def test_4comp():
proc = env.Processor(0)
prgm = env.Program()
code = prgm.get_stream()
inp = proc.alloc_remote('i', 1, 4, 1)
out = proc.alloc_remote('i', 4, 1, 1)
for i in xrange(0, 4):
inp[i] = i + 1
out[i] = 0
print "inp", inp[0:4]
print "out", out[0:4]
cal.set_active_code(code)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_resource(0, cal.pixtex_type.oned, cal.fmt.float,
UNNORM=True) # positions
r_cnt = prgm.acquire_register()
r = prgm.acquire_registers(4)
cal.mov(r_cnt, r_cnt('0000'))
for i in xrange(0, 4):
cal.sample(0, 0, r[i].x000, r_cnt.x)
cal.add(r_cnt, r_cnt, r_cnt('1111'))
cal.iadd(r[0], r[0], r[1]('0x00'))
cal.iadd(r[0], r[0], r[2]('00x0'))
cal.iadd(r[0], r[0], r[3]('000x'))
cal.iadd(r[0], r[0], r[0])
cal.mov(reg.o0, r[0])
prgm.set_binding(reg.i0, inp)
prgm.set_binding(reg.o0, out)
prgm.add(code)
prgm.print_code()
proc.execute(prgm, (0, 0, 1, 1))
print "inp", inp[0:4]
print "out", out[0:4]
for i in xrange(0, 4):
assert (out[i] == (i + 1) * 2)
return
def test_4comp():
proc = env.Processor(0)
prgm = env.Program()
code = prgm.get_stream()
inp = proc.alloc_remote('i', 1, 4, 1)
out = proc.alloc_remote('i', 4, 1, 1)
for i in xrange(0, 4):
inp[i] = i + 1
out[i] = 0
print "inp", inp[0:4]
print "out", out[0:4]
cal.set_active_code(code)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_resource(0, cal.pixtex_type.oned, cal.fmt.float, UNNORM=True) # positions
r_cnt = prgm.acquire_register()
r = prgm.acquire_registers(4)
cal.mov(r_cnt, r_cnt('0000'))
for i in xrange(0, 4):
cal.sample(0, 0, r[i].x000, r_cnt.x)
cal.add(r_cnt, r_cnt, r_cnt('1111'))
cal.iadd(r[0], r[0], r[1]('0x00'))
cal.iadd(r[0], r[0], r[2]('00x0'))
cal.iadd(r[0], r[0], r[3]('000x'))
cal.iadd(r[0], r[0], r[0])
cal.mov(reg.o0, r[0])
prgm.set_binding(reg.i0, inp)
prgm.set_binding(reg.o0, out)
prgm.add(code)
prgm.print_code()
proc.execute(prgm, (0, 0, 1, 1))
print "inp", inp[0:4]
print "out", out[0:4]
for i in xrange(0, 4):
assert(out[i] == (i + 1) * 2)
return
def TestSimpleKernelNPy():
import corepy.arch.cal.isa as isa
SIZE = 128
proc = Processor(0)
arr_input = proc.alloc_remote_npy('f', 4, SIZE, SIZE)
arr_output = proc.alloc_remote_npy('f', 4, SIZE, SIZE)
#for i in xrange(0, SIZE * SIZE * 4):
# arr_input[i] = float(i + 1)
# arr_output[i] = 0.0
#print arr_input.shape
#print arr_output.shape
#print type(arr_input.data)
val = 0.0
for i in xrange(0, SIZE):
for j in xrange(0, SIZE):
for k in xrange(0, 4):
arr_input[i][j][k] = val
arr_output[i][j][k] = 0.0
val += 1.0
# build and run the kernel
prgm = Program()
code = prgm.get_stream()
#code.add(isa.dcl_input('v0', USAGE=isa.usage.pos, INTERP='linear_noperspective'))
code.add("dcl_input_position_interp(constant) v0.xy__")
code.add(isa.dcl_output('o0', USAGE=isa.usage.generic))
code.add(isa.dcl_resource(0, '2d', isa.fmt.float, UNNORM=True))
code.add(isa.sample(0, 0, 'o0', 'v0.xy'))
#code.add(isa.load(0, 'o0', 'v0.g'))
domain = (0, 0, SIZE, SIZE)
prgm.set_binding("o0", arr_output)
prgm.set_binding("i0", arr_input)
prgm.add(code)
prgm.cache_code()
prgm.print_code()
proc.execute(prgm, domain)
# Check the output
val = 0.0
for i in xrange(0, SIZE):
for j in xrange(0, SIZE):
for k in xrange(0, 4):
if arr_output[i][j][k] != val:
print "ERROR index %d is %f, should be %f" % (i, arr_output[i], val)
val += 1.0
return
def TestSimpleKernel():
import corepy.arch.cal.isa as isa
SIZE = 128
proc = Processor(0)
ext_input = proc.alloc_remote('f', 4, SIZE, SIZE)
ext_output = proc.alloc_remote('f', 4, SIZE, SIZE)
for i in xrange(0, SIZE * SIZE * 4):
ext_input[i] = float(i + 1)
ext_output[i] = 0.0
# build and run the kernel
prgm = Program()
code = prgm.get_stream()
#code.add(isa.dcl_input('v0', USAGE=isa.usage.pos, INTERP='linear_noperspective'))
code.add("dcl_input_position_interp(constant) v0.xy__")
code.add(isa.dcl_output('o0', USAGE=isa.usage.generic))
code.add(isa.dcl_resource(0, '2d', isa.fmt.float, UNNORM=True))
code.add(isa.sample(0, 0, 'o0', 'v0.xy'))
#code.add(isa.load(0, 'o0', 'v0.g'))
domain = (0, 0, SIZE, SIZE)
prgm.set_binding("o0", ext_output)
prgm.set_binding("i0", ext_input)
prgm.add(code)
prgm.cache_code()
prgm.print_code()
proc.execute(prgm, domain)
# Check the output
for i in xrange(0, SIZE * SIZE * 4):
if ext_output[i] != float(i + 1):
print "ERROR index %d is %f, should be %f" % (i, ext_output[i], float(i + 1))
proc.free(ext_input)
proc.free(ext_output)
return
