Python pyopencl.Buffer方法代碼示例

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def ready_argument_list(self, arguments):
        """ready argument list to be passed to the kernel, allocates gpu mem

        :param arguments: List of arguments to be passed to the kernel.
            The order should match the argument list on the OpenCL kernel.
            Allowed values are numpy.ndarray, and/or numpy.int32, numpy.float32, and so on.
        :type arguments: list(numpy objects)

        :returns: A list of arguments that can be passed to an OpenCL kernel.
        :rtype: list( pyopencl.Buffer, numpy.int32, ... )
        """
        gpu_args = []
        for arg in arguments:
            # if arg i is a numpy array copy to device
            if isinstance(arg, numpy.ndarray):
                gpu_args.append(cl.Buffer(self.ctx, self.mf.READ_WRITE | self.mf.COPY_HOST_PTR, hostbuf=arg))
            else: # if not an array, just pass argument along
                gpu_args.append(arg)
        return gpu_args 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def run_kernel(self, func, gpu_args, threads, grid):
        """runs the OpenCL kernel passed as 'func'

        :param func: An OpenCL Kernel
        :type func: pyopencl.Kernel

        :param gpu_args: A list of arguments to the kernel, order should match the
            order in the code. Allowed values are either variables in global memory
            or single values passed by value.
        :type gpu_args: list( pyopencl.Buffer, numpy.int32, ...)

        :param threads: A tuple listing the number of work items in each dimension of
            the work group.
        :type threads: tuple(int, int, int)

        :param grid: A tuple listing the number of work groups in each dimension
            of the NDRange.
        :type grid: tuple(int, int)
        """
        global_size = (grid[0]*threads[0], grid[1]*threads[1], grid[2]*threads[2])
        local_size = threads
        event = func(self.queue, global_size, local_size, *gpu_args)
        event.wait() 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def memset(self, buffer, value, size):
        """set the memory in allocation to the value in value

        :param allocation: An OpenCL Buffer to fill
        :type allocation: pyopencl.Buffer

        :param value: The value to set the memory to
        :type value: a single 32-bit int

        :param size: The size of to the allocation unit in bytes
        :type size: int

        """
        if isinstance(buffer, cl.Buffer):
            try:
                cl.enqueue_fill_buffer(self.queue, buffer, numpy.uint32(value), 0, size)
            except AttributeError:
                src=numpy.zeros(size, dtype='uint8')+numpy.uint8(value)
                cl.enqueue_copy(self.queue, buffer, src) 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def test_ready_argument_list():

    size = 1000
    a = np.int32(75)
    b = np.random.randn(size).astype(np.float32)
    c = np.zeros_like(b)

    arguments = [c, a, b]

    dev = opencl.OpenCLFunctions(0)
    gpu_args = dev.ready_argument_list(arguments)

    assert isinstance(gpu_args[0], pyopencl.Buffer)
    assert isinstance(gpu_args[1], np.int32)
    assert isinstance(gpu_args[2], pyopencl.Buffer)

    gpu_args[0].release()
    gpu_args[2].release() 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def create_internal_buffer(self, ctx):
        cityxy = [(self.city_info[idx][0], self.city_info[idx][1]) for idx in range(len(self.city_info))]
        self.__np_cityxy = numpy.array(cityxy, dtype=numpy.float32)

        mf = cl.mem_flags

        self.__dev_cityxy = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                      hostbuf=self.__np_cityxy)

        self.__dev_cityxy = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                      hostbuf=self.__np_cityxy)

        self.__np_iterations = numpy.int32(self.iterations)
        self.__np_temperature = numpy.float32(self.temperature)
        self.__np_terminate_temperature = numpy.float32(self.terminate_temperature)
        self.__np_alpha = numpy.float32(self.alpha) 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def __init_cl_member(self):
        self.__np_costs = self.sas.get_cost_buffer()
        num_of_solution, self.__np_solution = self.sas.get_solution_info()

        mf = cl.mem_flags

        # Random number should be given by Host program because OpenCL doesn't have a random number
        # generator. We just include one, Noise.cl.
        rnum = [random.randint(0, 4294967295) for i in range(num_of_solution)]
        ## note: numpy.random.rand() gives us a list float32 and we cast it to uint32 at the calling
        ##       of kernel function. It just views the original byte order as uint32.
        self.__dev_rnum = cl.Buffer(self.__ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                    hostbuf=numpy.array(rnum, dtype=numpy.uint32))

        self.__dev_costs = cl.Buffer(self.__ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                     hostbuf=self.__np_costs)
        self.__dev_solution = cl.Buffer(self.__ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                        hostbuf=self.__np_solution)

        self.sas.create_internal_buffer(self.__ctx)

