path: root/mlir/utils/spirv/gen_spirv_dialect.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

# Copyright 2019 The MLIR Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Script for updating SPIR-V dialect by scraping information from SPIR-V
# HTML and JSON specs from the Internet.
#
# For example, to define the enum attribute for SPIR-V memory model:
#
# ./gen_spirv_dialect.py --base_td_path /path/to/SPIRVBase.td \
#                        --new-enum MemoryModel
#
# The 'operand_kinds' dict of spirv.core.grammar.json contains all supported
# SPIR-V enum classes.

import itertools
import re
import requests
import textwrap

SPIRV_HTML_SPEC_URL = 'https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html'
SPIRV_JSON_SPEC_URL = 'https://raw.githubusercontent.com/KhronosGroup/SPIRV-Headers/master/include/spirv/unified1/spirv.core.grammar.json'

AUTOGEN_OP_DEF_SEPARATOR = '\n// -----\n\n'
AUTOGEN_ENUM_SECTION_MARKER = 'enum section. Generated from SPIR-V spec; DO NOT MODIFY!'
AUTOGEN_OPCODE_SECTION_MARKER = (
    'opcode section. Generated from SPIR-V spec; DO NOT MODIFY!')


def get_spirv_doc_from_html_spec():
  """Extracts instruction documentation from SPIR-V HTML spec.

  Returns:
    - A dict mapping from instruction opcode to documentation.
  """
  response = requests.get(SPIRV_HTML_SPEC_URL)
  spec = response.content

  from bs4 import BeautifulSoup
  spirv = BeautifulSoup(spec, 'html.parser')

  section_anchor = spirv.find('h3', {'id': '_a_id_instructions_a_instructions'})

  doc = {}

  for section in section_anchor.parent.find_all('div', {'class': 'sect3'}):
    for table in section.find_all('table'):
      inst_html = table.tbody.tr.td.p
      opname = inst_html.a['id']
      # Ignore the first line, which is just the opname.
      doc[opname] = inst_html.text.split('\n', 1)[1].strip()

  return doc


def get_spirv_grammar_from_json_spec():
  """Extracts operand kind and instruction grammar from SPIR-V JSON spec.

  Returns:
    - A list containing all operand kinds' grammar
    - A list containing all instructions' grammar
  """
  response = requests.get(SPIRV_JSON_SPEC_URL)
  spec = response.content

  import json
  spirv = json.loads(spec)

  return spirv['operand_kinds'], spirv['instructions']


def split_list_into_sublists(items, offset):
  """Split the list of items into multiple sublists.

  This is to make sure the string composed from each sublist won't exceed
  80 characters.

  Arguments:
    - items: a list of strings
    - offset: the offset in calculating each sublist's length
  """
  chuncks = []
  chunk = []
  chunk_len = 0

  for item in items:
    chunk_len += len(item) + 2
    if chunk_len > 80:
      chuncks.append(chunk)
      chunk = []
      chunk_len = len(item) + 2
    chunk.append(item)

  if len(chunk) != 0:
    chuncks.append(chunk)

  return chuncks


def uniquify_enum_cases(lst):
  """Prunes duplicate enum cases from the list.

  Arguments:
   - lst: List whose elements are to be uniqued. Assumes each element is a
     (symbol, value) pair and elements already sorted according to value.

  Returns:
   - A list with all duplicates removed. The elements are sorted according to
     value and, for each value, uniqued according to symbol.
     original list,
  """
  cases = lst
  uniqued_cases = []

  # First sort according to the value
  cases.sort(key=lambda x: x[1])

  # Then group them according to the value
  for _, groups in itertools.groupby(cases, key=lambda x: x[1]):
    # For each value, sort according to the enumerant symbol.
    sorted_group = sorted(groups, key=lambda x: x[0])
    # Keep the "smallest" case, which is typically the symbol without extension
    # suffix. But we have special cases that we want to fix.
    case = sorted_group[0]
    if case[0] == 'HlslSemanticGOOGLE':
      case = sorted_group[1]
    uniqued_cases.append(case)

  return uniqued_cases


def gen_operand_kind_enum_attr(operand_kind):
  """Generates the TableGen EnumAttr definition for the given operand kind.

