[ET-VK][Ops] dequantize ops skeleton test framework

Differential Revision: D76267021

Pull Request resolved: #11480

Author: ahmtox

backends/vulkan/test/op_tests/dequantize_test.cpp

@@ -0,0 +1,182 @@
+/*
+ * 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.
+ */
+
+#include <gtest/gtest.h>
+
+#include <ATen/ATen.h>
+
+#include <executorch/backends/vulkan/runtime/api/api.h>
+#include <executorch/backends/vulkan/runtime/graph/ComputeGraph.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/OperatorRegistry.h>
+
+#include <executorch/extension/aten_util/make_aten_functor_from_et_functor.h>
+#include <executorch/extension/kernel_util/make_boxed_from_unboxed_functor.h>
+
+#include "test_utils.h"
+
+#include <cassert>
+#include <iostream>
+
+namespace torch {
+namespace executor {
+namespace native {
+
+// Forward declarations of the functions we're testing
+Tensor& dequantize_per_tensor_out(
+    const Tensor& input,
+    double scale,
+    int64_t zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    ScalarType dtype,
+    executorch::aten::optional<ScalarType> out_dtype,
+    Tensor& out);
+
+Tensor& dequantize_per_token_out(
+    const Tensor& input,
+    const Tensor& scale,
+    const Tensor& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    ScalarType dtype,
+    ScalarType out_dtype,
+    Tensor& out);
+
+// Wrapper function for dequantize_per_tensor_out without context
+Tensor& dequantize_per_tensor_out_no_context(
+    const Tensor& input,
+    double scale,
+    int64_t zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    ScalarType dtype,
+    executorch::aten::optional<ScalarType> out_dtype,
+    Tensor& out) {
+  return torch::executor::native::dequantize_per_tensor_out(
+      input, scale, zero_point, quant_min, quant_max, dtype, out_dtype, out);
+}
+
+// Wrapper function for dequantize_per_token_out without context
+Tensor& dequantize_per_token_out_no_context(
+    const Tensor& input,
+    const Tensor& scale,
+    const Tensor& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    ScalarType dtype,
+    ScalarType out_dtype,
+    Tensor& out) {
+  return torch::executor::native::dequantize_per_token_out(
+      input, scale, zero_points, quant_min, quant_max, dtype, out_dtype, out);
+}
+
+// ATen wrapper for dequantize_per_tensor
+at::Tensor dequantize_per_tensor_aten(
+    const at::Tensor& input,
+    double scale,
+    int64_t zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype,
+    at::ScalarType out_dtype) {
+  auto out = at::empty_like(input, out_dtype);
+  // Convert at::ScalarType to executorch::ScalarType
+  ScalarType et_dtype = at_scalartype_to_et_scalartype(dtype);
+  ScalarType et_out_dtype = at_scalartype_to_et_scalartype(out_dtype);
+
+  executorch::aten::optional<ScalarType> opt_et_out_dtype(et_out_dtype);
+
+  WRAP_TO_ATEN(dequantize_per_tensor_out_no_context, 7)
+  (input,
+   scale,
+   zero_point,
+   quant_min,
+   quant_max,
+   et_dtype,
+   opt_et_out_dtype,
+   out);
+  return out;
+}
+
+// ATen wrapper for dequantize_per_token
+at::Tensor dequantize_per_token_aten(
+    const at::Tensor& input,
+    const at::Tensor& scale,
+    const at::Tensor& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype,
+    at::ScalarType out_dtype) {
+  auto out = at::empty_like(input, out_dtype);
+  // Convert at::ScalarType to executorch::ScalarType
+  ScalarType et_dtype = at_scalartype_to_et_scalartype(dtype);
+  ScalarType et_out_dtype = at_scalartype_to_et_scalartype(out_dtype);
+
+  WRAP_TO_ATEN(dequantize_per_token_out_no_context, 7)
+  (input,
+   scale,
+   zero_points,
+   quant_min,
+   quant_max,
+   et_dtype,
+   et_out_dtype,
+   out);
+  return out;
+}
+
+} // namespace native
+} // namespace executor
+} // namespace torch
+
+void check_dequantize_args(
+    int64_t quant_min,
+    int64_t quant_max,
+    c10::ScalarType in_dtype,
+    c10::ScalarType out_dtype) {
+  using namespace vkcompute;
+
+  // Check that quant_min <= quant_max
+  VK_CHECK_COND(
+      quant_min <= quant_max,
+      "quant_min must be <= quant_max, got quant_min: ",
+      quant_min,
+      " quant_max: ",
+      quant_max);
+
+  // Check that input dtype is a quantized type
+  switch (in_dtype) {
+    case c10::kByte:
+    case c10::kChar:
+    case c10::kShort:
+    case c10::kInt:
+    case c10::kLong:
+      break;
+    default:
+      VK_THROW(
+          "Unsupported input dtype: ",
+          scalar_type_name(in_dtype),
+          " (",
+          static_cast<int>(in_dtype),
+          ")");
+  }
+
+  // Check that output dtype is a floating point type
+  switch (out_dtype) {
+    case c10::kHalf:
+    case c10::kFloat:
+    case c10::kDouble:
+      break;
+    default:
+      VK_THROW(
+          "Unsupported output dtype: ",
+          scalar_type_name(out_dtype),
+          " (",
+          static_cast<int>(out_dtype),
+          ")");
+  }
+}

backends/vulkan/test/op_tests/targets.bzl

@@ -186,6 +186,15 @@ def define_common_targets(is_fbcode = False):
             "//executorch/extension/aten_util:aten_bridge",
         ]
     )
+    define_test_targets(
+        "dequantize_test",
+        extra_deps = [
+            ":test_utils",
+            "//executorch/kernels/quantized/cpu:op_dequantize",
+            "//executorch/extension/tensor:tensor",
+            "//executorch/extension/aten_util:aten_bridge",
+        ]
+    )
     define_test_targets(
         "linear_weight_int4_test",
         extra_deps = [