    ## To save the annealing state 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def __prepare_fitness_args(self):
        mf = cl.mem_flags
        self.__fitness_args_list = [self.__dev_chromosomes, self.__dev_fitnesses]

        self.__extra_fitness_args_list = []

        if self.__fitness_args is not None:
            ## create buffers for fitness arguments
            for arg in self.__fitness_args:
                cl_buffer = cl.Buffer(self.__ctx,
                                mf.READ_WRITE | mf.COPY_HOST_PTR,
                                hostbuf=numpy.array(arg['v'],
                                dtype=self.__type_to_numpy_type(arg['t'])))
                self.__extra_fitness_args_list.append(cl_buffer)
        # concatenate two fitness args list
        self.__fitness_args_list = self.__fitness_args_list + self.__extra_fitness_args_list 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def benchmark(self, func, gpu_args, threads, grid):
        """runs the kernel and measures time repeatedly, returns average time

        Runs the kernel and measures kernel execution time repeatedly, number of
        iterations is set during the creation of OpenCLFunctions. Benchmark returns
        a robust average, from all measurements the fastest and slowest runs are
        discarded and the rest is included in the returned average. The reason for
        this is to be robust against initialization artifacts and other exceptional
        cases.

        :param func: A PyOpenCL kernel compiled for this specific kernel configuration
        :type func: pyopencl.Kernel

        :param gpu_args: A list of arguments to the kernel, order should match the
            order in the code. Allowed values are either variables in global memory
            or single values passed by value.
        :type gpu_args: list( pyopencl.Buffer, numpy.int32, ...)

        :param threads: A tuple listing the number of work items in each dimension of
            the work group.
        :type threads: tuple(int, int, int)

        :param grid: A tuple listing the number of work groups in each dimension
            of the NDRange.
        :type grid: tuple(int, int)

        :returns: All benchmark results.
        :rtype: dict()
        """
        result = dict()
        result["times"] = []
        global_size = (grid[0]*threads[0], grid[1]*threads[1], grid[2]*threads[2])
        local_size = threads
        time = []
        for _ in range(self.iterations):
            event = func(self.queue, global_size, local_size, *gpu_args)
            event.wait()
            result["times"].append((event.profile.end - event.profile.start)*1e-6)
        result["time"] = numpy.mean(result["times"])
        return result 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def memcpy_dtoh(self, dest, src):
        """perform a device to host memory copy

        :param dest: A numpy array in host memory to store the data
        :type dest: numpy.ndarray

        :param src: An OpenCL Buffer to copy data from
        :type src: pyopencl.Buffer
        """
        if isinstance(src, cl.Buffer):
            cl.enqueue_copy(self.queue, dest, src) 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def change_params(self, cl_platform_index=None, cl_device_index=None, **kargs):
        BasePeakDetector.change_params(self,  **kargs)
        OpenCL_Helper.initialize_opencl(self, cl_platform_index=cl_platform_index, cl_device_index=cl_device_index)

        if self.chunksize > self.max_wg_size:
            n = int(np.ceil(self.chunksize / self.max_wg_size))
            self.global_size =  (self.max_wg_size * n, )
            self.local_size = (self.max_wg_size,)
        else:
            self.global_size = (self.chunksize, )
            self.local_size = (self.chunksize, )


        chunksize2=self.chunksize+2*self.n_span
        
        self.sum_rectified = np.zeros((self.chunksize), dtype=self.dtype)
        self.peak_mask = np.zeros((self.chunksize), dtype='uint8')
        ring_sum = np.zeros((chunksize2), dtype=self.dtype)
        
        #GPU buffers
        self.sigs_cl = pyopencl.Buffer(self.ctx, mf.READ_WRITE, size=self.nb_channel*self.chunksize*self.dtype.itemsize)
        
        self.ring_sum_cl = pyopencl.Buffer(self.ctx, mf.READ_WRITE| mf.COPY_HOST_PTR, hostbuf=ring_sum)
        self.peak_mask_cl = pyopencl.Buffer(self.ctx, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=self.peak_mask)
        
        kernel = self.kernel%dict(chunksize=self.chunksize, nb_channel=self.nb_channel, n_span=self.n_span,
                    relative_threshold=self.relative_threshold, peak_sign={'+':1, '-':-1}[self.peak_sign])
        
        prg = pyopencl.Program(self.ctx, kernel)
        self.opencl_prg = prg.build(options='-cl-mad-enable')
        
        self.kern_detect_peaks = getattr(self.opencl_prg, 'detect_peaks') 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def _make_gpu_buffer(self):
        # TODO fifo size should be : chunksize+2*self.n_span
        self.fifo_size = self.chunksize + 2*self.n_span
        self.fifo_sigs = FifoBuffer((self.fifo_size, self.nb_channel), self.dtype)
        self.mask_peaks = np.zeros((self.chunksize, self.nb_channel), dtype='uint8')  # bool
        