  Returns:
    - The operand kind's name
    - A string containing the TableGen EnumAttr definition
  """
  if 'enumerants' not in operand_kind:
    return '', ''

  # Returns a symbol for the given case in the given kind. This function
  # handles Dim specially to avoid having numbers as the start of symbols,
  # which does not play well with C++ and the MLIR parser.
  def get_case_symbol(kind_name, case_name):
    if kind_name == 'Dim':
      if case_name == '1D' or case_name == '2D' or case_name == '3D':
        return 'Dim{}'.format(case_name)
    return case_name

  kind_name = operand_kind['kind']
  is_bit_enum = operand_kind['category'] == 'BitEnum'
  kind_category = 'Bit' if is_bit_enum else 'I32'
  kind_acronym = ''.join([c for c in kind_name if c >= 'A' and c <= 'Z'])
  kind_cases = [(case['enumerant'], case['value'])
                for case in operand_kind['enumerants']]
  kind_cases = uniquify_enum_cases(kind_cases)
  max_len = max([len(symbol) for (symbol, _) in kind_cases])

  # Generate the definition for each enum case
  fmt_str = 'def SPV_{acronym}_{case} {colon:>{offset}} '\
            '{category}EnumAttrCase<"{symbol}", {value}>;'
  case_defs = [
      fmt_str.format(
          category=kind_category,
          acronym=kind_acronym,
          case=case[0],
          symbol=get_case_symbol(kind_name, case[0]),
          value=case[1],
          colon=':',
          offset=(max_len + 1 - len(case[0]))) for case in kind_cases
  ]
  case_defs = '\n'.join(case_defs)

  # Generate the list of enum case names
  fmt_str = 'SPV_{acronym}_{symbol}';
  case_names = [fmt_str.format(acronym=kind_acronym,symbol=case[0])
                for case in kind_cases]

  # Split them into sublists and concatenate into multiple lines
  case_names = split_list_into_sublists(case_names, 6)
  case_names = ['{:6}'.format('') + ', '.join(sublist)
                for sublist in case_names]
  case_names = ',\n'.join(case_names)

  # Generate the enum attribute definition
  enum_attr = 'def SPV_{name}Attr :\n    '\
      '{category}EnumAttr<"{name}", "valid SPIR-V {name}", [\n{cases}\n'\
      '    ]> {{\n'\
      '  let cppNamespace = "::mlir::spirv";\n}}'.format(
          name=kind_name, category=kind_category, cases=case_names)
  return kind_name, case_defs + '\n\n' + enum_attr


def gen_opcode(instructions):
  """ Generates the TableGen definition to map opname to opcode

  Returns:
    - A string containing the TableGen SPV_OpCode definition
  """

  max_len = max([len(inst['opname']) for inst in instructions])
  def_fmt_str = 'def SPV_OC_{name} {colon:>{offset}} '\
            'I32EnumAttrCase<"{name}", {value}>;'
  opcode_defs = [
      def_fmt_str.format(
          name=inst['opname'],
          value=inst['opcode'],
          colon=':',
          offset=(max_len + 1 - len(inst['opname']))) for inst in instructions
  ]
  opcode_str = '\n'.join(opcode_defs)

  decl_fmt_str = 'SPV_OC_{name}'
  opcode_list = [
      decl_fmt_str.format(name=inst['opname']) for inst in instructions
  ]
  opcode_list = split_list_into_sublists(opcode_list, 6)
  opcode_list = [
      '{:6}'.format('') + ', '.join(sublist) for sublist in opcode_list
  ]
  opcode_list = ',\n'.join(opcode_list)
  enum_attr = 'def SPV_OpcodeAttr :\n'\
              '    I32EnumAttr<"{name}", "valid SPIR-V instructions", [\n'\
              '{lst}\n'\
              '      ]> {{\n'\
              '    let cppNamespace = "::mlir::spirv";\n}}'.format(
                  name='Opcode', lst=opcode_list)
  return opcode_str + '\n\n' + enum_attr


def update_td_opcodes(path, instructions, filter_list):
  """Updates SPIRBase.td with new generated opcode cases.

