Merge pull request #41 from xdslproject/compiler_more

compiler: more cleanup

Author: George Bisbas

devito/core/cpu.py

@@ -311,3 +311,32 @@ class Cpu64FsgCOperator(Cpu64FsgOperator):
 
 class Cpu64FsgOmpOperator(Cpu64FsgOperator):
     _Target = OmpTarget
+
+
+# -----------XDSL
+# This is a collection of xDSL optimization pipelines
+# Ideally they should follow the same type of subclassing as the rest of
+# the Devito Operatos
+
+
+MLIR_CPU_PIPELINE = '"builtin.module(canonicalize, cse, loop-invariant-code-motion, canonicalize, cse, loop-invariant-code-motion, cse, canonicalize, fold-memref-alias-ops, expand-strided-metadata, loop-invariant-code-motion, lower-affine, convert-scf-to-cf, convert-math-to-llvm, convert-func-to-llvm{use-bare-ptr-memref-call-conv}, finalize-memref-to-llvm, canonicalize, cse)"'  # noqa
+
+MLIR_OPENMP_PIPELINE = '"builtin.module(canonicalize, cse, loop-invariant-code-motion, canonicalize, cse, loop-invariant-code-motion,cse,canonicalize,fold-memref-alias-ops,expand-strided-metadata, loop-invariant-code-motion,lower-affine,finalize-memref-to-llvm,loop-invariant-code-motion,canonicalize,cse,convert-scf-to-openmp,finalize-memref-to-llvm,convert-scf-to-cf,convert-func-to-llvm{use-bare-ptr-memref-call-conv},convert-openmp-to-llvm,convert-math-to-llvm,reconcile-unrealized-casts,canonicalize,cse)"'   # noqa
+# gpu-launch-sink-index-computations seemed to have no impact
+MLIR_GPU_PIPELINE = lambda block_sizes: f'"builtin.module(test-math-algebraic-simplification,scf-parallel-loop-tiling{{parallel-loop-tile-sizes={block_sizes}}},func.func(gpu-map-parallel-loops),convert-parallel-loops-to-gpu,lower-affine, canonicalize,cse, fold-memref-alias-ops, gpu-launch-sink-index-computations, gpu-kernel-outlining, canonicalize{{region-simplify}},cse,fold-memref-alias-ops,expand-strided-metadata,lower-affine,canonicalize,cse,func.func(gpu-async-region),canonicalize,cse,convert-arith-to-llvm{{index-bitwidth=64}},convert-scf-to-cf,convert-cf-to-llvm{{index-bitwidth=64}},canonicalize,cse,convert-func-to-llvm{{use-bare-ptr-memref-call-conv}},gpu.module(convert-gpu-to-nvvm,reconcile-unrealized-casts,canonicalize,gpu-to-cubin),gpu-to-llvm,canonicalize,cse)"'   # noqa
+
+XDSL_CPU_PIPELINE = lambda nb_tiled_dims: f'"stencil-shape-inference,convert-stencil-to-ll-mlir{{{generate_tiling_arg(nb_tiled_dims)}}},printf-to-llvm"'  # noqa
+
+XDSL_GPU_PIPELINE = "stencil-shape-inference,convert-stencil-to-ll-mlir{target=gpu},reconcile-unrealized-casts,printf-to-llvm"  # noqa
+
+XDSL_MPI_PIPELINE = lambda decomp, nb_tiled_dims: f'"dmp-decompose{decomp},canonicalize-dmp,convert-stencil-to-ll-mlir{{{generate_tiling_arg(nb_tiled_dims)}}},dmp-to-mpi{{mpi_init=false}},lower-mpi,printf-to-llvm"'   # noqa
+
+
+def generate_tiling_arg(nb_tiled_dims: int):
+    """
+    Generate the tile-sizes arg for the convert-stencil-to-ll-mlir pass.
+    Generating no argument if the diled_dims arg is 0
+    """
+    if nb_tiled_dims == 0:
+        return ''
+    return "tile-sizes=" + ",".join(["64"]*nb_tiled_dims)

