manual merge

CMakeLists.txt

@@ -60,7 +60,11 @@ endif()
 set(MIGRAPHX_ENABLE_CPU Off CACHE BOOL "")
 
 set(CMAKE_CXX_STANDARD_DEFAULT "")
-add_compile_options(-std=c++14)
+if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+    add_compile_options(-std=c++17)
+else()
+    add_compile_options(-std=c++14)
+endif()
 
 list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake)
 include(EnableCompilerWarnings)
@@ -187,6 +191,8 @@ rocm_enable_cppcheck(
         definePrefix:*test/include/test.hpp
         useSmartPointer:*src/api/api.cpp
         useSmartPointer:*make_shared_array.hpp
+        constParameter:*src/targets/gpu/*.cpp
+        constParameter:*src/targets/gpu/*.hpp
     FORCE
     INCONCLUSIVE
     RULE_FILE

Dockerfile

@@ -74,7 +74,7 @@ RUN cget -p $PREFIX install facebook/zstd@v1.4.5 -X subdir -DCMAKE_DIR=build/cma
 RUN cget -p $PREFIX install ccache@v4.1
 
 # Install newer cmake for onnx runtime
-RUN cget -p /opt/cmake install kitware/cmake@v3.13.0
+RUN cget -p /opt/cmake install kitware/cmake@v3.13.4
 
 ARG ONNXRUNTIME_REPO=https://github.com/Microsoft/onnxruntime
 ARG ONNXRUNTIME_BRANCH=master
@@ -86,6 +86,8 @@ RUN git clone --single-branch --branch ${ONNXRUNTIME_BRANCH} --recursive ${ONNXR
 
 ADD tools/build_and_test_onnxrt.sh /onnxruntime/build_and_test_onnxrt.sh
 
+RUN PATH=/opt/cmake/bin:$PATH cget -p /usr/local install ROCmSoftwarePlatform/llvm-project-mlir@02078ce236ad90e3aec04c0c770ef5bfc99e49c2
+
 ENV MIOPEN_FIND_DB_PATH=/tmp/miopen/find-db
 ENV MIOPEN_USER_DB_PATH=/tmp/miopen/user-db
 ENV LD_LIBRARY_PATH=$PREFIX/lib

Jenkinsfile

@@ -94,6 +94,12 @@ rocmtest clang_debug: rocmnode('vega') { cmake_build ->
         cmake_build("/opt/rocm/llvm/bin/clang++", "-DCMAKE_BUILD_TYPE=release")
         stash includes: 'build/*.deb', name: 'migraphx-package'
     }
+}, mlir_debug: rocmnode('vega') { cmake_build ->
+    stage('MLIR Debug') {
+        def sanitizers = "undefined"
+        def debug_flags = "-g -O2 -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
+        cmake_build("/opt/rocm/llvm/bin/clang++", "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DMIGRAPHX_ENABLE_MLIR=On -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'")
+    }
 }
 
 def onnxnode(name, body) {

examples/python_bert_squad_example/BERT-Squad.ipynb

@@ -43,7 +43,7 @@
     "from os import path\n",
     "import sys\n",
     "\n",
-    "import tokenization\n",
+    "import tokenizers\n",
     "from run_onnx_squad import *\n",
     "\n",
     "import migraphx"
@@ -137,8 +137,7 @@
    "outputs": [],
    "source": [
     "vocab_file = os.path.join('uncased_L-12_H-768_A-12', 'vocab.txt')\n",
-    "tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,\n",
-    "                                       do_lower_case=True)"
+    "tokenizer = tokenizers.BertWordPieceTokenizer(vocab_file)"
    ]
   },
   {

examples/python_bert_squad_example/README.md

@@ -7,21 +7,25 @@ There are two ways to run the example:
 
 # Steps
 1) Install MIGraphX to your environment. Please follow the steps to build MIGraphX given at https://github.com/ROCmSoftwarePlatform/AMDMIGraphX
-2) Install the requirements file
+2) Upgrade your pip3 to latest version
 ```
-pip3 install -r requirements_migraphx.txt
+pip3 install --upgrade pip 
 ```
-3) Install `unzip` and fetch the uncased file (vocabulary):
+3) Install the requirements file
+```
+pip3 install -r requirements_bertsquad.txt
+```
+4) Install `unzip` and fetch the uncased file (vocabulary):
 ```
 apt-get install unzip
 wget -q https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-12_H-768_A-12.zip
 unzip uncased_L-12_H-768_A-12.zip
 ```
-4) Get BERT ONNX model (bertsquad-10.onnx):
+5) Get BERT ONNX model (bertsquad-10.onnx):
 ```
 wget https://github.com/onnx/models/raw/master/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx
 ```
-5) Run the inference, it will compile and run the model on three questions and small data provided in `inputs.json`:
+6) Run the inference, it will compile and run the model on three questions and small data provided in `inputs.json`:
 ```
 python3 bert-squad-migraphx.py
 ```

examples/python_bert_squad_example/bert-squad-migraphx.py

@@ -5,7 +5,7 @@ import os.path
 from os import path
 import sys
 
-import tokenization
+import tokenizers
 from run_onnx_squad import *
 
 import migraphx
@@ -30,8 +30,7 @@ n_best_size = 20
 max_answer_length = 30
 
 vocab_file = os.path.join('uncased_L-12_H-768_A-12', 'vocab.txt')
-tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,
-                                       do_lower_case=True)
+tokenizer = tokenizers.BertWordPieceTokenizer(vocab_file)
 
