def test_coarse_grain_svm(ctx_factory):
import sys
is_pypy = '__pypy__' in sys.builtin_module_names
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
dev = ctx.devices[0]
has_svm = (ctx._get_cl_version() >= (2, 0)
and ctx.devices[0]._get_cl_version() >= (2, 0)
and cl.get_cl_header_version() >= (2, 0))
if dev.platform.name == "Portable Computing Language":
has_svm = (get_pocl_version(dev.platform) >= (1, 0)
and cl.get_cl_header_version() >= (2, 0))
if not has_svm:
from pytest import skip
skip("SVM only available in OpenCL 2.0 and higher")
if ("AMD" in dev.platform.name and dev.type & cl.device_type.CPU):
pytest.xfail("AMD CPU doesn't do coarse-grain SVM")
n = 3000
svm_ary = cl.SVM(cl.csvm_empty(ctx, (n, ), np.float32, alignment=64))
if not is_pypy:
# https://bitbucket.org/pypy/numpy/issues/52
assert isinstance(svm_ary.mem.base, cl.SVMAllocation)
cl.enqueue_svm_memfill(queue, svm_ary, np.zeros((), svm_ary.mem.dtype))
with svm_ary.map_rw(queue) as ary:
ary.fill(17)
orig_ary = ary.copy()
prg = cl.Program(
ctx, """
__kernel void twice(__global float *a_g)
{
a_g[get_global_id(0)] *= 2;
}
""").build()
prg.twice(queue, svm_ary.mem.shape, None, svm_ary)
with svm_ary.map_ro(queue) as ary:
print(ary)
assert np.array_equal(orig_ary * 2, ary)
new_ary = np.empty_like(orig_ary)
new_ary.fill(-1)
if ctx.devices[0].platform.name != "Portable Computing Language":
# "Blocking memcpy is unimplemented (clEnqueueSVMMemcpy.c:61)"
# in pocl up to and including 1.0rc1.
cl.enqueue_copy(queue, new_ary, svm_ary)
assert np.array_equal(orig_ary * 2, new_ary)
def _may_have_svm(dev):
has_svm = (dev.platform._get_cl_version() >= (2, 0)
and cl.get_cl_header_version() >= (2, 0))
if dev.platform.name == "Portable Computing Language":
has_svm = (get_pocl_version(dev.platform) >= (1, 0)
and cl.get_cl_header_version() >= (2, 0))
return has_svm
def _may_have_svm(dev):
has_svm = (dev.platform._get_cl_version() >= (2, 0)
and cl.get_cl_header_version() >= (2, 0))
if dev.platform.name == "Portable Computing Language":
has_svm = (
get_pocl_version(dev.platform) >= (1, 0)
and cl.get_cl_header_version() >= (2, 0))
return has_svm
def print_info(self):
print("PyOpenCL Version:", cl.VERSION)
print("OpenCL Version:", cl.get_cl_header_version())
print()
print("Platform Name:", self._platform.get_info(cl.platform_info.NAME))
print("Platform Profile:",
self._platform.get_info(cl.platform_info.PROFILE))
print("Platform Vendor:",
self._platform.get_info(cl.platform_info.VENDOR))
print("Platform Version:",
self._platform.get_info(cl.platform_info.VERSION))
print()
print("GPU Name:", self._device.get_info(cl.device_info.NAME))
print("OpenCL Version:",
self._device.get_info(cl.device_info.OPENCL_C_VERSION))
print("GPU Vendor:", self._device.get_info(cl.device_info.VENDOR))
print("GPU Version:", self._device.get_info(cl.device_info.VERSION))
print("GPU Driver Version:",
self._device.get_info(cl.device_info.DRIVER_VERSION))
print("Max Work Group Size:",
self._device.get_info(cl.device_info.MAX_WORK_GROUP_SIZE))
print("Max Compute Units:",
self._device.get_info(cl.device_info.MAX_COMPUTE_UNITS))
print("Max Work Item Size:",
self._device.get_info(cl.device_info.MAX_WORK_ITEM_SIZES))
print("Local Memory Size:",
self._device.get_info(cl.device_info.LOCAL_MEM_SIZE) / 1024,
'KB')
def test_platform_get_devices(ctx_factory):
ctx = ctx_factory()
platform = ctx.devices[0].platform
if platform.name == "Apple":
pytest.xfail("Apple doesn't understand all the values we pass " "for dev_type")
dev_types = [
cl.device_type.ACCELERATOR,
cl.device_type.ALL,
cl.device_type.CPU,
cl.device_type.DEFAULT,
cl.device_type.GPU,
]
if (
platform._get_cl_version() >= (1, 2)
and cl.get_cl_header_version() >= (1, 2)
and not platform.name.lower().startswith("nvidia")
):
dev_types.append(cl.device_type.CUSTOM)
for dev_type in dev_types:
print(dev_type)
devs = platform.get_devices(dev_type)
if dev_type in (cl.device_type.DEFAULT, cl.device_type.ALL, getattr(cl.device_type, "CUSTOM", None)):
continue
for dev in devs:
assert dev.type & dev_type == dev_type
def test_sub_buffers(ctx_factory):
ctx = ctx_factory()
if (ctx._get_cl_version() < (1, 1) or
cl.get_cl_header_version() < (1, 1)):
from pytest import skip
skip("sub-buffers are only available in OpenCL 1.1")
alignment = ctx.devices[0].mem_base_addr_align
queue = cl.CommandQueue(ctx)
n = 30000
a = (np.random.rand(n) * 100).astype(np.uint8)
mf = cl.mem_flags
a_buf = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=a)
start = (5000 // alignment) * alignment
stop = start + 20 * alignment
a_sub_ref = a[start:stop]
a_sub = np.empty_like(a_sub_ref)
cl.enqueue_copy(queue, a_sub, a_buf[start:stop])
assert np.array_equal(a_sub, a_sub_ref)
def test_sub_buffers(ctx_factory):
ctx = ctx_factory()
if (ctx._get_cl_version() < (1, 1) and cl.get_cl_header_version() <
(1, 1)):
from pytest import skip
skip("sub-buffers are only available in OpenCL 1.1")
alignment = ctx.devices[0].mem_base_addr_align
queue = cl.CommandQueue(ctx)
n = 30000
a = (np.random.rand(n) * 100).astype(np.uint8)
mf = cl.mem_flags
a_buf = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=a)
start = (5000 // alignment) * alignment
stop = start + 20 * alignment
a_sub_ref = a[start:stop]
a_sub = np.empty_like(a_sub_ref)
cl.enqueue_copy(queue, a_sub, a_buf[start:stop])
assert np.array_equal(a_sub, a_sub_ref)
def get_info():
global selected_device, selected_platform, context, KERNELS_DEFS
info = {
"version" : {
"cl_header" : pyopencl.get_cl_header_version(),
},
"opengl" : pyopencl.have_gl(),
#"kernels" : KERNELS_DEFS.keys()
"pyopencl" : get_pyopencl_info(),
}
if selected_platform:
info["platform"] = {
"name" : selected_platform.name,
"vendor" : selected_platform.vendor,
"devices" : len(selected_platform.get_devices()),
}