  Arguments:
    - path: the path to SPIRBase.td
    - instructions: a list containing all SPIR-V instructions' grammar
    - filter_list: a list containing new opnames to add
  """

  with open(path, 'r') as f:
    content = f.read()

  content = content.split(AUTOGEN_OPCODE_SECTION_MARKER)
  assert len(content) == 3

  # Extend opcode list with existing list
  existing_opcodes = [k[11:] for k in re.findall('def SPV_OC_\w+', content[1])]
  filter_list.extend(existing_opcodes)
  filter_list = list(set(filter_list))

  # Generate the opcode for all instructions in SPIR-V
  filter_instrs = list(
      filter(lambda inst: (inst['opname'] in filter_list), instructions))
  # Sort instruction based on opcode
  filter_instrs.sort(key=lambda inst: inst['opcode'])
  opcode = gen_opcode(filter_instrs)

  # Substitute the opcode
  content = content[0] + AUTOGEN_OPCODE_SECTION_MARKER + '\n\n' + \
        opcode + '\n\n// End ' + AUTOGEN_OPCODE_SECTION_MARKER \
        + content[2]

  with open(path, 'w') as f:
    f.write(content)


def update_td_enum_attrs(path, operand_kinds, filter_list):
  """Updates SPIRBase.td with new generated enum definitions.

  Arguments:
    - path: the path to SPIRBase.td
    - operand_kinds: a list containing all operand kinds' grammar
    - filter_list: a list containing new enums to add
  """
  with open(path, 'r') as f:
    content = f.read()

  content = content.split(AUTOGEN_ENUM_SECTION_MARKER)
  assert len(content) == 3

  # Extend filter list with existing enum definitions
  existing_kinds = [
      k[8:-4] for k in re.findall('def SPV_\w+Attr', content[1])]
  filter_list.extend(existing_kinds)

  # Generate definitions for all enums in filter list
  defs = [gen_operand_kind_enum_attr(kind)
          for kind in operand_kinds if kind['kind'] in filter_list]
  # Sort alphabetically according to enum name
  defs.sort(key=lambda enum : enum[0])
  # Only keep the definitions from now on
  # Put Capability's definition at the very beginning because capability cases
  # will be referenced later
  defs = [enum[1] for enum in defs if enum[0] == 'Capability'
         ] + [enum[1] for enum in defs if enum[0] != 'Capability']

  # Substitute the old section
  content = content[0] + AUTOGEN_ENUM_SECTION_MARKER + '\n\n' + \
      '\n\n'.join(defs) + "\n\n// End " + AUTOGEN_ENUM_SECTION_MARKER  \
      + content[2];

  with open(path, 'w') as f:
    f.write(content)


def snake_casify(name):
  """Turns the given name to follow snake_case convention."""
  name = re.sub('\W+', '', name).split()
  name = [s.lower() for s in name]
  return '_'.join(name)


def map_spec_operand_to_ods_argument(operand):
  """Maps a operand in SPIR-V JSON spec to an op argument in ODS.

  Arguments:
    - A dict containing the operand's kind, quantifier, and name

  Returns:
    - A string containing both the type and name for the argument
  """
  kind = operand['kind']
  quantifier = operand.get('quantifier', '')

  # These instruction "operands" are for encoding the results; they should
  # not be handled here.
  assert kind != 'IdResultType', 'unexpected to handle "IdResultType" kind'
  assert kind != 'IdResult', 'unexpected to handle "IdResult" kind'

