(numba/dppy) Merged devel

Author: reazulhoque

CHANGE_LOG

@@ -0,0 +1,7 @@
+NUMBA Version 0.48.0 + DPPY Version 0.3.0 (June 29, 2020)
+--------------------------------------------------------
+
+This release includes:
+
+* Caching of dppy.kernels which will improve performance.
+* Addition of support for Intel Advisor which will help in profiling applications.

codegen.py

@@ -1,8 +1,8 @@
 from llvmlite import binding as ll
 from llvmlite.llvmpy import core as lc
 
-from numba.targets.codegen import BaseCPUCodegen, CodeLibrary
-from numba import utils
+from numba.core.codegen import BaseCPUCodegen, CodeLibrary
+from numba.core import utils
 
 
 SPIR_TRIPLE = {32: ' spir-unknown-unknown', 

compiler.py

@@ -3,23 +3,23 @@
 from collections import namedtuple
 
 from .dppy_passbuilder import DPPyPassBuilder
-from numba.typing.templates import ConcreteTemplate
-from numba import types, compiler, ir
-from numba.typing.templates import AbstractTemplate
-from numba import ctypes_support as ctypes
+from numba.core.typing.templates import ConcreteTemplate
+from numba.core import types, compiler, ir
+from numba.core.typing.templates import AbstractTemplate
+import ctypes
 from types import FunctionType
 from inspect import signature
 
 import dppy.ocldrv as driver
 from . import spirv_generator
 
 import os
-from numba.compiler import DefaultPassBuilder, CompilerBase
+from numba.core.compiler import DefaultPassBuilder, CompilerBase
 
 DEBUG=os.environ.get('NUMBA_DPPY_DEBUG', None)
-_NUMBA_PVC_READ_ONLY  = "read_only"
-_NUMBA_PVC_WRITE_ONLY = "write_only"
-_NUMBA_PVC_READ_WRITE = "read_write"
+_NUMBA_DPPY_READ_ONLY  = "read_only"
+_NUMBA_DPPY_WRITE_ONLY = "write_only"
+_NUMBA_DPPY_READ_WRITE = "read_write"
 
 def _raise_no_device_found_error():
     error_message = ("No OpenCL device specified. "
@@ -78,7 +78,7 @@ def compile_with_dppy(pyfunc, return_type, args, debug):
     # Do not compile (generate native code), just lower (to LLVM)
     flags.set('no_compile')
     flags.set('no_cpython_wrapper')
-    #flags.set('nrt')
+    flags.unset('nrt')
 
     # Run compilation pipeline
     if isinstance(pyfunc, FunctionType):
@@ -279,9 +279,9 @@ def __init__(self):
 
         # list of supported access types, stored in dict for fast lookup
         self.valid_access_types = {
-                _NUMBA_PVC_READ_ONLY: _NUMBA_PVC_READ_ONLY,
-                _NUMBA_PVC_WRITE_ONLY: _NUMBA_PVC_WRITE_ONLY,
-                _NUMBA_PVC_READ_WRITE: _NUMBA_PVC_READ_WRITE}
+                _NUMBA_DPPY_READ_ONLY: _NUMBA_DPPY_READ_ONLY,
+                _NUMBA_DPPY_WRITE_ONLY: _NUMBA_DPPY_WRITE_ONLY,
+                _NUMBA_DPPY_READ_WRITE: _NUMBA_DPPY_READ_WRITE}
 
     def copy(self):
         return copy.copy(self)
@@ -372,7 +372,7 @@ def _pack_argument(self, ty, val, device_env, device_arr, access_type):
         """
         if (device_arr and (access_type not in self.valid_access_types or
             access_type in self.valid_access_types and
-            self.valid_access_types[access_type] != _NUMBA_PVC_READ_ONLY)):
+            self.valid_access_types[access_type] != _NUMBA_DPPY_READ_ONLY)):
             # we get the date back to host if have created a
             # device_array or if access_type of this device_array
             # is not of type read_only and read_write
@@ -415,11 +415,11 @@ def _unpack_argument(self, ty, val, device_env, retr, kernelargs,
             dArr = None
 