 # Use convert_examples_to_features method from run_onnx_squad to get parameters from the input
 input_ids, input_mask, segment_ids, extra_data = convert_examples_to_features(

examples/python_bert_squad_example/requirements_bertsquad.txt

-tensorflow==1.14
-onnxruntime
+tensorflow==2.4.0
+onnxruntime
\ No newline at end of file
+tokenizers
\ No newline at end of file

examples/python_bert_squad_example/run_onnx_squad.py

@@ -38,7 +38,8 @@ from timeit import default_timer as timer
 import numpy as np
 import onnxruntime as onnxrt
 import six
-import tokenization
+from tokenizers import BertWordPieceTokenizer
+from tokenizers import pre_tokenizers
 
 RawResult = collections.namedtuple("RawResult",
                                    ["unique_id", "start_logits", "end_logits"])
@@ -70,9 +71,8 @@ class SquadExample(object):
 
     def __repr__(self):
         s = []
-        s.append("qas_id: %s" % (tokenization.printable_text(self.qas_id)))
-        s.append("question_text: %s" %
-                 (tokenization.printable_text(self.question_text)))
+        s.append("qas_id: %s" % (self.qas_id))
+        s.append("question_text: %s" % (self.question_text))
         s.append("doc_tokens: [%s]" % (" ".join(self.doc_tokens)))
         if self.start_position:
             s.append("start_position: %d" % (self.start_position))
@@ -130,7 +130,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
     unique_id = 0
 
     for (example_index, example) in enumerate(examples):
-        query_tokens = tokenizer.tokenize(example.question_text)
+        query_tokens = tokenizer.encode(example.question_text)
 
         if len(query_tokens) > max_query_length:
             query_tokens = query_tokens[0:max_query_length]
@@ -140,8 +140,8 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
         all_doc_tokens = []
         for (i, token) in enumerate(example.doc_tokens):
             orig_to_tok_index.append(len(all_doc_tokens))
-            sub_tokens = tokenizer.tokenize(token)
-            for sub_token in sub_tokens:
+            sub_tokens = tokenizer.encode(token, add_special_tokens=False)
+            for sub_token in sub_tokens.tokens:
                 tok_to_orig_index.append(i)
                 all_doc_tokens.append(sub_token)
 
@@ -172,7 +172,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
             segment_ids = []
             tokens.append("[CLS]")
             segment_ids.append(0)
-            for token in query_tokens:
+            for token in query_tokens.tokens:
                 tokens.append(token)
                 segment_ids.append(0)
             tokens.append("[SEP]")
@@ -192,7 +192,9 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
             tokens.append("[SEP]")
             segment_ids.append(1)
 
-            input_ids = tokenizer.convert_tokens_to_ids(tokens)
+            input_ids = []
+            for token in tokens:
+                input_ids.append(tokenizer.token_to_id(token))
 
             # The mask has 1 for real tokens and 0 for padding tokens. Only real
             # tokens are attended to.
@@ -437,9 +439,15 @@ def get_final_text(pred_text, orig_text, do_lower_case):
     # and `pred_text`, and check if they are the same length. If they are
     # NOT the same length, the heuristic has failed. If they are the same
     # length, we assume the characters are one-to-one aligned.
-    tokenizer = tokenization.BasicTokenizer(do_lower_case=do_lower_case)
+    tokenizer = pre_tokenizers.Sequence(
+        [pre_tokenizers.Whitespace(),
+         pre_tokenizers.Punctuation()])
 
-    tok_text = " ".join(tokenizer.tokenize(orig_text))
+    tok_text = []
+    for item in tokenizer.pre_tokenize_str(orig_text):
+        tok_text.append(item[0])
+
+    tok_text = " ".join(tok_text)
 
     start_position = tok_text.find(pred_text)
     if start_position == -1:
@@ -559,8 +567,7 @@ def main():
         sess_options = onnxrt.SessionOptions()
         sess_options.session_log_verbosity_level = args.log
 