if selected_device:
dinfo = {
"type" : device_type(selected_device),
"name" : selected_device.name.strip(),
"version" : selected_device.version,
"max_work_group_size" : selected_device.max_work_group_size,
"max_work_item_dimensions" : selected_device.max_work_item_dimensions,
"max_work_item_sizes" : selected_device.max_work_item_sizes,
"max-size" : selected_device_max_size}
if hasattr(selected_device, "opencl_c_version"):
dinfo["opencl_c_version"] = getattr(selected_device, "opencl_c_version")
info["device"] = dinfo
return info
def test_bitonic_sort(ctx_factory, size, dtype):
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
dev = ctx.devices[0]
if (dev.platform.name == "Apple" and dev.type & cl.device_type.CPU):
pytest.xfail("Bitonic sort won't work on Apple CPU: no workgroup "
"parallelism")
if (dev.platform.name == "Portable Computing Language"
and dtype == np.float64
and get_pocl_version(dev.platform) < (1, 0)):
pytest.xfail("Double precision bitonic sort doesn't work on POCL < 1.0")
if dtype == np.float64 and not has_double_support(dev):
from pytest import skip
skip("double precision not supported on %s" % dev)
import pyopencl.clrandom as clrandom
from pyopencl.bitonic_sort import BitonicSort
s = clrandom.rand(queue, (2, size, 3,), dtype, luxury=None, a=0, b=239482333)
sgs = s.copy()
# enqueue_marker crashes under CL 1.1 pocl if there is anything to wait for
# (no clEnqueueWaitForEvents) https://github.com/inducer/pyopencl/pull/237
if (dev.platform.name == "Portable Computing Language"
and cl.get_cl_header_version() < (1, 2)):
sgs.finish()
sorter = BitonicSort(ctx)
sgs, evt = sorter(sgs, axis=1)
assert np.array_equal(np.sort(s.get(), axis=1), sgs.get())
def get_info():
global selected_device, selected_platform, context, KERNELS_DEFS
info = {"version" : pyopencl.version.VERSION,
"version.text" : pyopencl.VERSION_TEXT,
"version.status" : pyopencl.VERSION_STATUS,
"version.cl_header" : pyopencl.get_cl_header_version(),
"opengl" : pyopencl.have_gl(),
#"kernels" : KERNELS_DEFS.keys()
}
if selected_platform:
info.update({
"platform.name" : selected_platform.name,
"platform.vendor" : selected_platform.vendor,
"platform.devices" : len(selected_platform.get_devices()),
})
if selected_device:
if hasattr(selected_device, "opencl_c_version"):
info["device.opencl_c_version"] = getattr(selected_device, "opencl_c_version")
info.update({
"device.type" : device_type(selected_device),
"device.name" : selected_device.name.strip(),
"device.version" : selected_device.version,
"device.max_work_group_size" : selected_device.max_work_group_size,
"device.max_work_item_dimensions" : selected_device.max_work_item_dimensions,
"device.max_work_item_sizes" : selected_device.max_work_item_sizes})
return info
def test_custom_type_zeros(ctx_factory):
context = ctx_factory()
queue = cl.CommandQueue(context)
if not (
queue._get_cl_version() >= (1, 2)
and cl.get_cl_header_version() >= (1, 2)):
pytest.skip("CL1.2 not available")
dtype = np.dtype([
("cur_min", np.int32),
("cur_max", np.int32),
("pad", np.int32),
])
from pyopencl.tools import get_or_register_dtype, match_dtype_to_c_struct
name = "mmc_type"
dtype, c_decl = match_dtype_to_c_struct(queue.device, name, dtype)
dtype = get_or_register_dtype(name, dtype)
n = 1000
z_dev = cl.array.zeros(queue, n, dtype=dtype)
z = z_dev.get()
assert np.array_equal(np.zeros(n, dtype), z)
def get_info():
global selected_device, selected_platform, context, KERNELS_DEFS
info = {
"version.cl_header" : pyopencl.get_cl_header_version(),
"opengl" : pyopencl.have_gl(),
#"kernels" : KERNELS_DEFS.keys()
}
updict(info, "pyopencl", get_pyopencl_info())
if selected_platform:
updict(info, "platform", {
"name" : selected_platform.name,
"vendor" : selected_platform.vendor,
"devices" : len(selected_platform.get_devices()),
})
if selected_device:
if hasattr(selected_device, "opencl_c_version"):
info["device.opencl_c_version"] = getattr(selected_device, "opencl_c_version")
updict(info, "device", {
"type" : device_type(selected_device),
"name" : selected_device.name.strip(),
"version" : selected_device.version,
"max_work_group_size" : selected_device.max_work_group_size,
"max_work_item_dimensions" : selected_device.max_work_item_dimensions,
"max_work_item_sizes" : selected_device.max_work_item_sizes,
"max-size" : selected_device_max_size})
return info
def test_custom_type_zeros(ctx_factory):
context = ctx_factory()
queue = cl.CommandQueue(context)
if not (queue._get_cl_version() >= (1, 2) and cl.get_cl_header_version() >=
(1, 2)):
pytest.skip("CL1.2 not available")
dtype = np.dtype([
("cur_min", np.int32),
("cur_max", np.int32),
("pad", np.int32),
])
from pyopencl.tools import get_or_register_dtype, match_dtype_to_c_struct
name = "mmc_type"
dtype, c_decl = match_dtype_to_c_struct(queue.device, name, dtype)
dtype = get_or_register_dtype(name, dtype)
n = 1000
z_dev = cl.array.zeros(queue, n, dtype=dtype)
z = z_dev.get()
assert np.array_equal(np.zeros(n, dtype), z)
def test_compile_link(ctx_factory):
ctx = ctx_factory()
if ctx._get_cl_version() < (1, 2) or cl.get_cl_header_version() < (1, 2):
pytest.skip(
"Context and ICD loader must understand CL1.2 for compile/link")
platform = ctx.devices[0].platform
if platform.name == "Apple":
pytest.skip("Apple doesn't like our compile/link test")
queue = cl.CommandQueue(ctx)
vsink_prg = cl.Program(
ctx, """//CL//
void value_sink(float x)
{
}
""").compile()
main_prg = cl.Program(
ctx, """//CL//
void value_sink(float x);
__kernel void experiment()
{
value_sink(3.1415f + get_global_id(0));
}
""").compile()
z = cl.link_program(ctx, [vsink_prg, main_prg], devices=ctx.devices)
z.experiment(queue, (128**2, ), (128, ))
queue.finish()
def test_spirv(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