code = prgm.get_stream()
inp = proc.alloc_remote('f', 4, 64)
out = proc.alloc_remote('f', 4, 64)
out.clear()
for i in xrange(0, 64):
inp[i] = float(i + 1)
cal.set_active_code(code)
cal.dcl_input(reg.v0.x, USAGE=cal.usage.pos)
cal.dcl_resource(0, cal.pixtex_type.oned, cal.fmt.float, UNNORM = True)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.sample(0, 0, reg.o0, reg.v0.x)
prgm.set_binding(reg.i0, inp)
prgm.set_binding(reg.o0, out)
prgm.add(code)
prgm.print_code()
proc.execute(prgm)
print "inp", inp
print "out", out
import corepy.lib.printer as printer
printer.PrintInstructionStream(code, printer.CAL_Asm())
def cal_nb_generate(n_bodies, dt):
code = env.InstructionStream()
cal.set_active_code(code)
fn_bodies = float(n_bodies)
r_count = code.acquire_register()
r_lpos = code.acquire_register()
r_rpos = code.acquire_register()
r_force = code.acquire_register()
r_diff = code.acquire_register()
r_dist_vec = code.acquire_register()
r_dist = code.acquire_register()
r_force_tmp = code.acquire_register()
r_force_vec = code.acquire_register()
r_vel = code.acquire_register()
#code.add("dcl_input_position_interp(linear_noperspective) v0.x___")
cal.dcl_input(reg.v0.x___, USAGE=cal.usage.pos, INTERP=cal.interp.linear_noperspective)
r_bodies = code.acquire_register((fn_bodies,) * 4)
r_G = code.acquire_register((G,) * 4)
r_dt = code.acquire_register((dt,) * 4)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_output(reg.o1, USAGE=cal.usage.generic)
cal.dcl_resource(0, cal.pixtex_type.oned, cal.fmt.float, UNNORM=True) # positions
cal.dcl_resource(1, cal.pixtex_type.oned, cal.fmt.float, UNNORM=True) # velocities
# Loop over all other points to calculate the force
cal.mov(r_count, r_count('0000')) # loop counter
cal.sample(0, 0, r_lpos, reg.v0.x) # Local position
cal.mov(r_force, r_force('0000')) # total force
# Compute force using input from every other point
cal.whileloop()
# Break if end of points reached
cal.breakc(cal.relop.ge, r_count, r_bodies)
cal.sample(0, 0, r_rpos, r_count.x) # Remote position
# d_xyz
cal.sub(r_diff, r_lpos.xyz0, r_rpos.xyz0) # local pos - remote pos
# dist_tmp
cal.mul(r_dist_vec, r_diff.xxxx, r_diff.xxxx)
cal.mad(r_dist_vec, r_diff.yyyy, r_diff.yyyy, r_dist_vec)
cal.mad(r_dist_vec, r_diff.zzzz, r_diff.zzzz, r_dist_vec)
# distance
# TODO - skip rest of force computation if distance is 0
cal.sqrt_vec(r_dist, r_dist_vec)
# force G * ((m[i] * m[j]) / dist_tmp)
cal.mul(r_force_tmp, r_lpos.wwww, r_rpos.wwww)
cal.div(cal.zeroop.zero, r_force_tmp, r_force_tmp, r_dist_vec)
cal.mul(r_force_tmp, r_force_tmp, r_G)
# f_xyz
cal.div(cal.zeroop.zero, r_force_vec, r_diff.xyz0, r_dist.xyz1)
cal.mul(r_force_vec, r_force_vec.xyz0, r_force_tmp.xyz0)
cal.sub(r_force, r_force.xyz0, r_force_vec.xyz0)
# Increment loop counter, end loop
cal.add(r_count, r_count, r_count('1111'))
cal.endloop()
# Acceleration
cal.div(cal.zeroop.zero, r_force, r_force.xyz0, r_lpos.wwww)
# Velocity
cal.sample(1, 1, r_vel, reg.v0.x) # Load velocity
cal.mad(r_vel, r_force, r_dt, r_vel)
cal.mov(reg.o1, r_vel)
# Position
cal.mad(reg.o0, r_vel.xyz0, r_dt.xyz0, r_lpos.xyzw)
return code
def cal_nb_generate_local(n_bodies, dt, steps):
code = env.InstructionStream()
cal.set_active_code(code)
fn_bodies = float(n_bodies)
steps = float(steps)
r_count = code.acquire_register()
r_step = code.acquire_register()
r_lpos = code.acquire_register()
r_rpos = code.acquire_register()
r_force = code.acquire_register()
r_diff = code.acquire_register()
r_dist_vec = code.acquire_register()
r_dist = code.acquire_register()
r_force_tmp = code.acquire_register()
r_force_vec = code.acquire_register()
r_vel = code.acquire_register()
print "fn_bodies", fn_bodies
code.add("dcl_input_position_interp(linear_noperspective) v0.xy__")
#cal.dcl_input(reg.v0.x___, USAGE=cal.usage.pos, INTERP=cal.interp.linear_noperspective)
r_numsteps = code.acquire_register((steps,) * 4)
r_bodies = code.acquire_register((fn_bodies,) * 4)
#r_bodiesquare = code.acquire_register((float(fn_bodies**2),) * 4)
r_G = code.acquire_register((G,) * 4)
r_dt = code.acquire_register((dt,) * 4)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_output(reg.o1, USAGE=cal.usage.generic)
cal.dcl_output(reg.o2, USAGE=cal.usage.generic)
cal.dcl_resource(0, cal.pixtex_type.twod, cal.fmt.float, UNNORM=True) # positions
cal.dcl_resource(1, cal.pixtex_type.twod, cal.fmt.float, UNNORM=True) # velocities
