Ms138 directives (numba#52)

* Fix bugs to make pranges work

* Update README

* Update README

* Only keep setitems to track buffers being modified inside kernel

Author: reazulhoque

README.rst

@@ -1,78 +1,80 @@
-                                ****************
-                                      DPPY
-                                ****************
-=========
+
+DPPY
+====
+
+========
 1. What?
-=========
+========
 
 DPPy proof-of-concept backend for NUMBA to support compilation for Intel CPU and
 GPU architectures. The present implementation of DPPy is based on OpenCL 2.1,
 but is likely to change in the future to rely on Sycl/DPC++ or Intel Level-0
 driver API.
 
-================
+===============
 2. Prequisites?
-================
-
-Bash                 : In the system and not as default Shell
-Tar                  : To extract files
-Git                  : To fetch required dependencies listed below
-C/C++ compiler       : To build the dependencies
-Cmake                : For managing build process of dependencies
-Python3              : Version 3 is required
-Conda or miniconda   : Can be found at https://docs.conda.io/en/latest/miniconda.html
-
-OpenCL 2.1 driver    : DPPy currently works for both Intel GPUs and CPUs is
-                       a correct OpenCL driver version is found on the system.
+===============
 
-                       Note. To use the GPU users should be added to "video"
-                       user group on Linux systems.
+- Bash                 : In the system and not as default Shell
+- Tar                  : To extract files
+- Git                  : To fetch required dependencies listed below
+- C/C++ compiler       : To build the dependencies
+- Cmake                : For managing build process of dependencies
+- Python3              : Version 3 is required
+- Conda or miniconda   : Can be found at https://docs.conda.io/en/latest/miniconda.html
+- OpenCL 2.1 driver    : DPPy currently works for both Intel GPUs and CPUs is a correct OpenCL driver version is found on the system. 
+Note. To use the GPU users should be added to "video" user group on Linux systems.
 
 
-The following requisites will be installed by the install script provided with
-this package.
+The following requisites will need to be present in the system. Refer to next section for more details.
+*******************************************************************************************************
 
-NUMBA v0.48          : The DPPy backend has only been tested for NUMBA v0.48.
-                       The included install script downloads and applies the
-                       DDPy patch to the correct NUMBA version.
+- NUMBA v0.48          : The DPPy backend has only been tested for NUMBA v0.48. The included install script downloads and applies the DDPy patch to the correct NUMBA version.
 
-LLVM-SPIRV translator: Used for SPIRV generation from LLVM IR.
+- LLVM-SPIRV translator: Used for SPIRV generation from LLVM IR.
 
-SPIRV-Tools          : Used internally for code-generation. The provided install
-                       script would handle downloading and installing the
-                       required version.
+- SPIRV-Tools          : Used internally for code-generation.
 
-LLVMDEV              : To support LLVM IR generation.
+- LLVMDEV              : To support LLVM IR generation.
 
-Others               : All existing dependecies for NUMBA, such as llvmlite,
-                       also apply to DPPy.
+- Others               : All existing dependecies for NUMBA, such as llvmlite, also apply to DPPy.
 
 ==================
 3. How to install?
 ==================
+Install Pre-requisites
+*************************
+Make sure the dependencies of NUMBA-DPPY are installed in the system, for convenience
+and to make sure the dependencies are installed with consistent version of LLVM we provide
+installation script that will create a CONDA environment and install LLVM-SPIRV translator,
+SPIRV-Tools and llvmlite in that environment. **To use this CONDA has to be available in the system**.
 
-Extract the archive:
+The above mentioned installation script can be found `here <https://github.intel.com/SAT/numba-pvc-build-scripts>`_. Please follow the README to run the installation script. 
 
-    tar -zxvf NUMBA-PVC-offline.tar.gz
+After successful installation the following message should be displayed:
 
-Run the installer script:
+    | #
+    | # Use the following to activate the correct environment
+    | #
+    | # `    $ ``conda activate numba-dppy-env`` `
+    | #
+    | #  Use the following to deactivate environment
+    | #
+    | # `    $ ``conda deactivate`` `
 
-    ./build_numba_dppy.sh --prefix $PATH_TO_INSTALL_NUMBA-DPPY
+The installer script creates a new conda environment called numba-dppy-env with
+all the needed dependencies already installed. **Please activate the numba-dppy-env before proceeding**.
 
