Commit cab18ca6 authored by Yu-Hang "Maxin" Tang's avatar Yu-Hang "Maxin" Tang
Browse files

Merge branch 'release/0.8a7'

parents 3aa4c310 612def1d

CHANGELOG.md

+5 −0
Original line number Diff line number Diff line 
# Change log of GraphDot



## 0.8a7 (2021-01-02)



- Gradient evaluation for the MaxiMin graph metric.

- Gradient evaluation for Gaussian field regressor prediction loss.



## 0.8a6 (2020-12-29)



- Optimized the evaluation of the gradient of the loss function for the

LICENSE

+1 −1
Original line number Diff line number Diff line 
GraphDot Copyright (c) 2019-2020, The Regents of the University of California,

GraphDot Copyright (c) 2019-2021, The Regents of the University of California,

through Lawrence Berkeley National Laboratory (subject to receipt of any

required approvals from the U.S. Dept. of Energy).  All rights reserved.

graphdot/__init__.py

+2 −2
Original line number Diff line number Diff line
@@ -5,14 +5,14 @@ from .graph import Graph 


__all__ = ['Graph']



__version__ = '0.8a6'

__version__ = '0.8a7'

__author__ = '''Yu-Hang "Maxin" Tang, Oguz Selvitopi, Chi-Feng Wang,

    Thomas H. Li'''

__maintainer__ = 'Yu-Hang "Maxin" Tang'

__email__ = 'Tang.Maxin@gmail.com'

__license__ = 'see LICENSE file'

__copyright__ = '''

GraphDot Copyright (c) 2019-2020, The Regents of the University of California,

GraphDot Copyright (c) 2019-2021, The Regents of the University of California,

through Lawrence Berkeley National Laboratory (subject to receipt of any

required approvals from the U.S. Dept. of Energy).  All rights reserved.

graphdot/codegen/sympy_printer.py

+1 −1
Original line number Diff line number Diff line 
#!/usr/bin/env python

# -*- coding: utf-8 -*-

from sympy.codegen import ast

from sympy.printing.cxxcode import CXX11CodePrinter

from sympy.printing.cxx import CXX11CodePrinter





class CUDACXX11CodePrinter(CXX11CodePrinter):

graphdot/kernel/marginalized/_backend_cuda.py

+22 −13
Original line number Diff line number Diff line
@@ -51,6 +51,7 @@ class CUDABackend(Backend): 
        self.ctx = kwargs.pop('cuda_context', graphdot.cuda.defctx)

        self.device = self.ctx.get_device()

        self.scratch_pcg = None

        self.scratch_pcg_d = None



        self.block_per_sm = kwargs.pop('block_per_sm', 8)

        self.block_size = kwargs.pop('block_size', 128)
@@ -74,7 +75,22 @@ class CUDABackend(Backend): 
                'try to normalize automatically with `Graph.normalize_types`.'

            )



    def _allocate_pcg_scratch(self, number, max_graph_size, traits):

    def _allocate_scratch(self, scratch, scratch_d, number, length,

                          n_temporaries):

        if (scratch is None or len(scratch) < number or

                scratch[0].nmax < length or

                scratch[0].ndim < n_temporaries):

            self.ctx.synchronize()

            scratch = [

                PCGScratch(length, n_temporaries) for _ in range(number)

            ]

            scratch_d = umarray(

                np.array([s.state for s in scratch], PCGScratch.dtype)

            )

            self.ctx.synchronize()

        return scratch, scratch_d



    def allocate_pcg_scratch(self, number, max_graph_size, traits):

        if traits.eval_gradient is True:

            if traits.nodal in [True, 'block']:

                length = max_graph_size**2
@@ -86,17 +102,10 @@ class CUDABackend(Backend): 
            length = max_graph_size**2

            n_temporaries = 5



        if (self.scratch_pcg is None or len(self.scratch_pcg) < number or

                self.scratch_pcg[0].nmax < length or

                self.scratch_pcg[0].ndim < n_temporaries):

            self.ctx.synchronize()

            self.scratch_pcg = [

                PCGScratch(length, n_temporaries) for _ in range(number)

            ]

            self.scratch_pcg_d = umarray(

                np.array([s.state for s in self.scratch_pcg], PCGScratch.dtype)

        self.scratch_pcg, self.scratch_pcg_d = self._allocate_scratch(

            self.scratch_pcg, self.scratch_pcg_d, number, length,

            n_temporaries

        )

            self.ctx.synchronize()

        return self.scratch_pcg_d



    def _register_graph(self, graph):
@@ -301,7 +310,7 @@ class CUDABackend(Backend): 


        ''' allocate scratch buffers '''

        max_graph_size = np.max([len(g.nodes) for g in graphs])

        scratch_pcg = self._allocate_pcg_scratch(

        scratch_pcg = self.allocate_pcg_scratch(

            launch_block_count, max_graph_size, traits

        )