r_foo = code.acquire_register()
cal.mov(r_foo, r_foo('0000'))
r_gpos = code.acquire_register()
cal.mad(r_gpos, reg.v0.y, r_bodies.x, reg.v0.x)
r_gvel = code.acquire_register()
cal.mad(r_gvel, r_bodies.x, r_bodies.x, r_gpos)
cal.ftoi(r_gpos, r_gpos)
cal.ftoi(r_gvel, r_gvel)
cal.sample(0, 0, r_lpos, reg.v0.xy) # Local position
cal.sample(1, 1, r_vel, reg.v0.xy) # Load velocity
cal.mov(reg.g[r_gpos.x], r_lpos)
cal.mov(reg.g[r_gvel.x], r_vel)
cal.mov(r_step, r_step('0000'))
cal.whileloop()
cal.breakc(cal.relop.ge, r_step.x, r_numsteps)
cal.mov(r_count, r_count('0000')) # loop counter
cal.whileloop()
cal.breakc(cal.relop.ge, r_count.x, r_bodies)
cal.add(r_foo, r_foo, r_foo('1111'))
# calculate force
r_tmp = code.acquire_register()
cal.ftoi(r_tmp, r_count)
cal.mov(r_rpos, reg.g[r_tmp.x])
# d_xyz
cal.sub(r_diff, r_lpos.xyz0, r_rpos.xyz0) # local pos - remote pos
# dist_tmp
cal.mul(r_dist_vec, r_diff.xxxx, r_diff.xxxx)
cal.mad(r_dist_vec, r_diff.yyyy, r_diff.yyyy, r_dist_vec)
cal.mad(r_dist_vec, r_diff.zzzz, r_diff.zzzz, r_dist_vec)
# distance
# TODO - skip rest of force computation if distance is 0
cal.sqrt_vec(r_dist, r_dist_vec)
# force G * ((m[i] * m[j]) / dist_tmp)
cal.mul(r_force_tmp, r_lpos.wwww, r_rpos.wwww)
cal.div(r_force_tmp, r_force_tmp, r_dist_vec, ZEROOP = cal.zeroop.zero)
cal.mul(r_force_tmp, r_force_tmp, r_G)
# f_xyz
# TODO - whats going on, is this right?
cal.div(r_force_vec, r_diff.xyz0, r_dist.xyz1, ZEROOP = cal.zeroop.zero)
cal.mul(r_force_vec, r_force_vec.xyz0, r_force_tmp.xyz0)
cal.sub(r_force, r_force.xyz0, r_force_vec.xyz0)
cal.add(r_count, r_count, r_count('1111'))
cal.endloop()
# Acceleration
cal.div(r_force, r_force.xyz0, r_lpos.wwww, ZEROOP = cal.zeroop.zero)
# Velocity
cal.mad(r_vel, r_force, r_dt, r_vel)
# Position
cal.mad(reg.o0, r_vel.xyz0, r_dt.xyz0, r_lpos.xyzw)
# store updated pos and vel
cal.mov(reg.g[r_gpos.x], r_lpos)
cal.mov(reg.g[r_gvel.x], r_vel)
cal.add(r_step, r_step, r_step('1111'))
cal.endloop()
cal.mov(reg.o0, r_lpos)
cal.mov(reg.o1, r_vel)
cal.mov(reg.o2, r_foo)
return code
def cal_nb_generate_2d(prgm, n_bodies, dt):
code = prgm.get_stream()
cal.set_active_code(code)
fn_bodies = float(n_bodies)
#r_cx = prgm.acquire_register()
#r_cy = prgm.acquire_register()
r_count = prgm.acquire_register()
r_lpos = prgm.acquire_register()
r_rpos = prgm.acquire_register()
r_force = prgm.acquire_register()
r_diff = prgm.acquire_register()
r_dist_vec = prgm.acquire_register()
r_dist = prgm.acquire_register()
r_force_tmp = prgm.acquire_register()
r_force_vec = prgm.acquire_register()
r_vel = prgm.acquire_register()
#code.add("dcl_input_position_interp(linear_noperspective) v0.xy__")
cal.dcl_input(reg.v0.xy__, USAGE=cal.usage.pos, INTERP=cal.interp.linear_noperspective)
r_bodies = prgm.acquire_register((fn_bodies,) * 4)
r_G = prgm.acquire_register((G,) * 4)
r_dt = prgm.acquire_register((dt,) * 4)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_output(reg.o1, USAGE=cal.usage.generic)
cal.dcl_resource(0, cal.pixtex_type.twod, cal.fmt.float, UNNORM=True) # positions
cal.dcl_resource(1, cal.pixtex_type.twod, cal.fmt.float, UNNORM=True) # velocities
# Loop over all other points to calculate the force
cal.mov(r_count, r_count('0000')) # loop counter
#cal.mov(r_cx, r_cx('0000')) # loop counter
#cal.mov(r_cy, r_cy('0000')) # loop counter
cal.sample(0, 0, r_lpos, reg.v0.xy) # Local position
cal.mov(r_force, r_force('0000')) # total force
# Compute force using input from every other point
cal.whileloop()
cal.breakc(cal.relop.ge, r_count.x, r_bodies)
cal.mov(r_count, r_count.x0zw)
cal.whileloop()
cal.breakc(cal.relop.ge, r_count.y, r_bodies)
#for i in xrange(0, 4):
#cal.add(r_count, r_cx('x000'), r_cy('0x00'))
cal.sample(0, 0, r_rpos, r_count.xy) # Remote position
# d_xyz
cal.sub(r_diff, r_lpos.xyz0, r_rpos.xyz0) # local pos - remote pos
# dist_tmp
#cal.mul(r_dist_vec, r_diff.xxxx, r_diff.xxxx)
#cal.mad(r_dist_vec, r_diff.yyyy, r_diff.yyyy, r_dist_vec)
#cal.mad(r_dist_vec, r_diff.zzzz, r_diff.zzzz, r_dist_vec)
cal.dp3(r_dist_vec, r_diff, r_diff, IEEE = False)
# distance
# TODO - skip rest of force computation if distance is 0
cal.sqrt_vec(r_dist, r_dist_vec)
# force G * ((m[i] * m[j]) / dist_tmp)
cal.mul(r_force_tmp, r_lpos.wwww, r_rpos.wwww, IEEE = False)
cal.div(r_force_tmp, r_force_tmp, r_dist_vec, ZEROOP = cal.zeroop.zero)
cal.mul(r_force_tmp, r_force_tmp, r_G, IEEE = False)