             if (default_behavior or
-                self.valid_access_types[access_type] == _NUMBA_PVC_READ_ONLY or
-                self.valid_access_types[access_type] == _NUMBA_PVC_READ_WRITE):
+                self.valid_access_types[access_type] == _NUMBA_DPPY_READ_ONLY or
+                self.valid_access_types[access_type] == _NUMBA_DPPY_READ_WRITE):
                 # default, read_only and read_write case
                 dArr = device_env.copy_array_to_device(val)
-            elif self.valid_access_types[access_type] == _NUMBA_PVC_WRITE_ONLY:
+            elif self.valid_access_types[access_type] == _NUMBA_DPPY_WRITE_ONLY:
                 # write_only case, we do not copy the host data
                 dArr = device_env.create_device_array(val)
 

decorators.py

@@ -1,5 +1,5 @@
 from __future__ import print_function, absolute_import, division
-from numba import sigutils, types
+from numba.core import sigutils, types
 from .compiler import (compile_kernel, JitDPPyKernel, compile_dppy_func_template,
                        compile_dppy_func, get_ordered_arg_access_types)
 

descriptor.py

@@ -1,11 +1,11 @@
 from __future__ import print_function, division, absolute_import
-from numba.targets.descriptors import TargetDescriptor
-from numba.targets.options import TargetOptions
+from numba.core.descriptors import TargetDescriptor
+from numba.core.options import TargetOptions
 
-from .. import dispatcher, utils, typing
+from numba.core import dispatcher, utils, typing
 from .target import DPPyTargetContext, DPPyTypingContext
 
-from numba.targets.cpu import CPUTargetOptions
+from numba.core.cpu import CPUTargetOptions
 
 
 class DPPyTarget(TargetDescriptor):

dppy_host_fn_call_gen.py

@@ -5,9 +5,9 @@
 import llvmlite.ir.values as liv
 import llvmlite.ir as lir
 import llvmlite.binding as lb
-from .. import types, cgutils
+from numba.core import types, cgutils
 
-from numba.ir_utils import legalize_names
+from numba.core.ir_utils import legalize_names
 
 class DPPyHostFunctionCallsGenerator(object):
     def __init__(self, lowerer, cres, num_inputs):

dppy_lowerer.py

@@ -9,15 +9,11 @@
 import numpy as np
 
 import numba
-from .. import (compiler,
-                ir,
-                types,
-                six,
-                sigutils,
-                lowering,
-                parfor,
-                funcdesc)
-from numba.ir_utils import (add_offset_to_labels,
+from numba.core import (compiler, ir, types, sigutils, lowering,
+                funcdesc, config)
+from numba.parfors import parfor
+from numba import dppy
+from numba.core.ir_utils import (add_offset_to_labels,
                             replace_var_names,
                             remove_dels,
                             legalize_names,
@@ -35,14 +31,11 @@
                             get_unused_var_name,
                             find_potential_aliases,
                             is_const_call)
-from ..typing import signature
-from numba import config, dppy
-from numba.targets.cpu import ParallelOptions
-from numba.six import exec_
-import operator
+
+from numba.core.typing import signature
 
 import warnings
-from ..errors import NumbaParallelSafetyWarning
+from numba.core.errors import NumbaParallelSafetyWarning
 
 from .target import SPIR_GENERIC_ADDRSPACE
 from .dufunc_inliner import dufunc_inliner
@@ -289,12 +282,12 @@ def _create_gufunc_for_parfor_body(
             parfor_params.add(stop.name)
 
     # Get just the outputs of the parfor.
-    parfor_outputs = numba.parfor.get_parfor_outputs(parfor, parfor_params)
+    parfor_outputs = numba.parfors.parfor.get_parfor_outputs(parfor, parfor_params)
 
