sglang_v0.5.2/pytorch_2.8.0/third_party/XNNPACK/test/rope.cc

// Copyright 2023 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <algorithm>  // For std::generate, std::min.
#include <array>      // For std::array.
#include <cmath>      // For std::lrintf.
#include <cstddef>    // For size_t.
#include <cstdint>    // For uint32_t.
#include <memory>     // For std::unique_ptr.
#include <random>     // For std::uniform_real_distribution.
#include <vector>     // For std::vector.

#include <gtest/gtest.h>
#include "xnnpack.h"
#include "xnnpack/math.h"
#include "xnnpack/node-type.h"
#include "xnnpack/operator.h"
#include "xnnpack/subgraph.h"
#include "xnnpack/buffer.h"
#include "replicable_random_device.h"

template <class T> class RoPETestBase : public ::testing::Test {
 protected:
  RoPETestBase() {
    f32dist = std::uniform_real_distribution<float>(0.1f, 1.0f);
    dim_dist = std::uniform_int_distribution<size_t>(5, 15);

    batch_size = dim_dist(rng);
    tokens = dim_dist(rng);
    do {
      max_tokens = dim_dist(rng);
    } while (max_tokens < tokens);
    heads = dim_dist(rng);
    channels = dim_dist(rng) * 2;  // ensure the number of channels is even

    input = xnnpack::Buffer<T>(XNN_EXTRA_BYTES / sizeof(T) +
                           batch_size * tokens * heads * channels);
    weights =
        xnnpack::Buffer<T>(XNN_EXTRA_BYTES / sizeof(T) + max_tokens * channels);
    operator_output = xnnpack::Buffer<T>(batch_size * tokens * heads * channels);
    subgraph_output = xnnpack::Buffer<T>(operator_output.size());
  }

  xnnpack::ReplicableRandomDevice rng;
  std::uniform_real_distribution<float> f32dist;
  std::uniform_int_distribution<size_t> dim_dist;

  size_t batch_size;
  size_t max_tokens;
  size_t tokens;
  size_t heads;
  size_t channels;

  xnnpack::Buffer<T> input;
  xnnpack::Buffer<T> weights;
  xnnpack::Buffer<T> operator_output;
  xnnpack::Buffer<T> subgraph_output;
};

using RoPETestF16 = RoPETestBase<xnn_float16>;
using RoPETestF32 = RoPETestBase<float>;

TEST_F(RoPETestF16, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(3, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> input_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(),
                            /*data=*/nullptr, /*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t weights_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 2> weights_dims{{max_tokens, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp16, weights_dims.size(), weights_dims.data(),
                            weights.data(), /*external_id=*/1, /*flags=*/0, &weights_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> output_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/2, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success,
    xnn_define_rope(subgraph, max_tokens, input_id, weights_id, output_id, /*flags=*/0));

  ASSERT_EQ(subgraph->num_nodes, 1);
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_rope);
  ASSERT_EQ(node->num_inputs, 2);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], weights_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(RoPETestF32, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(3, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> input_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(),
                            /*data=*/nullptr, /*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t weights_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 2> weights_dims{{max_tokens, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp32, weights_dims.size(), weights_dims.data(),
                            weights.data(), /*external_id=*/1, /*flags=*/0, &weights_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> output_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/2, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success,
    xnn_define_rope(subgraph, max_tokens, input_id, weights_id, output_id, /*flags=*/0));

  ASSERT_EQ(subgraph->num_nodes, 1);
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_rope);
  ASSERT_EQ(node->num_inputs, 2);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], weights_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(RoPETestF16, matches_operator_api)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
  std::generate(weights.begin(), weights.end(), [&]() { return f32dist(rng); });

  const xnn_status status = xnn_create_rope_nthc_f16(/*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);

  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(xnn_status_success,
    xnn_reshape_rope_nthc_f16(op,
      batch_size, tokens, heads, channels,
      /*threadpool=*/nullptr));

