summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--aten/src/ATen/native/quantized/QTensor.cpp12
-rw-r--r--aten/src/ATen/quantized/QTensorImpl.h7
-rw-r--r--aten/src/ATen/quantized/Quantizer.cpp16
-rw-r--r--aten/src/ATen/quantized/Quantizer.h20
-rw-r--r--aten/src/ATen/test/quantized_test.cpp2
5 files changed, 30 insertions, 27 deletions
diff --git a/aten/src/ATen/native/quantized/QTensor.cpp b/aten/src/ATen/native/quantized/QTensor.cpp
index 84243b1d59..85c4481b74 100644
--- a/aten/src/ATen/native/quantized/QTensor.cpp
+++ b/aten/src/ATen/native/quantized/QTensor.cpp
@@ -7,32 +7,32 @@
namespace at {
namespace native {
-QTensor quantize_linear_cpu(const RealTensor& self, double scale, int64_t zero_point) {
+Tensor quantize_linear_cpu(const Tensor& self, double scale, int64_t zero_point) {
auto quantizer = make_per_tensor_affine_quantizer(scale, zero_point);
return quantizer->quantize(self);
}
-RealTensor dequantize_quant(const QTensor& self) {
+Tensor dequantize_quant(const Tensor& self) {
return get_qtensorimpl(self)->quantizer()->dequantize(self);
}
-Scalar q_scale_quant(const QTensor& self) {
+Scalar q_scale_quant(const Tensor& self) {
auto quantizer = get_qtensorimpl(self)->quantizer();
AT_ASSERT(quantizer->qscheme() == kPerTensorAffine);
return Scalar(static_cast<PerTensorAffineQuantizer*>(quantizer.get())->scale());
}
-Scalar q_zero_point_quant(const QTensor& self) {
+Scalar q_zero_point_quant(const Tensor& self) {
auto quantizer = get_qtensorimpl(self)->quantizer();
AT_ASSERT(quantizer->qscheme() == kPerTensorAffine);
return Scalar(static_cast<PerTensorAffineQuantizer*>(quantizer.get())->zero_point());
}
-Quantizer* quantizer(const QTensor& self) {
+Quantizer* quantizer(const Tensor& self) {
return get_qtensorimpl(self)->quantizer().get();
}
-Tensor int_repr_quant(const QTensor& self) {
+Tensor int_repr_quant(const Tensor& self) {
Tensor dst = at::empty(self.sizes(), self.options().dtype(at::kByte));
uint8_t* self_data = reinterpret_cast<uint8_t *>(self.data<qint8>());
uint8_t* dst_data = dst.data<uint8_t>();
diff --git a/aten/src/ATen/quantized/QTensorImpl.h b/aten/src/ATen/quantized/QTensorImpl.h
index 6e9af90e76..3aeeb40257 100644
--- a/aten/src/ATen/quantized/QTensorImpl.h
+++ b/aten/src/ATen/quantized/QTensorImpl.h
@@ -6,6 +6,13 @@
namespace at {
+/**
+ * QTensorImpl is a TensorImpl for Quantized Tensors, it stores Quantizer which
+ * specifies the quantization scheme and parameters, for more information please
+ * see ATen/quantized/Quantizer.h
+ *
+ * We'll use QTensor in code or documentation to refer to a Tensor with QTensorImpl.
+ */
struct CAFFE2_API QTensorImpl : public c10::TensorImpl {
public:
QTensorImpl(
diff --git a/aten/src/ATen/quantized/Quantizer.cpp b/aten/src/ATen/quantized/Quantizer.cpp
index 1ec6ce9f8d..c0784a549e 100644
--- a/aten/src/ATen/quantized/Quantizer.cpp
+++ b/aten/src/ATen/quantized/Quantizer.cpp
@@ -18,7 +18,7 @@ QuantizerPtr make_per_tensor_affine_quantizer(
static_cast<float>(scale), static_cast<uint8_t>(zero_point));
}
-QTensorImpl* get_qtensorimpl(const QTensor& self) {
+QTensorImpl* get_qtensorimpl(const Tensor& self) {
// TODO: remove this when Variable and Tensor are merged
AT_ASSERTM(
!self.is_variable(),
@@ -29,7 +29,7 @@ QTensorImpl* get_qtensorimpl(const QTensor& self) {
return static_cast<QTensorImpl*>(self.unsafeGetTensorImpl());
}
-inline QTensor new_qtensor_cpu(
+inline Tensor new_qtensor_cpu(
IntArrayRef sizes,
const TensorOptions& options,
QuantizerPtr quantizer) {
@@ -40,7 +40,7 @@ inline QTensor new_qtensor_cpu(
int64_t nelements = at::prod_intlist(sizes);
auto dtype = options.dtype();
AT_CHECK(isQIntType(typeMetaToScalarType(dtype)),
- "ScalarType not supported for QTensor in new_qtensor_cpu.");
+ "ScalarType is not supported in new_qtensor_cpu.");
auto storage = c10::make_intrusive<StorageImpl>(
dtype,
nelements,
@@ -76,14 +76,14 @@ qint8 quantize_uint8(float scale, uint8_t zero_point, float value) {
return static_cast<qint8>(qvalue);
}
-QTensor PerTensorAffineQuantizer::quantize(RealTensor tensor) {
+Tensor PerTensorAffineQuantizer::quantize(Tensor tensor) {
IntArrayRef sizes = tensor.sizes();
// Here we need a std::intrusive_ptr<Quantizer>.. but actually "this" is the