if (ctx._get_cl_version() < (2, 1) or cl.get_cl_header_version() < (2, 1)):
pytest.skip("SPIR-V program creation only available "
"in OpenCL 2.1 and higher")
n = 50000
a_dev = cl.clrandom.rand(queue, n, np.float32)
b_dev = cl.clrandom.rand(queue, n, np.float32)
dest_dev = cl_array.empty_like(a_dev)
with open("add-vectors-%d.spv" % queue.device.address_bits,
"rb") as spv_file:
spv = spv_file.read()
prg = cl.Program(ctx, spv).build()
if (not prg.all_kernels()
and queue.device.platform.name.startswith("AMD Accelerated")):
pytest.skip(
"SPIR-V program creation on AMD did not result in any kernels")
prg.sum(queue, a_dev.shape, None, a_dev.data, b_dev.data, dest_dev.data)
assert la.norm((dest_dev - (a_dev + b_dev)).get()) < 1e-7
def test_bitonic_argsort(ctx_factory, size, dtype):
import sys
is_pypy = "__pypy__" in sys.builtin_module_names
if not size and is_pypy:
# https://bitbucket.org/pypy/numpy/issues/53/specifying-strides-on-zero-sized-array
pytest.xfail("pypy doesn't seem to handle as_strided "
"on zero-sized arrays very well")
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
device = queue.device
if device.platform.vendor == "The pocl project" \
and device.type & cl.device_type.GPU:
pytest.xfail("bitonic argsort fails on POCL + Nvidia,"
"at least the K40, as of pocl 1.6, 2021-01-20")
dev = ctx.devices[0]
if (dev.platform.name == "Portable Computing Language"
and sys.platform == "darwin"):
pytest.xfail("Bitonic sort crashes on Apple POCL")
if (dev.platform.name == "Apple" and dev.type & cl.device_type.CPU):
pytest.xfail("Bitonic sort won't work on Apple CPU: no workgroup "
"parallelism")
if (dev.platform.name == "Portable Computing Language"
and dtype == np.float64
and get_pocl_version(dev.platform) < (1, 0)):
pytest.xfail("Double precision bitonic sort doesn't work on POCL < 1.0")
if (dev.platform.name == "Intel(R) OpenCL" and size == 0):
pytest.xfail("size-0 arange fails on Intel CL")
if dtype == np.float64 and not has_double_support(dev):
from pytest import skip
skip("double precision not supported on %s" % dev)
import pyopencl.clrandom as clrandom
from pyopencl.bitonic_sort import BitonicSort
index = cl_array.arange(queue, 0, size, 1, dtype=np.int32)
m = clrandom.rand(queue, (size,), dtype, luxury=None, a=0, b=239432234)
sorterm = BitonicSort(ctx)
ms = m.copy()
# enqueue_marker crashes under CL 1.1 pocl if there is anything to wait for
# (no clEnqueueWaitForEvents) https://github.com/inducer/pyopencl/pull/237
if (dev.platform.name == "Portable Computing Language"
and cl.get_cl_header_version() < (1, 2)):
ms.finish()
index.finish()
ms, evt = sorterm(ms, idx=index, axis=0)
assert np.array_equal(np.sort(m.get()), ms.get())
# may be False because of identical values in array
# assert np.array_equal(np.argsort(m.get()), index.get())
# Check values by indices
assert np.array_equal(m.get()[np.argsort(m.get())], m.get()[index.get()])
def test_unload_compiler(platform):
if (platform._get_cl_version() < (1, 2) or
cl.get_cl_header_version() < (1, 2)):
from pytest import skip
skip("clUnloadPlatformCompiler is only available in OpenCL 1.2")
_skip_if_pocl(platform, 'pocl does not support unloading compiler')
if platform.vendor == "Intel(R) Corporation":
from pytest import skip
skip("Intel proprietary driver does not support unloading compiler")
cl.unload_platform_compiler(platform)
def test_unload_compiler(platform):
if (platform._get_cl_version() < (1, 2) or cl.get_cl_header_version() <
(1, 2)):
from pytest import skip
skip("clUnloadPlatformCompiler is only available in OpenCL 1.2")
_skip_if_pocl(platform, 'pocl does not support unloading compiler')
if platform.vendor == "Intel(R) Corporation":
from pytest import skip
skip("Intel proprietary driver does not support unloading compiler")
cl.unload_platform_compiler(platform)
def test_fine_grain_svm(ctx_factory):
ctx = ctx_factory()
# queue = cl.CommandQueue(ctx)
if (ctx._get_cl_version() < (2, 0) or cl.get_cl_header_version() < (2, 0)):
from pytest import skip
skip("SVM only available in OpenCL 2.0 and higher")
svm_ary = cl.fsvm_empty(ctx, (100, 100), np.float32, alignment=64)
assert isinstance(svm_ary.base, cl.SVMAllocation)
def test_fine_grain_svm(ctx_factory):
ctx = ctx_factory()
# queue = cl.CommandQueue(ctx)
if (ctx._get_cl_version() < (2, 0) or
cl.get_cl_header_version() < (2, 0)):
from pytest import skip
skip("SVM only available in OpenCL 2.0 and higher")
svm_ary = cl.fsvm_empty(ctx, (100, 100), np.float32, alignment=64)
assert isinstance(svm_ary.base, cl.SVMAllocation)
def test_coarse_grain_svm(ctx_factory):
import sys
is_pypy = '__pypy__' in sys.builtin_module_names
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
if (ctx._get_cl_version() < (2, 0) or
cl.get_cl_header_version() < (2, 0)):
from pytest import skip
skip("SVM only available in OpenCL 2.0 and higher")
dev = ctx.devices[0]
if ("AMD" in dev.platform.name
and dev.type & cl.device_type.CPU):
pytest.xfail("AMD CPU doesn't do coarse-grain SVM")
n = 3000
svm_ary = cl.SVM(cl.csvm_empty(ctx, (n,), np.float32, alignment=64))
if not is_pypy:
# https://bitbucket.org/pypy/numpy/issues/52
assert isinstance(svm_ary.mem.base, cl.SVMAllocation)
if dev.platform.name != "Portable Computing Language":
# pocl 0.13 has a bug misinterpreting the size parameter
cl.enqueue_svm_memfill(queue, svm_ary, np.zeros((), svm_ary.mem.dtype))
with svm_ary.map_rw(queue) as ary:
ary.fill(17)
orig_ary = ary.copy()
prg = cl.Program(ctx, """
__kernel void twice(__global float *a_g)
{
a_g[get_global_id(0)] *= 2;
}
""").build()
prg.twice(queue, svm_ary.mem.shape, None, svm_ary)
with svm_ary.map_ro(queue) as ary:
print(ary)
assert np.array_equal(orig_ary*2, ary)
new_ary = np.empty_like(orig_ary)
new_ary.fill(-1)
if ctx.devices[0].platform.name != "Portable Computing Language":