# f_xyz
# TODO - whats going on, is this right?
cal.div(r_force_vec, r_diff.xyz0, r_dist.xyz1, ZEROOP = cal.zeroop.zero)
cal.mul(r_force_vec, r_force_vec.xyz0, r_force_tmp.xyz0, IEEE = False)
cal.sub(r_force, r_force.xyz0, r_force_vec.xyz0)
#cal.add(r_cy, r_cy, r_count('1111'))
#cal.add(r_count, r_count, r_count('0100'))
#cal.ifc(cal.relop.ge, r_count.y, r_bodies.y)
## TODO - can I merge these two?
#cal.mov(r_count('_y__'), r_count('x0zw'))
#cal.add(r_count, r_count, r_count('1000'))
#cal.endif()
# Increment loop counter, end loop
cal.add(r_count, r_count, r_count('0100'))
cal.endloop()
cal.add(r_count, r_count, r_count('1000'))
#cal.add(r_cx, r_cx, r_cx('1111'))
cal.endloop()
# Acceleration
cal.div(r_force, r_force.xyz0, r_lpos.wwww, ZEROOP = cal.zeroop.zero)
# Velocity
cal.sample(1, 1, r_vel, reg.v0.xy) # Load velocity
cal.mad(r_vel, r_force, r_dt, r_vel, IEEE = False)
cal.mov(reg.o1, r_vel)
# Position
cal.mad(reg.o0, r_vel.xyz0, r_dt.xyz0, r_lpos.xyzw, IEEE = False)
#cal.mov(reg.g[0], r_vel)
return code
def ParMD5Transform(parcontext, parblock, blocki):
num = parcontext.number
temp_block = extarray.extarray('I', 16 * num)
ParDecode(num, temp_block, parblock, blocki, 64)
proc = env.Processor(0)
N = int(math.sqrt(num / 4))
#print "N = ", N
def address_4_1d(i, pitch=64):
x = i % N
y = i // 64 * 4
#return x*4 + y*pitch*4*4
return i
def address_4_2d(x, y, pitch=64):
return x * 4 + y * pitch * 4
input_statea = proc.alloc_remote('I', 4, N, N)
input_stateb = proc.alloc_remote('I', 4, N, N)
input_statec = proc.alloc_remote('I', 4, N, N)
input_stated = proc.alloc_remote('I', 4, N, N)
input_block = [proc.alloc_remote('I', 4, N, N) for i in range(16)]
outputa = proc.alloc_remote('I', 4, N, N)
outputb = proc.alloc_remote('I', 4, N, N)
outputc = proc.alloc_remote('I', 4, N, N)
outputd = proc.alloc_remote('I', 4, N, N)
for j in range(N):
for i in range(N):
for k in range(4):
input_statea[address_4_2d(i, j) +
k] = parcontext.statea[k + (i + j * N) * 4]
input_stateb[address_4_2d(i, j) +
k] = parcontext.stateb[k + (i + j * N) * 4]
input_statec[address_4_2d(i, j) +
k] = parcontext.statec[k + (i + j * N) * 4]
input_stated[address_4_2d(i, j) +
k] = parcontext.stated[k + (i + j * N) * 4]
for k in range(N):
for j in range(N):
for l in range(4):
for i in range(16):
input_block[i][address_4_2d(j, k) +
l] = temp_block[i + (j + k * N) * 4 * 16 +
l * 16]
global xcode
if xcode == None:
xcode = env.InstructionStream()
cal.set_active_code(xcode)
S11 = xcode.acquire_register((7, 7, 7, 7))
S12 = xcode.acquire_register((12, 12, 12, 12))
S13 = xcode.acquire_register((17, 17, 17, 17))
S14 = xcode.acquire_register((22, 22, 22, 22))
S21 = xcode.acquire_register((5, 5, 5, 5))
S22 = xcode.acquire_register((9, 9, 9, 9))
S23 = xcode.acquire_register((14, 14, 14, 14))
S24 = xcode.acquire_register((20, 20, 20, 20))
S31 = xcode.acquire_register((4, 4, 4, 4))
S32 = xcode.acquire_register((11, 11, 11, 11))
S33 = xcode.acquire_register((16, 16, 16, 16))
S34 = xcode.acquire_register((23, 23, 23, 23))
S41 = xcode.acquire_register((6, 6, 6, 6))
S42 = xcode.acquire_register((10, 10, 10, 10))
S43 = xcode.acquire_register((15, 15, 15, 15))
S44 = xcode.acquire_register((21, 21, 21, 21))
a = xcode.acquire_register()
b = xcode.acquire_register()
c = xcode.acquire_register()
d = xcode.acquire_register()
x = [xcode.acquire_register() for i in range(16)]
r = xcode.acquire_register()
cal.dcl_resource(0, cal.pixtex_type.twod, cal.fmt.uint,
UNNORM=True) # statea
cal.dcl_resource(1, cal.pixtex_type.twod, cal.fmt.uint,
UNNORM=True) # stateb