devito/ir/ietxdsl/__init__.py

@@ -1,2 +1,5 @@
-from devito.ir.ietxdsl.lowering import LowerIetForToScfFor, LowerIetForToScfParallel, DropIetComments, iet_to_standard_mlir # noqa
-from devito.ir.ietxdsl.cluster_to_ssa import finalize_module_with_globals, convert_devito_stencil_to_xdsl_stencil  # noqa
+from devito.ir.ietxdsl.lowering import (LowerIetForToScfFor, LowerIetForToScfParallel)
+from devito.ir.ietxdsl.cluster_to_ssa import (finalize_module_with_globals,
+                                              convert_devito_stencil_to_xdsl_stencil)
+
+# flake8: noqa

devito/ir/ietxdsl/cluster_to_ssa.py

@@ -1,22 +1,33 @@
 # ------------- General imports -------------#
 
 from typing import Any
+from dataclasses import dataclass, field
 from sympy import Add, Expr, Float, Indexed, Integer, Mod, Mul, Pow, Symbol
 
 # ------------- xdsl imports -------------#
-from xdsl.dialects import arith, builtin, func, memref, scf, stencil, gpu
+from xdsl.dialects import (arith, builtin, func, memref, scf,
+                           stencil, gpu, llvm)
 from xdsl.dialects.experimental import math
 from xdsl.ir import Block, Operation, OpResult, Region, SSAValue
+from xdsl.pattern_rewriter import (
+    GreedyRewritePatternApplier,
+    PatternRewriter,
+    PatternRewriteWalker,
+    RewritePattern,
+    op_type_rewrite_pattern,
+)
 
 # ------------- devito imports -------------#
 from devito import Grid, SteppingDimension
 from devito.ir.equations import LoweredEq
-from devito.symbolics import retrieve_indexed
+from devito.symbolics import retrieve_indexed, retrieve_function_carriers
 from devito.logger import perf
 
 # ------------- devito-xdsl SSA imports -------------#
 from devito.ir.ietxdsl import iet_ssa
+from devito.ir.ietxdsl.utils import is_int, is_float
 from devito.ir.ietxdsl.ietxdsl_functions import dtypes_to_xdsltypes
+from devito.ir.ietxdsl.lowering import LowerIetForToScfFor
 
 # flake8: noqa
 
@@ -50,34 +61,34 @@ def _convert_eq(self, eq: LoweredEq):
         function = eq.lhs.function
         mlir_type = dtypes_to_xdsltypes[function.dtype]
         grid: Grid = function.grid
-        # get the halo of the space dimensions only e.g [(2, 2), (2, 2)] for the 2d case
+
+        # Get the halo of the grid.dimensions
+        # e.g [(2, 2), (2, 2)] for the 2D case
         # Do not forget the issue with Devito adding an extra point!
+        # Check 'def halo_setup' for more
         # (for derivative regions)
-        halo = [function.halo[function.dimensions.index(d)] for d in grid.dimensions]
+        halo = [function.halo[d] for d in grid.dimensions]
 
         # Shift all time values so that for all accesses at t + n, n>=0.
         self.time_offs = min(
             int(idx.indices[0] - grid.stepping_dim) for idx in retrieve_indexed(eq)
         )
 
-        # Calculate the actual size of our time dimension
-        actual_time_size = (
-            max(int(idx.indices[0] - grid.stepping_dim) for idx in retrieve_indexed(eq))
-            - self.time_offs
-            + 1
-        )
 
+        # Get the time_size       
+        time_size = max(d.function.time_size for d in retrieve_function_carriers(eq))
+        
         # Build the for loop
         perf("Build Time Loop")
-        loop = self._build_iet_for(grid.stepping_dim, actual_time_size)
+        loop = self._build_iet_for(grid.stepping_dim, time_size)
 