# "Blocking memcpy is unimplemented (clEnqueueSVMMemcpy.c:61)"
# in pocl 0.13.
cl.enqueue_copy(queue, new_ary, svm_ary)
assert np.array_equal(orig_ary*2, new_ary)
def test_platform_get_devices(platform):
dev_types = [cl.device_type.ACCELERATOR, cl.device_type.ALL,
cl.device_type.CPU, cl.device_type.DEFAULT, cl.device_type.GPU]
if (platform._get_cl_version() >= (1, 2) and
cl.get_cl_header_version() >= (1, 2)):
dev_types.append(cl.device_type.CUSTOM)
for dev_type in dev_types:
devs = platform.get_devices(dev_type)
if dev_type in (cl.device_type.DEFAULT,
cl.device_type.ALL,
getattr(cl.device_type, 'CUSTOM', None)):
continue
for dev in devs:
assert dev.type == dev_type
def test_platform_get_devices(platform):
dev_types = [
cl.device_type.ACCELERATOR, cl.device_type.ALL, cl.device_type.CPU,
cl.device_type.DEFAULT, cl.device_type.GPU
]
if (platform._get_cl_version() >= (1, 2) and cl.get_cl_header_version() >=
(1, 2)):
dev_types.append(cl.device_type.CUSTOM)
for dev_type in dev_types:
devs = platform.get_devices(dev_type)
if dev_type in (cl.device_type.DEFAULT, cl.device_type.ALL,
getattr(cl.device_type, 'CUSTOM', None)):
continue
for dev in devs:
assert dev.type == dev_type
def test_enqueue_barrier_marker(ctx_factory):
ctx = ctx_factory()
# Still relevant on pocl 1.0RC1.
_xfail_if_pocl(
ctx.devices[0].platform, (1, 0), "pocl crashes on enqueue_barrier")
queue = cl.CommandQueue(ctx)
if queue._get_cl_version() >= (1, 2) and cl.get_cl_header_version() <= (1, 1):
pytest.skip("CL impl version >= 1.2, header version <= 1.1--cannot be sure "
"that clEnqueueWaitForEvents is implemented")
cl.enqueue_barrier(queue)
evt1 = cl.enqueue_marker(queue)
evt2 = cl.enqueue_marker(queue, wait_for=[evt1])
cl.enqueue_barrier(queue, wait_for=[evt1, evt2])
def test_empty_ndrange(ctx_factory, empty_shape):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
if ctx._get_cl_version() < (1, 2) or cl.get_cl_header_version() < (1, 2):
pytest.skip("OpenCL 1.2 required for empty NDRange suuport")
a = cl_array.zeros(queue, empty_shape, dtype=np.float32)
prg = cl.Program(ctx, """
__kernel void add_two(__global float *a_g)
{
a_g[get_global_id(0)] += 2;
}
""").build()
prg.add_two(queue, a.shape, None, a.data, allow_empty_ndrange=True)
def test_user_event(ctx_factory):
ctx = ctx_factory()
if ctx._get_cl_version() < (1, 1) and cl.get_cl_header_version() < (1, 1):
from pytest import skip
skip("UserEvent is only available in OpenCL 1.1")
if ctx.devices[0].platform.name == "Portable Computing Language":
# https://github.com/pocl/pocl/issues/201
pytest.xfail("POCL's user events don't work right")
status = {}
def event_waiter1(e, key):
e.wait()
status[key] = True
def event_waiter2(e, key):
cl.wait_for_events([e])
status[key] = True
from threading import Thread
from time import sleep
evt = cl.UserEvent(ctx)
Thread(target=event_waiter1, args=(evt, 1)).start()
sleep(0.05)
if status.get(1, False):
raise RuntimeError("UserEvent triggered before set_status")
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(0.05)
if not status.get(1, False):
raise RuntimeError("UserEvent.wait timeout")
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
evt = cl.UserEvent(ctx)
Thread(target=event_waiter2, args=(evt, 2)).start()
sleep(0.05)
if status.get(2, False):
raise RuntimeError("UserEvent triggered before set_status")
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(0.05)
if not status.get(2, False):
raise RuntimeError("cl.wait_for_events timeout on UserEvent")
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
def test_platform_get_devices(platform):
if platform.name == "Apple":
pytest.xfail("Apple doesn't understand all the values we pass "
"for dev_type")
dev_types = [cl.device_type.ACCELERATOR, cl.device_type.ALL,
cl.device_type.CPU, cl.device_type.DEFAULT, cl.device_type.GPU]
if (platform._get_cl_version() >= (1, 2) and
cl.get_cl_header_version() >= (1, 2)):
dev_types.append(cl.device_type.CUSTOM)
for dev_type in dev_types:
devs = platform.get_devices(dev_type)
if dev_type in (cl.device_type.DEFAULT,
cl.device_type.ALL,
getattr(cl.device_type, 'CUSTOM', None)):
continue
for dev in devs:
assert dev.type & dev_type == dev_type
def test_user_event(ctx_factory):
ctx = ctx_factory()
if (ctx._get_cl_version() < (1, 1) and cl.get_cl_header_version() <
(1, 1)):
from pytest import skip
skip("UserEvent is only available in OpenCL 1.1")
if ctx.devices[0].platform.name == "Portable Computing Language":
# https://github.com/pocl/pocl/issues/201
pytest.xfail("POCL's user events don't work right")
status = {}
def event_waiter1(e, key):
e.wait()
status[key] = True
def event_waiter2(e, key):
cl.wait_for_events([e])
status[key] = True
from threading import Thread
from time import sleep
evt = cl.UserEvent(ctx)
Thread(target=event_waiter1, args=(evt, 1)).start()
sleep(.05)
if status.get(1, False):
raise RuntimeError('UserEvent triggered before set_status')
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(.05)
if not status.get(1, False):
raise RuntimeError('UserEvent.wait timeout')
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
evt = cl.UserEvent(ctx)
Thread(target=event_waiter2, args=(evt, 2)).start()
sleep(.05)
if status.get(2, False):
raise RuntimeError('UserEvent triggered before set_status')
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(.05)
if not status.get(2, False):
raise RuntimeError('cl.wait_for_events timeout on UserEvent')
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
def test_user_event(ctx_factory):
ctx = ctx_factory()
if (ctx._get_cl_version() < (1, 1) and
cl.get_cl_header_version() < (1, 1)):
from pytest import skip
skip("UserEvent is only available in OpenCL 1.1")
status = {}
def event_waiter1(e, key):
e.wait()
status[key] = True
def event_waiter2(e, key):
cl.wait_for_events([e])
status[key] = True
from threading import Thread
from time import sleep
evt = cl.UserEvent(ctx)
Thread(target=event_waiter1, args=(evt, 1)).start()
sleep(.05)
if status.get(1, False):
raise RuntimeError('UserEvent triggered before set_status')
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(.05)
if not status.get(1, False):
raise RuntimeError('UserEvent.wait timeout')
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
evt = cl.UserEvent(ctx)
Thread(target=event_waiter2, args=(evt, 2)).start()
sleep(.05)
if status.get(2, False):
raise RuntimeError('UserEvent triggered before set_status')
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(.05)