-After successful installation the following message should be displayed:
 
-    #
-    #  Use the following to activate the correct environment
-    #
-    #    $ conda activate numba-dppy-env
-    #
-    #  Use the following to deactivate environment
-    #
-    #    $ conda deactivate
+Install DPPY backend
+***********************
+NUMBA-DPPY also depend on DPPY backend. It can be found `here <https://github.intel.com/SAT/dppy>`_. Please run 
+`build_for_conda.sh` to install DPPY backend.
 
-The installer script creates a new conda environment called numba-dppy-env with
-all the needed dependencies already installed. To use the DPPy backend, please
-activate the numba-dppy-env
+Install NUMBA-DPPY
+*********************
+After all the dependencies are installed please run ``build_for_develop.sh`` to get a local installation of NUMBA-DPPY. **Both step 2 and 3 assumes CONDA environment with
+the dependencies of NUMBA-DPPY installed in it, was activated**.
 
 ================
 4. Running tests
@@ -81,12 +83,11 @@ activate the numba-dppy-env
 To make sure the installation was successful, try running the examples and the
 test suite:
 
-    $PATH_TO_INSTALL_NUMBA-DPPY/numba/dppy/examples/
-    $PATH_TO_INSTALL_NUMBA-DPPY/numba/dppy/tests/dppy/
+    $PATH_TO_NUMBA-DPPY/numba/dppy/examples/
 
 To run the test suite execute the following:
 
-    $ python -m numba.runtests numba.dppy.tests
+    $ ``python -m numba.runtests numba.dppy.tests``
 
 ===========================
 5. How Tos and Known Issues
@@ -101,5 +102,3 @@ examples, supported functionalities, and known issues.
 ===================
 
 Please email diptorup.deb@intel.com to report issues and bugs.
-
-

numba/dppy/dppy_host_fn_call_gen.py

@@ -7,6 +7,8 @@
 import llvmlite.binding as lb
 from .. import types, cgutils
 
+from numba.ir_utils import legalize_names
+
 class DPPyHostFunctionCallsGenerator(object):
     def __init__(self, lowerer, cres, num_inputs):
         self.lowerer = lowerer
@@ -126,7 +128,7 @@ def _call_dppy_kernel_arg_fn(self, args):
         self.builder.store(self.builder.load(kernel_arg), dst)
 
 
-    def process_kernel_arg(self, var, llvm_arg, arg_type, gu_sig, val_type, index):
+    def process_kernel_arg(self, var, llvm_arg, arg_type, gu_sig, val_type, index, modified_arrays):
 
         if isinstance(arg_type, types.npytypes.Array):
             if llvm_arg is None:
@@ -176,6 +178,13 @@ def process_kernel_arg(self, var, llvm_arg, arg_type, gu_sig, val_type, index):
                     buffer_ptr]
             self.builder.call(self.create_dppy_rw_mem_buffer, args)
 
+            # names are replaces usig legalize names, we have to do the same for them to match
+            legal_names = legalize_names([var])
+
+            if legal_names[var] in modified_arrays:
+                self.read_bufs_after_enqueue.append((buffer_ptr, total_size, data_member))
+
+            # We really need to detect when an array needs to be copied over
             if index < self.num_inputs:
                 args = [self.builder.inttoptr(self.gpu_device_int_const, self.void_ptr_t),
                         self.builder.load(buffer_ptr),
@@ -185,8 +194,6 @@ def process_kernel_arg(self, var, llvm_arg, arg_type, gu_sig, val_type, index):
                         self.builder.bitcast(self.builder.load(data_member), self.void_ptr_t)]
 
                 self.builder.call(self.write_mem_buffer_to_device, args)
-            else:
-                self.read_bufs_after_enqueue.append((buffer_ptr, total_size, data_member))
 
             self.builder.call(self.create_dppy_kernel_arg_from_buffer, [buffer_ptr, kernel_arg])
             dst = self.builder.gep(self.kernel_arg_array, [self.context.get_constant(types.intp, self.cur_arg)])
@@ -215,7 +222,7 @@ def process_kernel_arg(self, var, llvm_arg, arg_type, gu_sig, val_type, index):
             for this_stride in range(arg_type.ndim):
                 stride_entry = self.builder.gep(stride_member,
                                 [self.context.get_constant(types.int32, 0),
-                                 self.context.get_constant(types.int32, this_dim)])
+                                 self.context.get_constant(types.int32, this_stride)])
 
