def test_reg_counter_reduction():
knl = lp.make_kernel(
"{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
],
name="matmul_serial", assumptions="n,m,l >= 1")
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32, b=np.float32))
regs = estimate_regs_per_thread(knl)
assert regs == 6
def test_reg_counter_logic():
knl = lp.make_kernel(
"{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
[
"""
e[i,k] = if(not(k<l-2) and k>6 or k/2==l, g[i,k]*2, g[i,k]+h[i,k]/2)
"""
],
name="logic", assumptions="n,m,l >= 1")
knl = lp.add_and_infer_dtypes(knl, dict(g=np.float32, h=np.float64))
regs = estimate_regs_per_thread(knl)
assert regs == 6
def test_reg_counter_specialops():
knl = lp.make_kernel(
"{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
[
"""
c[i, j, k] = (2*a[i,j,k])%(2+b[i,j,k]/3.0)
e[i, k] = (1+g[i,k])**(1+h[i,k+1])
"""
],
name="specialops", assumptions="n,m,l >= 1")
knl = lp.add_and_infer_dtypes(knl,
dict(a=np.float32, b=np.float32, g=np.float64, h=np.float64))
regs = estimate_regs_per_thread(knl)
assert regs == 6
def test_reg_counter_basic():
knl = lp.make_kernel(
"[n,m,l] -> {[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
[
"""
c[i, j, k] = a[i,j,k]*b[i,j,k]/3.0+a[i,j,k]
e[i, k+1] = g[i,k]*h[i,k+1]
"""
],
name="basic", assumptions="n,m,l >= 1")
knl = lp.add_and_infer_dtypes(knl,
dict(a=np.float32, b=np.float32, g=np.float64, h=np.float64))
regs = estimate_regs_per_thread(knl)
assert regs == 6
def test_reg_counter_bitwise():
knl = lp.make_kernel(
"{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
[
"""
c[i, j, k] = (a[i,j,k] | 1) + (b[i,j,k] & 1)
e[i, k] = (g[i,k] ^ k)*(~h[i,k+1]) + (g[i, k] << (h[i,k] >> k))
"""
],
name="bitwise", assumptions="n,m,l >= 1")
knl = lp.add_and_infer_dtypes(
knl, dict(
a=np.int32, b=np.int32,
g=np.int64, h=np.int64))
regs = estimate_regs_per_thread(knl)
assert regs == 6
def run_fd_trials(ctx, queue, nvals, configs_t,
Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all, train_test_config):
A = []
HK_predict = []
actual = []
dtype = np.float32
for n in nvals:
u_mat_dev = cl.clrandom.rand(queue, (n+2, n+2), dtype=dtype)
knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n}",
"result[i,j] = u[i, j]**2 + -1 + (-4)*u[i + 1, j + 1] \
+ u[i + 1 + 1, j + 1] + u[i + 1 + -1, j + 1] \
+ u[i + 1, j + 1 + 1] + u[i + 1, j + 1 + -1]",
name="finite_diff")
knl = lp.add_and_infer_dtypes(knl, {"u": dtype})
ref_knl = knl
for BSIZEx, BSIZEy in configs_t:
knl = ref_knl
knl = lp.split_iname(knl,
"i", BSIZEx, outer_tag="g.1", inner_tag="l.1")
knl = lp.split_iname(knl,
"j", BSIZEy, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "u",
["i_inner", "j_inner"],
fetch_bounding_box=True)
#check = lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=n),
# print_code=True)
#print "Correctness check: \n", check
# use ptx src to determine resource usage
#ptx_dump(ctx, knl, n, BSIZEx, BSIZEy)
params = {'n': n}
barrier_poly = get_barrier_poly(knl)
barrier_ct = barrier_poly.eval_with_dict(params)
op_map = get_op_poly(knl)
op_map2 = get_op_poly2(knl)
flops, iops = get_32b_ops(op_map, params)
amd_op32 = get_32b_amd_ops(op_map2, params)
other_op32 = get_32b_ops_all(op_map2, params) - sum(amd_op32)
if flops + iops != sum(amd_op32) + other_op32: #TODO remove after debug
print("<debug> PROBLEM!, ops don't add up: ",
flops, iops, sum(amd_op32), other_op32)
sub_map = get_DRAM_access_poly(knl) # noqa
f32coal_l, f32coal_s, f32uncoal_l, f32uncoal_s = get_DRAM_f32_accesses(
sub_map, params)
f32coal = f32coal_l + f32coal_s
f32uncoal = f32uncoal_l + f32uncoal_s
# execute
#print "="*40+"TIMING RESULTS"
print("running kernel...")
#knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
trial_times = []
for i in range(averaging_trials+warmup_trials):
evt, (out,) = knl(queue, u=u_mat_dev)
evt.wait()
trial_times.append((evt.profile.END - evt.profile.START)*1e-9)
avg_time = np.average(trial_times[warmup_trials:])
gstats = GPUStats('TeslaC2070')
#'''
if n % BSIZEx == 0 and n % BSIZEy == 0:
reg32_per_thread = 14
else:
reg32_per_thread = 15 # 16 for k20
#'''
#reg32_per_thread = 1 #estimate_regs_per_thread(knl)
#print(reg32_per_thread, estimate_regs_per_thread(knl))
reg32_per_thread = estimate_regs_per_thread(knl)
shared_mem_per_block = 4*(BSIZEx+2)*(BSIZEy+2)
total_blocks = math.ceil(n/BSIZEx)*math.ceil(n/BSIZEy)
total_threads = total_blocks*BSIZEx*BSIZEy # TODO unused
kstats = KernelStats(flops/(n*n), f32uncoal/(n*n), f32coal/(n*n),
barrier_ct, reg32_per_thread, shared_mem_per_block)
tconfig = ThreadConfig(BSIZEx*BSIZEy, total_blocks)
model = PerfModel(gstats, kstats, tconfig,
np.dtype(dtype))
cycles = model.compute_total_cycles()
actual.append(avg_time)
#for time in trial_times: #!!!!!
# actual.append(time)
HK_predict.append(cycles/(gstats.sm_clock_freq*10**9))
#for time in trial_times:
'''
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
ops = copy.deepcopy(amd_op32)
ops.append(other_op32)
update_LS_matrix2(A, ops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
update_lstsq_mats(Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all,
A, actual, HK_predict, train_test_config)
def run_empt_trials(ctx, queue, nvals, configs_t,
Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all, train_test_config):
A = []
HK_predict = []
actual = []
dtype = np.float32
for n in nvals:
knl = lp.make_kernel(
"{[i,j]: 0<=i,j<%d}" % n,
[
""
],
name="empty")
for BSIZEx, BSIZEy in configs_t:
#check = lp.auto_test_vs_ref(ref_knl, ctx, knl, print_code=True)
#print "Correctness check: \n", check
# use ptx src to determine resource usage
#ptx_dump(ctx, knl, n, BSIZEx, BSIZEy)
params = {'n': n}
barrier_poly = get_barrier_poly(knl)
barrier_ct = barrier_poly.eval_with_dict(params)
#op_map = get_op_poly(knl) #TODO figure out error
op_map = {}
op_map2 = {}
flops, iops = get_32b_ops(op_map, params)
amd_op32 = get_32b_amd_ops(op_map2, params)
other_op32 = get_32b_ops_all(op_map2, params) - sum(amd_op32)
if flops + iops != sum(amd_op32) + other_op32: #TODO remove after debug
print("<debug> PROBLEM!, ops don't add up: ",
flops, iops, sum(amd_op32), other_op32)
#sub_map = get_DRAM_access_poly(knl) #TODO figure out error
sub_map = {}
f32coal_l, f32coal_s, f32uncoal_l, f32uncoal_s = get_DRAM_f32_accesses(
sub_map, params)
# execute
#print "="*40+"TIMING RESULTS"
print("running kernel...")
#knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
trial_times = []
for i in range(averaging_trials+warmup_trials):
evt, out = knl(queue)
evt.wait()
trial_times.append((evt.profile.END - evt.profile.START)*1e-9)
avg_time = np.average(trial_times[warmup_trials:])
gstats = GPUStats('TeslaC2070')
reg32_per_thread = 2
#print(reg32_per_thread, estimate_regs_per_thread(knl))
reg32_per_thread = estimate_regs_per_thread(knl)
#reg32_per_thread = 1 #estimate_regs_per_thread(knl)
shared_mem_per_block = 0
total_blocks = math.ceil(n/BSIZEx)*math.ceil(n/BSIZEy)
total_threads = total_blocks*BSIZEx*BSIZEy # TODO unused
# TODO actually increase threads/blocks but expect 0 result
kstats = KernelStats(0, 0, 0, barrier_ct, reg32_per_thread,
shared_mem_per_block)
tconfig = ThreadConfig(BSIZEx*BSIZEy, total_blocks)
model = PerfModel(gstats, kstats, tconfig,
np.dtype(dtype))
cycles = model.compute_total_cycles()
actual.append(avg_time)
#for time in trial_times: #!!!!!
# actual.append(time)
HK_predict.append(cycles/(gstats.sm_clock_freq*10**9))
#for time in trial_times:
'''
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
ops = copy.deepcopy(amd_op32)
ops.append(other_op32)
update_LS_matrix2(A, ops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
update_lstsq_mats(Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all,
A, actual, HK_predict, train_test_config)
def run_conv_trials(ctx, queue, nvals, configs_t,
Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all, train_test_config):
A = []
HK_predict = []
actual = []
dtype = np.float32
ncolors = 3
for n in nvals:
knl = lp.make_kernel(
"{ [iimg, ifeat, icolor, im_x, im_y, f_x, f_y]: \
-f_w <= f_x,f_y <= f_w \
and 0 <= im_x < im_w and 0 <= im_y < im_h \
and 0<=iimg<=nimgs and 0<=ifeat<nfeats and 0<=icolor<ncolors \
}",
"""
out[iimg, ifeat, im_x, im_y] = sum((f_x, f_y, icolor), \
img[iimg, f_w+im_x-f_x, f_w+im_y-f_y, icolor] \
* f[ifeat, f_w+f_x, f_w+f_y, icolor])
""",
[
lp.GlobalArg("f", dtype, shape=lp.auto),
lp.GlobalArg("img", dtype, shape=lp.auto),
lp.GlobalArg("out", dtype, shape=lp.auto),
"..."
],
assumptions="f_w>=1 and im_w, im_h >= 2*f_w+1 and nfeats>=1 and nimgs>=0",
flags="annotate_inames",
defines=dict(ncolors=ncolors),
name="conv")
f_w = 3
knl = lp.fix_parameters(knl, f_w=f_w)
ref_knl = knl
for BSIZEx, BSIZEy in configs_t:
knl = ref_knl
im_w = n
im_h = n
nfeats = 3
nimgs = 3
f_dev = cl.clrandom.rand(queue, (nfeats, 2*f_w+1, 2*f_w+1, ncolors),
dtype=dtype)
img_dev = cl.clrandom.rand(queue, (nimgs+1, n+2*f_w, n+2*f_w, ncolors),
dtype=dtype)
knl = lp.split_iname(knl, "im_x", BSIZEx,
outer_tag="g.0", inner_tag="l.0")
knl = lp.split_iname(knl, "im_y", BSIZEy,
outer_tag="g.1", inner_tag="l.1")
knl = lp.tag_inames(knl, dict(ifeat="g.2"))
knl = lp.add_prefetch(knl, "f[ifeat,:,:,:]")
knl = lp.add_prefetch(knl, "img", "im_x_inner, im_y_inner, f_x, f_y")
params = dict(im_w=im_w, im_h=im_h, f_w=f_w, nfeats=nfeats, nimgs=nimgs)
#check = lp.auto_test_vs_ref(ref_knl, ctx, knl, print_code=True,
# parameters=params)
#print "Correctness check: \n", check
# use ptx src to determine resource usage
#ptx_dump(ctx, knl, n, BSIZEx, BSIZEy)
barrier_poly = get_barrier_poly(knl)
barrier_ct = barrier_poly.eval_with_dict(params)
op_map = get_op_poly(knl)
op_map2 = get_op_poly2(knl)
flops, iops = get_32b_ops(op_map, params)
amd_op32 = get_32b_amd_ops(op_map2, params)
other_op32 = get_32b_ops_all(op_map2, params) - sum(amd_op32)
if flops + iops != sum(amd_op32) + other_op32: #TODO remove after debug
print("<debug> PROBLEM!, ops don't add up: ",
flops, iops, sum(amd_op32), other_op32)
#TODO why do blk sizes that don't fit perfecty increase total flops/iops
sub_map = get_DRAM_access_poly(knl) # noqa
f32coal_l, f32coal_s, f32uncoal_l, f32uncoal_s = get_DRAM_f32_accesses(
sub_map, params)
f32coal = f32coal_l + f32coal_s
f32uncoal = f32uncoal_l + f32uncoal_s
# execute
print("running kernel...")
#knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
trial_times = []
for i in range(averaging_trials+warmup_trials):
evt, (out,) = knl(queue, f=f_dev, img=img_dev, im_w=im_w, im_h=im_h,
nfeats=nfeats, nimgs=nimgs)
evt.wait()
trial_times.append((evt.profile.END - evt.profile.START)*1e-9)
avg_time = np.average(trial_times[warmup_trials:])
gstats = GPUStats('TeslaC2070')
reg32_per_thread = 20 #20 for c2070, 33 for k20
#print(reg32_per_thread, estimate_regs_per_thread(knl))
reg32_per_thread = estimate_regs_per_thread(knl)
#reg32_per_thread = 1 #estimate_regs_per_thread(knl)
shared_mem_per_block = (ncolors * (f_w*2+1) * (f_w*2+1) +
(BSIZEx+f_w*2) * (BSIZEy+f_w*2)
) * np.dtype(dtype).itemsize
total_blocks = math.ceil(n/BSIZEx)*math.ceil(n/BSIZEy)
total_threads = total_blocks*BSIZEx*BSIZEy # TODO unused
kstats = KernelStats(flops/(n*n), f32uncoal/(n*n), f32coal/(n*n),
barrier_ct, reg32_per_thread, shared_mem_per_block)
tconfig = ThreadConfig(BSIZEx*BSIZEy, total_blocks)
model = PerfModel(gstats, kstats, tconfig, np.dtype(dtype))
cycles = model.compute_total_cycles()
actual.append(avg_time)
#for time in trial_times: #!!!!!
# actual.append(time)
HK_predict.append(cycles/(gstats.sm_clock_freq*10**9))
#for time in trial_times:
'''
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
ops = copy.deepcopy(amd_op32)
ops.append(other_op32)
update_LS_matrix2(A, ops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
#TODO try total_threads for n*n
update_lstsq_mats(Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all,
A, actual, HK_predict, train_test_config)
def run_tp_trials(ctx, queue, nvals, configs_t,
Atrain_all, Atest_all, ytrain_all, ytest_all, actual_times_all,
HK_predict_all, train_test_config, prefetch=True):
A = []
HK_predict = []
actual = []
dtype = np.float32
for n in nvals:
a_mat_dev = cl.clrandom.rand(queue, (n, n), dtype=dtype)
b_mat_dev = cl.clrandom.rand(queue, (n, n), dtype=dtype)
order = "C"
knl = lp.make_kernel(
"{[i,j]: 0<=i,j<%d}" % n,
[
"b[i, j] = a[j, i]"
], [
lp.GlobalArg("a", dtype, shape=(n, n), order=order),
lp.GlobalArg("b", dtype, shape=(n, n), order=order),
],
name="transpose")
ref_knl = knl
for BSIZEx, BSIZEy in configs_t:
knl = ref_knl
knl = lp.split_iname(knl, "i", BSIZEy, outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", BSIZEx, outer_tag="g.1", inner_tag="l.0")
if prefetch:
knl = lp.add_prefetch(knl, 'a', ["i_inner", "j_inner"])
#check = lp.auto_test_vs_ref(ref_knl, ctx, knl, print_code=True)
#print "Correctness check: \n", check
# use ptx src to determine resource usage
#ptx_dump(ctx, knl, n, BSIZEx, BSIZEy)
params = {'n': n}
barrier_poly = get_barrier_poly(knl)
barrier_ct = barrier_poly.eval_with_dict(params)
op_map = get_op_poly(knl)
op_map2 = get_op_poly2(knl)
flops, iops = get_32b_ops(op_map, params)
amd_op32 = get_32b_amd_ops(op_map2, params)
other_op32 = get_32b_ops_all(op_map2, params) - sum(amd_op32)
if flops + iops != sum(amd_op32) + other_op32: #TODO remove after debug
print("<debug> PROBLEM!, ops don't add up: ",
flops, iops, sum(amd_op32), other_op32)
sub_map = get_DRAM_access_poly(knl) # noqa
f32coal_l, f32coal_s, f32uncoal_l, f32uncoal_s = get_DRAM_f32_accesses(
sub_map, params)
f32coal = f32coal_l + f32coal_s
f32uncoal = f32uncoal_l + f32uncoal_s
# execute
#print "="*40+"TIMING RESULTS"
print("running kernel...")
#knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
#if not prefetch:
# knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
trial_times = []
for i in range(averaging_trials+warmup_trials):
evt, (out,) = knl(queue, a=a_mat_dev, b=b_mat_dev)
evt.wait()
trial_times.append((evt.profile.END - evt.profile.START)*1e-9)
avg_time = np.average(trial_times[warmup_trials:])
#if not prefetch:
# 1/0
gstats = GPUStats('TeslaC2070')
''' for k20
if n % BSIZEx == 0 and n % BSIZEy == 0:
if prefetch:
reg32_per_thread = 10
else:
reg32_per_thread = 8
else:
if prefetch:
reg32_per_thread = 8
else:
reg32_per_thread = 9
'''
reg32_per_thread = 8 # for c2070
#print(reg32_per_thread, estimate_regs_per_thread(knl))
reg32_per_thread = estimate_regs_per_thread(knl)
#reg32_per_thread = 1 #estimate_regs_per_thread(knl)
if prefetch:
shared_mem_per_block = 4*BSIZEx*BSIZEy
else:
shared_mem_per_block = 0
# TODO why is HK way off on the non-prefetch version?
total_blocks = math.ceil(n/BSIZEx)*math.ceil(n/BSIZEy)
total_threads = total_blocks*BSIZEx*BSIZEy # TODO unused
kstats = KernelStats(flops/(n*n), f32uncoal/(n*n), f32coal/(n*n),
barrier_ct, reg32_per_thread, shared_mem_per_block)
tconfig = ThreadConfig(BSIZEx*BSIZEy, total_blocks)
model = PerfModel(gstats, kstats, tconfig,
np.dtype(dtype))
cycles = model.compute_total_cycles()
actual.append(avg_time)
#for time in trial_times: #!!!!!
# actual.append(time)
HK_predict.append(cycles/(gstats.sm_clock_freq*10**9))
#update_LS_matrix(A, flops, f32coal_l, f32coal_s, f32uncoal_l,
#for time in trial_times:
'''
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
ops = copy.deepcopy(amd_op32)
ops.append(other_op32)
update_LS_matrix2(A, ops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
update_lstsq_mats(Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all,
A, actual, HK_predict, train_test_config)
def run_axpy_trials(ctx, queue, nvals, configs_t,
Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all, train_test_config):
A = []
HK_predict = []
actual = []
dtype = np.float32
#TODO figure out smem usage issue
for n in nvals:
x_vec_dev = cl.clrandom.rand(queue, n, dtype=dtype)
y_vec_dev = cl.clrandom.rand(queue, n, dtype=dtype)
z_vec_dev = cl.clrandom.rand(queue, n, dtype=dtype)
knl = lp.make_kernel(
"[n] -> {[i]: 0<=i<%d}" % n,
[
"z[i] = 5.0*x[i]+7.0*y[i]"
], [
lp.GlobalArg("x", dtype, shape=n),
lp.GlobalArg("y", dtype, shape=n),
lp.GlobalArg("z", dtype, shape=n),
], name="axpy")
ref_knl = knl
for BSIZEx, BSIZEy in configs_t:
knl = ref_knl
unroll = 4
knl = lp.split_iname(knl, "i", unroll*BSIZEx,
outer_tag="g.0", slabs=(0, 1))
knl = lp.split_iname(knl, "i_inner", BSIZEx,
outer_tag="unr", inner_tag="l.0")
#check = lp.auto_test_vs_ref(ref_knl, ctx, knl, print_code=False)
#print "Correctness check: \n", check
# use ptx src to determine resource usage
#ptx_dump(ctx, knl, n, BSIZEx, BSIZEy)
params = {'n': n}
barrier_poly = get_barrier_poly(knl)
barrier_ct = barrier_poly.eval_with_dict(params)
op_map = get_op_poly(knl)
op_map2 = get_op_poly2(knl)
flops, iops = get_32b_ops(op_map, params)
amd_op32 = get_32b_amd_ops(op_map2, params)
other_op32 = get_32b_ops_all(op_map2, params) - sum(amd_op32)
if flops + iops != sum(amd_op32) + other_op32: #TODO remove after debug
print("<debug> PROBLEM!, ops don't add up: ",
flops, iops, sum(amd_op32), other_op32)
sub_map = get_DRAM_access_poly(knl) # noqa
f32coal_l, f32coal_s, f32uncoal_l, f32uncoal_s = get_DRAM_f32_accesses(
sub_map, params)
f32coal = f32coal_l + f32coal_s
f32uncoal = f32uncoal_l + f32uncoal_s
'''
print_ptx_src_msg(knl.name)
print "="*40+"KERNEL STATS"
print "barrier count: ", barrier_ct
print "flops: ", flops
print(sub_map)
print "="*40
'''
# execute
print("running kernel...")
#knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
trial_times = []
for i in range(averaging_trials+warmup_trials):
evt, (out,) = knl(queue, x=x_vec_dev, y=y_vec_dev, z=z_vec_dev)
evt.wait()
trial_times.append((evt.profile.END - evt.profile.START)*1e-9)
avg_time = np.average(trial_times[warmup_trials:])
gstats = GPUStats('TeslaC2070')
reg32_per_thread = 18 #18 for c2070, 20 for k20
#reg32_per_thread = 1 #estimate_regs_per_thread(knl)
#print(reg32_per_thread, estimate_regs_per_thread(knl))
reg32_per_thread = estimate_regs_per_thread(knl)
shared_mem_per_block = 0
total_blocks = math.ceil(n/(BSIZEx*unroll))
kstats = KernelStats(flops*unroll/n, f32uncoal*unroll/n,
f32coal*unroll/n, barrier_ct, reg32_per_thread,
shared_mem_per_block)
tconfig = ThreadConfig(BSIZEx*BSIZEy, total_blocks)
model = PerfModel(gstats, kstats, tconfig, np.dtype(dtype))
cycles = model.compute_total_cycles()
actual.append(avg_time)
#for time in trial_times: #!!!!!
# actual.append(time)
HK_predict.append(cycles/(gstats.sm_clock_freq*10**9))
#for time in trial_times:
'''
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n/unroll,
np.dtype(dtype).itemsize, model)
'''
ops = copy.deepcopy(amd_op32)
ops.append(other_op32)
update_LS_matrix2(A, ops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n/unroll,
np.dtype(dtype).itemsize, model)
update_lstsq_mats(Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all,
A, actual, HK_predict, train_test_config)
def run_mm_trials(ctx, queue, nvals, configs_t,
Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all, train_test_config, version):
A = []
HK_predict = []
actual = []
dtype = np.float32
#TODO figure out smem usage issue
for n in nvals:
a_mat_dev = cl.clrandom.rand(queue, (n, n), dtype=dtype)
b_mat_dev = cl.clrandom.rand(queue, (n, n), dtype=dtype)
c_mat_dev = cl.clrandom.rand(queue, (n, n), dtype=dtype)
order = "C"
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
], [
lp.GlobalArg("a", dtype, shape=(n, n), order=order),
lp.GlobalArg("b", dtype, shape=(n, n), order=order),
lp.GlobalArg("c", dtype, shape=(n, n), order=order),
], name="matmul")
ref_knl = knl
for BSIZEx, BSIZEy in configs_t:
knl = ref_knl
if version == "allcoal":
knl = lp.split_iname(knl, "i", BSIZEy,
outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", BSIZEx,
outer_tag="g.1", inner_tag="l.0")
elif version == "partcoal":
knl = lp.split_iname(knl, "i", BSIZEy,
outer_tag="g.0", inner_tag="l.0")
knl = lp.split_iname(knl, "j", BSIZEx,
outer_tag="g.1", inner_tag="l.1")
else:
1/0
# TODO error
ksplit = BSIZEy
knl = lp.split_iname(knl, "k", ksplit)
knl = lp.add_prefetch(knl, "a", ["k_inner", "i_inner"])
knl = lp.add_prefetch(knl, "b", ["j_inner", "k_inner", ])
#check = lp.auto_test_vs_ref(ref_knl, ctx, knl, print_code=True)
#print "Correctness check: \n", check
# use ptx src to determine resource usage
#ptx_dump(ctx, knl, n, BSIZEx, BSIZEy)
params = {'n': n}
barrier_poly = get_barrier_poly(knl)