-    tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file,
-                                           do_lower_case=True)
+    tokenizer = BertWordPieceTokenizer(vocab_file)
 
     eval_examples = read_squad_examples(input_file=args.predict_file)
     input_ids, input_mask, segment_ids, extra_data = \

examples/python_bert_squad_example/tokenization.py

-# coding=utf-8
-# Copyright 2018 The Google AI Language Team 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.
-"""Tokenization classes."""
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import collections
-import re
-import unicodedata
-import six
-import tensorflow as tf
-
-
-def validate_case_matches_checkpoint(do_lower_case, init_checkpoint):
-    """Checks whether the casing config is consistent with the checkpoint name."""
-
-    # The casing has to be passed in by the user and there is no explicit check
-    # as to whether it matches the checkpoint. The casing information probably
-    # should have been stored in the bert_config.json file, but it's not, so
-    # we have to heuristically detect it to validate.
-
-    if not init_checkpoint:
-        return
-
-    m = re.match("^.*?([A-Za-z0-9_-]+)/bert_model.ckpt", init_checkpoint)
-    if m is None:
-        return
-
-    model_name = m.group(1)
-
-    lower_models = [
-        "uncased_L-24_H-1024_A-16", "uncased_L-12_H-768_A-12",
-        "multilingual_L-12_H-768_A-12", "chinese_L-12_H-768_A-12"
-    ]
-
-    cased_models = [
-        "cased_L-12_H-768_A-12", "cased_L-24_H-1024_A-16",
-        "multi_cased_L-12_H-768_A-12"
-    ]
-
-    is_bad_config = False
-    if model_name in lower_models and not do_lower_case:
-        is_bad_config = True
-        actual_flag = "False"
-        case_name = "lowercased"
-        opposite_flag = "True"
-
-    if model_name in cased_models and do_lower_case:
-        is_bad_config = True
-        actual_flag = "True"
-        case_name = "cased"
-        opposite_flag = "False"
-
-    if is_bad_config:
-        raise ValueError(
-            "You passed in `--do_lower_case=%s` with `--init_checkpoint=%s`. "
-            "However, `%s` seems to be a %s model, so you "
-            "should pass in `--do_lower_case=%s` so that the fine-tuning matches "
-            "how the model was pre-training. If this error is wrong, please "
-            "just comment out this check." %
-            (actual_flag, init_checkpoint, model_name, case_name,
-             opposite_flag))
-
-
-def convert_to_unicode(text):
-    """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
-    if six.PY3:
-        if isinstance(text, str):
-            return text
-        elif isinstance(text, bytes):
-            return text.decode("utf-8", "ignore")
-        else:
-            raise ValueError("Unsupported string type: %s" % (type(text)))
-    elif six.PY2:
-        if isinstance(text, str):
-            return text.decode("utf-8", "ignore")
-        elif isinstance(text, unicode):
-            return text
-        else:
-            raise ValueError("Unsupported string type: %s" % (type(text)))
-    else:
-        raise ValueError("Not running on Python2 or Python 3?")
-
-
-def printable_text(text):
-    """Returns text encoded in a way suitable for print or `tf.logging`."""
-
-    # These functions want `str` for both Python2 and Python3, but in one case
-    # it's a Unicode string and in the other it's a byte string.
-    if six.PY3:
-        if isinstance(text, str):
-            return text
-        elif isinstance(text, bytes):
-            return text.decode("utf-8", "ignore")
-        else:
-            raise ValueError("Unsupported string type: %s" % (type(text)))
-    elif six.PY2:
-        if isinstance(text, str):
-            return text
-        elif isinstance(text, unicode):
-            return text.encode("utf-8")
-        else:
-            raise ValueError("Unsupported string type: %s" % (type(text)))
-    else:
-        raise ValueError("Not running on Python2 or Python 3?")
-
-
-def load_vocab(vocab_file):
-    """Loads a vocabulary file into a dictionary."""
-    vocab = collections.OrderedDict()
-    index = 0
-    with tf.gfile.GFile(vocab_file, "r") as reader:
-        while True:
-            token = convert_to_unicode(reader.readline())
-            if not token:
-                break
-            token = token.strip()
-            vocab[token] = index
-            index += 1
-    return vocab
-
-
-def convert_by_vocab(vocab, items):
-    """Converts a sequence of [tokens|ids] using the vocab."""
-    output = []
-    for item in items:
-        output.append(vocab[item])
-    return output
-
-
-def convert_tokens_to_ids(vocab, tokens):
-    return convert_by_vocab(vocab, tokens)
-
-
-def convert_ids_to_tokens(inv_vocab, ids):
-    return convert_by_vocab(inv_vocab, ids)
-
-
-def whitespace_tokenize(text):
-    """Runs basic whitespace cleaning and splitting on a piece of text."""
-    text = text.strip()
-    if not text:
-        return []
-    tokens = text.split()
-    return tokens
-
-
-class FullTokenizer(object):
-    """Runs end-to-end tokenziation."""
-    def __init__(self, vocab_file, do_lower_case=True):
-        self.vocab = load_vocab(vocab_file)
-        self.inv_vocab = {v: k for k, v in self.vocab.items()}
-        self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
-        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
-
-    def tokenize(self, text):
-        split_tokens = []
-        for token in self.basic_tokenizer.tokenize(text):
-            for sub_token in self.wordpiece_tokenizer.tokenize(token):
-                split_tokens.append(sub_token)
-
-        return split_tokens
-
-    def convert_tokens_to_ids(self, tokens):
-        return convert_by_vocab(self.vocab, tokens)
-
-    def convert_ids_to_tokens(self, ids):
-        return convert_by_vocab(self.inv_vocab, ids)
-
-
-class BasicTokenizer(object):
-    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
-    def __init__(self, do_lower_case=True):
-        """Constructs a BasicTokenizer.
-
-    Args:
-      do_lower_case: Whether to lower case the input.
-    """
-        self.do_lower_case = do_lower_case
-
-    def tokenize(self, text):
-        """Tokenizes a piece of text."""
-        text = convert_to_unicode(text)
-        text = self._clean_text(text)
-
-        # This was added on November 1st, 2018 for the multilingual and Chinese
-        # models. This is also applied to the English models now, but it doesn't
-        # matter since the English models were not trained on any Chinese data
-        # and generally don't have any Chinese data in them (there are Chinese
-        # characters in the vocabulary because Wikipedia does have some Chinese