  if kind == 'IdRef':
    if quantifier == '':
      arg_type = 'SPV_Type'
    elif quantifier == '?':
      arg_type = 'SPV_Optional<SPV_Type>'
    else:
      arg_type = 'Variadic<SPV_Type>'
  elif kind == 'IdMemorySemantics' or kind == 'IdScope':
    # TODO(antiagainst): Need to further constrain 'IdMemorySemantics'
    # and 'IdScope' given that they should be generated from OpConstant.
    assert quantifier == '', ('unexpected to have optional/variadic memory '
                              'semantics or scope <id>')
    arg_type = 'SPV_' + kind[2:] + 'Attr'
  elif kind == 'LiteralInteger':
    if quantifier == '':
      arg_type = 'I32Attr'
    elif quantifier == '?':
      arg_type = 'OptionalAttr<I32Attr>'
    else:
      arg_type = 'OptionalAttr<I32ArrayAttr>'
  elif kind == 'LiteralString' or \
      kind == 'LiteralContextDependentNumber' or \
      kind == 'LiteralExtInstInteger' or \
      kind == 'LiteralSpecConstantOpInteger' or \
      kind == 'PairLiteralIntegerIdRef' or \
      kind == 'PairIdRefLiteralInteger' or \
      kind == 'PairIdRefIdRef':
    assert False, '"{}" kind unimplemented'.format(kind)
  else:
    # The rest are all enum operands that we represent with op attributes.
    assert quantifier != '*', 'unexpected to have variadic enum attribute'
    arg_type = 'SPV_{}Attr'.format(kind)
    if quantifier == '?':
      arg_type = 'OptionalAttr<{}>'.format(arg_type)

  name = operand.get('name', '')
  name = snake_casify(name) if name else kind.lower()

  return '{}:${}'.format(arg_type, name)


def get_description(text, assembly):
  """Generates the description for the given SPIR-V instruction.

  Arguments:
    - text: Textual description of the operation as string.
    - assembly: Custom Assembly format with example as string.

  Returns:
    - A string that corresponds to the description of the Tablegen op.
  """
  fmt_str = ('{text}\n\n    ### Custom assembly ' 'form\n{assembly}\n  ')
  return fmt_str.format(
      text=text, assembly=assembly)


def get_op_definition(instruction, doc, existing_info):
  """Generates the TableGen op definition for the given SPIR-V instruction.

  Arguments:
    - instruction: the instruction's SPIR-V JSON grammar
    - doc: the instruction's SPIR-V HTML doc
    - existing_info: a dict containing potential manually specified sections for
      this instruction

  Returns:
    - A string containing the TableGen op definition
  """
  fmt_str = ('def SPV_{opname}Op : '
             'SPV_{inst_category}<"{opname}"{category_args}[{traits}]> '
             '{{\n  let summary = {summary};\n\n  let description = '
             '[{{\n{description}}}];\n')
  inst_category = existing_info.get('inst_category', 'Op')
  if inst_category == 'Op':
    fmt_str +='\n  let arguments = (ins{args});\n\n'\
              '  let results = (outs{results});\n'

  fmt_str +='{extras}'\
            '}}\n'

  opname = instruction['opname'][2:]
  category_args = existing_info.get('category_args', '')
  # Make sure we have ', ' to separate the category arguments from traits
  category_args = category_args.rstrip(', ') + ', '

  if '\n' in doc:
    summary, text = doc.split('\n', 1)
  else:
    summary = doc
    text = ''
  wrapper = textwrap.TextWrapper(
      width=76, initial_indent='    ', subsequent_indent='    ')

  # Format summary. If the summary can fit in the same line, we print it out
  # as a "-quoted string; otherwise, wrap the lines using "[{...}]".
  summary = summary.strip();
  if len(summary) + len('  let summary = "";') <= 80:
    summary = '"{}"'.format(summary)
  else:
    summary = '[{{\n{}\n  }}]'.format(wrapper.fill(summary))

  # Wrap text
  text = text.split('\n')
  text = [wrapper.fill(line) for line in text if line]
  text = '\n\n'.join(text)

  operands = instruction.get('operands', [])

  # Set op's result
  results = ''
  if len(operands) > 0 and operands[0]['kind'] == 'IdResultType':
    results = '\n    SPV_Type:$result\n  '
    operands = operands[1:]
  if 'results' in existing_info:
    results = existing_info['results']

  # Ignore the operand standing for the result <id>
  if len(operands) > 0 and operands[0]['kind'] == 'IdResult':
    operands = operands[1:]