cal.dcl_resource(2, cal.pixtex_type.twod, cal.fmt.uint,
UNNORM=True) # statec
cal.dcl_resource(3, cal.pixtex_type.twod, cal.fmt.uint,
UNNORM=True) # stated
for i in range(16):
cal.dcl_resource(i + 4,
cal.pixtex_type.twod,
cal.fmt.uint,
UNNORM=True)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_output(reg.o1, USAGE=cal.usage.generic)
cal.dcl_output(reg.o2, USAGE=cal.usage.generic)
cal.dcl_output(reg.o3, USAGE=cal.usage.generic)
cal.sample(0, 0, a, reg.v0.xy)
cal.sample(1, 0, b, reg.v0.xy)
cal.sample(2, 0, c, reg.v0.xy)
cal.sample(3, 0, d, reg.v0.xy)
for i in range(16):
cal.sample(i + 4, 0, x[i], reg.v0.xy)
# Round 1
FF(a, b, c, d, x[0], S11, 0xd76aa478)
# 1
FF(d, a, b, c, x[1], S12, 0xe8c7b756)
# 2
FF(c, d, a, b, x[2], S13, 0x242070db)
# 3
FF(b, c, d, a, x[3], S14, 0xc1bdceee)
# 4
FF(a, b, c, d, x[4], S11, 0xf57c0faf)
# 5
FF(d, a, b, c, x[5], S12, 0x4787c62a)
# 6
FF(c, d, a, b, x[6], S13, 0xa8304613)
# 7
FF(b, c, d, a, x[7], S14, 0xfd469501)
# 8
FF(a, b, c, d, x[8], S11, 0x698098d8)
# 9
FF(d, a, b, c, x[9], S12, 0x8b44f7af)
# 10
FF(c, d, a, b, x[10], S13, 0xffff5bb1)
# 11
FF(b, c, d, a, x[11], S14, 0x895cd7be)
# 12
FF(a, b, c, d, x[12], S11, 0x6b901122)
# 13
FF(d, a, b, c, x[13], S12, 0xfd987193)
# 14
FF(c, d, a, b, x[14], S13, 0xa679438e)
# 15
FF(b, c, d, a, x[15], S14, 0x49b40821)
# 16
# Round 2
GG(a, b, c, d, x[1], S21, 0xf61e2562)
# 17
GG(d, a, b, c, x[6], S22, 0xc040b340)
# 18
GG(c, d, a, b, x[11], S23, 0x265e5a51)
# 19
GG(b, c, d, a, x[0], S24, 0xe9b6c7aa)
# 20
GG(a, b, c, d, x[5], S21, 0xd62f105d)
# 21
GG(d, a, b, c, x[10], S22, 0x2441453)
# 22
GG(c, d, a, b, x[15], S23, 0xd8a1e681)
# 23
GG(b, c, d, a, x[4], S24, 0xe7d3fbc8)
# 24
GG(a, b, c, d, x[9], S21, 0x21e1cde6)
# 25
GG(d, a, b, c, x[14], S22, 0xc33707d6)
# 26
GG(c, d, a, b, x[3], S23, 0xf4d50d87)
# 27
GG(b, c, d, a, x[8], S24, 0x455a14ed)
# 28
GG(a, b, c, d, x[13], S21, 0xa9e3e905)
# 29
GG(d, a, b, c, x[2], S22, 0xfcefa3f8)
# 30
GG(c, d, a, b, x[7], S23, 0x676f02d9)
# 31
GG(b, c, d, a, x[12], S24, 0x8d2a4c8a)
# 32
# Round 3
HH(a, b, c, d, x[5], S31, 0xfffa3942)
# 33
HH(d, a, b, c, x[8], S32, 0x8771f681)
# 34
HH(c, d, a, b, x[11], S33, 0x6d9d6122)
# 35
HH(b, c, d, a, x[14], S34, 0xfde5380c)
# 36
HH(a, b, c, d, x[1], S31, 0xa4beea44)
# 37
HH(d, a, b, c, x[4], S32, 0x4bdecfa9)
# 38
HH(c, d, a, b, x[7], S33, 0xf6bb4b60)
# 39
HH(b, c, d, a, x[10], S34, 0xbebfbc70)
# 40
HH(a, b, c, d, x[13], S31, 0x289b7ec6)
# 41
HH(d, a, b, c, x[0], S32, 0xeaa127fa)
# 42
HH(c, d, a, b, x[3], S33, 0xd4ef3085)
# 43
HH(b, c, d, a, x[6], S34, 0x4881d05)
# 44
HH(a, b, c, d, x[9], S31, 0xd9d4d039)
# 45
HH(d, a, b, c, x[12], S32, 0xe6db99e5)
# 46
HH(c, d, a, b, x[15], S33, 0x1fa27cf8)
# 47
HH(b, c, d, a, x[2], S34, 0xc4ac5665)
# 48
# Round 4
II(a, b, c, d, x[0], S41, 0xf4292244)
# 49
II(d, a, b, c, x[7], S42, 0x432aff97)
# 50
II(c, d, a, b, x[14], S43, 0xab9423a7)
# 51
II(b, c, d, a, x[5], S44, 0xfc93a039)
# 52
II(a, b, c, d, x[12], S41, 0x655b59c3)
# 53
II(d, a, b, c, x[3], S42, 0x8f0ccc92)
# 54
II(c, d, a, b, x[10], S43, 0xffeff47d)
# 55
II(b, c, d, a, x[1], S44, 0x85845dd1)
# 56
II(a, b, c, d, x[8], S41, 0x6fa87e4f)
# 57
II(d, a, b, c, x[15], S42, 0xfe2ce6e0)
# 58
II(c, d, a, b, x[6], S43, 0xa3014314)
# 59
II(b, c, d, a, x[13], S44, 0x4e0811a1)
# 60
II(a, b, c, d, x[4], S41, 0xf7537e82)
# 61
II(d, a, b, c, x[11], S42, 0xbd3af235)
# 62
II(c, d, a, b, x[2], S43, 0x2ad7d2bb)
# 63
II(b, c, d, a, x[9], S44, 0xeb86d391)
# 64
cal.mov('o0', a)
cal.mov('o1', b)
cal.mov('o2', c)
cal.mov('o3', d)
xcode.release_register(a)
xcode.release_register(b)
xcode.release_register(c)
