Support int16 in HiFi quantized_layer_norm (#24091)

Summary:

The HiFi quantized_layer_norm kernels dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 inputs fell through to ET_DCHECK_MSG. That check compiles out in release builds, which left the output tensor unwritten and produced garbage numerics for int16 models.

Simply widening the dispatch is not enough because the statistics used int32 accumulators. For int16, a single squared value can reach 2^30, so sq_sum overflows after two elements. The variance was also computed as E[x^2] - E[x]^2 in fp32, which cancels catastrophically when the mean is large relative to the spread.

Changes:
- Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16.
- Accumulate sum and sq_sum in int64 for 16-bit types. 8-bit types keep int32 in the inner loop.
- Widen to the accumulator type before squaring, so the product cannot overflow int32.
- Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers before converting to float. For 8-bit this is equal or more accurate than before.
- Declare quantized_layer_norm_out and quantized_layer_norm_per_tensor_out in operators.h and add a gtest.

Reviewed By: DrJessop

Differential Revision: D126939431

Author: mvartani-meta

backends/cadence/hifi/operators/op_quantized_layer_norm.cpp

@@ -11,7 +11,9 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <algorithm>
 #include <cmath>
+#include <cstdint>
 #include <tuple>
+#include <type_traits>
 
 using ::executorch::aten::IntArrayRef;
 using ::executorch::aten::ScalarType;
@@ -55,32 +57,41 @@ void quantized_layer_norm_per_tensor_(
     const T* __restrict__ x = in_data + i * last_dim;
     T* __restrict__ y = out_data + i * last_dim;
 
-    // compute sum and squared sum. The fp32 sum can be approximated as:
-    // (X_1 - in_zero_point) * in_scale + (X_2 - in_zero_point) * in_scale + ...
-    // (X_N - in_zero_point) * in_scale.
-    int32_t sum = 0;
-    int32_t sq_sum = last_dim * input_zero_point * input_zero_point;
+    // 16-bit squares overflow an int32 sum after two elements.
+    using acc_t = std::conditional_t<sizeof(T) == 1, int32_t, int64_t>;
+    acc_t sum = 0;
+    acc_t sq_sum = 0;
 #pragma simd
     for (size_t j = 0; j < last_dim; ++j) {
-      int32_t val = x[j];
+      const int32_t val = x[j];
       sum += val;
-      sq_sum += val * val;
+      sq_sum += static_cast<acc_t>(val) * val;
     }
-    sq_sum -= (2 * sum * input_zero_point);
-    sum -= (last_dim * input_zero_point);
-
-    float mean = XT_DIV_S(XT_MUL_S(input_scale, sum), last_dim);
-    float variance =
-        XT_DIV_S(
-            XT_MUL_S(sq_sum, XT_MUL_S(input_scale, input_scale)), last_dim) -
-        XT_MUL_S(mean, mean);
+
+    // Shift raw sums to zero-point-centered sums, then form the variance
+    // numerator N * sum(v^2) - sum(v)^2 exactly in integers: doing
+    // E[x^2] - E[x]^2 in fp32 cancels catastrophically for 16-bit inputs.
+    const int64_t n = static_cast<int64_t>(last_dim);
+    const int64_t zp = input_zero_point;
+    const int64_t centered_sum = static_cast<int64_t>(sum) - n * zp;
+    const int64_t centered_sq_sum =
+        static_cast<int64_t>(sq_sum) - 2 * zp * sum + n * zp * zp;
+    const int64_t var_numerator =
+        n * centered_sq_sum - centered_sum * centered_sum;
+
+    const float n_f = static_cast<float>(n);
+    float mean =
+        XT_DIV_S(XT_MUL_S(input_scale, static_cast<float>(centered_sum)), n_f);
+    float variance = XT_MUL_S(
+        XT_DIV_S(static_cast<float>(var_numerator), XT_MUL_S(n_f, n_f)),
+        XT_MUL_S(input_scale, input_scale));
     float inv_std = XT_RECIP_S(XT_SQRT_S(XT_ADD_S(variance, (float)eps)));
 
     // y = (x - mean) / std * kGamma + kBeta
 #pragma simd
     for (size_t j = 0; j < last_dim; ++j) {
       // Since X is quantized, we dequantize it, compute fp32 result, and
-      // quantize the result to an int8/uint8 value.
+      // quantize the result back to T.
       float val = dequantize<T>(x[j], input_scale, input_zero_point);
       val = (val - mean) * inv_std * weight_data[j] + bias_data[j];
       y[j] = quantize<T>(val, output_inv_scale, output_zero_point);
@@ -147,7 +158,7 @@ void quantized_layer_norm_out(
 