         # build stencil
         perf("Initialize a stencil Op")
         stencil_op = iet_ssa.Stencil.get(
             loop.subindice_ssa_vals(),
             grid.shape_local,
             halo,
-            actual_time_size,
+            time_size,
             mlir_type,
             eq.lhs.function._C_name,
         )
@@ -87,7 +98,7 @@ def _convert_eq(self, eq: LoweredEq):
         # dims -> ssa vals
         perf("Apply time offsets")
         time_offset_to_field: dict[str, SSAValue] = {
-            i: stencil_op.block.args[i] for i in range(actual_time_size - 1)
+            i: stencil_op.block.args[i] for i in range(time_size - 1)
         }
 
         # reset loaded values
@@ -103,8 +114,9 @@ def _convert_eq(self, eq: LoweredEq):
 
         # emit return
         offsets = _get_dim_offsets(eq.lhs, self.time_offs)
+
         assert (
-            offsets[0] == actual_time_size - 1
+            offsets[0] == time_size - 1
         ), "result should be written to last time buffer"
         assert all(
             o == 0 for o in offsets[1:]
@@ -118,51 +130,47 @@ def _convert_eq(self, eq: LoweredEq):
         )
 
     def _visit_math_nodes(self, node: Expr) -> SSAValue:
+        # Handle Indexeds
         if isinstance(node, Indexed):
             offsets = _get_dim_offsets(node, self.time_offs)
             return self.loaded_values[offsets]
-        if isinstance(node, Integer):
+        # Handle Integers
+        elif isinstance(node, Integer):
             cst = arith.Constant.from_int_and_width(int(node), builtin.i64)
             self.block.add_op(cst)
             return cst.result
-        if isinstance(node, Float):
+        # Handle Floats
+        elif isinstance(node, Float):
             cst = arith.Constant.from_float_and_width(float(node), builtin.f32)
             self.block.add_op(cst)
             return cst.result
-        # if isinstance(math, Constant):
-        #    symb = iet_ssa.LoadSymbolic.get(math.name, dtypes_to_xdsltypes[math.dtype])
-        #    self.block.add_op(symb)
-        #    return symb.result
-        if isinstance(node, Symbol):
+        # Handle Symbols
+        elif isinstance(node, Symbol):
             symb = iet_ssa.LoadSymbolic.get(node.name, builtin.f32)
             self.block.add_op(symb)
-            return symb.result
-
-        # handle all of the math
-        if not isinstance(node, (Add, Mul, Pow, Mod)):
-            raise ValueError(f"Unknown math: {node}", node)
-
-        args = [self._visit_math_nodes(arg) for arg in node.args]
-
-        # make sure all args are the same type:
-        if isinstance(node, (Add, Mul)):
+            return symb.result     
+        # Handle Add Mul
+        elif isinstance(node, (Add, Mul)):
+            args = [self._visit_math_nodes(arg) for arg in node.args]
             # add casts when necessary
             # get first element out, store the rest in args
             # this makes the reduction easier
             carry, *args = self._ensure_same_type(*args)
             # select the correct op from arith.addi, arith.addf, arith.muli, arith.mulf
             if isinstance(carry.type, builtin.IntegerType):
                 op_cls = arith.Addi if isinstance(node, Add) else arith.Muli
-            else:
+            elif isinstance(carry.type, builtin.Float32Type):
                 op_cls = arith.Addf if isinstance(node, Add) else arith.Mulf
-
+            else:
+                raise("Add support for another type")
             for arg in args:
                 op = op_cls(carry, arg)
                 self.block.add_op(op)
                 carry = op.result
             return carry
-
-        if isinstance(node, Pow):
+        # Handle Pow
+        elif isinstance(node, Pow):
+            args = [self._visit_math_nodes(arg) for arg in node.args]
             assert len(args) == 2, "can't pow with != 2 args!"
             base, ex = args
             if is_int(base):
@@ -183,11 +191,12 @@ def _visit_math_nodes(self, node: Expr) -> SSAValue:
             op = op_cls.get(base, ex)
             self.block.add_op(op)
             return op.result
+        # Handle Mod
+        elif isinstance(node, Mod):
+            raise NotImplementedError("Go away, no mod here. >:(")
+        else:
+            raise NotImplementedError(f"Unknown math: {node}", node)
 
