def test_enqueue_task(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
prg = cl.Program(
ctx, """
__kernel void
reverse(__global const float *in, __global float *out, int n)
{
for (int i = 0;i < n;i++) {
out[i] = in[n - 1 - i];
}
}
""").build()
knl = prg.reverse
n = 100
a = np.random.rand(n).astype(np.float32)
b = np.empty_like(a)
buf1 = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a)
buf2 = cl.Buffer(ctx, mf.WRITE_ONLY, b.nbytes)
knl.set_args(buf1, buf2, np.int32(n))
cl.enqueue_task(queue, knl)
cl.enqueue_copy(queue, b, buf2).wait()
assert la.norm(a[::-1] - b) == 0
def test_enqueue_task(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
prg = cl.Program(ctx, """
__kernel void
reverse(__global const float *in, __global float *out, int n)
{
for (int i = 0;i < n;i++) {
out[i] = in[n - 1 - i];
}
}
""").build()
knl = prg.reverse
n = 100
a = np.random.rand(n).astype(np.float32)
b = np.empty_like(a)
buf1 = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a)
buf2 = cl.Buffer(ctx, mf.WRITE_ONLY, b.nbytes)
knl.set_args(buf1, buf2, np.int32(n))
cl.enqueue_task(queue, knl)
cl.enqueue_copy(queue, b, buf2).wait()
assert la.norm(a[::-1] - b) == 0
def fpga_function(self, benchmark=False):
for layer in range(0, self.layers):
rows_in = self.layer_height[layer]
rows_out = self.layer_height[layer + 1]
is_last_layer = layer + 1 == self.layers
print("Running layer {} ({}->{}) kernel, with relu: {}".format(
layer, rows_in, rows_out, 1 if not is_last_layer else 0))
self.kForward.set_args(self.act_buffers[layer],
self.weights_buffers[layer],
self.bias_buffers[layer],
self.act_buffers[layer + 1],
self.minibatch_size, rows_in, rows_out,
1 if not is_last_layer else 0)
cl.enqueue_task(self.queue, self.kForward).wait()
if is_last_layer:
print("Running layer {} softmax kernel".format(layer))
self.kForwardSoftMax.set_args(self.act_buffers[layer + 1],
self.minibatch_size, rows_out)
cl.enqueue_task(self.queue, self.kForwardSoftMax).wait()
def bw_function(self, benchmark=False):
for layer in range(self.layers, -1, -1):
rows_in = self.layer_height[layer]
is_last_layer = layer == self.layers
if is_last_layer:
#print("Running layer {} first_delta kernel".format(layer))
self.kBackwardFirstDelta.set_args(self.act_buffers[layer],
self.ground_truth_buffer,
self.delta_buffers[layer],
self.minibatch_size,
self.layer_height[layer])
cl.enqueue_task(self.queue, self.kBackwardFirstDelta).wait()
#print("Running layer {} first_delta kernel. COMPLETED".format(layer))
else:
rows_out = self.layer_height[layer + 1]
#print("Running layer {} backwards kernel".format(layer))
self.kBackward.set_args(
self.act_buffers[layer], self.weights_buffers[layer],
self.dW_buffers[layer], self.bias_buffers[layer],
self.delta_buffers[layer], self.delta_buffers[layer + 1],
self.learn_rate, self.minibatch_size, rows_in, rows_out,
layer)
cl.enqueue_task(self.queue, self.kBackward).wait()
def main():
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
buf = cl.Buffer(ctx, mf.READ_ONLY, 1000)
buf2 = cl.Buffer(ctx, mf.READ_WRITE, 8)
prg = cl.Program(ctx, """
__kernel void
get_addr(__global const int *in, __global long *out)
{
*out = (long)in;
}
""").build()
knl = prg.get_addr
knl.set_args(buf, buf2)
cl.enqueue_task(queue, knl)
b = np.empty([1], dtype=np.int64)
cl.enqueue_copy(queue, b, buf2).wait()
print(b[0])
prg = cl.Program(ctx, """
__kernel void
get_addr(__global const int *in, __global long *out)
{
*out = (long)in;
}
""").build()
knl = prg.get_addr
knl.set_args(buf, buf2)
cl.enqueue_task(queue, knl)
b = np.empty([1], dtype=np.int64)
cl.enqueue_copy(queue, b, buf2).wait()
print(b[0])
context = cl.Context(devices=[dev])
queue = cl.CommandQueue(context, dev)
# Build program in the specified context using the kernel source code
prog = cl.Program(context, kernel_src)
try:
prog.build(options=['-Werror'], devices=[dev])
except:
print('Build log:')
print(prog.get_build_info(dev, cl.program_build_info.LOG))
raise
# Data and device buffers
data = np.arange(start=0, stop=NUM_SHORTS, dtype=np.uint16)
np.random.shuffle(data)
print('Input: ' + str(data))
mf = cl.mem_flags
data_buffer = cl.Buffer(context,
mf.READ_WRITE | mf.COPY_HOST_PTR,
hostbuf=data)
# Execute kernel
# radix_sort8(__global ushort8 *global_data)
kernel = prog.radix_sort8
kernel.set_arg(0, data_buffer)
cl.enqueue_task(queue, kernel)
cl.enqueue_copy(queue, dest=data, src=data_buffer, is_blocking=True)
print('Output: ' + str(data))
def f(t, y_in, y_out, wait_for=None):