   ScalarType dtype = input.scalar_type();
   switch (dtype) {
-    ET_FORALL_CADENCE_QUANTIZED_TYPES(typed_quantized_layer_norm)
+    ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16(typed_quantized_layer_norm)
     default:
       ET_DCHECK_MSG(
           false, "Unhandled dtype %s", torch::executor::toString(dtype));
@@ -185,7 +196,7 @@ void quantized_layer_norm_per_tensor_out(
 
   ScalarType dtype = input.scalar_type();
   switch (dtype) {
-    ET_FORALL_CADENCE_QUANTIZED_TYPES(typed_quantized_layer_norm)
+    ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16(typed_quantized_layer_norm)
     default:
       ET_DCHECK_MSG(
           false, "Unhandled dtype %s", torch::executor::toString(dtype));

backends/cadence/hifi/operators/operators.h

@@ -62,6 +62,32 @@ void quantized_relu_out(
     const ::executorch::aten::Tensor& out_shift,
     ::executorch::aten::Tensor& output);
 
+void quantized_layer_norm_out(
+    ::executorch::runtime::KernelRuntimeContext& ctx,
+    const ::executorch::aten::Tensor& input,
+    const ::executorch::aten::Tensor& in_scale,
+    const ::executorch::aten::Tensor& in_zero_point,
+    const ::executorch::aten::IntArrayRef normalized_shape,
+    const ::executorch::aten::Tensor& weight,
+    const ::executorch::aten::Tensor& bias,
+    double eps,
+    double output_scale,
+    int64_t output_zero_point,
+    ::executorch::aten::Tensor& out);
+
+void quantized_layer_norm_per_tensor_out(
+    ::executorch::runtime::KernelRuntimeContext& ctx,
+    const ::executorch::aten::Tensor& input,
+    double in_scale,
+    int64_t in_zero_point,
+    const ::executorch::aten::IntArrayRef normalized_shape,
+    const ::executorch::aten::Tensor& weight,
+    const ::executorch::aten::Tensor& bias,
+    double eps,
+    double output_scale,
+    int64_t output_zero_point,
+    ::executorch::aten::Tensor& out);
+
 void quantized_linear_out(
     ::executorch::runtime::KernelRuntimeContext& ctx,
     const ::executorch::aten::Tensor& in,

backends/cadence/hifi/operators/tests/BUCK

@@ -27,6 +27,27 @@ jarvis_wrapper.cxx_test(
 )
 
 
+jarvis_wrapper.cxx_test(
+    name = "test_op_quantized_layer_norm",
+    srcs = [
+        "test_op_quantized_layer_norm.cpp",
+    ],
+    compatible_backends = ["hifi"],
+    labels = [ci.linux(ci.mode("fbsource//arvr/mode/platform010/dev"))],
+    platforms = CXX,
+    remote_execution = XTENSA_TEST_REMOTE_EXECUTION_ASIC,
+    visibility = [
+        "fbsource//xplat/executorch/backends/cadence/...",
+        "fbcode//executorch/backends/cadence/...",
+    ],
+    deps = [
+        "fbsource//xplat/executorch/kernels/test:gtest_utils",
+        "fbsource//xplat/executorch/backends/cadence/hifi/operators:op_quantized_layer_norm",
+        "fbsource//xplat/executorch/backends/cadence/runtime:et_pal",
+        "fbsource//xplat/executorch/runtime/core/exec_aten/testing_util:tensor_util",
+    ],
+)
+
 jarvis_wrapper.cxx_test(
     name = "test_op_quantize_per_tensor",
     srcs = [

backends/cadence/hifi/operators/tests/CMakeLists.txt

@@ -56,6 +56,7 @@ add_executable(
   test_op_permute_copy.cpp
   test_op_quantize_per_tensor.cpp
   test_op_quantized_conv2d_out.cpp
+  test_op_quantized_layer_norm.cpp
   test_op_quantized_matmul_out.cpp
   test_op_quantized_relu_out.cpp
   test_op_relu.cpp