// quantizer that can be reused, so I'm using intrusive_from_this here
AT_CHECK(
tensor.options().device() == kCPU,
"quantize only works for CPU backend right now.");
- QTensor qv = new_qtensor_cpu(
+ Tensor qv = new_qtensor_cpu(
sizes,
tensor.options().dtype(at::kQInt8),
intrusive_from_this());
@@ -110,11 +110,11 @@ QTensor PerTensorAffineQuantizer::quantize(RealTensor tensor) {
return qv;
}
-RealTensor PerTensorAffineQuantizer::dequantize(QTensor tensor) {
+Tensor PerTensorAffineQuantizer::dequantize(Tensor tensor) {
std::vector<int64_t> sizes = tensor.sizes().vec();
- at::TensorOptions real_options = tensor.options().dtype(at::kFloat);
+ at::TensorOptions options = tensor.options().dtype(at::kFloat);
- RealTensor rv = at::empty(sizes, real_options);
+ Tensor rv = at::empty(sizes, options);
float* rvd = rv.data<float>();
tensor = tensor.contiguous();
diff --git a/aten/src/ATen/quantized/Quantizer.h b/aten/src/ATen/quantized/Quantizer.h
index f4badcfcf7..e735f8f33c 100644
--- a/aten/src/ATen/quantized/Quantizer.h
+++ b/aten/src/ATen/quantized/Quantizer.h
@@ -14,10 +14,6 @@
namespace at {
struct QTensorImpl;
-
-using QTensor = Tensor;
-using RealTensor = Tensor;
-
struct Quantizer;
using QuantizerPtr = c10::intrusive_ptr<Quantizer>;
@@ -42,7 +38,7 @@ using QuantizerPtr = c10::intrusive_ptr<Quantizer>;
* they should have one to one mapping.
*
* Note about intrusive_ptr:
- * QTensor holds an intrusive_ptr to Quantizer, and multiple Tensor can
+ * Quantized Tensor holds an intrusive_ptr to Quantizer, and multiple Tensor can
* share the same Quantizer. Quantizer should be immutable.
*/
struct CAFFE2_API Quantizer : public c10::intrusive_ptr_target {
@@ -66,12 +62,12 @@ struct CAFFE2_API Quantizer : public c10::intrusive_ptr_target {
/**
* quantize a float Tensor into a quantized Tensor.
*/
- virtual QTensor quantize(RealTensor t) = 0;
+ virtual Tensor quantize(Tensor t) = 0;
/**
* dequantize a quantized Tensor into a float Tensor.
*/
- virtual RealTensor dequantize(QTensor t) = 0;
+ virtual Tensor dequantize(Tensor t) = 0;
};
/**
@@ -173,8 +169,8 @@ struct CAFFE2_API PerTensorAffineQuantizer : public AffineQuantizer {
scale_(scale),
zero_point_(zero_point) {}
- QTensor quantize(RealTensor tensor) override;
- RealTensor dequantize(QTensor tensor) override;
+ Tensor quantize(Tensor tensor) override;
+ Tensor dequantize(Tensor tensor) override;
float scale() const {
return scale_;
@@ -231,7 +227,7 @@ struct CAFFE2_API PerChannelAffineQuantizer : public AffineQuantizer {
// setters/getters for QTensorImpl fields; otherwise, you should use
// the low level setters/getters that were implemented using this.
// This may be called repeatedly, so make sure it's pretty cheap.
-CAFFE2_API QTensorImpl* get_qtensorimpl(const QTensor& self);
+CAFFE2_API QTensorImpl* get_qtensorimpl(const Tensor& self);
// Quantize a float value into a uint8 value given scale and zero_point
CAFFE2_API qint8 quantize_uint8(float scale, uint8_t zero_point, float value);
@@ -241,8 +237,8 @@ CAFFE2_API qint8 quantize_uint8(float scale, uint8_t zero_point, float value);
CAFFE2_API QuantizerPtr
make_per_tensor_affine_quantizer(double scale, int64_t zero_point);
-// Create a QTensor given arguments for normal Tensor and a quantizer
-CAFFE2_API QTensor new_qtensor_cpu(
+// Create a Quantized Tensor given arguments for normal Tensor and a quantizer
+CAFFE2_API Tensor new_qtensor_cpu(
IntArrayRef sizes,
const TensorOptions& options,
QuantizerPtr quantizer);
diff --git a/aten/src/ATen/test/quantized_test.cpp b/aten/src/ATen/test/quantized_test.cpp
index 1c62698729..7163ef7250 100644
--- a/aten/src/ATen/test/quantized_test.cpp
+++ b/aten/src/ATen/test/quantized_test.cpp
@@ -60,7 +60,7 @@ TEST(TestQTensor, EmptyQuantized) {
int zero_point = 10;
int val = 100;
int numel = 10;
- QTensor q = at::_empty_affine_quantized({numel}, at::device(at::kCPU).dtype(kQInt8), scale, zero_point);
+ Tensor q = at::_empty_affine_quantized({numel}, at::device(at::kCPU).dtype(kQInt8), scale, zero_point);
// Assigning to QTensor
auto* q_data = q.data<qint8>();
for (int i = 0; i < numel; ++i) {
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-25 02:43:52 +0000