xcode.release_register(d)
for xi in x:
xcode.release_register(xi)
xcode.set_remote_binding('i0', input_statea)
xcode.set_remote_binding('i1', input_stateb)
xcode.set_remote_binding('i2', input_statec)
xcode.set_remote_binding('i3', input_stated)
for i in range(16): #range(len(input_block)):
xcode.set_remote_binding('i' + str(i + 4), input_block[i])
xcode.set_remote_binding('o0', outputa)
xcode.set_remote_binding('o1', outputb)
xcode.set_remote_binding('o2', outputc)
xcode.set_remote_binding('o3', outputd)
domain = (0, 0, N, N)
global TIME
start_time = time.time()
proc.execute(xcode, domain)
end_time = time.time()
TIME += (end_time - start_time)
for j in range(N):
for i in range(N):
for k in range(4):
parcontext.statea[k + (i + j * N) *
4] += outputa[address_4_2d(i, j) + k]
parcontext.stateb[k + (i + j * N) *
4] += outputb[address_4_2d(i, j) + k]
parcontext.statec[k + (i + j * N) *
4] += outputc[address_4_2d(i, j) + k]
parcontext.stated[k + (i + j * N) *
4] += outputd[address_4_2d(i, j) + k]
proc.free_remote(input_statea)
proc.free_remote(input_stateb)
proc.free_remote(input_statec)
proc.free_remote(input_stated)
for block in input_block:
proc.free_remote(block)
proc.free_remote(outputa)
proc.free_remote(outputb)
proc.free_remote(outputc)
proc.free_remote(outputd)
code = prgm.get_stream()
inp = proc.alloc_remote('f', 4, 64)
out = proc.alloc_remote('f', 4, 64)
out.clear()
for i in xrange(0, 64):
inp[i] = float(i + 1)
cal.set_active_code(code)
cal.dcl_input(reg.v0.x, USAGE=cal.usage.pos)
cal.dcl_resource(0, cal.pixtex_type.oned, cal.fmt.float, UNNORM=True)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.sample(0, 0, reg.o0, reg.v0.x)
prgm.set_binding(reg.i0, inp)
prgm.set_binding(reg.o0, out)
prgm.add(code)
prgm.print_code()
proc.execute(prgm)
print "inp", inp
print "out", out
import corepy.lib.printer as printer
printer.PrintInstructionStream(code, printer.CAL_Asm())
def ParMD5Transform(parcontext, parblock, blocki):
num = parcontext.number
temp_block = extarray.extarray('I', 16*num)
ParDecode(num, temp_block, parblock, blocki, 64)
proc = env.Processor(0)
N = int(math.sqrt(num/4))
#print "N = ", N
def address_4_1d(i, pitch=64):
x = i % N
y = i // 64*4
#return x*4 + y*pitch*4*4
return i
def address_4_2d(x, y, pitch=64):
return x*4 + y*pitch*4
input_statea = proc.alloc_remote('I', 4, N, N)
input_stateb = proc.alloc_remote('I', 4, N, N)
input_statec = proc.alloc_remote('I', 4, N, N)
input_stated = proc.alloc_remote('I', 4, N, N)
input_block = proc.alloc_remote('I', 4, N*4*4, N)
outputa = proc.alloc_remote('I', 4, N, N)
outputb = proc.alloc_remote('I', 4, N, N)
outputc = proc.alloc_remote('I', 4, N, N)
outputd = proc.alloc_remote('I', 4, N, N)
for j in range(N):
for i in range(N):
for k in range(4):
input_statea[address_4_2d(i, j) + k] = parcontext.statea[k + (i + j*N)*4]
input_stateb[address_4_2d(i, j) + k] = parcontext.stateb[k + (i + j*N)*4]
input_statec[address_4_2d(i, j) + k] = parcontext.statec[k + (i + j*N)*4]
input_stated[address_4_2d(i, j) + k] = parcontext.stated[k + (i + j*N)*4]
for k in range(N):
for j in range(0, N*4):
for i in range(16):
input_block[address_4_2d(j*4, k) + i] = temp_block[i + (j + k*N)*16]
#print address_4_2d(j*4, k) + i, i + (j + k*N)*16
#print "N = ", N
#for i in range(num):
# print i, map(hex, [input_block[i*16+j] for j in range(16)])
global xcode
if xcode == None:
xcode = env.InstructionStream()
cal.set_active_code(xcode)
S11 = xcode.acquire_register((7, 7, 7, 7))
S12 = xcode.acquire_register((12, 12, 12, 12))
S13 = xcode.acquire_register((17, 17, 17, 17))
