Add embedding ops aten (pytorch#1129)

Summary:
Pull Request resolved: pytorch#1129

Adds embedding ops for aten

Reviewed By: digantdesai, Jack-Khuu

Differential Revision: D64477035

torchao/experimental/CMakeLists.txt

@@ -33,11 +33,13 @@ if(CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
     # Defines target torchao_kernels_aarch64
     add_subdirectory(kernels/cpu/aarch64)
     add_subdirectory(ops/linear_8bit_act_xbit_weight)
+    add_subdirectory(ops/embedding_xbit)
 
     add_library(torchao_ops_aten SHARED)
     target_link_libraries(
         torchao_ops_aten PRIVATE
         torchao_ops_linear_8bit_act_xbit_weight_aten
+        torchao_ops_embedding_xbit_aten
     )
     install(
         TARGETS torchao_ops_aten

torchao/experimental/ops/embedding_xbit/CMakeLists.txt

@@ -0,0 +1,19 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+cmake_minimum_required(VERSION 3.19)
+
+include(${CMAKE_CURRENT_SOURCE_DIR}/../../Utils.cmake)
+
+find_package(Torch REQUIRED)
+add_library(torchao_ops_embedding_xbit_aten OBJECT
+    op_embedding_xbit_aten.cpp
+)
+target_link_torchao_parallel_backend(torchao_ops_embedding_xbit_aten "aten_openmp")
+target_link_libraries(torchao_ops_embedding_xbit_aten PRIVATE torchao_kernels_aarch64)
+target_include_directories(torchao_ops_embedding_xbit_aten PRIVATE "${TORCH_INCLUDE_DIRS}")
+target_link_libraries(torchao_ops_embedding_xbit_aten PRIVATE "${TORCH_LIBRARIES}")
+target_compile_definitions(torchao_ops_embedding_xbit_aten PRIVATE USE_ATEN=1)