if not status.get(2, False):
raise RuntimeError('cl.wait_for_events timeout on UserEvent')
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
def test_user_event(ctx_factory):
ctx = ctx_factory()
if (ctx._get_cl_version() < (1, 1) and cl.get_cl_header_version() <
(1, 1)):
from pytest import skip
skip("UserEvent is only available in OpenCL 1.1")
status = {}
def event_waiter1(e, key):
e.wait()
status[key] = True
def event_waiter2(e, key):
cl.wait_for_events([e])
status[key] = True
from threading import Thread
from time import sleep
evt = cl.UserEvent(ctx)
Thread(target=event_waiter1, args=(evt, 1)).start()
sleep(.05)
if status.get(1, False):
raise RuntimeError('UserEvent triggered before set_status')
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(.05)
if not status.get(1, False):
raise RuntimeError('UserEvent.wait timeout')
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
evt = cl.UserEvent(ctx)
Thread(target=event_waiter2, args=(evt, 2)).start()
sleep(.05)
if status.get(2, False):
raise RuntimeError('UserEvent triggered before set_status')
evt.set_status(cl.command_execution_status.COMPLETE)
sleep(.05)
if not status.get(2, False):
raise RuntimeError('cl.wait_for_events timeout on UserEvent')
assert evt.command_execution_status == cl.command_execution_status.COMPLETE
def test_spirv(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
if (ctx._get_cl_version() < (2, 1) or cl.get_cl_header_version() < (2, 1)):
from pytest import skip
skip("SPIR-V program creation only available in OpenCL 2.1 and higher")
n = 50000
a_dev = cl.clrandom.rand(queue, n, np.float32)
b_dev = cl.clrandom.rand(queue, n, np.float32)
dest_dev = cl_array.empty_like(a_dev)
with open("add-vectors.spv", "rb") as spv_file:
spv = spv_file.read()
prg = cl.Program(ctx, spv)
prg.sum(queue, a_dev.shape, None, a_dev.data, b_dev.data, dest_dev.data)
assert la.norm((dest_dev - (a_dev + b_dev)).get()) < 1e-7
def print_info():
print("PyOpenCL Version:", cl.VERSION)
print("OpenCL Head Version:", cl.get_cl_header_version())
print()
platforms = cl.get_platforms()
print("Platforms Amount:", len(platforms))
for plat in platforms:
print("Platform:", plat.get_info(cl.platform_info.NAME))
print("--Platform Profile:",
plat.get_info(cl.platform_info.PROFILE))
print("--Platform Vendor:", plat.get_info(cl.platform_info.VENDOR))
print("--Platform Version:",
plat.get_info(cl.platform_info.VERSION))
devices = plat.get_devices(cl.device_type.GPU)
print("--GPU Amount:", len(devices))
for device in devices:
print("--GPU:", device.get_info(cl.device_info.NAME))
print("----OpenCL Version:",
device.get_info(cl.device_info.OPENCL_C_VERSION))
print("----GPU Vendor:",
device.get_info(cl.device_info.VENDOR))
print("----GPU Version:",
device.get_info(cl.device_info.VERSION))
print("----GPU Driver Version:",
device.get_info(cl.device_info.DRIVER_VERSION))
print("----Max Work Group Size:",
device.get_info(cl.device_info.MAX_WORK_GROUP_SIZE))
print("----Max Compute Units:",
device.get_info(cl.device_info.MAX_COMPUTE_UNITS))
print("----Max Work Item Size:",
device.get_info(cl.device_info.MAX_WORK_ITEM_SIZES))
print("----Local Memory Size:",
device.get_info(cl.device_info.LOCAL_MEM_SIZE) / 1024,
'KB')
print()
def test_fine_grain_svm(ctx_factory):
import sys
is_pypy = '__pypy__' in sys.builtin_module_names
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
from pytest import skip
if (ctx._get_cl_version() < (2, 0) or
cl.get_cl_header_version() < (2, 0)):
skip("SVM only available in OpenCL 2.0 and higher")
if not (ctx.devices[0].svm_capabilities
& cl.device_svm_capabilities.FINE_GRAIN_BUFFER):
skip("device does not support fine-grain SVM")
n = 3000
ary = cl.fsvm_empty(ctx, n, np.float32, alignment=64)
if not is_pypy:
# https://bitbucket.org/pypy/numpy/issues/52
assert isinstance(ary.base, cl.SVMAllocation)
ary.fill(17)
orig_ary = ary.copy()
prg = cl.Program(ctx, """
__kernel void twice(__global float *a_g)
{
a_g[get_global_id(0)] *= 2;
}
""").build()
prg.twice(queue, ary.shape, None, cl.SVM(ary))
queue.finish()
print(ary)
assert np.array_equal(orig_ary*2, ary)
def test_fine_grain_svm(ctx_factory):
import sys
is_pypy = '__pypy__' in sys.builtin_module_names
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
from pytest import skip
if (ctx._get_cl_version() < (2, 0) or cl.get_cl_header_version() < (2, 0)):
skip("SVM only available in OpenCL 2.0 and higher")
if not (ctx.devices[0].svm_capabilities
& cl.device_svm_capabilities.FINE_GRAIN_BUFFER):
skip("device does not support fine-grain SVM")
n = 3000
ary = cl.fsvm_empty(ctx, n, np.float32, alignment=64)
if not is_pypy:
# https://bitbucket.org/pypy/numpy/issues/52
assert isinstance(ary.base, cl.SVMAllocation)
ary.fill(17)
orig_ary = ary.copy()
prg = cl.Program(
ctx, """
__kernel void twice(__global float *a_g)
{
a_g[get_global_id(0)] *= 2;
}
""").build()
prg.twice(queue, ary.shape, None, cl.SVM(ary))
queue.finish()
print(ary)
assert np.array_equal(orig_ary * 2, ary)
def test_spirv(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
if (ctx._get_cl_version() < (2, 1) or
cl.get_cl_header_version() < (2, 1)):
from pytest import skip
skip("SPIR-V program creation only available in OpenCL 2.1 and higher")
n = 50000
a_dev = cl.clrandom.rand(queue, n, np.float32)
b_dev = cl.clrandom.rand(queue, n, np.float32)
dest_dev = cl_array.empty_like(a_dev)
with open("add-vectors-%d.spv" % queue.device.address_bits, "rb") as spv_file:
spv = spv_file.read()
prg = cl.Program(ctx, spv)
prg.sum(queue, a_dev.shape, None, a_dev.data, b_dev.data, dest_dev.data)
assert la.norm((dest_dev - (a_dev+b_dev)).get()) < 1e-7
def test_platform_get_devices(ctx_factory):
ctx = ctx_factory()
platform = ctx.devices[0].platform
if platform.name == "Apple":
pytest.xfail("Apple doesn't understand all the values we pass "
"for dev_type")
dev_types = [
cl.device_type.ACCELERATOR, cl.device_type.ALL, cl.device_type.CPU,
cl.device_type.DEFAULT, cl.device_type.GPU
]
if (platform._get_cl_version() >= (1, 2) and cl.get_cl_header_version() >=
(1, 2) and not platform.name.lower().startswith("nvidia")):
dev_types.append(cl.device_type.CUSTOM)
for dev_type in dev_types:
print(dev_type)
devs = platform.get_devices(dev_type)
if dev_type in (cl.device_type.DEFAULT, cl.device_type.ALL,
getattr(cl.device_type, 'CUSTOM', None)):
continue
for dev in devs:
assert dev.type & dev_type == dev_type
def __init__(self, kernel_file):
if self.layers + 1 != len(self.layer_height):
print("Bad network config.")
