# Why does this still matter? It affects ties in the heuristic..
else:
# d depends on inst, and at least 1 other inst not in start.
# look at the insts d depends on. if any are in start, move
# them to the front of the start set.
# how does this help? helps insts closer to getting into start,
# get in sooner. get a larger start set for choosing best inst
for e in g_in[d]:
if e[0] in start:
start.remove(e[0])
start.insert(0, e[0])
# end while len(start) > 0
if block.branch is not None:
fcode.add(block.branch)
# end for block in blocks
spu.set_active_code(old_active_code)
return fcode
def _get_active_code(self):
return spu.get_active_code()
def synthesize(self, code):
self._check_inputs()
old_code = spu.get_active_code()
spu.set_active_code(code)
zero = var.Word(reg=code.r_zero)
one = self.log.consts['ONE']
two = self.consts['TWO']
x = var.Word(self.x0)
r = var.Word(0)
cmp = var.Word(0)
x_neg = var.Word(0)
fmax = var.Word(self.max_init)
temp = var.SingleFloat()
fmax.v = spu.cuflt.ex(fmax, 155)
# Init
for i in spuiter.syn_iter(code, self.max_init):
# x = r[i % r_max] * x * (1.0 - x)
self._next_r(r)
temp.v = spu.fs.ex(one, x)
x.v = spu.fm.ex(x, temp)
x.v = spu.fm.ex(r, x)
# if x == float('-infinity'):
# return -10.0
# Derive Exponent
total = var.Word(0)
logx = var.SingleFloat()
for i in spuiter.syn_iter(code, self.max_n):
# x = ri * x * (1.0 - x)
self._next_r(r)
temp.v = spu.fs.ex(one, x)
x.v = spu.fm.ex(x, temp)
x.v = spu.fm.ex(r, x)
# logx = ri - 2.0 * ri * x
logx.v = spu.fm.ex(two, x)
logx.v = spu.fm.ex(r, logx)
logx.v = spu.fs.ex(r, logx)
# abs(logx)
x_neg.v = spu.fs.ex(zero, logx)
cmp.v = spu.fcgt.ex(logx, zero)
logx.v = spu.selb.ex(x_neg, logx, cmp)
# logx.v = spu.selb.ex(logx, x_neg, cmp)
# log(logx)
self.log.set_result(logx)
self.log.set_x(logx)
self.log.synthesize(code)
# total = total + x
total.v = spu.fa.ex(total, logx)
# return total / float(max_n)
fdiv(code, self.result, total, fmax, one)
spu.set_active_code(code)
return