torchao/experimental/ops/embedding_xbit/op_embedding_xbit-impl.h

@@ -0,0 +1,268 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the license found in the
+// LICENSE file in the root directory of this source tree.
+
+#pragma once
+
+#if defined(__aarch64__) || defined(__ARM_NEON)
+#include <torchao/experimental/kernels/cpu/aarch64/embedding/embedding.h>
+#endif // defined(__aarch64__) || defined(__ARM_NEON)
+
+#include <torchao/experimental/ops/embedding_xbit/packed_weights_header.h>
+#include <torchao/experimental/ops/library.h>
+#include <torchao/experimental/ops/packed_weights_header.h>
+#include <torchao/experimental/ops/parallel.h>
+
+template <int weight_nbit>
+void check_embedding_inputs(
+    const Tensor& packed_weight_qvals,
+    int num_embeddings,
+    int embedding_dim,
+    const Tensor& weight_scales,
+    const Tensor& weight_zeros,
+    const Tensor& indices,
+    int& group_size) {
+  TORCHAO_CHECK(
+      packed_weight_qvals.dim() == 1, "packed_weight_qvals must be 1D");
+#ifdef USE_ATEN
+  TORCHAO_CHECK(
+      packed_weight_qvals.dtype() == torch::kInt8,
+      "packed_weight_qvals must be byte");
+#endif // USE_ATEN
+  TORCHAO_CHECK(
+      (embedding_dim * weight_nbit) % 8 == 0,
+      "embedding_dim * weight_nbit must be a multiple of 8");
+  int packed_embedding_dim = (embedding_dim * weight_nbit) / 8;
+  TORCHAO_CHECK(
+      packed_weight_qvals.size(0) ==
+          (torchao::ops::PackedWeightsHeader::size() +
+           (num_embeddings * packed_embedding_dim)),
+      "packed_weight_qvals is not the correct size");
+
+  // Check header
+  auto header = torchao::ops::PackedWeightsHeader::read(
+      packed_weight_qvals.const_data_ptr());
+  TORCHAO_CHECK(
+      header ==
+          torchao::ops::embedding_xbit::get_packed_weights_header_universal(
+              weight_nbit,
+              /*min_value_chunk_size=*/32,
+              /*max_value_chunk_size=*/128),
+      "packed_weights are not compatible with the kernel");
+
+#ifdef USE_ATEN
+  TORCHAO_CHECK(
+      weight_scales.dtype() == torch::kFloat32,
+      "weight_scales must be float32");
+#endif // USE_ATEN
+  TORCHAO_CHECK(weight_scales.dim() == 2, "weight_scales must be 2D");
+  TORCHAO_CHECK(
+      weight_scales.size(0) == num_embeddings,
+      "weight_scales must be same shape as packed_weight_qvals in dim0 (num_embeddings)");
+  int num_groups = weight_scales.size(1);
+  TORCHAO_CHECK(
+      num_groups >= 1, "weight_scales must be at least 1 in dim1 (num_groups)");
+  TORCHAO_CHECK(
+      embedding_dim % num_groups == 0,
+      "embedding_dim must be a multiple of num_groups");
+  group_size = embedding_dim / num_groups;
+  TORCHAO_CHECK(group_size % 32 == 0, "group_size must be a multiple of 32");
+
+#ifdef USE_ATEN
+  TORCHAO_CHECK(
+      weight_zeros.dtype() == torch::kInt8, "weight_zeros must be int8");
+#endif // USE_ATEN
+  TORCHAO_CHECK(weight_zeros.dim() == 2, "weight_zeros must be 2D");
+  TORCHAO_CHECK(
+      weight_zeros.size(0) == weight_scales.size(0) &&
+          weight_zeros.size(1) == weight_scales.size(1),
+      "zeros must be same shape as scales");
+
+  TORCHAO_CHECK(indices.dim() == 1, "indices must be 1D");
+  TORCHAO_CHECK(
+      (indices.dtype() == Tensor_dtype_kInt32) ||
+          (indices.dtype() == Tensor_dtype_kInt64),
+      "indices must be int32 or int64");
+}
+
+#if defined(USE_ATEN) || defined(USE_EXECUTORCH)
+template <int weight_nbit>
+Tensor embedding_out_cpu(
+    const Tensor& packed_weight_qvals,
+    // TODO(T200095131): convert to
+    // int64_t when supported by AOTI
+    // Currently they are tensors with size
+    // equal to (0, the int they wrap)
+    const Tensor& num_embeddings_tensor,
+    const Tensor& embedding_dim_tensor,
+    const Tensor& weight_scales,
+    const Tensor& weight_zeros,
+    const Tensor& indices,
+    Tensor& out) {
+  int num_embeddings = num_embeddings_tensor.size(1);
+  int embedding_dim = embedding_dim_tensor.size(1);
+  int group_size;
+  check_embedding_inputs<weight_nbit>(
+      packed_weight_qvals,
+      num_embeddings,
+      embedding_dim,
+      weight_scales,
+      weight_zeros,
+      indices,
+      group_size);
+
+  int num_out = indices.size(0);
+  const int8_t* weight_zeros_ptr = weight_zeros.const_data_ptr<int8_t>();
+
+#ifdef USE_ATEN
+  TORCHAO_CHECK(out.dtype() == torch::kFloat32, "out must be float32");
+  out.resize_({num_out, embedding_dim});
+#endif // USE_ATEN
+
+#ifdef USE_EXECUTORCH
+  TORCHAO_CHECK(out.dim() == 2, "out must be 2D");
+  TORCHAO_CHECK(out.size(0) == num_out, "out shape is incorrect");
+  TORCHAO_CHECK(out.size(1) == embedding_dim, "out shape is incorrect");
+#endif // USE_EXECUTORCH
+
+  const int32_t* index32_ptr = nullptr;