knl.set_args(y_in.base_data, y_out.base_data)
return cl.enqueue_task(queue, knl, wait_for=wait_for + [start_evt])
def _post_kernel_reduction_task(self, nelems, reduction_operator):
assert reduction_operator in [INC, MIN, MAX]
def generate_code():
def headers():
if self.dtype == np.dtype('float64'):
return """
#if defined(cl_khr_fp64)
#if defined(cl_amd_fp64)
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#endif
#elif defined(cl_amd_fp64)
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
"""
else:
return ""
op = {INC: 'INC', MIN: 'min', MAX: 'max'}[reduction_operator]
return """
%(headers)s
#define INC(a,b) ((a)+(b))
__kernel
void global_%(type)s_%(dim)s_post_reduction (
__global %(type)s* dat,
__global %(type)s* tmp,
__private int count
)
{
__private %(type)s accumulator[%(dim)d];
for (int j = 0; j < %(dim)d; ++j)
{
accumulator[j] = dat[j];
}
for (int i = 0; i < count; ++i)
{
for (int j = 0; j < %(dim)d; ++j)
{
accumulator[j] = %(op)s(accumulator[j], *(tmp + i * %(dim)d + j));
}
}
for (int j = 0; j < %(dim)d; ++j)
{
dat[j] = accumulator[j];
}
}
""" % {'headers': headers(), 'dim': self.cdim, 'type': self._cl_type, 'op': op}
src, kernel = _reduction_task_cache.get(
(self.dtype, self.cdim, reduction_operator), (None, None))
if src is None:
src = generate_code()
prg = cl.Program(_ctx, src).build(options="-Werror")
name = "global_%s_%s_post_reduction" % (self._cl_type, self.cdim)
kernel = prg.__getattr__(name)
_reduction_task_cache[
(self.dtype, self.cdim, reduction_operator)] = (src, kernel)
kernel.set_arg(0, self._array.data)
kernel.set_arg(1, self._d_reduc_array.data)
kernel.set_arg(2, np.int32(nelems))
cl.enqueue_task(_queue, kernel).wait()
self._array.get(queue=_queue, ary=self._data)
self.state = DeviceDataMixin.BOTH
del self._d_reduc_array
def _post_kernel_reduction_task(self, nelems, reduction_operator):
assert reduction_operator in [INC, MIN, MAX]
def generate_code():
def headers():
if self.dtype == np.dtype('float64'):
return """
#if defined(cl_khr_fp64)
#if defined(cl_amd_fp64)
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#endif
#elif defined(cl_amd_fp64)
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
"""
else:
return ""
op = {INC: 'INC', MIN: 'min', MAX: 'max'}[reduction_operator]
return """
%(headers)s
#define INC(a,b) ((a)+(b))
__kernel
void global_%(type)s_%(dim)s_post_reduction (
__global %(type)s* dat,
__global %(type)s* tmp,
__private int count
)
{
__private %(type)s accumulator[%(dim)d];
for (int j = 0; j < %(dim)d; ++j)
{
accumulator[j] = dat[j];
}
for (int i = 0; i < count; ++i)
{
for (int j = 0; j < %(dim)d; ++j)
{
accumulator[j] = %(op)s(accumulator[j], *(tmp + i * %(dim)d + j));
}
}
for (int j = 0; j < %(dim)d; ++j)
{
dat[j] = accumulator[j];
}
}
""" % {'headers': headers(), 'dim': self.cdim, 'type': self._cl_type, 'op': op}
src, kernel = _reduction_task_cache.get(
(self.dtype, self.cdim, reduction_operator), (None, None))
if src is None:
src = generate_code()
prg = cl.Program(_ctx, src).build(options="-Werror")
name = "global_%s_%s_post_reduction" % (self._cl_type, self.cdim)
kernel = prg.__getattr__(name)
_reduction_task_cache[
(self.dtype, self.cdim, reduction_operator)] = (src, kernel)
kernel.set_arg(0, self._array.data)
kernel.set_arg(1, self._d_reduc_array.data)
kernel.set_arg(2, np.int32(nelems))
cl.enqueue_task(_queue, kernel).wait()
self._array.get(queue=_queue, ary=self._data)
self.state = DeviceDataMixin.BOTH
del self._d_reduc_array
# Get device and context, create command queue and program
dev = utility.get_default_device()
context = cl.Context(devices=[dev], properties=None, dev_type=None, cache_dir=None)
queue = cl.CommandQueue(context, dev, properties=None)
# Build program in the specified context using the kernel source code
prog = cl.Program(context, kernel_src)
try:
prog.build(options=["-Werror"], devices=[dev], cache_dir=None)
except:
print("Build log:")
print(prog.get_build_info(dev, cl.program_build_info.LOG))
raise
# Data and device buffers
data = np.arange(start=0, stop=NUM_SHORTS, dtype=np.uint16)
np.random.shuffle(data)
print("Input: " + str(data))
mf = cl.mem_flags
data_buffer = cl.Buffer(context, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=data)
# Execute kernel
# radix_sort8(__global ushort8 *global_data)
kernel = prog.radix_sort8
kernel.set_arg(0, data_buffer)
cl.enqueue_task(queue, kernel)
cl.enqueue_copy(queue, dest=data, src=data_buffer, is_blocking=True)
print("Output: " + str(data))