[quant] PerChannelFloatQParams support for quint4x2 dtype

aten/src/ATen/native/quantized/affine_quantizer.cpp

@@ -174,7 +174,7 @@ Tensor quantize_tensor_per_channel_float_qparams(
   checkSameDevice(fn_name, rtensor, qtensor);
   checkSameSize(fn_name, qtensor, rtensor);
 
-  AT_DISPATCH_QINT_TYPES(qtensor.scalar_type(), fn_name, [&]() {
+  AT_DISPATCH_QINT_AND_SUB_BYTE_TYPES(qtensor.scalar_type(), fn_name, [&]() {
     checkQuantizedTensor<scalar_t>(fn_name, qtensor);
   });
 
@@ -269,7 +269,7 @@ Tensor dequantize_tensor_per_channel_float_qparams(
   checkSameDevice(fn_name, rtensor, qtensor);
   checkSameSize(fn_name, qtensor, rtensor);
 
-  AT_DISPATCH_QINT_TYPES(qtensor.scalar_type(), fn_name, [&]() {
+  AT_DISPATCH_QINT_AND_SUB_BYTE_TYPES(qtensor.scalar_type(), fn_name, [&]() {
     checkQuantizedTensor<scalar_t>(fn_name, qtensor);
   });
 
@@ -410,17 +410,13 @@ CAFFE2_API float dequantize_val(double scale, int64_t zero_point, T value) {
 * Note: For the case of embedding quantization we will set zero_point
 * to (-Xmin/scale), where Xmin is the min value in input tensor row.
 */
-template <typename T>
-T quantize_val_float_qparams(float scale, float zero_point, float value) {
-  int64_t qvalue;
+int quantize_val_float_qparams(float scale, float zero_point, float value, int qmin, int qmax) {
+  int qvalue;
 
-  // TODO make sure qmax and qmin for dtypes other than int8, uint8 is correctly defined.
-  constexpr int64_t qmin = std::numeric_limits<typename T::underlying>::min();
-  constexpr int64_t qmax = std::numeric_limits<typename T::underlying>::max();
   float inv_scale = scale == 0 ? 1.0f : 1.0f / scale;
   qvalue = lrintf(value * inv_scale + zero_point);
   qvalue = std::max(qmin, std::min(qvalue, qmax));
-  return static_cast<T>(qvalue);
+  return qvalue;
 }
 
 template <typename SRC_T, typename DST_T>
@@ -507,11 +503,5 @@ requantize_from_int<quint8>(double, int64_t, int64_t);
 template CAFFE2_API qint32
 requantize_from_int<qint32>(double, int64_t, int64_t);
 
-template CAFFE2_API qint8
-quantize_val_float_qparams<qint8>(float scale, float zero_point, float value);
-template CAFFE2_API quint8
-quantize_val_float_qparams<quint8>(float scale, float zero_point, float value);
-template CAFFE2_API qint32
-quantize_val_float_qparams<qint32>(float scale, float zero_point, float value);
 } // namespace native
 } // namespace at

aten/src/ATen/native/quantized/affine_quantizer.h

@@ -158,8 +158,7 @@ template <typename DST_T>
 CAFFE2_API DST_T
 requantize_from_int(double multiplier, int64_t zero_point, int64_t src);
 
-template <typename T>
-CAFFE2_API T quantize_val_float_qparams(float scale, float zero_point, float value);
+int quantize_val_float_qparams(float scale, float zero_point, float value, int qmin, int qmax);
 
 } // namespace native
 } // namespace at

aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp

@@ -2592,7 +2592,8 @@ void dequantize_per_channel_affine_kernel(
       Tensor rtensor,
       Tensor scales,
       Tensor zero_points,
-      int64_t axis) {
+      int64_t axis,
+      int bit_width=8) {
 