exit()
print("Running with {} hidden layers and {} layers total.".format(
self.hidden_layers, self.layers))
print("OpenCL Version v{}".format(".".join(
[str(i) for i in cl.get_cl_header_version()])))
print("Finding platform....")
platform = self.findPlatform(VENDOR_NAME)
if not platform:
print("ERROR: Platform not found for name {0}".format(VENDOR_NAME))
exit(1)
print("Getting devices...")
devices = platform.get_devices(device_type=DEVICE_TYPE)
if len(devices) < 1:
print("ERROR: No device found for type {0}.".format(DEVICE_TYPE))
exit(1)
devices = [devices[1]]
self.ctx = cl.Context(devices=devices)
if DEVICE_TYPE == cl.device_type.ACCELERATOR:
print("Reading binary...")
binary = kernel_file.read()
binaries = [binary] * len(devices)
print("Building...")
program = cl.Program(self.ctx, devices, binaries)
else:
print("Reading program...")
binary = kernel_file.read()
program = cl.Program(self.ctx, binary.decode('utf-8')).build()
self.kForward = program.forward
self.kForwardSoftMax = program.forward_softmax
# self.kBackwardFirstDelta = program.backward_first_delta
# self.kBackward = program.backward
self.kForward.set_scalar_arg_dtypes(
[None, None, None, None, np.int32, np.int32, np.int32, np.int32])
self.kForwardSoftMax.set_scalar_arg_dtypes([None, np.int32, np.int32])
# self.kBackwardFirstDelta.set_scalar_arg_dtypes([None, None, None, np.int32, np.int32])
# self.kBackward.set_scalar_arg_dtypes(
# [None, None, None, None, NN_T, NN_T, np.int32, np.int32, np.int32])
self.queue = cl.CommandQueue(self.ctx)
print("Loading data...")
_, (self.x_test, self.y_test) = input_data.load_data()
self.y_test = self.y_test.reshape((10000, ))
self.x_test = self.x_test.reshape(10000, self.layer_height[0])
self.x_test = self.x_test.astype('float32')
self.x_test /= 255
self.correct_pred_fpga = 0
self.wrong_pred_fpga = 0
self.correct_pred_cpu = 0
self.wrong_pred_cpu = 0
def log_version_info():
log.info("PyOpenCL loaded, header version: %s, GL support: %s",
".".join([str(x) for x in pyopencl.get_cl_header_version()]), pyopencl.have_gl())
def test_get_info(ctx_factory):
ctx = ctx_factory()
device, = ctx.devices
platform = device.platform
failure_count = [0]
pocl_quirks = [
(cl.Buffer, cl.mem_info.OFFSET),
(cl.Program, cl.program_info.BINARIES),
(cl.Program, cl.program_info.BINARY_SIZES),
]
if ctx._get_cl_version() >= (1, 2) and cl.get_cl_header_version() >= (1,
2):
pocl_quirks.extend([
(cl.Program, cl.program_info.KERNEL_NAMES),
(cl.Program, cl.program_info.NUM_KERNELS),
])
CRASH_QUIRKS = [ # noqa
(("NVIDIA Corporation", "NVIDIA CUDA", "OpenCL 1.0 CUDA 3.0.1"), [
(cl.Event, cl.event_info.COMMAND_QUEUE),
]),
(("NVIDIA Corporation", "NVIDIA CUDA", "OpenCL 1.2 CUDA 7.5"), [
(cl.Buffer, getattr(cl.mem_info, "USES_SVM_POINTER", None)),
]),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.8-pre"), pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.8"), pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.9-pre"), pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.9"), pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.10-pre"), pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.10"), pocl_quirks),
(("Apple", "Apple", "OpenCL 1.2"), [
(cl.Program, cl.program_info.SOURCE),
]),
]
QUIRKS = [] # noqa
def find_quirk(quirk_list, cl_obj, info):
for (vendor, name, version), quirks in quirk_list:
if (vendor == platform.vendor and name == platform.name
and platform.version.startswith(version)):
for quirk_cls, quirk_info in quirks:
if (isinstance(cl_obj, quirk_cls) and quirk_info == info):
return True
return False
def do_test(cl_obj, info_cls, func=None, try_attr_form=True):
if func is None:
def func(info):
cl_obj.get_info(info)
for info_name in dir(info_cls):
if not info_name.startswith("_") and info_name != "to_string":
print(info_cls, info_name)
info = getattr(info_cls, info_name)
if find_quirk(CRASH_QUIRKS, cl_obj, info):
print("not executing get_info", type(cl_obj), info_name)
print("(known crash quirk for %s)" % platform.name)
continue
try:
func(info)
except:
msg = "failed get_info", type(cl_obj), info_name
if find_quirk(QUIRKS, cl_obj, info):
msg += ("(known quirk for %s)" % platform.name)
else:
failure_count[0] += 1
if try_attr_form:
try:
getattr(cl_obj, info_name.lower())
except:
print("failed attr-based get_info", type(cl_obj),
info_name)
if find_quirk(QUIRKS, cl_obj, info):
print("(known quirk for %s)" % platform.name)
else:
failure_count[0] += 1
do_test(platform, cl.platform_info)
do_test(device, cl.device_info)
do_test(ctx, cl.context_info)
props = 0
if (device.queue_properties
& cl.command_queue_properties.PROFILING_ENABLE):
profiling = True
props = cl.command_queue_properties.PROFILING_ENABLE
queue = cl.CommandQueue(ctx, properties=props)
do_test(queue, cl.command_queue_info)
prg = cl.Program(
ctx, """
__kernel void sum(__global float *a)
{ a[get_global_id(0)] *= 2; }
""").build()
do_test(prg, cl.program_info)
do_test(prg,
cl.program_build_info,
lambda info: prg.get_build_info(device, info),
try_attr_form=False)
n = 2000
a_buf = cl.Buffer(ctx, 0, n * 4)
do_test(a_buf, cl.mem_info)
kernel = prg.sum
do_test(kernel, cl.kernel_info)
evt = kernel(queue, (n, ), None, a_buf)
do_test(evt, cl.event_info)
if profiling:
evt.wait()
do_test(evt,
cl.profiling_info,
lambda info: evt.get_profiling_info(info),
try_attr_form=False)
