[Executorch][quant] Optimize per channel dequantize

Pull Request resolved: #5670 When using quantized kv cache, dequantization routine takes significantly long. This diff just vectorizes dequant per channel for common case. ghstack-source-id: 255730818 @exported-using-ghexport Differential Revision: [D63338858](https://our.internmc.facebook.com/intern/diff/D63338858/) Co-authored-by: Kimish Patel <[email protected]>
pytorch · Dec 2, 2024 · ddec0c7 · ddec0c7
1 parent 9d084c4
commit ddec0c7
Show file tree

Hide file tree

Showing 2 changed files with 238 additions and 21 deletions.
diff --git a/kernels/quantized/cpu/op_dequantize.cpp b/kernels/quantized/cpu/op_dequantize.cpp
@@ -11,6 +11,9 @@
 #include <algorithm>
 #include <cinttypes>
 #include <cmath>
+#if defined(__aarch64__) || defined(__ARM_NEON)
+#include <arm_neon.h>
+#endif
 
 /**
  * For an input tensor, use the scale and zero_point arguments to quantize it.
@@ -22,6 +25,8 @@ namespace native {
 using Tensor = exec_aten::Tensor;
 using Scalar = exec_aten::Scalar;
 using ScalarType = exec_aten::ScalarType;
+using StridesType = exec_aten::StridesType;
+using SizesType = exec_aten::SizesType;
 