S14 = xcode.acquire_register((22, 22, 22, 22))
S21 = xcode.acquire_register((5, 5, 5, 5))
S22 = xcode.acquire_register((9, 9, 9, 9))
S23 = xcode.acquire_register((14, 14, 14, 14))
S24 = xcode.acquire_register((20, 20, 20, 20))
S31 = xcode.acquire_register((4, 4, 4, 4))
S32 = xcode.acquire_register((11, 11, 11, 11))
S33 = xcode.acquire_register((16, 16, 16, 16))
S34 = xcode.acquire_register((23, 23, 23, 23))
S41 = xcode.acquire_register((6, 6, 6, 6))
S42 = xcode.acquire_register((10, 10, 10, 10))
S43 = xcode.acquire_register((15, 15, 15, 15))
S44 = xcode.acquire_register((21, 21, 21, 21))
a = xcode.acquire_register()
b = xcode.acquire_register()
c = xcode.acquire_register()
d = xcode.acquire_register()
# TODO: Ensure these are all contiguous - necessary for the transposes
x = [xcode.acquire_register() for i in range(16)]
r = xcode.acquire_register()
cal.dcl_resource(0, cal.pixtex_type.twod, cal.fmt.uint, UNNORM=True) # statea
cal.dcl_resource(1, cal.pixtex_type.twod, cal.fmt.uint, UNNORM=True) # stateb
cal.dcl_resource(2, cal.pixtex_type.twod, cal.fmt.uint, UNNORM=True) # statec
cal.dcl_resource(3, cal.pixtex_type.twod, cal.fmt.uint, UNNORM=True) # stated
cal.dcl_resource(4, cal.pixtex_type.twod, cal.fmt.uint, UNNORM=True)
cal.dcl_output(reg.o0, USAGE=cal.usage.generic)
cal.dcl_output(reg.o1, USAGE=cal.usage.generic)
cal.dcl_output(reg.o2, USAGE=cal.usage.generic)
cal.dcl_output(reg.o3, USAGE=cal.usage.generic)
cal.sample(0, 0, a, reg.v0.xy)
cal.sample(1, 0, b, reg.v0.xy)
cal.sample(2, 0, c, reg.v0.xy)
cal.sample(3, 0, d, reg.v0.xy)
cal.dclpi(('0', '0', '-', '-'), reg.vWinCoord0, CENTERED=True)
buffer_index = xcode.acquire_register()
temp = xcode.acquire_register()
consts = xcode.acquire_register((1.0, 2.0, 3.0, 4.0))
one = consts.x
two = consts.y
three = consts.z
four = consts.w
sixteen = xcode.acquire_register((16.0,)*4)
cal.mov(buffer_index, reg.vWinCoord0.xy)
cal.mul(buffer_index.x, buffer_index, sixteen)
for i in range(4):
cal.mov(temp.xy, buffer_index.xy)
cal.sample(4, 0, x[i*4], buffer_index.xy)
cal.add(buffer_index.x, buffer_index, four)
cal.sample(4, 0, x[i*4+1], buffer_index.xy)
cal.add(buffer_index.x, buffer_index, four)
cal.sample(4, 0, x[i*4+2], buffer_index.xy)
cal.add(buffer_index.x, buffer_index, four)
cal.sample(4, 0, x[i*4+3], buffer_index.xy)
cal.mov(buffer_index.xy, temp.xy)
cal.add(buffer_index.x, buffer_index, one)
cal.transpose(x[0], x[0])
cal.transpose(x[4], x[4])
cal.transpose(x[8], x[8])
cal.transpose(x[12], x[12])
# Round 1
FF (a, b, c, d, x[ 0], S11, 0xd76aa478); # 1
FF (d, a, b, c, x[ 1], S12, 0xe8c7b756); # 2
FF (c, d, a, b, x[ 2], S13, 0x242070db); # 3
FF (b, c, d, a, x[ 3], S14, 0xc1bdceee); # 4
FF (a, b, c, d, x[ 4], S11, 0xf57c0faf); # 5
FF (d, a, b, c, x[ 5], S12, 0x4787c62a); # 6
FF (c, d, a, b, x[ 6], S13, 0xa8304613); # 7
FF (b, c, d, a, x[ 7], S14, 0xfd469501); # 8
FF (a, b, c, d, x[ 8], S11, 0x698098d8); # 9
FF (d, a, b, c, x[ 9], S12, 0x8b44f7af); # 10
FF (c, d, a, b, x[10], S13, 0xffff5bb1); # 11
FF (b, c, d, a, x[11], S14, 0x895cd7be); # 12
FF (a, b, c, d, x[12], S11, 0x6b901122); # 13
FF (d, a, b, c, x[13], S12, 0xfd987193); # 14
FF (c, d, a, b, x[14], S13, 0xa679438e); # 15
FF (b, c, d, a, x[15], S14, 0x49b40821); # 16
# Round 2
GG (a, b, c, d, x[ 1], S21, 0xf61e2562); # 17
GG (d, a, b, c, x[ 6], S22, 0xc040b340); # 18
GG (c, d, a, b, x[11], S23, 0x265e5a51); # 19
GG (b, c, d, a, x[ 0], S24, 0xe9b6c7aa); # 20
GG (a, b, c, d, x[ 5], S21, 0xd62f105d); # 21
GG (d, a, b, c, x[10], S22, 0x2441453); # 22
GG (c, d, a, b, x[15], S23, 0xd8a1e681); # 23