        #GPU buffers
        self.fifo_sigs_cl = pyopencl.Buffer(self.ctx, mf.READ_WRITE| mf.COPY_HOST_PTR, hostbuf=self.fifo_sigs.buffer)
        self.neighbours_cl = pyopencl.Buffer(self.ctx, mf.READ_WRITE| mf.COPY_HOST_PTR, hostbuf=self.neighbours)
        self.mask_peaks_cl = pyopencl.Buffer(self.ctx, mf.READ_WRITE| mf.COPY_HOST_PTR, hostbuf=self.mask_peaks) 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def printfullinfo(self):
        print('\n' + '=' * 60 + '\nOpenCL Platforms and Devices')
        i=0
        for platform in cl.get_platforms():
            print('=' * 60)
            print('Platform %d - Name: ' %i + platform.name)
            print('Platform %d - Vendor: ' %i + platform.vendor)
            print('Platform %d - Version: ' %i + platform.version)
            print('Platform %d - Profile: ' %i + platform.profile)

            for device in platform.get_devices():
                print(' ' + '-' * 56)
                print(' Device - Name: ' \
                      + device.name)
                print(' Device - Type: ' \
                      + cl.device_type.to_string(device.type))
                print(' Device - Max Clock Speed: {0} Mhz' \
                      .format(device.max_clock_frequency))
                print(' Device - Compute Units: {0}' \
                      .format(device.max_compute_units))
                print(' Device - Local Memory: {0:.0f} KB' \
                      .format(device.local_mem_size / 1024.0))
                print(' Device - Constant Memory: {0:.0f} KB' \
                      .format(device.max_constant_buffer_size / 1024.0))
                print(' Device - Global Memory: {0:.0f} GB' \
                      .format(device.global_mem_size / 1073741824.0))
                print(' Device - Max Buffer/Image Size: {0:.0f} MB' \
                      .format(device.max_mem_alloc_size / 1048576.0))
                print(' Device - Max Work Group Size: {0:.0f}' \
                      .format(device.max_work_group_size))
                print('\n')
        i+=1 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def test_gpu_vector_sum(a, b):
    #define the PyOpenCL Context
    platform = cl.get_platforms()[0]
    device = platform.get_devices()[0]
    context = cl.Context([device])
    queue = cl.CommandQueue(context, \
                            properties=cl.command_queue_properties.PROFILING_ENABLE)   
    #prepare the data structure
    a_buffer = cl.Buffer\
               (context, cl.mem_flags.READ_ONLY | cl.mem_flags.COPY_HOST_PTR, hostbuf=a)
    b_buffer = cl.Buffer\
               (context, cl.mem_flags.READ_ONLY | cl.mem_flags.COPY_HOST_PTR, hostbuf=b)
    c_buffer = cl.Buffer\
               (context, cl.mem_flags.WRITE_ONLY, b.nbytes)   
    program = cl.Program(context, """
    __kernel void sum(__global const float *a, __global const float *b, __global float *c)
    {
        int i = get_global_id(0);
        int j;
        for(j = 0; j < 10000; j++)
        {
            c[i] = a[i] + b[i];
        }
    }""").build()
    #start the gpu test
    gpu_start_time = time()   
    event = program.sum(queue, a.shape, None, a_buffer, b_buffer, c_buffer)   
    event.wait()   
    elapsed = 1e-9*(event.profile.end - event.profile.start)   
    print("GPU Kernel evaluation Time: {0} s".format(elapsed))   
    c_gpu = np.empty_like(a)  
    cl._enqueue_read_buffer(queue, c_buffer, c_gpu).wait()  
    gpu_end_time = time()  
    print("GPU Time: {0} s".format(gpu_end_time - gpu_start_time))   
    return c_gpu   

#start the test 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def create_bytearray(ctx, size):
    mf = cl.mem_flags
    py_buffer = numpy.zeros(size, dtype=numpy.int32)
    cl_buffer = cl.Buffer(ctx,
                          mf.READ_WRITE | mf.COPY_HOST_PTR,
                          hostbuf=py_buffer)
    return py_buffer, cl_buffer 

# 需要導入模塊: import pyopencl [as 別名]
# 或者: from pyopencl import Buffer [as 別名]
def preexecute_kernels(self, ctx, queue, population):
        ## initialize global variables for kernel execution
        total_dna_size = population * self.dna_total_length

        other_chromosomes = numpy.zeros(total_dna_size, dtype=numpy.int32)
        cross_map = numpy.zeros(total_dna_size, dtype=numpy.int32)
        ratios = numpy.zeros(population, dtype=numpy.float32)

        mf = cl.mem_flags

        self.__dev_ratios = cl.Buffer(ctx, mf.WRITE_ONLY, ratios.nbytes)
        self.__dev_other_chromosomes = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                                 hostbuf=other_chromosomes)
        self.__dev_cross_map = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR,
                                         hostbuf=cross_map)