-        # words in the English Wikipedia.).
-        text = self._tokenize_chinese_chars(text)
-
-        orig_tokens = whitespace_tokenize(text)
-        split_tokens = []
-        for token in orig_tokens:
-            if self.do_lower_case:
-                token = token.lower()
-                token = self._run_strip_accents(token)
-            split_tokens.extend(self._run_split_on_punc(token))
-
-        output_tokens = whitespace_tokenize(" ".join(split_tokens))
-        return output_tokens
-
-    def _run_strip_accents(self, text):
-        """Strips accents from a piece of text."""
-        text = unicodedata.normalize("NFD", text)
-        output = []
-        for char in text:
-            cat = unicodedata.category(char)
-            if cat == "Mn":
-                continue
-            output.append(char)
-        return "".join(output)
-
-    def _run_split_on_punc(self, text):
-        """Splits punctuation on a piece of text."""
-        chars = list(text)
-        i = 0
-        start_new_word = True
-        output = []
-        while i < len(chars):
-            char = chars[i]
-            if _is_punctuation(char):
-                output.append([char])
-                start_new_word = True
-            else:
-                if start_new_word:
-                    output.append([])
-                start_new_word = False
-                output[-1].append(char)
-            i += 1
-
-        return ["".join(x) for x in output]
-
-    def _tokenize_chinese_chars(self, text):
-        """Adds whitespace around any CJK character."""
-        output = []
-        for char in text:
-            cp = ord(char)
-            if self._is_chinese_char(cp):
-                output.append(" ")
-                output.append(char)
-                output.append(" ")
-            else:
-                output.append(char)
-        return "".join(output)
-
-    def _is_chinese_char(self, cp):
-        """Checks whether CP is the codepoint of a CJK character."""
-        # This defines a "chinese character" as anything in the CJK Unicode block:
-        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
-        #
-        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
-        # despite its name. The modern Korean Hangul alphabet is a different block,
-        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
-        # space-separated words, so they are not treated specially and handled
-        # like the all of the other languages.
-        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
-            (cp >= 0x3400 and cp <= 0x4DBF) or  #
-            (cp >= 0x20000 and cp <= 0x2A6DF) or  #
-            (cp >= 0x2A700 and cp <= 0x2B73F) or  #
-            (cp >= 0x2B740 and cp <= 0x2B81F) or  #
-            (cp >= 0x2B820 and cp <= 0x2CEAF) or
-            (cp >= 0xF900 and cp <= 0xFAFF) or  #
-            (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
-            return True
-
-        return False
-
-    def _clean_text(self, text):
-        """Performs invalid character removal and whitespace cleanup on text."""
-        output = []
-        for char in text:
-            cp = ord(char)
-            if cp == 0 or cp == 0xfffd or _is_control(char):
-                continue
-            if _is_whitespace(char):
-                output.append(" ")
-            else:
-                output.append(char)
-        return "".join(output)
-
-
-class WordpieceTokenizer(object):
-    """Runs WordPiece tokenziation."""
-    def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=200):
-        self.vocab = vocab
-        self.unk_token = unk_token
-        self.max_input_chars_per_word = max_input_chars_per_word
-
-    def tokenize(self, text):
-        """Tokenizes a piece of text into its word pieces.
-
-    This uses a greedy longest-match-first algorithm to perform tokenization
-    using the given vocabulary.
-
-    For example:
-      input = "unaffable"
-      output = ["un", "##aff", "##able"]
-
-    Args:
-      text: A single token or whitespace separated tokens. This should have
-        already been passed through `BasicTokenizer.
-
-    Returns:
-      A list of wordpiece tokens.
-    """
-
-        text = convert_to_unicode(text)
-
-        output_tokens = []
-        for token in whitespace_tokenize(text):
-            chars = list(token)
-            if len(chars) > self.max_input_chars_per_word:
-                output_tokens.append(self.unk_token)
-                continue
-
-            is_bad = False
-            start = 0
-            sub_tokens = []
-            while start < len(chars):
-                end = len(chars)
-                cur_substr = None
-                while start < end:
-                    substr = "".join(chars[start:end])
-                    if start > 0:
-                        substr = "##" + substr
-                    if substr in self.vocab:
-                        cur_substr = substr
-                        break
-                    end -= 1
-                if cur_substr is None:
-                    is_bad = True
-                    break
-                sub_tokens.append(cur_substr)
-                start = end
-
-            if is_bad:
-                output_tokens.append(self.unk_token)
-            else:
-                output_tokens.extend(sub_tokens)
-        return output_tokens
-
-
-def _is_whitespace(char):
-    """Checks whether `chars` is a whitespace character."""
-    # \t, \n, and \r are technically contorl characters but we treat them
-    # as whitespace since they are generally considered as such.
-    if char == " " or char == "\t" or char == "\n" or char == "\r":
-        return True
-    cat = unicodedata.category(char)
-    if cat == "Zs":
-        return True
-    return False
-
-
-def _is_control(char):
-    """Checks whether `chars` is a control character."""
-    # These are technically control characters but we count them as whitespace
-    # characters.
-    if char == "\t" or char == "\n" or char == "\r":
-        return False
-    cat = unicodedata.category(char)
-    if cat in ("Cc", "Cf"):
-        return True
-    return False
-
-
-def _is_punctuation(char):
-    """Checks whether `chars` is a punctuation character."""
-    cp = ord(char)
-    # We treat all non-letter/number ASCII as punctuation.
-    # Characters such as "^", "$", and "`" are not in the Unicode
-    # Punctuation class but we treat them as punctuation anyways, for
-    # consistency.
-    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64)
-            or (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
-        return True
-    cat = unicodedata.category(char)
-    if cat.startswith("P"):
-        return True
-    return False