  # Set op' argument
  arguments = existing_info.get('arguments', None)
  if arguments is None:
    arguments = [map_spec_operand_to_ods_argument(o) for o in operands]
    arguments = ',\n    '.join(arguments)
    if arguments:
      # Prepend and append whitespace for formatting
      arguments = '\n    {}\n  '.format(arguments)

  description = existing_info.get('description', None)
  if description is None:
    assembly = '\n    ```\n'\
               '    [TODO]\n'\
               '    ```\n\n'\
               '    For example:\n\n'\
               '    ```\n'\
               '    [TODO]\n' \
               '    ```'
    description = get_description(text, assembly)

  return fmt_str.format(
      opname=opname,
      category_args=category_args,
      inst_category=inst_category,
      traits=existing_info.get('traits', ''),
      summary=summary,
      description=description,
      args=arguments,
      results=results,
      extras=existing_info.get('extras', ''))


def get_string_between(base, start, end):
  """Extracts a substring with a specified start and end from a string.

  Arguments:
    - base: string to extract from.
    - start: string to use as the start of the substring.
    - end: string to use as the end of the substring.

  Returns:
    - The substring if found
    - The part of the base after end of the substring. Is the base string itself
      if the substring wasnt found.
  """
  split = base.split(start, 1)
  if len(split) == 2:
    rest = split[1].split(end, 1)
    assert len(rest) == 2, \
           'cannot find end "{end}" while extracting substring '\
           'starting with {start}'.format(start=start, end=end)
    return rest[0].rstrip(end), rest[1]
  return '', split[0]


def get_string_between_nested(base, start, end):
  """Extracts a substring with a nested start and end from a string.

  Arguments:
    - base: string to extract from.
    - start: string to use as the start of the substring.
    - end: string to use as the end of the substring.

  Returns:
    - The substring if found
    - The part of the base after end of the substring. Is the base string itself
      if the substring wasnt found.
  """
  split = base.split(start, 1)
  if len(split) == 2:
    # Handle nesting delimiters
    rest = split[1]
    unmatched_start = 1
    index = 0
    while unmatched_start > 0 and index < len(rest):
      if rest[index:].startswith(end):
        unmatched_start -= 1
        index += len(end)
      elif rest[index:].startswith(start):
        unmatched_start += 1
        index += len(start)
      else:
        index += 1

    assert index < len(rest), \
           'cannot find end "{end}" while extracting substring '\
           'starting with "{start}"'.format(start=start, end=end)
    return rest[:index - len(end)].rstrip(end), rest[index:]
  return '', split[0]


def extract_td_op_info(op_def):
  """Extracts potentially manually specified sections in op's definition.

  Arguments: - A string containing the op's TableGen definition
    - doc: the instruction's SPIR-V HTML doc

  Returns:
    - A dict containing potential manually specified sections
  """
  # Get opname
  opname = [o[8:-2] for o in re.findall('def SPV_\w+Op', op_def)]
  assert len(opname) == 1, 'more than one ops in the same section!'
  opname = opname[0]

  # Get instruction category
  inst_category = [
      o[4:] for o in re.findall('SPV_\w+Op',
                                op_def.split(':', 1)[1])
  ]
  assert len(inst_category) <= 1, 'more than one ops in the same section!'
  inst_category = inst_category[0] if len(inst_category) == 1 else 'Op'

  # Get category_args
  op_tmpl_params = get_string_between_nested(op_def, '<', '>')[0]
  opstringname, rest = get_string_between(op_tmpl_params, '"', '"')
  category_args = rest.split('[', 1)[0]

  # Get traits
  traits, _ = get_string_between(rest, '[', ']')

  # Get description
  description, rest = get_string_between(op_def, 'let description = [{\n',
                                         '}];\n')

  # Get arguments
  args, rest = get_string_between(rest, '  let arguments = (ins', ');\n')

  # Get results
  results, rest = get_string_between(rest, '  let results = (outs', ');\n')

  extras = rest.strip(' }\n')
  if extras:
    extras = '\n  {}\n'.format(extras)

  return {
      # Prefix with 'Op' to make it consistent with SPIR-V spec
      'opname': 'Op{}'.format(opname),
      'inst_category': inst_category,
      'category_args': category_args,
      'traits': traits,
      'description': description,
      'arguments': args,
      'results': results,
      'extras': extras
  }


def update_td_op_definitions(path, instructions, docs, filter_list,
                             inst_category):
  """Updates SPIRVOps.td with newly generated op definition.