 namespace {
 
@@ -63,6 +68,183 @@ void check_dequantize_per_tensor_args(
       quant_max);
 }
 
+/**
+ * Useful to reduce a tensor `in` over a given dimension `dim` using the
+ * reduce function `fn`, which should have the following signature:
+ * void fn(const size_t size, const size_t stride, const size_t base_ix)
+ * where `size` and `stride` are the size and stride of the dimension being
+ * reduced and `base_ix` is the index of the first element of the reduction.
+ */
+template <typename Fn>
+void apply_over_unpacked_dim(
+    const Fn& fn,
+    const exec_aten::Tensor& in,
+    const int64_t& dim) {
+  if (in.numel() == 0) {
+    return;
+  }
+
+  ET_CHECK_MSG(in.dim() > 0, "Input tensor must have at least one dimension");
+  ET_CHECK_VALID_DIM(dim, in.dim());
+
+  const size_t d = ET_NORMALIZE_IX(dim, in.dim());
+  const size_t dim_size = in.size(d);
+  const size_t outer_size = getLeadingDims(in, d);
+  const size_t inner_size = getTrailingDims(in, d);
+  // Loop through all outer dimensions
+  for (size_t outer_idx = 0; outer_idx < outer_size; ++outer_idx) {
+    // Loop through dim
+    for (size_t unpacked_dim_idx = 0; unpacked_dim_idx < dim_size;
+         ++unpacked_dim_idx) {
+      fn(inner_size, outer_idx, unpacked_dim_idx);
+    }
+  }
+}
+
+void dequantize_optimized(
+    const int8_t* in,
+    const double scale,
+    const int64_t zero_point,
+    float* out,
+    int64_t quant_min,
+    int64_t quant_max,
+    size_t numel) {
+  ET_CHECK_MSG(
+      zero_point >= quant_min,
+      "zero_point must be %" PRId64 " <= quant_min %" PRId64,
+      zero_point,
+      quant_min);
+  ET_CHECK_MSG(
+      zero_point <= quant_max,
+      "zero_point must be %" PRId64 " >= quant_max %" PRId64,
+      zero_point,
+      quant_max);
+  size_t i = 0;
+#if defined(__aarch64__) || defined(__ARM_NEON)
+  int8x8_t zero_point_vec = vdup_n_s8(zero_point);
+  float32x4_t scales = vdupq_n_f32(static_cast<float>(scale));
+  constexpr int32_t kVecSize = 16;
+  const size_t num_vecs = numel / kVecSize;
+  const int8_t* in_copy = in;
+  float* out_copy = out;
+  for (; i < num_vecs; i++) {
+    int8x16_t in_vec = vld1q_s8(in_copy);
+    int16x8_t sub_vec_0_7 = vsubl_s8(vget_low_s8(in_vec), zero_point_vec);
+    int32x4_t sub_vec_0_3 = vmovl_s16(vget_low_s16(sub_vec_0_7));
+    int32x4_t sub_vec_4_7 = vmovl_s16(vget_high_s16(sub_vec_0_7));
+    float32x4_t out_vec_0_3 = vmulq_f32(vcvtq_f32_s32(sub_vec_0_3), scales);
+    float32x4_t out_vec_4_7 = vmulq_f32(vcvtq_f32_s32(sub_vec_4_7), scales);
+
+    int16x8_t sub_vec_8_15 = vsubl_s8(vget_high_s8(in_vec), zero_point_vec);
+    int32x4_t sub_vec_8_11 = vmovl_s16(vget_low_s16(sub_vec_8_15));
+    int32x4_t sub_vec_12_15 = vmovl_s16(vget_high_s16(sub_vec_8_15));
+    float32x4_t out_vec_8_11 = vmulq_f32(vcvtq_f32_s32(sub_vec_8_11), scales);
+    float32x4_t out_vec_12_15 = vmulq_f32(vcvtq_f32_s32(sub_vec_12_15), scales);
+    vst1q_f32(out_copy + 0, out_vec_0_3);
+    vst1q_f32(out_copy + 4, out_vec_4_7);
+    vst1q_f32(out_copy + 8, out_vec_8_11);
+    vst1q_f32(out_copy + 12, out_vec_12_15);
+    in_copy += kVecSize;
+    out_copy += kVecSize;
+  }
+  i = i * kVecSize;
+#endif
+  for (; i < numel; i++) {
+    out[i] = (in[i] - zero_point) * scale;
+  }
+}
+
+float get_scale(const Tensor& scale, size_t channel_ix) {
+  ET_CHECK_MSG(
+      (scale.scalar_type() == ScalarType::Double) ||
+          (scale.scalar_type() == ScalarType::Float),
+      "scale.scalar_type() %" PRId8 " is not double or float type",
+      static_cast<int8_t>(scale.scalar_type()));
+  if (scale.scalar_type() == ScalarType::Double) {
+    return static_cast<float>(scale.const_data_ptr<double>()[channel_ix]);
+  } else {
+    return scale.const_data_ptr<float>()[channel_ix];
+  }
+}
+
+bool can_use_optimized_dequantize_per_channel(
+    const Tensor& in,
+    const ScalarType in_dtype,
+    exec_aten::optional<ScalarType>& out_dtype) {
+  bool is_contiguous = false;
+#ifdef USE_ATEN_LIB
+  is_contiguous = in.is_contiguous();
+#else
+  is_contiguous = executorch::runtime::is_contiguous_dim_order(
+      in.dim_order().data(), in.dim());
+#endif
+  if (!is_contiguous || (in_dtype != ScalarType::Char) ||
+      (out_dtype.has_value() && out_dtype.value() != ScalarType::Float)) {
+    return false;
+  }
+  return true;
+}
+
+void dequantize_per_channel_optimized(
+    const Tensor& in,
+    const Tensor& scales,
+    const optional<Tensor>& opt_zero_points,
+    Tensor& out,
+    int64_t axis,
+    int64_t quant_min,
+    int64_t quant_max,
+    ScalarType in_dtype,
+    exec_aten::optional<ScalarType>& out_dtype) {
+  check_dequantize_per_tensor_args(
+      in, quant_min, quant_max, in_dtype, out_dtype, out);
+  ET_CHECK_MSG(
+      in_dtype == ScalarType::Char,
+      "in.scalar_type() %" PRId8 " is not supported:",
+      static_cast<int8_t>(in.scalar_type()));
+  if (out_dtype.has_value()) {
+    ET_CHECK_MSG(
+        out_dtype.value() == ScalarType::Float,
+        "Only float output is supported");
+  }
+  const int8_t* in_data = in.const_data_ptr<int8_t>();
+  float* out_data = out.mutable_data_ptr<float>();
+  const int64_t* zero_points_data = nullptr;
+  if (opt_zero_points.has_value()) {
+    zero_points_data = opt_zero_points.value().const_data_ptr<int64_t>();
+  }
+  const StridesType axis_stride = in.strides()[axis];
+  const StridesType outer_stride = in.size(axis) * axis_stride;
+  apply_over_unpacked_dim(
+      [in_data,
+       out_data,
+       &scales,
+       zero_points_data,
+       axis_stride,
+       outer_stride,
+       quant_min,
+       quant_max](
+          SizesType numel, SizesType outer_idx, SizesType unpacked_dim_idx) {
+        const int8_t* in_data_local =
+            in_data + outer_idx * outer_stride + unpacked_dim_idx * axis_stride;
+        const double scale = get_scale(scales, unpacked_dim_idx);
+        const int64_t zero_point = zero_points_data != nullptr
+            ? zero_points_data[unpacked_dim_idx]
+            : 0;
+        float* out_data_local = out_data + outer_idx * outer_stride +
+            unpacked_dim_idx * axis_stride;
+        dequantize_optimized(
+            in_data_local,
+            scale,
+            zero_point,
+            out_data_local,
+            quant_min,
+            quant_max,
+            numel);
+      },
+      in,
+      axis);
+}
+
 } // namespace
 
 /**
@@ -172,19 +354,6 @@ Tensor& dequantize_per_tensor_tensor_args_out(
   return out;
 }
 