requirements.txt

@@ -6,4 +6,4 @@ nlohmann/json@v3.8.0
 blaze,https://bitbucket.org/blaze-lib/blaze/get/f0755dea0e03.tar.gz -X header -DHEADER_DIR=blaze
 half,https://github.com/pfultz2/half/archive/1.12.0.tar.gz -X header -H sha256:0a08660b68abb176ebc2a0cdf8de46e3182a7f46c66443bb80dbfaaec98cf969
 pybind/pybind11@d159a563383d10c821ba7b2a71905d1207db6de4 --build
-msgpack/msgpack-c@cpp-3.3.0 -DMSGPACK_BUILD_TESTS=Off
+msgpack/msgpack-c@cpp-3.3.0 -DMSGPACK_BUILD_TESTS=Off
\ No newline at end of file

src/CMakeLists.txt

@@ -7,6 +7,7 @@ include(CheckCXXLinkerFlag)
 add_library(migraphx 
     adjust_allocation.cpp
     analyze_streams.cpp
+    argument.cpp
     auto_contiguous.cpp
     eliminate_common_subexpression.cpp
     decompose.cpp
@@ -121,6 +122,7 @@ register_migraphx_ops(
     pad
     pooling
     pow
+    prefix_scan_sum
     prelu
     quant_convolution
     quant_dot

src/api/api.cpp

@@ -49,6 +49,7 @@ shape::type_t to_shape_type(migraphx_shape_datatype_t t)
 {
     switch(t)
     {
+    case migraphx_shape_tuple_type: return shape::tuple_type;
 #define MIGRAPHX_DETAIL_SHAPE_CASE_CONVERT(x, y) \
     case migraphx_shape_##x: return shape::x;
         MIGRAPHX_SHAPE_VISIT_TYPES(MIGRAPHX_DETAIL_SHAPE_CASE_CONVERT)
@@ -61,6 +62,7 @@ migraphx_shape_datatype_t to_shape_type(shape::type_t t)
 {
     switch(t)
     {
+    case shape::tuple_type: return migraphx_shape_tuple_type;
 #define MIGRAPHX_DETAIL_SHAPE_CASE_CONVERT(x, y) \
     case shape::x: return migraphx_shape_##x;
         MIGRAPHX_SHAPE_VISIT_TYPES(MIGRAPHX_DETAIL_SHAPE_CASE_CONVERT)

src/api/include/migraphx/migraphx.h

@@ -36,6 +36,7 @@ typedef enum {
 #define MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES(x, t) migraphx_shape_##x,
 /// An enum to represent the different data type inputs
 typedef enum {
+    migraphx_shape_tuple_type,
     MIGRAPHX_SHAPE_VISIT_TYPES(MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES)
 } migraphx_shape_datatype_t;
 #undef MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES

src/argument.cpp

+#include <migraphx/argument.hpp>
+#include <migraphx/functional.hpp>
+#include <unordered_map>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+argument::argument(const shape& s) : m_shape(s)
+{
+    auto buffer = make_shared_array<char>(s.bytes());
+    m_data      = {[=]() mutable { return buffer.get(); }};
+}
+
+argument::argument(shape s, std::nullptr_t)
+    : m_shape(std::move(s)), m_data({[] { return nullptr; }})
+{
+}
+
+argument::argument(const shape& s, const argument::data_t& d) : m_shape(s), m_data(d) {}
+
+argument argument::load(const shape& s, char* buffer)
+{
+    if(s.type() != shape::tuple_type)
+        return argument{s, buffer};
+    // Collect all shapes
+    std::unordered_map<std::size_t, shape> shapes;
+    {
+        // cppcheck-suppress variableScope
+        std::size_t i = 0;
+        fix([&](auto self, auto ss) {
+            if(ss.sub_shapes().empty())
+            {
+                shapes[i] = ss;
+                i++;
+            }
+            else
+            {
+                for(auto&& child : ss.sub_shapes())
+                    self(child);
+            }
+        })(s);
+    }
+    // Sort by type size
+    std::vector<std::size_t> order(shapes.size());
+    std::iota(order.begin(), order.end(), 0);
+    std::sort(order.begin(), order.end(), by(std::greater<>{}, [&](auto i) {
+                  return shapes[i].type_size();
+              }));
+    // Compute offsets
+    std::unordered_map<std::size_t, std::size_t> offsets;
+    std::size_t offset = 0;
+    for(auto i : order)
+    {
+        offsets[i] = offset;
+        offset += shapes[i].bytes();
+    }
+    assert(offset == s.bytes());
+
+    // cppcheck-suppress variableScope
+    std::size_t i = 0;
+    return fix<argument>([&](auto self, auto ss) {
+        if(ss.sub_shapes().empty())
+        {
+            argument r{shapes[i], buffer + offsets[i]};
+            i++;
+            return r;
+        }
+        std::vector<argument> subs;
+        std::transform(ss.sub_shapes().begin(),
+                       ss.sub_shapes().end(),
+                       std::back_inserter(subs),
+                       [&](auto child) { return self(child); });
+        return argument{subs};
+    })(s);
+}
+
+std::vector<shape> to_shapes(const std::vector<argument>& args)
+{
+    std::vector<shape> shapes;
+    std::transform(args.begin(), args.end(), std::back_inserter(shapes), [](auto&& arg) {
+        return arg.get_shape();
+    });
+    return shapes;
+}
+
+argument::argument(const std::vector<argument>& args)
+    : m_shape(to_shapes(args)), m_data(data_t::from_args(args))
+{
+}
+
+char* argument::data() const
+{
+    assert(m_shape.type() != shape::tuple_type);
+    assert(not this->empty());
+    return m_data.get();
+}
+
+bool argument::empty() const { return not m_data.get and m_data.sub.empty(); }
+
+const shape& argument::get_shape() const { return this->m_shape; }
+
+argument argument::reshape(const shape& s) const { return {s, this->m_data}; }
+
+argument::data_t argument::data_t::share() const
+{
+    data_t result;
+    if(this->get)
+    {
+        auto self  = std::make_shared<data_t>(*this);
+        result.get = [self]() mutable { return self->get(); };
+    }
+    std::transform(sub.begin(), sub.end(), std::back_inserter(result.sub), [](const auto& d) {
+        return d.share();
+    });
+    return result;
+}
+
+argument::data_t argument::data_t::from_args(const std::vector<argument>& args)
+{
+    data_t result;
+    std::transform(args.begin(), args.end(), std::back_inserter(result.sub), [](auto&& arg) {
+        return arg.m_data;
+    });
+    return result;
+}
+
+argument argument::share() const { return {m_shape, m_data.share()}; }
+
+std::vector<argument> argument::get_sub_objects() const
+{
+    std::vector<argument> result;
+    assert(m_shape.sub_shapes().size() == m_data.sub.size());
+    std::transform(m_shape.sub_shapes().begin(),
+                   m_shape.sub_shapes().end(),
+                   m_data.sub.begin(),
+                   std::back_inserter(result),
+                   [](auto&& s, auto&& d) {
+                       return argument{s, d};
+                   });
+    return result;
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/dead_code_elimination.cpp

@@ -29,14 +29,16 @@ std::ptrdiff_t bidistance(const Range& r, Iterator start, Iterator last)
         return -n;
 }
 