     # Get all parfor reduction vars, and operators.
     typemap = lowerer.fndesc.typemap
 
-    parfor_redvars, parfor_reddict = numba.parfor.get_parfor_reductions(
+    parfor_redvars, parfor_reddict = numba.parfors.parfor.get_parfor_reductions(
                                         lowerer.func_ir,
                                         parfor,
                                         parfor_params,
@@ -453,7 +446,7 @@ def print_arg_with_addrspaces(args):
     # Force gufunc outline into existence.
     globls = {"np": np, "numba": numba, "dppy": dppy}
     locls = {}
-    exec_(gufunc_txt, globls, locls)
+    exec(gufunc_txt, globls, locls)
     gufunc_func = locls[gufunc_name]
 
     if config.DEBUG_ARRAY_OPT:
@@ -679,7 +672,7 @@ def _lower_parfor_gufunc(lowerer, parfor):
 
     alias_map = {}
     arg_aliases = {}
-    numba.parfor.find_potential_aliases_parfor(parfor, parfor.params, typemap,
+    numba.parfors.parfor.find_potential_aliases_parfor(parfor, parfor.params, typemap,
                                         lowerer.func_ir, alias_map, arg_aliases)
     if config.DEBUG_ARRAY_OPT:
         print("alias_map", alias_map)
@@ -690,7 +683,7 @@ def _lower_parfor_gufunc(lowerer, parfor):
     # dict will become invalid
     assert parfor.params != None
 
-    parfor_output_arrays = numba.parfor.get_parfor_outputs(
+    parfor_output_arrays = numba.parfors.parfor.get_parfor_outputs(
         parfor, parfor.params)
 
 
@@ -705,7 +698,7 @@ def _lower_parfor_gufunc(lowerer, parfor):
     for l in parfor.loop_nests[1:]:
         assert typemap[l.index_variable.name] == index_var_typ
 
-    numba.parfor.sequential_parfor_lowering = True
+    numba.parfors.parfor.sequential_parfor_lowering = True
     loop_ranges = [(l.start, l.stop, l.step) for l in parfor.loop_nests]
 
     func, func_args, func_sig, func_arg_types, modified_arrays =(
@@ -722,7 +715,7 @@ def _lower_parfor_gufunc(lowerer, parfor):
         index_var_typ,
         parfor.races))
 
-    numba.parfor.sequential_parfor_lowering = False
+    numba.parfors.parfor.sequential_parfor_lowering = False
 
     # get the shape signature
     get_shape_classes = parfor.get_shape_classes
@@ -961,7 +954,7 @@ def load_range(v):
     dppy_cpu_lowerer.enqueue_kernel_and_read_back(loop_ranges)
 
 
-from numba.lowering import Lower
+from numba.core.lowering import Lower
 
 class DPPyLower(Lower):
     def __init__(self, context, library, fndesc, func_ir, metadata=None):

dppy_passbuilder.py

@@ -1,20 +1,21 @@
 from __future__ import print_function, division, absolute_import
 
-from numba.compiler_machinery import PassManager
+from numba.core.compiler_machinery import PassManager
 
-from numba.untyped_passes import (ExtractByteCode, TranslateByteCode, FixupArgs,
+from numba.core.untyped_passes import (ExtractByteCode, TranslateByteCode, FixupArgs,
                                   IRProcessing, DeadBranchPrune,
                                   RewriteSemanticConstants, InlineClosureLikes,
                                   GenericRewrites, WithLifting,
                                   InlineInlinables, FindLiterallyCalls,
                                   MakeFunctionToJitFunction,
                                   CanonicalizeLoopExit, CanonicalizeLoopEntry,
+                                  ReconstructSSA,
                                   LiteralUnroll)
 
-from numba.typed_passes import (NopythonTypeInference, AnnotateTypes,
+from numba.core.typed_passes import (NopythonTypeInference, AnnotateTypes,
                                 NopythonRewrites, PreParforPass, ParforPass,
                                 DumpParforDiagnostics, IRLegalization,
-                                InlineOverloads)
+                                InlineOverloads, PreLowerStripPhis)
 
 from .dppy_passes import (
         DPPyConstantSizeStaticLocalMemoryPass,
@@ -68,15 +69,18 @@ def default_numba_nopython_pipeline(state, pm):
         pm.add_pass(FindLiterallyCalls, "find literally calls")
         pm.add_pass(LiteralUnroll, "handles literal_unroll")
 