GG (b, c, d, a, x[ 4], S24, 0xe7d3fbc8); # 24
GG (a, b, c, d, x[ 9], S21, 0x21e1cde6); # 25
GG (d, a, b, c, x[14], S22, 0xc33707d6); # 26
GG (c, d, a, b, x[ 3], S23, 0xf4d50d87); # 27
GG (b, c, d, a, x[ 8], S24, 0x455a14ed); # 28
GG (a, b, c, d, x[13], S21, 0xa9e3e905); # 29
GG (d, a, b, c, x[ 2], S22, 0xfcefa3f8); # 30
GG (c, d, a, b, x[ 7], S23, 0x676f02d9); # 31
GG (b, c, d, a, x[12], S24, 0x8d2a4c8a); # 32
# Round 3
HH (a, b, c, d, x[ 5], S31, 0xfffa3942); # 33
HH (d, a, b, c, x[ 8], S32, 0x8771f681); # 34
HH (c, d, a, b, x[11], S33, 0x6d9d6122); # 35
HH (b, c, d, a, x[14], S34, 0xfde5380c); # 36
HH (a, b, c, d, x[ 1], S31, 0xa4beea44); # 37
HH (d, a, b, c, x[ 4], S32, 0x4bdecfa9); # 38
HH (c, d, a, b, x[ 7], S33, 0xf6bb4b60); # 39
HH (b, c, d, a, x[10], S34, 0xbebfbc70); # 40
HH (a, b, c, d, x[13], S31, 0x289b7ec6); # 41
HH (d, a, b, c, x[ 0], S32, 0xeaa127fa); # 42
HH (c, d, a, b, x[ 3], S33, 0xd4ef3085); # 43
HH (b, c, d, a, x[ 6], S34, 0x4881d05); # 44
HH (a, b, c, d, x[ 9], S31, 0xd9d4d039); # 45
HH (d, a, b, c, x[12], S32, 0xe6db99e5); # 46
HH (c, d, a, b, x[15], S33, 0x1fa27cf8); # 47
HH (b, c, d, a, x[ 2], S34, 0xc4ac5665); # 48
# Round 4
II (a, b, c, d, x[ 0], S41, 0xf4292244); # 49
II (d, a, b, c, x[ 7], S42, 0x432aff97); # 50
II (c, d, a, b, x[14], S43, 0xab9423a7); # 51
II (b, c, d, a, x[ 5], S44, 0xfc93a039); # 52
II (a, b, c, d, x[12], S41, 0x655b59c3); # 53
II (d, a, b, c, x[ 3], S42, 0x8f0ccc92); # 54
II (c, d, a, b, x[10], S43, 0xffeff47d); # 55
II (b, c, d, a, x[ 1], S44, 0x85845dd1); # 56
II (a, b, c, d, x[ 8], S41, 0x6fa87e4f); # 57
II (d, a, b, c, x[15], S42, 0xfe2ce6e0); # 58
II (c, d, a, b, x[ 6], S43, 0xa3014314); # 59
II (b, c, d, a, x[13], S44, 0x4e0811a1); # 60
II (a, b, c, d, x[ 4], S41, 0xf7537e82); # 61
II (d, a, b, c, x[11], S42, 0xbd3af235); # 62
II (c, d, a, b, x[ 2], S43, 0x2ad7d2bb); # 63
II (b, c, d, a, x[ 9], S44, 0xeb86d391); # 64
#cal.mov(buffer_index, reg.vWinCoord0.xy)
#cal.mul(buffer_index.x, buffer_index, sixteen)
#cal.add(buffer_index, buffer_index, reg.v0)
#cal.mov('o0', buffer_index)
cal.mov('o0', a)
cal.mov('o1', b)
cal.mov('o2', c)
cal.mov('o3', d)
xcode.release_register(a)
xcode.release_register(b)
xcode.release_register(c)
xcode.release_register(d)
for xi in x:
xcode.release_register(xi)
#xcode.cache_code()
#print xcode.render_string
xcode.set_remote_binding('i0', input_statea)
xcode.set_remote_binding('i1', input_stateb)
xcode.set_remote_binding('i2', input_statec)
xcode.set_remote_binding('i3', input_stated)
xcode.set_remote_binding('i4', input_block)
xcode.set_remote_binding('o0', outputa)
xcode.set_remote_binding('o1', outputb)
xcode.set_remote_binding('o2', outputc)
xcode.set_remote_binding('o3', outputd)
domain = (0, 0, N, N)
global TIME
start_time = time.time()
proc.execute(xcode, domain)
end_time = time.time()
TIME += (end_time - start_time)
#print map(hex, [outputa[i] for i in range(4)])
#print map(hex, outputa)
#print outputa
for j in range(N):
for i in range(N):
for k in range(4):
parcontext.statea[k + (i + j*N)*4] += outputa[address_4_2d(i, j) + k]
parcontext.stateb[k + (i + j*N)*4] += outputb[address_4_2d(i, j) + k]
parcontext.statec[k + (i + j*N)*4] += outputc[address_4_2d(i, j) + k]
parcontext.stated[k + (i + j*N)*4] += outputd[address_4_2d(i, j) + k]
proc.free_remote(input_statea)
proc.free_remote(input_stateb)
proc.free_remote(input_statec)
proc.free_remote(input_stated)
proc.free_remote(input_block)
proc.free_remote(outputa)
proc.free_remote(outputb)
proc.free_remote(outputc)
proc.free_remote(outputd)