-void dead_code_elimination::apply(module& p) const
+void dead_code_elimination::apply(program& p) const { p.remove_unused_modules(); }
+
+void dead_code_elimination::apply(module& m) const
 {
-    auto last = std::prev(p.end());
-    for(auto ins : iterator_for(p))
+    auto last = std::prev(m.end());
+    for(auto ins : iterator_for(m))
     {
         // Skip the first instruction, since we always process the previous
         // instruction
-        if(ins == p.begin())
+        if(ins == m.begin())
             continue;
         const auto i = std::prev(ins);
         // Skip the last instruction
@@ -46,9 +48,9 @@ void dead_code_elimination::apply(module& p) const
         if(i->get_shape().elements() == 0 and i->name().front() != '@' and
            i->name() != "undefined" and i->name() != "identity")
             continue;
-        assert(bidistance(p, i, last) > 0);
+        assert(bidistance(m, i, last) > 0);
         fix([&](auto self, auto leaf) {
-            if(not p.has_instruction(leaf))
+            if(not m.has_instruction(leaf))
                 return;
 
             if(leaf->outputs().empty())
@@ -56,15 +58,15 @@ void dead_code_elimination::apply(module& p) const
                 std::unordered_set<instruction_ref> args(leaf->inputs().begin(),
                                                          leaf->inputs().end());
                 leaf->clear_arguments();
-                assert(bidistance(p, last, leaf) < 0);
+                assert(bidistance(m, last, leaf) < 0);
                 assert(leaf != ins);
-                p.move_instruction(leaf, p.end());
+                m.move_instruction(leaf, m.end());
                 for(auto arg : args)
                     self(arg);
             }
         })(i);
     }
-    p.remove_instructions(std::next(last), p.end());
+    m.remove_instructions(std::next(last), m.end());
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/decompose.cpp

@@ -12,35 +12,44 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace {
+
+struct alpha_beta
+{
+    float alpha = 0.0;
+    float beta  = 0.0;
+};
+
+alpha_beta get_alpha_beta(const operation& op)
+{
+    auto v = op.to_value();
+    return {v.at("alpha").to<float>(), v.at("beta").to<float>()};
+}
+
 struct find_dot_add
 {
-    auto matcher() const { return match::name("dot")(match::nargs(3)); }
+    auto matcher() const { return match::name("dot", "quant_dot")(match::nargs(3)); }
 
     void apply(module& p, const match::matcher_result& r) const
     {
-        auto ins = r.result;
-        auto dot = any_cast<op::dot>(ins->get_operator());
-        if(not float_equal(dot.beta, 1) and
-           not contains({shape::float_type, shape::half_type, shape::double_type},
-                        ins->get_shape().type()))
-            return;
+        auto ins   = r.result;
+        auto dot   = get_alpha_beta(ins->get_operator());
         auto a_ins = ins->inputs()[0];
         auto b_ins = ins->inputs()[1];
         if(not float_equal(dot.alpha, 1))
         {
-            auto alpha = p.add_literal(literal{shape{ins->get_shape().type()}, {dot.alpha}});
+            auto alpha = p.add_literal(literal{shape{a_ins->get_shape().type()}, {dot.alpha}});
             auto alpha_broadcast = p.insert_instruction(
                 ins,
                 make_op("multibroadcast", {{"output_lens", a_ins->get_shape().lens()}}),
                 alpha);
             a_ins = p.insert_instruction(ins, make_op("mul"), a_ins, alpha_broadcast);
         }
-        auto dot_ins = p.insert_instruction(ins, make_op("dot", {{"beta", 0}}), a_ins, b_ins);
+        auto dot_ins = p.insert_instruction(ins, make_op(ins->name(), {{"beta", 0}}), a_ins, b_ins);
 
         auto c_ins = ins->inputs()[2];
         if(not float_equal(dot.beta, 1))
         {
-            auto beta = p.add_literal(literal{shape{ins->get_shape().type()}, {dot.beta}});
+            auto beta = p.add_literal(literal{shape{c_ins->get_shape().type()}, {dot.beta}});
             auto beta_broadcast = p.insert_instruction(
                 ins, make_op("multibroadcast", {{"output_lens", ins->get_shape().lens()}}), beta);
             c_ins = p.insert_instruction(ins, make_op("mul"), c_ins, beta_broadcast);
@@ -51,24 +60,24 @@ struct find_dot_add
 
 struct find_dot_alpha
 {
-    auto matcher() const { return match::name("dot")(match::nargs(2)); }
+    auto matcher() const { return match::name("dot", "quant_dot")(match::nargs(2)); }
 
     void apply(module& p, const match::matcher_result& r) const
     {
         auto ins   = r.result;
-        auto dot   = any_cast<op::dot>(ins->get_operator());
+        auto dot   = get_alpha_beta(ins->get_operator());
         auto a_ins = ins->inputs()[0];
         auto b_ins = ins->inputs()[1];
         if(not float_equal(dot.alpha, 1))
         {
-            auto alpha = p.add_literal(literal{shape{ins->get_shape().type()}, {dot.alpha}});
+            auto alpha = p.add_literal(literal{shape{a_ins->get_shape().type()}, {dot.alpha}});
             auto alpha_broadcast = p.insert_instruction(
                 ins,
                 make_op("multibroadcast", {{"output_lens", a_ins->get_shape().lens()}}),
                 alpha);
             a_ins = p.insert_instruction(ins, make_op("mul"), a_ins, alpha_broadcast);
         }
-        p.replace_instruction(ins, make_op("dot", {{"beta", 0}}), a_ins, b_ins);
+        p.replace_instruction(ins, make_op(ins->name(), {{"beta", 0}}), a_ins, b_ins);
     }
 };
 