                 args = [self.builder.bitcast(stride_entry, self.void_ptr_ptr_t),
                         self.context.get_constant(types.uintp, self.sizeof_intp)]
@@ -275,4 +282,4 @@ def enqueue_kernel_and_read_back(self, loop_ranges):
                     self.zero,
                     self.builder.load(array_size_member),
                     self.builder.bitcast(self.builder.load(data_member), self.void_ptr_t)]
-            self.builder.call(self.read_mem_buffer_from_device, args)
+            self.builder.call(self.read_mem_buffer_from_device, args)

numba/dppy/dppy_lowerer.py

@@ -50,6 +50,18 @@
 import dppy.core as driver
 
 
+def _print_block(block):
+    for i, inst in enumerate(block.body):
+        print("    ", i, inst)
+
+def _print_body(body_dict):
+    '''Pretty-print a set of IR blocks.
+    '''
+    for label, block in body_dict.items():
+        print("label: ", label)
+        _print_block(block)
+
+
 # This loop scheduler is pretty basic, there is only
 # 3 dimension allowed in OpenCL, so to make the backend
 # functional we will schedule the first 3 dimensions
@@ -201,6 +213,20 @@ def to_scalar_from_0d(x):
             return x.dtype
     return x
 
+def find_setitems_block(setitems, block, typemap):
+    for inst in block.body:
+        if isinstance(inst, ir.StaticSetItem) or isinstance(inst, ir.SetItem):
+            setitems.add(inst.target.name)
+        elif isinstance(inst, parfor.Parfor):
+            find_setitems_block(setitems, inst.init_block, typemap)
+            find_setitems_body(setitems, inst.loop_body, typemap)
+
+def find_setitems_body(setitems, loop_body, typemap):
+    """
+      Find the arrays that are written into (goes into setitems)
+    """
+    for label, block in loop_body.items():
+        find_setitems_block(setitems, block, typemap)
 
 def _create_gufunc_for_regular_parfor():
     #TODO
@@ -487,6 +513,9 @@ def print_arg_with_addrspaces(args):
     wrapped_blocks = wrap_loop_body(loop_body)
     #hoisted, not_hoisted = hoist(parfor_params, loop_body,
     #                             typemap, wrapped_blocks)
+    setitems = set()
+    find_setitems_body(setitems, loop_body, typemap)
+
     hoisted = []
     not_hoisted = []
 
@@ -599,7 +628,7 @@ def print_arg_with_addrspaces(args):
     if config.DEBUG_ARRAY_OPT:
         print("kernel_sig = ", kernel_sig)
 
-    return kernel_func, parfor_args, kernel_sig, func_arg_types
+    return kernel_func, parfor_args, kernel_sig, func_arg_types, setitems
 
 
 def _lower_parfor_gufunc(lowerer, parfor):
@@ -680,7 +709,7 @@ def _lower_parfor_gufunc(lowerer, parfor):
     numba.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 =(
+    func, func_args, func_sig, func_arg_types, modified_arrays =(
     _create_gufunc_for_parfor_body(
         lowerer,
         parfor,
@@ -731,7 +760,8 @@ def _lower_parfor_gufunc(lowerer, parfor):
         loop_ranges,
         parfor.init_block,
         index_var_typ,
-        parfor.races)
+        parfor.races,
+        modified_arrays)
 
     if config.DEBUG_ARRAY_OPT:
         sys.stdout.flush()
@@ -820,7 +850,8 @@ def generate_dppy_host_wrapper(lowerer,
                                loop_ranges,
                                init_block,
                                index_var_typ,
-                               races):
+                               races,
+                               modified_arrays):
     '''
     Adds the call to the gufunc function from the main function.
     '''
@@ -841,6 +872,7 @@ def generate_dppy_host_wrapper(lowerer,
         print("sin", sin)
         print("sout", sout)
         print("cres", cres, type(cres))
+        print("modified_arrays", modified_arrays)
 #        print("cres.library", cres.library, type(cres.library))
 #        print("cres.fndesc", cres.fndesc, type(cres.fndesc))
 
@@ -908,7 +940,7 @@ def val_type_or_none(context, lowerer, x):
                   "\n\tindex:", index)
 
         dppy_cpu_lowerer.process_kernel_arg(var, llvm_arg, arg_type, gu_sig,
-                                            val_type, index)
+                                            val_type, index, modified_arrays)
     # -----------------------------------------------------------------------
 
     # loadvars for loop_ranges