-float get_scale(const Tensor& scale, size_t channel_ix) {
-  ET_CHECK_MSG(
-      (scale.scalar_type() == ScalarType::Double) ||
-          (scale.scalar_type() == ScalarType::Float),
-      "scale.scalar_type() %" PRId8 " is not double or float type",
-      static_cast<int8_t>(scale.scalar_type()));
-  if (scale.scalar_type() == ScalarType::Double) {
-    return static_cast<float>(scale.const_data_ptr<double>()[channel_ix]);
-  } else {
-    return scale.const_data_ptr<float>()[channel_ix];
-  }
-}
-
 Tensor& dequantize_per_channel_out(
     const Tensor& input,
     const Tensor& scale,
@@ -229,6 +398,20 @@ Tensor& dequantize_per_channel_out(
   check_dequantize_per_tensor_args(
       input, quant_min, quant_max, dtype, out_dtype, out);
 
+  if (can_use_optimized_dequantize_per_channel(input, dtype, out_dtype)) {
+    dequantize_per_channel_optimized(
+        input,
+        scale,
+        opt_zero_points,
+        out,
+        axis,
+        quant_min,
+        quant_max,
+        dtype,
+        out_dtype);
+    return out;
+  }
+
   // a list contains all dimensions except axis
   int64_t dims[kTensorDimensionLimit];
   for (int64_t i = 0; i < input.dim() - 1; i++) {

diff --git a/kernels/quantized/test/op_dequantize_test.cpp b/kernels/quantized/test/op_dequantize_test.cpp
@@ -123,13 +123,13 @@ TEST(OpDequantizeOutTest, TensorArgOverload) {
   EXPECT_TENSOR_EQ(out, expected);
 }
 
-TEST(OpDequantizeOutTest, DequantizePerChannel) {
-  et_pal_init();
-  TensorFactory<ScalarType::Byte> tf_byte;
+template <ScalarType DTYPE>
+void test_per_channel_dtype() {
+  TensorFactory<DTYPE> tf;
   TensorFactory<ScalarType::Double> tf_double;
   TensorFactory<ScalarType::Long> tf_long;
 
-  Tensor input = tf_byte.full({3, 2}, 100);
+  Tensor input = tf.full({3, 2}, 100);
   Tensor scale = tf_double.make({2}, {0.5, 1});
   Tensor zero_point = tf_long.make({2}, {30, 60});
   int64_t quant_min = 0;
@@ -147,7 +147,7 @@ TEST(OpDequantizeOutTest, DequantizePerChannel) {
       /*axis=*/1,
       quant_min,
       quant_max,
-      ScalarType::Byte,
+      DTYPE,
       optional<ScalarType>(),
       out);
 
@@ -168,15 +168,15 @@ TEST(OpDequantizeOutTest, DequantizePerChannel) {
       /*axis=*/0,
       quant_min,
       quant_max,
-      ScalarType::Byte,
+      DTYPE,
       optional<ScalarType>(),
       out);
 
   EXPECT_TENSOR_EQ(out, expected);
 
   // Test with a different axis
   out = tfo.zeros({3});
-  input = tf_byte.make({3}, {100, 100, 100});
+  input = tf.make({3}, {100, 100, 100});
   scale = tf_double.make({3}, {0.5, 0.75, 1});
   zero_point = tf_long.make({3}, {30, 50, 60});
   // (100 - 30) * 0.5
@@ -190,8 +190,42 @@ TEST(OpDequantizeOutTest, DequantizePerChannel) {
       /*axis=*/0,
       quant_min,
       quant_max,
-      ScalarType::Byte,
+      DTYPE,
+      optional<ScalarType>(),
+      out);
+  EXPECT_TENSOR_EQ(out, expected);
+
+  // Test with a different axis
+  input = tf.full({3, 19}, 100);
+  out = tfo.zeros({3, 19});
+  scale = tf_double.make({3}, {0.5, 0.75, 1});
+  zero_point = tf_long.make({3}, {30, 50, 60});
+  // (100 - 30) * 0.5
+  // (100 - 50) * 0.75
+  // (100 - 60) * 1
+  expected = tfo.make(
+      {3, 19},
+      {35,   35,   35,   35,   35,   35,   35,   35,   35,   35,   35,   35,
+       35,   35,   35,   35,   35,   35,   35,   37.5, 37.5, 37.5, 37.5, 37.5,
+       37.5, 37.5, 37.5, 37.5, 37.5, 37.5, 37.5, 37.5, 37.5, 37.5, 37.5, 37.5,
+       37.5, 37.5, 40,   40,   40,   40,   40,   40,   40,   40,   40,   40,
+       40,   40,   40,   40,   40,   40,   40,   40,   40});
+  dequantize_per_channel_out(
+      input,
+      scale,
+      zero_point,
+      /*axis=*/0,
+      quant_min,
+      quant_max,
+      DTYPE,
       optional<ScalarType>(),
       out);
+
   EXPECT_TENSOR_EQ(out, expected);
 }
+
+TEST(OpDequantizeOutTest, DequantizePerChannel) {
+  et_pal_init();
+  test_per_channel_dtype<ScalarType::Byte>();
+  test_per_channel_dtype<ScalarType::Char>();
+}