backends/cadence/hifi/operators/tests/test_op_quantized_layer_norm.cpp

@@ -0,0 +1,133 @@
+/*
+ * 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 <executorch/kernels/test/TestUtil.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <executorch/runtime/core/exec_aten/testing_util/tensor_factory.h>
+#include <executorch/runtime/core/exec_aten/testing_util/tensor_util.h>
+
+#include <executorch/backends/cadence/hifi/operators/operators.h>
+
+namespace impl {
+namespace HiFi {
+namespace native {
+namespace {
+
+using ::executorch::aten::IntArrayRef;
+using ::executorch::aten::ScalarType;
+using ::executorch::aten::Tensor;
+using ::executorch::runtime::testing::TensorFactory;
+
+// Scales are powers of two and inputs are symmetric around their mean, so
+// every row normalizes to exactly +/-1 (eps is negligible).
+constexpr double kEps = 1e-9;
+
+class HiFiQuantizedLayerNormTest : public OperatorTest {};
+
+TEST_F(HiFiQuantizedLayerNormTest, Int16LargeMagnitudePerTensor) {
+  TensorFactory<ScalarType::Short> tf_short;
+  TensorFactory<ScalarType::Float> tf_float;
+  const std::vector<int64_t> normalized_shape{8};
+
+  // Per-row sum of squares is 8 * 30000^2, well past int32 range.
+  const Tensor input = tf_short.make(
+      {2, 8},
+      {30000, -30000, 30000, -30000, 30000, -30000, 30000, -30000,
+       -30000, 30000, -30000, 30000, -30000, 30000, -30000, 30000});
+  const Tensor weight = tf_float.ones({8});
+  const Tensor bias = tf_float.zeros({8});
+  Tensor out = tf_short.zeros({2, 8});
+
+  quantized_layer_norm_per_tensor_out(
+      context_,
+      input,
+      /*in_scale=*/1.0 / 1024,
+      /*in_zero_point=*/0,
+      IntArrayRef(normalized_shape.data(), normalized_shape.size()),
+      weight,
+      bias,
+      kEps,
+      /*output_scale=*/1.0 / 16384,
+      /*output_zero_point=*/0,
+      out);
+
+  const Tensor expected = tf_short.make(
+      {2, 8},
+      {16384, -16384, 16384, -16384, 16384, -16384, 16384, -16384,
+       -16384, 16384, -16384, 16384, -16384, 16384, -16384, 16384});
+  EXPECT_TENSOR_EQ(out, expected);
+}
+
+TEST_F(HiFiQuantizedLayerNormTest, Int16LargeMeanSmallSpreadWithZeroPoints) {
+  TensorFactory<ScalarType::Short> tf_short;
+  TensorFactory<ScalarType::Float> tf_float;
+  TensorFactory<ScalarType::Long> tf_long;
+  const std::vector<int64_t> normalized_shape{8};
+
+  // Centered values are 21000 +/- 4: the mean dwarfs the spread, which breaks
+  // an fp32 E[x^2] - E[x]^2 variance.
+  const Tensor input = tf_short.make(
+      {1, 8}, {20004, 19996, 20004, 19996, 20004, 19996, 20004, 19996});
+  const Tensor in_scale = tf_float.make({1}, {1.0f / 64});
+  const Tensor in_zero_point = tf_long.make({1}, {-1000});
+  const Tensor weight = tf_float.full({8}, 2.0f);
+  const Tensor bias = tf_float.full({8}, 0.5f);
+  Tensor out = tf_short.zeros({1, 8});
+
+  quantized_layer_norm_out(
+      context_,
+      input,
+      in_scale,
+      in_zero_point,
+      IntArrayRef(normalized_shape.data(), normalized_shape.size()),
+      weight,
+      bias,
+      kEps,
+      /*output_scale=*/1.0 / 4096,
+      /*output_zero_point=*/100,
+      out);
+
+  // (+1 * 2 + 0.5) * 4096 + 100 and (-1 * 2 + 0.5) * 4096 + 100.
+  const Tensor expected = tf_short.make(
+      {1, 8}, {10340, -6044, 10340, -6044, 10340, -6044, 10340, -6044});
+  EXPECT_TENSOR_EQ(out, expected);
+}
+
+TEST_F(HiFiQuantizedLayerNormTest, Int8WithZeroPoint) {
+  TensorFactory<ScalarType::Char> tf_char;
+  TensorFactory<ScalarType::Float> tf_float;
+  const std::vector<int64_t> normalized_shape{4};
+
+  const Tensor input = tf_char.make({1, 4}, {10, 30, 10, 30});
+  const Tensor weight = tf_float.ones({4});
+  const Tensor bias = tf_float.zeros({4});
+  Tensor out = tf_char.zeros({1, 4});
+
+  quantized_layer_norm_per_tensor_out(
+      context_,
+      input,
+      /*in_scale=*/1.0 / 8,
+      /*in_zero_point=*/20,
+      IntArrayRef(normalized_shape.data(), normalized_shape.size()),
+      weight,
+      bias,
+      kEps,
+      /*output_scale=*/1.0 / 64,
+      /*output_zero_point=*/0,
+      out);
+
+  const Tensor expected = tf_char.make({1, 4}, {-64, 64, -64, 64});
+  EXPECT_TENSOR_EQ(out, expected);
+}
+
+} // namespace
+} // namespace native
+} // namespace HiFi
+} // namespace impl