src/eliminate_data_type.cpp

@@ -13,6 +13,8 @@ void eliminate_data_type::apply(module& m) const
     {
         if(ins->name()[0] == '@')
             continue;
+        if(ins->name() == "convert")
+            continue;
         auto inputs = ins->inputs();
         std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](auto i) {
             if(types.count(i->get_shape().type()) == 0)

src/include/migraphx/algorithm.hpp

@@ -7,6 +7,20 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
+template <class Iterator, class Output, class Predicate, class F>
+void transform_if(Iterator start, Iterator last, Output out, Predicate pred, F f)
+{
+    while(start != last)
+    {
+        if(pred(*start))
+        {
+            *out = f(*start);
+            ++out;
+        }
+        ++start;
+    }
+}
+
 template <class Iterator, class Output, class Predicate>
 void group_by(Iterator start, Iterator last, Output out, Predicate pred)
 {

src/include/migraphx/argument.hpp

@@ -8,6 +8,7 @@
 #include <functional>
 #include <utility>
 
+// clang-format off
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
@@ -20,57 +21,61 @@ inline namespace MIGRAPHX_INLINE_NS {
  */
 struct argument : raw_data<argument>
 {
-    argument() {}
+    argument() = default;
 
-    argument(const shape& s) : m_shape(s)
-    {
-        auto buffer = make_shared_array<char>(s.bytes());
-        data        = [=]() mutable { return buffer.get(); };
-    }
+    argument(const shape& s);
 
     template <class F, MIGRAPHX_REQUIRES(std::is_pointer<decltype(std::declval<F>()())>{})>
     argument(shape s, F d)
-        : data([f = std::move(d)]() mutable { return reinterpret_cast<char*>(f()); }),
-          m_shape(std::move(s))
+        : m_shape(std::move(s)),
+          m_data({[f = std::move(d)]() mutable { return reinterpret_cast<char*>(f()); }})
+
     {
     }
     template <class T>
     argument(shape s, T* d)
-        : data([d] { return reinterpret_cast<char*>(d); }), m_shape(std::move(s))
+        : m_shape(std::move(s)), m_data({[d] { return reinterpret_cast<char*>(d); }})
     {
     }
 
     template <class T>
     argument(shape s, std::shared_ptr<T> d)
-        : data([d] { return reinterpret_cast<char*>(d.get()); }), m_shape(std::move(s))
+        : m_shape(std::move(s)), m_data({[d] { return reinterpret_cast<char*>(d.get()); }})
     {
     }
 
-    argument(shape s, std::nullptr_t) : data([] { return nullptr; }), m_shape(std::move(s)) {}
+    argument(shape s, std::nullptr_t);
+    
+    argument(const std::vector<argument>& args);
+
+    static argument load(const shape& s, char* buffer);
 
     /// Provides a raw pointer to the data
-    std::function<char*()> data = nullptr;
+    char* data() const;
 
     /// Whether data is available
-    bool empty() const { return not data; }
+    bool empty() const;
 
-    const shape& get_shape() const { return this->m_shape; }
+    const shape& get_shape() const;
 
-    argument reshape(const shape& s) const
-    {
-        argument self = *this;
-        return {s, [=]() mutable { return self.data(); }};
-    }
+    argument reshape(const shape& s) const;
 
     /// Make copy of the argument that is always sharing the data
-    argument share() const
-    {
-        auto self = std::make_shared<argument>(*this);
-        return {m_shape, [self]() mutable { return self->data(); }};
-    }
+    argument share() const;
+
+    std::vector<argument> get_sub_objects() const;
 
     private:
+    struct data_t
+    {
+        std::function<char*()> get = nullptr;
+        std::vector<data_t> sub = {};
+        data_t share() const;
+        static data_t from_args(const std::vector<argument>& args);
+    };
+    argument(const shape& s, const data_t& d);
     shape m_shape;
+    data_t m_data{};
 };
 
 void migraphx_to_value(value& v, const argument& a);
@@ -78,5 +83,6 @@ void migraphx_from_value(const value& v, argument& a);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
+// clang-format on
 
 #endif

src/include/migraphx/dead_code_elimination.hpp

@@ -9,6 +9,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 struct module;
+struct program;
 
 /**
  * Remove instructions where the output is not used.
@@ -16,7 +17,8 @@ struct module;
 struct dead_code_elimination
 {
     std::string name() const { return "dead_code_elimination"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
+    void apply(program& p) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS