Update OSS repo (#2033)

Summary:
Pull Request resolved: #2033

Update the OSS Xtensa repo with more up to date compiler and quantizer things. Introduce a test folder and a conv1d test.

Reviewed By: tarun292, cccclai

Differential Revision: D54034581

fbshipit-source-id: c2bf0c43897a2ef7dff291698370d2583433a6ba

Author: mcremon-meta

backends/vulkan/test/op_tests/utils/codegen.py

@@ -165,7 +165,7 @@ def prepack_ref(self, ref: ValueRef) -> bool:
         else:
             return ref.supports_prepack and self.should_prepack
 
-    def create_value_for(self, ref: ValueRefList) -> str:
+    def create_value_for(self, ref: ValueRefList) -> str:  # noqa: C901
         if isinstance(ref, list):
             ret_str = ""
             for r in ref:

docs/source/build-run-xtensa.md

@@ -68,13 +68,14 @@ examples/xtensa/
 ├── aot
 ├── kernels
 ├── ops
+├── tests
 ├── third-party
 └── utils
 ```
 
 ***AoT (Ahead-of-Time) Components***:
 
-The AoT folder contains all of the python scripts and functions needed to export the model to an ExecuTorch `.pte` file. In our case, [export_example.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/export_example.py) defines a model and some example inputs (set to a vector of ones), and runs it through the quantizer (from [quantizer.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/quantizer.py)). Then a few compiler passes, also defined in [quantizer.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/quantizer.py), will replace operators with custom ones that are supported and optimized on the chip. Any operator needed to compute things should be defined in [meta_registrations.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/meta_registrations.py) and have corresponding implemetations in the other folders.
+The AoT folder contains all of the python scripts and functions needed to export the model to an ExecuTorch `.pte` file. In our case, [export_example.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/export_example.py) is an API that takes a model (nn.Module) and representative inputs and runs it through the quantizer (from [quantizer.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/quantizer.py)). Then a few compiler passes, also defined in [quantizer.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/quantizer.py), will replace operators with custom ones that are supported and optimized on the chip. Any operator needed to compute things should be defined in [meta_registrations.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/meta_registrations.py) and have corresponding implemetations in the other folders.
 
 ***Operators***:
 
@@ -97,17 +98,31 @@ cd executorch
 python3 -m examples.portable.scripts.export --model_name="add"
 ```
 
-***Quantized Linear***:
+***Quantized Operators***:
 
-The second, more complex model is a quantized [linear](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html) operation. The model is defined [here](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/export_example.py#L88). Linear is the backbone of most Automatic Speech Recognition (ASR) models.
+The other, more complex model are custom operators, including:
+  - a quantized [linear](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html) operation. The model is defined [here](https://github.com/pytorch/executorch/blob/main/examples/xtensa/tests/quantized_linear_example.py#L28). Linear is the backbone of most Automatic Speech Recognition (ASR) models.
+  - a quantized [conv1d](https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html) operation. The model is defined [here](https://github.com/pytorch/executorch/blob/main/examples/xtensa/tests/quantized_conv1d_example.py#L36). Convolutions are important in wake word and many denoising models.
 
-The generated file is called `XtensaDemoModel.pte`.
+In both cases the generated file is called `XtensaDemoModel.pte`.
+
+```bash
+cd executorch
+python3 -m examples.xtensa.tests.quantized_<linear,conv1d>_example
+```
+
+***Small Model: RNNT predictor***:
+
+The torchaudio [RNNT-emformer](https://pytorch.org/audio/stable/tutorials/online_asr_tutorial.html) model is an Automatic Speech Recognition (ASR) model, comprised of three different submodels: an encoder, a predictor and a joiner.
+The predictor is a sequence of basic ops (embedding, ReLU, linear, layer norm) and can be exported using:
 
 ```bash
 cd executorch
-python3 -m examples.xtensa.aot.export_example
+python3 -m examples.xtensa.tests.rnnt_predictor_quantized_example
 ```
 
+The generated file is called `XtensaDemoModel.pte`.
+
 ### Runtime
 
 **Building the DSP firmware image**
@@ -139,12 +154,14 @@ cmake -DBUCK2=buck2 \
     -DCMAKE_TOOLCHAIN_FILE=<path_to_executorch>/examples/xtensa/xtensa.cmake \
     -DCMAKE_INSTALL_PREFIX=cmake-out \
     -DCMAKE_BUILD_TYPE=Debug \
+    -DPYTHON_EXECUTABLE=python3 \
+    -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
     -DEXECUTORCH_BUILD_HOST_TARGETS=ON \
     -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
+    -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+    -DEXECUTORCH_BUILD_CPUINFO=OFF \
     -DEXECUTORCH_BUILD_FLATC=OFF \
     -DFLATC_EXECUTABLE="$(which flatc)" \
-    -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
-    -DPYTHON_EXECUTABLE=python3 \
     -Bcmake-out .
 
 cmake --build cmake-out -j8 --target install --config Debug
@@ -196,6 +213,6 @@ First 20 elements of output 0
 
 In this tutorial, you have learned how to export a quantized operation, build the ExecuTorch runtime and run this model on the Xtensa HiFi4 DSP chip.
 
-The model in this tutorial is a typical operation appearing in ASR models, and can be extended to a complete ASR model by creating the model in [export_example.py](https://github.com/pytorch/executorch/blob/main/examples/xtensa/aot/export_example.py) and adding the needed operators/kernels to [operators](https://github.com/pytorch/executorch/blob/main/examples/xtensa/ops) and [kernels](https://github.com/pytorch/executorch/blob/main/examples/xtensa/kernels).
+The (quantized linear) model in this tutorial is a typical operation appearing in ASR models, and can be extended to a complete ASR model by creating the model as a new test and adding the needed operators/kernels to [operators](https://github.com/pytorch/executorch/blob/main/examples/xtensa/ops) and [kernels](https://github.com/pytorch/executorch/blob/main/examples/xtensa/kernels).
 
 Other models can be created following the same structure, always assuming that operators and kernels are available.

examples/xtensa/aot/compiler.py

@@ -0,0 +1,68 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from typing import Any, Callable
+
+import torch
+
+from executorch.exir import EdgeCompileConfig, EdgeProgramManager, to_edge
+
+from torch.export import export
+from torch.export.exported_program import ExportedProgram
+
+
+def export_program(
+    model: Callable,
+    inputs: Any,
+    pt2_quant: bool = False,
+) -> ExportedProgram:
+    # we don't support training mode. Make it eval
+    if hasattr(model, "eval"):
+        if pt2_quant:
+            # pyre-fixme[6]: Incompatible parameter type.
+            torch.ao.quantization.move_exported_model_to_eval(model)
+        else:
+            # pyre-fixme[16]: Anonymous callable has no attribute `eval`.
+            model.eval()
+
+    # if it's already an ExportedProgram, just return it
+    if isinstance(model, ExportedProgram):
+        return model
+
+    assert isinstance(model, torch.nn.Module), "model should be an nn.Module"
+
+    # Prevent mkldnn decompositions
+    torch._C._set_mkldnn_enabled(False)
+
+    # else: capture the model and return it.
+    return export(model, inputs)
+
+
+# Export the model and lower it it edge IR.
+def export_to_edge(
+    model: Callable,
+    inputs: Any,
+    pt2_quant: bool = False,
+    dump_graphs: bool = False,
+) -> EdgeProgramManager:
+    # Export the model into an ExportedProgram.
+    expo_program = export_program(model, inputs, pt2_quant)
+
+    if dump_graphs:
+        logging.info(f"Exported graph:\n{expo_program.graph_module.graph}")
+
+    # Call to_edge to convert the graph to edge IR.
+    edge_prog_manager = to_edge(
+        expo_program, compile_config=EdgeCompileConfig(_check_ir_validity=False)
+    )
+
+    if dump_graphs:
+        logging.info(
+            f"Edge graph:\n{edge_prog_manager.exported_program().graph_module.graph}"
+        )
+
+    return edge_prog_manager

examples/xtensa/aot/export_example.py

@@ -10,47 +10,27 @@
 
 from .meta_registrations import *  # noqa
 
-import torch
-from executorch.exir import EdgeCompileConfig
 from torch._export import capture_pre_autograd_graph
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
-from ...portable.utils import export_to_edge, save_pte_program
+from ...portable.utils import save_pte_program
 
+from .compiler import export_to_edge
 from .quantizer import (
     QuantFusion,
     ReplacePT2DequantWithXtensaDequant,
     ReplacePT2QuantWithXtensaQuant,
-    XtensaQuantizer,
+    XtensaBaseQuantizer,
 )
 
 
 FORMAT = "[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s"
 logging.basicConfig(level=logging.INFO, format=FORMAT)
 
 
-if __name__ == "__main__":
-    in_features = 32
-    out_features = 16
-    bias = True
-    shape = [64, in_features]
-
-    class QuantizedLinear(torch.nn.Module):
-        def __init__(self, in_features: int, out_features: int, bias: bool):
-            super().__init__()
-            self.output_linear = torch.nn.Linear(in_features, out_features, bias=bias)
-
-        def forward(self, x: torch.Tensor):
-            output_linear_out = self.output_linear(x)
-            return output_linear_out
-
-    model = QuantizedLinear(in_features, out_features, bias)
-    model.eval()
-
-    example_inputs = (torch.ones(shape),)
-
+def export_xtensa_model(model, example_inputs):
     # Quantizer
-    quantizer = XtensaQuantizer()
+    quantizer = XtensaBaseQuantizer()
 
     # Export
     model_exp = capture_pre_autograd_graph(model, example_inputs)
@@ -66,29 +46,20 @@ def forward(self, x: torch.Tensor):
     patterns = [q.pattern for q in quantizer.quantizers]
     QuantFusion(patterns)(converted_model)
 
-    # pre-autograd export. eventually this will become torch.export
-    converted_model_exp = capture_pre_autograd_graph(converted_model, example_inputs)
+    # Get edge program (note: the name will change to export_to_xtensa in future PRs)
+    edge_prog_manager = export_to_edge(converted_model, example_inputs, pt2_quant=True)
 
-    converted_model_exp = torch.ao.quantization.move_exported_model_to_eval(
-        converted_model_exp
+    # Run a couple required passes for quant/dequant ops
+    xtensa_prog_manager = edge_prog_manager.transform(
+        [ReplacePT2QuantWithXtensaQuant(), ReplacePT2DequantWithXtensaDequant()],
+        check_ir_validity=False,
     )
 
-    exec_prog = (
-        export_to_edge(
-            converted_model_exp,
-            example_inputs,
-            edge_compile_config=EdgeCompileConfig(
-                _check_ir_validity=False,
-            ),
-        )
-        .transform(
-            [ReplacePT2QuantWithXtensaQuant(), ReplacePT2DequantWithXtensaDequant()],
-            check_ir_validity=False,
-        )
-        .to_executorch()
-    )
+    exec_prog = xtensa_prog_manager.to_executorch()
 
-    logging.info(f"Final exported graph:\n{exec_prog.exported_program().graph}")
+    logging.info(
+        f"Final exported graph module:\n{exec_prog.exported_program().graph_module}"
+    )
 
     # Save the program as XtensaDemoModel.pte
     save_pte_program(exec_prog, "XtensaDemoModel")

examples/xtensa/aot/meta_registrations.py

@@ -4,10 +4,14 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from typing import Optional, Tuple
+
 import torch
 from executorch.exir.scalar_type import ScalarType
 from torch.library import impl, Library
 
+from .utils import get_conv1d_output_size
+
 lib = Library("xtensa", "DEF")
 
 lib.define(
@@ -25,10 +29,31 @@
 )
 
 lib.define(
-    "quantized_linear_pt2(Tensor src, Tensor weight, Tensor bias, float src_scale, int src_zero_point, float weight_scale, int weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point) -> (Tensor Z)"
+    "quantized_layer_norm(Tensor X, Tensor X_scale, Tensor X_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point) -> (Tensor Y)"
+)
+
+lib.define(
+    "quantized_layer_norm.out(Tensor X, Tensor X_scale, Tensor X_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point, *, Tensor(a!) out) -> Tensor (a!)"
+)
+
+lib.define(
+    "quantized_linear(Tensor src, Tensor weight, Tensor bias, int src_zero_point, Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset) -> (Tensor Z)"
+)
+lib.define(
+    "quantized_linear.out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) ->  Tensor(a!)"
+)
+
+lib.define("quantized_relu(Tensor X, Tensor X_zero_point) -> (Tensor Y)")
+
+lib.define(
+    "quantized_relu.out(Tensor X, Tensor X_zero_point, *, Tensor(a!) out) -> Tensor (a!)"
+)
+
+lib.define(
+    "quantized_conv(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, bool channel_last=False) -> (Tensor Z)"
 )
 lib.define(
-    "quantized_linear_pt2.out(Tensor src, Tensor weight, Tensor bias, float src_scale, int src_zero_point, float weight_scale, int weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, *, Tensor(a!) out) ->  Tensor(a!)"
+    "quantized_conv.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)"
 )
 
 m = Library("xtensa", "IMPL", "Meta")
@@ -58,18 +83,17 @@ def dequantize_per_tensor_meta(
     return input.new_empty(input.size(), dtype=torch.float)
 
 
-@impl(m, "quantized_linear_pt2")
-def quantized_linear_pt2_meta(
+@impl(m, "quantized_linear")
+def quantized_linear_meta(
     src: torch.Tensor,
     weight: torch.Tensor,
     bias: torch.Tensor,
-    in_scale: float,
     in_zero_point: int,
-    weight_scale: float,
-    weight_zero_point: int,
-    out_multiplier: int,
-    out_shift: int,
+    weight_zero_point: torch.Tensor,
+    out_multiplier: torch.Tensor,
+    out_shift: torch.Tensor,
     out_zero_point: int,
+    offset: Optional[torch.Tensor],
 ):
     # src comes in shape [leading_dims, in_dim]
     # weight comes in shape [out_dim, in_dim]
@@ -79,3 +103,58 @@ def quantized_linear_pt2_meta(
     assert len(weight_size) == 2
     out_size[-1] = weight_size[0]
     return src.new_empty(out_size, dtype=torch.uint8)
+
+
+@impl(m, "quantized_conv")
+def quantized_conv_meta(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    stride: Tuple[int],
+    padding: Tuple[int],
+    dilation: Tuple[int],
+    groups: int,
+    in_zero_point: int,
+    weight_zero_point: torch.Tensor,
+    bias_scale: torch.Tensor,
+    output_scale: float,
+    output_zero_point: int,
+    out_multiplier: torch.Tensor,
+    out_shift: torch.Tensor,
+    channel_last: bool = False,
+):
+    out_channels, _in_channels, *kernel_size = weight.shape
+    in_size = input.shape
+    # Assert that the input tensor has at least 3 dimensions, and at most 6
+    assert len(in_size) > 2
+    assert len(in_size) < 6
+
+    # Compute the output tensor size
+    output_size = get_conv1d_output_size(
+        in_size, out_channels, stride[0], padding[0], dilation[0], kernel_size[0]
+    )
+
+    return input.new_empty(output_size, dtype=input.dtype)
+
+
+@impl(m, "quantized_layer_norm")
+def quantized_layer_norm_meta(
+    input: torch.Tensor,
+    X_scale: torch.Tensor,
+    X_zero_point: torch.Tensor,
+    normalized_shape: int,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float,
+    output_scale: float,
+    output_zero_point: int,
+):
+    return input.new_empty(input.size(), dtype=torch.uint8)
+
+
+@impl(m, "quantized_relu")
+def quantized_relu_meta(
+    X: torch.Tensor,
+    X_zero_point: torch.Tensor,
+):
+    return X.new_empty(X.size(), dtype=torch.uint8)