-        if isinstance(node, Mod):
-            raise ValueError("Go away, no mod here. >:(")
-
-        raise ValueError("Unknown math!")
 
     def _add_access_ops(
         self, reads: list[Indexed], time_offset_to_field: dict[int, SSAValue]
@@ -202,10 +211,11 @@ def _add_access_ops(
             """
             # get the compile time constant offsets for this read
             offsets = _get_dim_offsets(read, self.time_offs)
+            
             if offsets in self.loaded_values:
                 continue
 
-            # assume time dimension is first dimension
+            # Assume time dimension is first dimension
             t_offset = offsets[0]
             space_offsets = offsets[1:]
 
@@ -251,10 +261,10 @@ def _ensure_same_type(self, *vals: SSAValue):
         if all(is_float(val) for val in vals):
             return vals
         # not everything homogeneous
-        new_vals = []
+        processed = []
         for val in vals:
             if is_float(val):
-                new_vals.append(val)
+                processed.append(val)
                 continue
             # if the val is the result of a arith.constant with no uses,
             # we change the type of the arith.constant to our desired type
@@ -267,19 +277,20 @@ def _ensure_same_type(self, *vals: SSAValue):
                 val.op.attributes["value"] = builtin.FloatAttr(
                     float(val.op.value.value.data), builtin.f32
                 )
-                new_vals.append(val)
+                processed.append(val)
                 continue
             # insert an integer to float cast op
             conv = arith.SIToFPOp(val, builtin.f32)
             self.block.add_op(conv)
-            new_vals.append(conv.result)
-        return new_vals
+            processed.append(conv.result)
+        return processed
 
 
 def _get_dim_offsets(idx: Indexed, t_offset: int) -> tuple:
     # shift all time values so that for all accesses at t + n, n>=0.
     # time_offs = min(int(i - d) for i, d in zip(idx.indices, idx.function.dimensions))
     halo = ((t_offset, 0), *idx.function.halo[1:])
+
     try:
         return tuple(
             int(i - d - halo_offset)
@@ -291,36 +302,13 @@ def _get_dim_offsets(idx: Indexed, t_offset: int) -> tuple:
         raise ValueError("Indices must be constant offset from dimension!") from ex
 
 
-def is_int(val: SSAValue):
-    return isinstance(val.type, builtin.IntegerType)
-
-
-def is_float(val: SSAValue):
-    return val.type in (builtin.f32, builtin.f64)
-
 
 # -------------------------------------------------------- ####
 #                                                          ####
 #           devito.stencil  ---> stencil dialect           ####
 #                                                          ####
 # -------------------------------------------------------- ####
 
-from dataclasses import dataclass, field
-
-from xdsl.pattern_rewriter import (
-    GreedyRewritePatternApplier,
-    PatternRewriter,
-    PatternRewriteWalker,
-    RewritePattern,
-    op_type_rewrite_pattern,
-)
-
-from devito.ir.ietxdsl.lowering import (
-    LowerIetForToScfFor,
-)
-
-from xdsl.dialects import llvm
-
 @dataclass
 class WrapFunctionWithTransfers(RewritePattern):
     func_name: str
@@ -545,8 +533,10 @@ def finalize_module_with_globals(module: builtin.ModuleOp, known_symbols: dict[s
         _InsertSymbolicConstants(known_symbols),
         _LowerLoadSymbolidToFuncArgs(),
     ]
-    grpa = GreedyRewritePatternApplier(patterns)
-    PatternRewriteWalker(grpa).rewrite_module(module)
+    rewriter = GreedyRewritePatternApplier(patterns)
+    PatternRewriteWalker(rewriter).rewrite_module(module)
+
+    # GPU boilerplate
     if gpu_boilerplate:
         walker = PatternRewriteWalker(GreedyRewritePatternApplier([WrapFunctionWithTransfers('apply_kernel')]))
         walker.rewrite_module(module)