  ASSERT_EQ(xnn_status_success,
    xnn_setup_rope_nthc_f16(op,
      input.data(), weights.data(), operator_output.data()));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(3, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> input_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(),
                            /*data=*/nullptr, /*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t weights_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 2> weights_dims{{max_tokens, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp16, weights_dims.size(), weights_dims.data(),
                            weights.data(), /*external_id=*/1, /*flags=*/0, &weights_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> output_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(),
                          /*data=*/nullptr, /*external_id=*/2, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success,
    xnn_define_rope(subgraph, max_tokens, input_id, weights_id, output_id, /*flags=*/0));

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);

  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);

  const std::array<xnn_external_value, 2> external{{
    xnn_external_value{input_id, input.data()},
    xnn_external_value{output_id, subgraph_output.data()}
  }};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));

  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  // Check outputs match.
  for (size_t i = 0; i < operator_output.size(); i++) {
    ASSERT_EQ(subgraph_output[i], operator_output[i]);
  }
}

TEST_F(RoPETestF32, matches_operator_api)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
  std::generate(weights.begin(), weights.end(), [&]() { return f32dist(rng); });

  const xnn_status status = xnn_create_rope_nthc_f32(/*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);

  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(xnn_status_success,
    xnn_reshape_rope_nthc_f32(op,
      batch_size, tokens, heads, channels,
      /*threadpool=*/nullptr));

  ASSERT_EQ(xnn_status_success,
    xnn_setup_rope_nthc_f32(op,
      input.data(), weights.data(), operator_output.data()));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(3, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> input_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(),
                            /*data=*/nullptr, /*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t weights_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 2> weights_dims{{max_tokens, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp32, weights_dims.size(), weights_dims.data(),
                            weights.data(), /*external_id=*/1, /*flags=*/0, &weights_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> output_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(),
                          /*data=*/nullptr, /*external_id=*/2, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success,
    xnn_define_rope(subgraph, max_tokens, input_id, weights_id, output_id, /*flags=*/0));

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);

  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);

  const std::array<xnn_external_value, 2> external{{
    xnn_external_value{input_id, input.data()},
    xnn_external_value{output_id, subgraph_output.data()}
  }};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));

  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  // Check outputs match.
  for (size_t i = 0; i < operator_output.size(); i++) {
    ASSERT_EQ(subgraph_output[i], operator_output[i]);
  }
}

TEST_F(RoPETestF32, reshape_output)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(3, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  std::array<size_t, 4> input_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(),
                            /*data=*/nullptr, /*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t weights_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 2> weights_dims{{max_tokens, channels}};
  ASSERT_EQ(xnn_status_success,
    xnn_define_tensor_value(subgraph, xnn_datatype_fp32, weights_dims.size(), weights_dims.data(),
                            weights.data(), /*external_id=*/1, /*flags=*/0, &weights_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  const std::array<size_t, 4> output_dims{{batch_size, tokens, heads, channels}};
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(),
                          /*data=*/nullptr, /*external_id=*/2, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success,
    xnn_define_rope(subgraph, max_tokens, input_id, weights_id, output_id, /*flags=*/0));

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);

  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);

  const std::array<xnn_external_value, 2> external{{
    xnn_external_value{input_id, input.data()},
    xnn_external_value{output_id, subgraph_output.data()}
  }};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));

  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  input_dims[0] += 4;
  input_dims[2] += 4;
  input_dims[3] += 4;
  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, input_dims.size(), input_dims.data()));
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->reshape(&runtime->opdata[0], runtime->values, runtime->num_values, /*threadpool=*/nullptr), xnn_status_reallocation_required);
  const xnn_shape* output_shape = &runtime->values[node->outputs[0]].shape;
  for (size_t i = 0; i < input_dims.size(); ++i) {
    ASSERT_EQ(output_shape->dim[i], input_dims[i]);
  }

  input_dims[3] -= 4;
  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, input_dims.size(), input_dims.data()));
  ASSERT_EQ(node->reshape(&runtime->opdata[0], runtime->values, runtime->num_values, /*threadpool=*/nullptr), xnn_status_success);
  for (size_t i = 0; i < input_dims.size(); ++i) {
    ASSERT_EQ(output_shape->dim[i], input_dims[i]);
  }
}