def fdiv(code, d, x, y, one = None):
"""
Single-precision floating point division for x / y
"""
Y = code.acquire_registers(3)
t = code.acquire_register()
regs = Y[:]
regs.append(t)
if one is None:
one = code.acquire_register()
spu.xor(one, one, one)
spu.ai(one, one, 1)
spu.cuflt(one, one, 155)
regs.append(one)
# Compute 1/y (from SPU ISA 1.1, p208, Normal case)
spu.frest(Y[0], y)
spu.fi(Y[1], y, Y[0])
spu.fnms(t, y, Y[1], one)
spu.fma(Y[2], t, Y[1], Y[1])
# Compute x * (1/y)
spu.fm(d, x, Y[2])
code.release_registers(regs)
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 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 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
y1 = 121 y2 = 122 t1 = 119 result = 118 ione = 110 fone = 111 insts = [ # Create fone = 1.0, fa = 2.0 and fb = 4.0 spu.ai(ione, 0, 1), spu.ai(ia, 0, 2), spu.ai(ib, 0, 4), spu.cuflt(fone, ione, 155), spu.cuflt(fa, ia, 155), spu.cuflt(fb, ib, 155), # Compute 1/fb spu.frest(y0, fb), spu.fi(y1, fb, y0), spu.fnms(t1, fb, y1, fone), spu.fma(y2, t1, y1, y1), spu.fm(result, fa, y2) ] for inst in insts: cli.execute(inst)