+        if state.flags.enable_ssa:
+            pm.add_pass(ReconstructSSA, "ssa")
         # typing
         pm.add_pass(NopythonTypeInference, "nopython frontend")
         pm.add_pass(AnnotateTypes, "annotate types")
 
+        # strip phis
+        pm.add_pass(PreLowerStripPhis, "remove phis nodes")
+
         # optimisation
         pm.add_pass(InlineOverloads, "inline overloaded functions")
 
-        # legalise
-        pm.add_pass(IRLegalization, "ensure IR is legal prior to lowering")
 
 
     @staticmethod
@@ -91,7 +95,9 @@ def define_nopython_pipeline(state, name='dppy_nopython'):
         if not state.flags.no_rewrites:
             pm.add_pass(NopythonRewrites, "nopython rewrites")
         pm.add_pass(DPPyParforPass, "convert to parfors")
-        #pm.add_pass(InlineParforVectorize, "inline vectorize inside parfors ")
+
+        # legalise
+        pm.add_pass(IRLegalization, "ensure IR is legal prior to lowering")
 
         # lower
         pm.add_pass(SpirvFriendlyLowering, "SPIRV-friendly lowering pass")

dppy_passes.py

@@ -2,37 +2,32 @@
 from contextlib import contextmanager
 import warnings
 
-from numba import ir
+from numba.core import ir
 import weakref
 from collections import namedtuple, deque
 import operator
 
-from numba.lowering import Lower, _VarArgItem
-
-from llvmlite.llvmpy.core import Constant, Type, Builder
-from numba.stencilparfor import StencilPass
-
-from numba import (
+from numba.core import (
     config,
     errors,
-    types,
-    rewrites,
-    typeinfer,
     funcdesc,
     utils,
-    typing
+    typing,
+    types,
     )
 
-from numba.errors import (LoweringError, new_error_context, TypingError,
+from numba.core.ir_utils import remove_dels
+
+from numba.core.errors import (LoweringError, new_error_context, TypingError,
                      LiteralTypingError)
 
-from numba.compiler_machinery import FunctionPass, LoweringPass, register_pass
+from numba.core.compiler_machinery import FunctionPass, LoweringPass, register_pass
 
 from .dppy_lowerer import DPPyLower
 
-from numba.parfor import PreParforPass as _parfor_PreParforPass
-from numba.parfor import ParforPass as _parfor_ParforPass
-from numba.parfor import Parfor
+from numba.parfors.parfor import PreParforPass as _parfor_PreParforPass
+from numba.parfors.parfor import ParforPass as _parfor_ParforPass
+from numba.parfors.parfor import Parfor
 
 
 @register_pass(mutates_CFG=True, analysis_only=False)
@@ -168,6 +163,8 @@ def run_pass(self, state):
 
         parfor_pass.run()
 
+        remove_dels(state.func_ir.blocks)
+
         if config.DEBUG or config.DUMP_IR:
             name = state.func_ir.func_id.func_qualname
             print(("IR DUMP: %s" % name).center(80, "-"))
@@ -244,11 +241,12 @@ def run_pass(self, state):
                 lower.lower()
                 if not flags.no_cpython_wrapper:
                     lower.create_cpython_wrapper(flags.release_gil)
+
                 env = lower.env
                 call_helper = lower.call_helper
                 del lower
 
-            from numba.compiler import _LowerResult  # TODO: move this
+            from numba.core.compiler import _LowerResult  # TODO: move this
             if flags.no_compile:
                 state['cr'] = _LowerResult(fndesc, call_helper,
                                            cfunc=None, env=env)
@@ -278,7 +276,7 @@ def run_pass(self, state):
         lowered = state['cr']
         signature = typing.signature(state.return_type, *state.args)
 
-        from numba.compiler import compile_result
+        from numba.core.compiler import compile_result
         state.cr = compile_result(
             typing_context=state.typingctx,
             target_context=state.targetctx,
@@ -296,4 +294,7 @@ def run_pass(self, state):
             metadata=state.metadata,
             reload_init=state.reload_init,
         )
+
+        remove_dels(state.func_ir.blocks)
+
         return True