# crashes on intel...
# and pocl does not support CL_ADDRESS_CLAMP
if device.image_support and platform.vendor not in [
"Intel(R) Corporation",
"The pocl project",
]:
smp = cl.Sampler(ctx, False, cl.addressing_mode.CLAMP,
cl.filter_mode.NEAREST)
do_test(smp, cl.sampler_info)
img_format = cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D)[0]
img = cl.Image(ctx, cl.mem_flags.READ_ONLY, img_format, (128, 256))
assert img.shape == (128, 256)
img.depth
img.image.depth
do_test(img, cl.image_info, lambda info: img.get_image_info(info))
def test_get_info(ctx_factory):
ctx = ctx_factory()
device, = ctx.devices
platform = device.platform
failure_count = [0]
pocl_quirks = [
(cl.Buffer, cl.mem_info.OFFSET),
(cl.Program, cl.program_info.BINARIES),
(cl.Program, cl.program_info.BINARY_SIZES),
]
if ctx._get_cl_version() >= (1, 2) and cl.get_cl_header_version() >= (1, 2):
pocl_quirks.extend([
(cl.Program, cl.program_info.KERNEL_NAMES),
(cl.Program, cl.program_info.NUM_KERNELS),
])
CRASH_QUIRKS = [ # noqa
(("NVIDIA Corporation", "NVIDIA CUDA",
"OpenCL 1.0 CUDA 3.0.1"),
[
(cl.Event, cl.event_info.COMMAND_QUEUE),
]),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.8-pre"),
pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.8"),
pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.9-pre"),
pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.9"),
pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.10-pre"),
pocl_quirks),
(("The pocl project", "Portable Computing Language",
"OpenCL 1.2 pocl 0.10"),
pocl_quirks),
(("Apple", "Apple",
"OpenCL 1.2"),
[
(cl.Program, cl.program_info.SOURCE),
]),
]
QUIRKS = [] # noqa
def find_quirk(quirk_list, cl_obj, info):
for (vendor, name, version), quirks in quirk_list:
if (
vendor == platform.vendor
and name == platform.name
and platform.version.startswith(version)):
for quirk_cls, quirk_info in quirks:
if (isinstance(cl_obj, quirk_cls)
and quirk_info == info):
return True
return False
def do_test(cl_obj, info_cls, func=None, try_attr_form=True):
if func is None:
def func(info):
cl_obj.get_info(info)
for info_name in dir(info_cls):
if not info_name.startswith("_") and info_name != "to_string":
print(info_cls, info_name)
info = getattr(info_cls, info_name)
if find_quirk(CRASH_QUIRKS, cl_obj, info):
print("not executing get_info", type(cl_obj), info_name)
print("(known crash quirk for %s)" % platform.name)
continue
try:
func(info)
except:
msg = "failed get_info", type(cl_obj), info_name
if find_quirk(QUIRKS, cl_obj, info):
msg += ("(known quirk for %s)" % platform.name)
else:
failure_count[0] += 1
if try_attr_form:
try:
getattr(cl_obj, info_name.lower())
except:
print("failed attr-based get_info", type(cl_obj), info_name)
if find_quirk(QUIRKS, cl_obj, info):
print("(known quirk for %s)" % platform.name)
else:
failure_count[0] += 1
do_test(platform, cl.platform_info)
do_test(device, cl.device_info)
do_test(ctx, cl.context_info)
props = 0
if (device.queue_properties
& cl.command_queue_properties.PROFILING_ENABLE):
profiling = True
props = cl.command_queue_properties.PROFILING_ENABLE
queue = cl.CommandQueue(ctx,
properties=props)
do_test(queue, cl.command_queue_info)
prg = cl.Program(ctx, """
__kernel void sum(__global float *a)
{ a[get_global_id(0)] *= 2; }
""").build()
do_test(prg, cl.program_info)
do_test(prg, cl.program_build_info,
lambda info: prg.get_build_info(device, info),
try_attr_form=False)
n = 2000
a_buf = cl.Buffer(ctx, 0, n*4)
do_test(a_buf, cl.mem_info)
kernel = prg.sum
do_test(kernel, cl.kernel_info)
evt = kernel(queue, (n,), None, a_buf)
do_test(evt, cl.event_info)
if profiling:
evt.wait()
do_test(evt, cl.profiling_info,
lambda info: evt.get_profiling_info(info),
try_attr_form=False)