barrier_ct = barrier_poly.eval_with_dict(params)
op_map = get_op_poly(knl)
op_map2 = get_op_poly2(knl)
flops, iops = get_32b_ops(op_map, params)
amd_op32 = get_32b_amd_ops(op_map2, params)
other_op32 = get_32b_ops_all(op_map2, params) - sum(amd_op32)
if flops + iops != sum(amd_op32) + other_op32: #TODO remove after debug
print("<debug> PROBLEM!, ops don't add up: ",
flops, iops, sum(amd_op32), other_op32)
sub_map = get_DRAM_access_poly(knl) # noqa
f32coal_l, f32coal_s, f32uncoal_l, f32uncoal_s = get_DRAM_f32_accesses(
sub_map, params)
f32coal = f32coal_l + f32coal_s
f32uncoal = f32uncoal_l + f32uncoal_s
'''
print_ptx_src_msg(knl.name)
print "="*40+"KERNEL STATS"
print "barrier count: ", barrier_ct
print "flops: ", flops
print(sub_map)
print "="*40
'''
# execute
#print "="*40+"TIMING RESULTS"
print("running kernel...")
#knl = lp.set_options(knl, write_cl=True, highlight_cl=True)
trial_times = []
for i in range(averaging_trials+warmup_trials):
evt, (out,) = knl(queue, a=a_mat_dev, b=b_mat_dev, c=c_mat_dev)
evt.wait()
trial_times.append((evt.profile.END - evt.profile.START)*1e-9)
avg_time = np.average(trial_times[warmup_trials:])
gstats = GPUStats('TeslaC2070')
''' for k20:
if BSIZEx == 8 or BSIZEx == 32: # TODO fix hack
reg32_per_thread = 25
elif BSIZEx == 24:
reg32_per_thread = 18
elif BSIZEx == 16:
reg32_per_thread = 22
'''
# for C2070
if BSIZEx == 8 or BSIZEx == 16: # TODO fix hack
reg32_per_thread = 20
elif BSIZEx == 32:
reg32_per_thread = 19
elif BSIZEx == 24:
reg32_per_thread = 12
#reg32_per_thread = 1 #estimate_regs_per_thread(knl)
#print(reg32_per_thread, estimate_regs_per_thread(knl))
reg32_per_thread = estimate_regs_per_thread(knl)
shared_mem_per_block = 4*ksplit*(BSIZEx+BSIZEy)
total_blocks = math.ceil(n/BSIZEx)*math.ceil(n/BSIZEy)
total_threads = total_blocks*BSIZEx*BSIZEy # TODO never used
kstats = KernelStats(flops/(n*n), f32uncoal/(n*n), f32coal/(n*n),
barrier_ct, reg32_per_thread, shared_mem_per_block)
tconfig = ThreadConfig(BSIZEx*BSIZEy, total_blocks)
model = PerfModel(gstats, kstats, tconfig,
np.dtype(dtype))
cycles = model.compute_total_cycles()
actual.append(avg_time)
#for time in trial_times: #!!!!!
# actual.append(time)
HK_predict.append(cycles/(gstats.sm_clock_freq*10**9))
'''
print "actual runtime: ", actual[-1]
print "total predicted time: ", predicted[-1]
print "total predicted execution cycles: ", cycles
print "="*40
'''
#''' #!!!!!
'''
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
ops = copy.deepcopy(amd_op32)
ops.append(other_op32)
update_LS_matrix2(A, ops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
for time in trial_times:
update_LS_matrix(A, flops, iops, f32coal_l, f32coal_s, f32uncoal_l,
f32uncoal_s, barrier_ct, total_blocks, n*n,
np.dtype(dtype).itemsize, model)
'''
update_lstsq_mats(Atrain_all, Atest_all, ytrain_all, ytest_all,
actual_times_all, HK_predict_all,
A, actual, HK_predict, train_test_config)