  Arguments:
    - path: path to SPIRVOps.td
    - instructions: SPIR-V JSON grammar for all instructions
    - docs: SPIR-V HTML doc for all instructions
    - filter_list: a list containing new opnames to include

  Returns:
    - A string containing all the TableGen op definitions
  """
  with open(path, 'r') as f:
    content = f.read()

  # Split the file into chuncks, each containing one op.
  ops = content.split(AUTOGEN_OP_DEF_SEPARATOR)
  header = ops[0]
  footer = ops[-1]
  ops = ops[1:-1]

  # For each existing op, extract the manually-written sections out to retain
  # them when re-generating the ops. Also append the existing ops to filter
  # list.
  name_op_map = {}  # Map from opname to its existing ODS definition
  op_info_dict = {}
  for op in ops:
    info_dict = extract_td_op_info(op)
    opname = info_dict['opname']
    name_op_map[opname] = op
    op_info_dict[opname] = info_dict
    filter_list.append(opname)
  filter_list = sorted(list(set(filter_list)))

  op_defs = []
  for opname in filter_list:
    # Find the grammar spec for this op
    try:
      instruction = next(
          inst for inst in instructions if inst['opname'] == opname)
      op_defs.append(
          get_op_definition(
              instruction, docs[opname],
              op_info_dict.get(opname, {'inst_category': inst_category})))
    except StopIteration:
      # This is an op added by us; use the existing ODS definition.
      op_defs.append(name_op_map[opname])

  # Substitute the old op definitions
  op_defs = [header] + op_defs + [footer]
  content = AUTOGEN_OP_DEF_SEPARATOR.join(op_defs)

  with open(path, 'w') as f:
    f.write(content)


if __name__ == '__main__':
  import argparse

  cli_parser = argparse.ArgumentParser(
      description='Update SPIR-V dialect definitions using SPIR-V spec')

  cli_parser.add_argument(
      '--base-td-path',
      dest='base_td_path',
      type=str,
      default=None,
      help='Path to SPIRVBase.td')
  cli_parser.add_argument(
      '--op-td-path',
      dest='op_td_path',
      type=str,
      default=None,
      help='Path to SPIRVOps.td')

  cli_parser.add_argument(
      '--new-enum',
      dest='new_enum',
      type=str,
      default=None,
      help='SPIR-V enum to be added to SPIRVBase.td')
  cli_parser.add_argument(
      '--new-opcodes',
      dest='new_opcodes',
      type=str,
      default=None,
      nargs='*',
      help='update SPIR-V opcodes in SPIRVBase.td')
  cli_parser.add_argument(
      '--new-inst',
      dest='new_inst',
      type=str,
      default=None,
      nargs='*',
      help='SPIR-V instruction to be added to ops file')
  cli_parser.add_argument(
      '--inst-category',
      dest='inst_category',
      type=str,
      default='Op',
      help='SPIR-V instruction category used for choosing '\
           'the TableGen base class to define this op')

  args = cli_parser.parse_args()

  operand_kinds, instructions = get_spirv_grammar_from_json_spec()

  # Define new enum attr
  if args.new_enum is not None:
    assert args.base_td_path is not None
    filter_list = [args.new_enum] if args.new_enum else []
    update_td_enum_attrs(args.base_td_path, operand_kinds, filter_list)

  # Define new opcode
  if args.new_opcodes is not None:
    assert args.base_td_path is not None
    update_td_opcodes(args.base_td_path, instructions, args.new_opcodes)

  # Define new op
  if args.new_inst is not None:
    assert args.op_td_path is not None
    docs = get_spirv_doc_from_html_spec()
    update_td_op_definitions(args.op_td_path, instructions, docs, args.new_inst,
                             args.inst_category)
    print('Done. Note that this script just generates a template; ', end='')
    print('please read the spec and update traits, arguments, and ', end='')
    print('results accordingly.')