# crashes on intel...
# and pocl does not support CL_ADDRESS_CLAMP
if device.image_support and platform.vendor not in [
"Intel(R) Corporation",
"The pocl project",
]:
smp = cl.Sampler(ctx, False,
cl.addressing_mode.CLAMP,
cl.filter_mode.NEAREST)
do_test(smp, cl.sampler_info)
img_format = cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D)[0]
img = cl.Image(ctx, cl.mem_flags.READ_ONLY, img_format, (128, 256))
assert img.shape == (128, 256)
img.depth
img.image.depth
do_test(img, cl.image_info,
lambda info: img.get_image_info(info))
'''
Prints relevant information regarding the capabilities of the current OpenCL runtime and devices
Note that pyopencl has a script that prints all properties in its examples folder
'''
import pyopencl as cl
print('PyOpenCL version: ' + cl.VERSION_TEXT)
print('OpenCL header version: ' + '.'.join(map(str, cl.get_cl_header_version())) + '\n')
# Get installed platforms (SDKs)
print('- Installed platforms (SDKs) and available devices:')
platforms = cl.get_platforms()
for plat in platforms:
indent = ''
# Get and print platform info
print(indent + '{} ({})'.format(plat.name, plat.vendor))
indent = '\t'
print(indent + 'Version: ' + plat.version)
print(indent + 'Profile: ' + plat.profile)
print(indent + 'Extensions: ' + str(plat.extensions.strip().split(' ')))
# Get and print device info
devices = plat.get_devices(cl.device_type.ALL)
print(indent + 'Available devices: ')
if not devices:
print(indent + '\tNone')
from npsolve import runner
from math import ceil, log2
pp = pprint.PrettyPrinter(depth=5)
mf = cl.mem_flags
def dec2str(num):
k = []
s = str(num)
for a in s:
k.append(ascii_lowercase[int(a)])
return 'qq' + ''.join(k)
print(cl.get_cl_header_version())
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
set = 'random'
traindata = idxs("train-images-idx3-ubyte.idx", "train-labels-idx1-ubyte.idx")
testdata = idxs("t10k-images-idx3-ubyte.idx", "t10k-labels-idx1-ubyte.idx")
result = 1.0
ninpt = traindata.count #Samples count ( 60000 for set )
nvarsd = traindata.rows * traindata.cols #Count of equations members ( 28*28 for set)
topology = [nvarsd, 5, 4, 3, 1]
nvarsg = genn.countcns(topology) #Count of equations members
print("Total connections is", nvarsg)
nsamp = 64 #ctx.get_info(cl.context_info.DEVICES)[0].max_work_group_size #Genome samples count (current sort limitation to local_size)
print("Population count is", nsamp)
def ensure_required_version():
ver = cl.get_cl_header_version()
assert ver[0] >= 2 and ver[1] >= 0, "OpenCL must be version 2.0 or greater."
import pprint
from idxread import idxs
from npsolve import runner
from math import ceil, log2
pp = pprint.PrettyPrinter(depth=5)
mf = cl.mem_flags
def dec2str(num):
k = []
s = str(num)
for a in s:
k.append(ascii_lowercase[int(a)])
return 'qq'+''.join(k)
print( cl.get_cl_header_version() )
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
set = 'random'
traindata = idxs("train-images-idx3-ubyte.idx", "train-labels-idx1-ubyte.idx")
testdata = idxs("t10k-images-idx3-ubyte.idx", "t10k-labels-idx1-ubyte.idx")
result = 1.0
ninpt = traindata.count #Samples count ( 60000 for set )
nvarsd = traindata.rows*traindata.cols #Count of equations members ( 28*28 for set)
topology = [nvarsd, 5, 4, 3, 1]
nvarsg = genn.countcns(topology) #Count of equations members
print("Total connections is", nvarsg)
nsamp = 64#ctx.get_info(cl.context_info.DEVICES)[0].max_work_group_size #Genome samples count (current sort limitation to local_size)
print("Population count is", nsamp)
def pyOpenCLInfo(self, output_info=True):
if (output_info):
print('PyOpenCL version: ' + cl.VERSION_TEXT)
print('OpenCL header version: ' +
'.'.join(map(str, cl.get_cl_header_version())) + '\n')
# Get installed platforms (SDKs)
print('- Installed platforms (SDKs) and available devices:')
platforms = cl.get_platforms()
for plat in platforms:
indent = ''
# Get and print platform info
print(indent + '{} ({})'.format(plat.name, plat.vendor))
indent = '\t'
print(indent + 'Version: ' + plat.version)
print(indent + 'Profile: ' + plat.profile)
print(indent + 'Extensions: ' +
str(plat.extensions.strip().split(' ')))
# Get and print device info
devices = plat.get_devices(cl.device_type.ALL)
print(indent + 'Available devices: ')
if not devices:
print(indent + '\tNone')
for dev in devices:
indent = '\t\t'
print(indent + '{} ({})'.format(dev.name, dev.vendor))
indent = '\t\t\t'
flags = [('Version', dev.version),
('Type', cl.device_type.to_string(dev.type)),
('Extensions',
str(dev.extensions.strip().split(' '))),
('Memory (global)', str(dev.global_mem_size)),
('Memory (local)', str(dev.local_mem_size)),
('Address bits', str(dev.address_bits)),
('Max work item dims',
str(dev.max_work_item_dimensions)),
('Max work group size',
str(dev.max_work_group_size)),
('Max compute units', str(dev.max_compute_units)),
('Driver version', dev.driver_version),
('Image support', str(bool(dev.image_support))),
('Little endian', str(bool(dev.endian_little))),
('Device available', str(bool(dev.available))),
('Compiler available',
str(bool(dev.compiler_available)))]
[
print(indent +
'{0:<25}{1:<10}'.format(name + ':', flag))
for name, flag in flags
]
# Device version string has the following syntax, extract the number like this
# OpenCL<space><major_version.minor_version><space><vendor-specific information>
version_number = float(dev.version.split(' ')[1])
print('')
else:
platform = cl.get_platforms()[self.platform_id]
device = platform.get_devices(cl.device_type.ALL)[0]
self.recommend_CU = device.max_compute_units
import pyopencl as cl
print("CL_VERSION:", cl.VERSION)
print("CL_HEADER_VERSION:", cl.get_cl_header_version())
print()
platforms = cl.get_platforms()
print("Platform num:", len(platforms))
for plat in platforms:
print("--Platform Name:", plat.get_info(cl.platform_info.NAME))
# print("--Platform Extensions:",plat.get_info(cl.platform_info.EXTENSIONS))
print("--Platform Profile:", plat.get_info(cl.platform_info.PROFILE))
print("--Platform Vendor:", plat.get_info(cl.platform_info.VENDOR))
print("--Platform Version:", plat.get_info(cl.platform_info.VERSION))
devices = plat.get_devices(cl.device_type.ALL)
print("--device num:", len(devices))
for device in devices:
print("----Name:", device.get_info(cl.device_info.NAME))
print("----OpenCL_C_Version:",
device.get_info(cl.device_info.OPENCL_C_VERSION))
print("----Vendor:", device.get_info(cl.device_info.VENDOR))
print("----Version:", device.get_info(cl.device_info.VERSION))
print("----Driver Version:",
device.get_info(cl.device_info.DRIVER_VERSION))
print("----MAX_WORK_GROUP_SIZE:",
device.get_info(cl.device_info.MAX_WORK_GROUP_SIZE))
print("----MAX_COMPUTE_UNITS:",