+  const int64_t* index64_ptr = nullptr;
+  if (indices.dtype() == Tensor_dtype_kInt32) {
+    index32_ptr = indices.const_data_ptr<int32_t>();
+  } else {
+    TORCHAO_CHECK(
+        indices.dtype() == Tensor_dtype_kInt64,
+        "indices must be int32 or int64");
+    index64_ptr = indices.const_data_ptr<int64_t>();
+  }
+  torchao::parallel_1d(0, num_out, [&](int64_t idx) {
+    int index = -1;
+    if (index32_ptr != nullptr) {
+      index = index32_ptr[idx];
+    } else {
+      index = index64_ptr[idx];
+    }
+    TORCHAO_CHECK(index >= 0 && index < num_embeddings, "index out of bounds");
+#if defined(__aarch64__) || defined(__ARM_NEON)
+    torchao::kernels::cpu::aarch64::embedding::embedding<weight_nbit>(
+        out.mutable_data_ptr<float>() + idx * embedding_dim,
+        embedding_dim,
+        group_size,
+        packed_weight_qvals.const_data_ptr<int8_t>() +
+            torchao::ops::PackedWeightsHeader::size(),
+        weight_scales.const_data_ptr<float>(),
+        weight_zeros_ptr,
+        index);
+#else
+    TORCHAO_CHECK(false, "Unsupported platform");
+#endif // defined(__aarch64__) || defined(__ARM_NEON)
+  });
+
+  return out;
+}
+#endif // defined(USE_ATEN) || defined(USE_EXECUTORCH)
+
+#ifdef USE_ATEN
+template <int weight_nbit>
+Tensor embedding_cpu(
+    const Tensor& packed_weight_qvals,
+    // TODO(T200095131): convert to
+    // int64_t when supported by AOTI
+    // Currently they are tensors with size
+    // equal to (0, the int they wrap)
+    const Tensor& num_embeddings_tensor,
+    const Tensor& embedding_dim_tensor,
+    const Tensor& weight_scales,
+    const Tensor& weight_zeros,
+    const Tensor& indices) {
+  Tensor output_tensor = torch::empty({}, torch::kFloat32);
+  embedding_out_cpu<weight_nbit>(
+      packed_weight_qvals,
+      num_embeddings_tensor,
+      embedding_dim_tensor,
+      weight_scales,
+      weight_zeros,
+      indices,
+      output_tensor);
+  return output_tensor;
+}
+#endif // USE_ATEN
+
+#ifdef USE_ATEN
+template <int weight_nbit>
+Tensor embedding_meta(
+    const Tensor& packed_weight_qvals,
+    // TODO(T200095131): convert to
+    // int64_t when supported by AOTI
+    // Currently they are tensors with size
+    // equal to (0, the int they wrap)
+    const Tensor& num_embeddings_tensor,
+    const Tensor& embedding_dim_tensor,
+    const Tensor& weight_scales,
+    const Tensor& weight_zeros,
+    const Tensor& indices) {
+  int embedding_dim = embedding_dim_tensor.size(1);
+  int num_out = indices.size(0);
+  return torch::empty({num_out, embedding_dim}).to("meta");
+}
+#endif // USE_ATEN
+
+#ifdef USE_ATEN
+template <int weight_nbit>
+Tensor pack_embedding_cpu(const Tensor& weight_qvals) {
+  TORCHAO_CHECK(weight_qvals.dim() == 2, "weight_qvals must be 2D");
+  int num_embeddings = weight_qvals.size(0);
+  int embedding_dim = weight_qvals.size(1);
+  TORCHAO_CHECK(
+      embedding_dim % 8 == 0, "embedding_dim must be a multiple of 8 to pack");
+  int packed_embedding_dim = embedding_dim * weight_nbit / 8;
+  TORCHAO_CHECK(
+      weight_qvals.dtype() == torch::kInt8, "weight_qvals must be int8");
+
+  auto out = torch::empty(
+                 torchao::ops::PackedWeightsHeader::size() +
+                 (num_embeddings * packed_embedding_dim))
+                 .to(torch::kInt8);
+
+  auto header =
+      torchao::ops::embedding_xbit::get_packed_weights_header_universal(
+          weight_nbit,
+          /*min_value_chunk_size=*/32,
+          /*max_value_chunk_size=*/128);
+  header.write(out.mutable_data_ptr());
+
+  torchao::parallel_1d(0, num_embeddings, [&](int64_t idx) {
+#if defined(__aarch64__) || defined(__ARM_NEON)
+    torchao::kernels::cpu::aarch64::embedding::pack_embedding_weight_qvals<
+        weight_nbit>(
+        out.mutable_data_ptr<int8_t>() +
+            torchao::ops::PackedWeightsHeader::size(),
+        embedding_dim,
+        weight_qvals.const_data_ptr<int8_t>(),
+        idx);
+#else
+    TORCHAO_CHECK(false, "Unsupported platform");
+#endif // defined(__aarch64__) || defined(__ARM_NEON)
+  });
+
+  return out;
+}
+#endif // USE_ATEN
+
+#ifdef USE_ATEN
+template <int weight_nbit>
+Tensor pack_embedding_meta(const Tensor& weight_qvals) {
+  TORCHAO_CHECK(weight_qvals.dim() == 2, "weight_qvals must be 2D");
+  int num_embeddings = weight_qvals.size(0);
+  int embedding_dim = weight_qvals.size(1);
+  TORCHAO_CHECK(
+      embedding_dim % 8 == 0, "embedding_dim must be a multiple of 8 to pack");
+  int packed_embedding_dim = embedding_dim * weight_nbit / 8;
+  return torch::empty(
+             torchao::ops::PackedWeightsHeader::size() +
+             (num_embeddings * packed_embedding_dim))
+      .to("meta");
+}
+#endif // USE_ATEN

torchao/experimental/ops/embedding_xbit/op_embedding_xbit_aten.cpp

@@ -0,0 +1,56 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <torchao/experimental/ops/embedding_xbit/op_embedding_xbit-impl.h>
+
+#define DEFINE_OP(weight_nbit)                                                                                                                                            \
+  m.def("_pack_embedding_" #weight_nbit "bit(Tensor weight_qvals) -> Tensor");                                                                                            \
+  m.def(                                                                                                                                                                  \
+      "_embedding_" #weight_nbit                                                                                                                                          \
+      "bit(Tensor packed_weight_qvals, Tensor num_embeddings_tensor, Tensor embedding_dim_tensor, Tensor weight_scales, Tensor weight_zeros, Tensor indices) -> Tensor"); \
+  m.def(                                                                                                                                                                  \
+      "_embedding_" #weight_nbit                                                                                                                                          \
+      "bit.out(Tensor packed_weight_qvals, Tensor num_embeddings_tensor, Tensor embedding_dim_tensor, Tensor weight_scales, Tensor weight_zeros, Tensor indices, *, Tensor(a!) out) -> Tensor(a!)");
+
+#define DEFINE_CPU_IMPL(weight_nbit)                                     \
+  m.impl(                                                                \
+      "_pack_embedding_" #weight_nbit "bit",                             \
+      &pack_embedding_cpu<weight_nbit>);                                 \
+  m.impl("_embedding_" #weight_nbit "bit", &embedding_cpu<weight_nbit>); \
+  m.impl("_embedding_" #weight_nbit "bit.out", &embedding_out_cpu<weight_nbit>);
+
+#define DEFINE_META_IMPL(weight_nbit)        \
+  m.impl(                                    \
+      "_pack_embedding_" #weight_nbit "bit", \
+      &pack_embedding_meta<weight_nbit>);    \
+  m.impl("_embedding_" #weight_nbit "bit", &embedding_meta<weight_nbit>);
+
+TORCH_LIBRARY_FRAGMENT(torchao, m) {
+  DEFINE_OP(1);
+  DEFINE_OP(2);
+  DEFINE_OP(3);
+  DEFINE_OP(4);
+  DEFINE_OP(5);
+  DEFINE_OP(6);
+}
+
+TORCH_LIBRARY_IMPL(torchao, CPU, m) {
+  DEFINE_CPU_IMPL(1);
+  DEFINE_CPU_IMPL(2);
+  DEFINE_CPU_IMPL(3);
+  DEFINE_CPU_IMPL(4);
+  DEFINE_CPU_IMPL(5);
+  DEFINE_CPU_IMPL(6);
+}
+
+TORCH_LIBRARY_IMPL(torchao, Meta, m) {
+  DEFINE_META_IMPL(1);
+  DEFINE_META_IMPL(2);
+  DEFINE_META_IMPL(3);
+  DEFINE_META_IMPL(4);
+  DEFINE_META_IMPL(5);
+  DEFINE_META_IMPL(6);
+}

torchao/experimental/ops/embedding_xbit/packed_weights_header.h

@@ -0,0 +1,36 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the license found in the
+// LICENSE file in the root directory of this source tree.
+
+#pragma once
+#include <torchao/experimental/ops/library.h>
+#include <torchao/experimental/ops/packed_weights_header.h>
+
+namespace torchao::ops::embedding_xbit {
+
+inline torchao::ops::PackedWeightsHeader get_packed_weights_header_universal(
+    int weight_nbit,
+    int min_value_chunk_size,
+    int max_value_chunk_size,
+    int version = 1) {
+  return torchao::ops::PackedWeightsHeader(
+      torchao::ops::PackedWeightsFormat::embedding_xbit_universal,
+      {version,
+       weight_nbit,
+       min_value_chunk_size,
+       max_value_chunk_size,
+       0,
+       0,
+       0,
+       0,
+       0,
+       0,
+       0,
+       0,
+       0,
+       0});
+}
+
+} // namespace torchao::ops::embedding_xbit