   // For contiguous tensors, e.g. NCHW, arbitrary axis can be used.
   // For channels_last/3d however axis == 0 or 1.
@@ -2611,6 +2612,7 @@ void dequantize_per_channel_affine_kernel(
   check_tensor_memory_format(qtensor, rtensor);
   const auto* qd = qtensor.data_ptr<Q>();
   float* rd = rtensor.data_ptr<float>();
+  const auto elem_per_byte = 8 / bit_width;
   if (axis == 1 && (rtensor.is_contiguous(MemoryFormat::ChannelsLast) ||
       rtensor.is_contiguous(MemoryFormat::ChannelsLast3d))) {
     for (auto b = 0; b < batches; ++b) {
@@ -2619,8 +2621,12 @@ void dequantize_per_channel_affine_kernel(
           auto i = b * channel * elements_per_channel + e * channel + c;
           // We need to convert the qint8 value to float to ensure the
           // subtraction subexpression returns a float
-          rd[i] = (static_cast<float>(qd[i].val_) - zero_points_data[c]) *
-              scales_data[c];
+          auto qvalue = qd[i / elem_per_byte].val_;
+          if (bit_width < 8) {
+            qvalue >>= (i % elem_per_byte) * bit_width;
+            qvalue &= (1 << bit_width) - 1;
+          }
+          rd[i] = (static_cast<float>(qvalue) - zero_points_data[c]) * scales_data[c];
         }
       }
     }
@@ -2632,8 +2638,12 @@ void dequantize_per_channel_affine_kernel(
               c * elements_per_channel + e;
           // We need to convert the qint8 value to float to ensure the
           // subtraction subexpression returns a float
-          rd[i] = (static_cast<float>(qd[i].val_) - zero_points_data[c]) *
-              scales_data[c];
+          auto qvalue = qd[i / elem_per_byte].val_;
+          if (bit_width < 8) {
+            qvalue >>= (i % elem_per_byte) * bit_width;
+            qvalue &= (1 << bit_width) - 1;
+          }
+          rd[i] = (static_cast<float>(qvalue) - zero_points_data[c]) * scales_data[c];
         }
       }
     }
@@ -2667,7 +2677,7 @@ void quantize_tensor_per_channel_float_qparams_cpu(
   TORCH_CHECK(rtensor.is_contiguous() || (axis <=1),
       "If tensor is channels_last contig then per channel quantization "
       "is supported only for axis = 0 or 1.");
-  AT_DISPATCH_QINT_TYPES(
+  AT_DISPATCH_QINT_AND_SUB_BYTE_TYPES(
       qtensor.scalar_type(), "quantize_tensor_per_channel_float_qparams_cpu", [&]() {
         int64_t batches = size_to_dim_(axis, rtensor.sizes());
         int64_t elements_per_channel =
@@ -2677,15 +2687,22 @@ void quantize_tensor_per_channel_float_qparams_cpu(
         auto zero_points_data = zero_points.data_ptr<float>();
         check_tensor_memory_format(rtensor, qtensor);
         const float* rdata = rtensor.data_ptr<float>();
-        auto qdata = qtensor.data_ptr<scalar_t>();
+        auto qdata = reinterpret_cast<underlying_t*>(qtensor.data_ptr<scalar_t>());
+        const auto elem_per_byte = CHAR_BIT / bit_width;
+        int qvalue = 0;
         if (axis == 1 && (rtensor.is_contiguous(MemoryFormat::ChannelsLast) ||
             rtensor.is_contiguous(MemoryFormat::ChannelsLast3d))) {
           for (auto b = 0; b < batches; ++b) {
             for (auto e = 0; e < elements_per_channel; ++e) {
               for (auto c = 0; c < channel; ++c) {
                 auto i = b * channel * elements_per_channel + e * channel + c;
-                qdata[i] = quantize_val_float_qparams<scalar_t>(
-                    scales_data[c], zero_points_data[c], rdata[i]);
+                qvalue = quantize_val_float_qparams(
+                    scales_data[c], zero_points_data[c], rdata[i], quant_min, quant_max);
+                if (i % elem_per_byte == 0) {
+                  qdata[i / elem_per_byte] = static_cast<underlying_t>(qvalue);
+                } else {
+                  qdata[i / elem_per_byte] |= static_cast<underlying_t>(qvalue << ((i % elem_per_byte) * bit_width));
+                }
               }
             }
           }
@@ -2695,8 +2712,13 @@ void quantize_tensor_per_channel_float_qparams_cpu(
               for (auto e = 0; e < elements_per_channel; ++e) {
                 auto i = b * channel * elements_per_channel +
                     c * elements_per_channel + e;
-                qdata[i] = quantize_val_float_qparams<scalar_t>(
-                    scales_data[c], zero_points_data[c], rdata[i]);
+                qvalue = quantize_val_float_qparams(
+                    scales_data[c], zero_points_data[c], rdata[i], quant_min, quant_max);
+                if (i % elem_per_byte == 0) {
+                  qdata[i / elem_per_byte] = static_cast<underlying_t>(qvalue);
+                } else {
+                  qdata[i / elem_per_byte] |= static_cast<underlying_t>(qvalue << ((i % elem_per_byte) * bit_width));
+                }
               }
             }
           }
@@ -2710,9 +2732,9 @@ void dequantize_tensor_per_channel_float_qparams_cpu(
     Tensor scales,
     Tensor zero_points,
     int64_t axis) {
-  AT_DISPATCH_QINT_TYPES(
+  AT_DISPATCH_QINT_AND_SUB_BYTE_TYPES(
       qtensor.scalar_type(), "dequantize_tensor_per_channel_float_qparams_cpu", [&]() {
-        dequantize_per_channel_affine_kernel<float, float, scalar_t>(qtensor, rtensor, scales, zero_points, axis);
+        dequantize_per_channel_affine_kernel<float, float, scalar_t>(qtensor, rtensor, scales, zero_points, axis, bit_width);
       });
 }
 

test/quantization/test_quantized_tensor.py

@@ -429,6 +429,52 @@ def quantize_ref(data, scales, zero_points):
         zero_points = torch.tensor([0.1, 0.2, 1.], dtype=torch.float)
         self._test_quantize_per_channel(r, scales, zero_points, 0, True)
 
+    def test_quantize_per_channel_sub_byte(self):
+        """ Tests the per channel quantization scheme for 4-bit qtensors.
+        The scale and zero point for this have to be in floating point. """
+        r = torch.rand(3, 2, dtype=torch.float) * 4
+        scales = torch.tensor([0.2, 0.3, 0.1], dtype=torch.float)
+        zero_points = torch.tensor([0.1, 0.2, 0.3], dtype=torch.float)
+        qr = torch.quantize_per_channel(r, scales, zero_points, 0, torch.quint4x2)
+        dequant_tensor = qr.dequantize()
+
+        def _get_qranges(bit_width):
+            if bit_width == 4:
+                return 0, 15
+
+        def _quantize_per_channel_sub_byte_ref(data, scales, zero_points, axis, bit_width):
+            dims = data.size()
+            data = data.view(-1, dims[axis], np.prod(dims[axis + 1:]))
+            qtensor_size = math.ceil(data.numel() / 2)
+            res = torch.empty(qtensor_size, dtype=torch.uint8)
+            elem_per_byte = 8 / bit_width
+            quant_min, quant_max = _get_qranges(bit_width)
+            for i in range(data.size()[0]):
+                for j in range(data.size()[1]):
+                    for k in range(data.size()[2]):
+                        inv_scale = 1.0 / scales[j]
+                        index = i * data.size()[1] * data.size()[2] + j * data.size()[2] + k
+                        qvalue = np.clip(
+                            np.round(data[i][j][k] * inv_scale + zero_points[j]), quant_min, quant_max).to(dtype=torch.int)
+                        res_idx = int(index / elem_per_byte)
+                        if (index % elem_per_byte == 0):
+                            res[res_idx] = qvalue
+                        else:
+                            res[res_idx] |= (qvalue << ((index % elem_per_byte) * bit_width))
+            return res
+
+        ref_res = _quantize_per_channel_sub_byte_ref(r, scales, zero_points, 0, 4)
+        self.assertTrue(np.allclose(qr.int_repr(), ref_res))
+        self.assertTrue(np.allclose(r.numpy(), dequant_tensor.numpy(), atol=1 / np.min(scales.numpy())))
+
+        # Check 4D tensor with non-zero axis.
+        r = torch.rand(3, 2, 4, 5, dtype=torch.float) * 4
+        scales = torch.tensor([0.2, 0.03], dtype=torch.float)
+        zero_points = torch.tensor([0.1, 0.2], dtype=torch.float)
+        qr = torch.quantize_per_channel(r, scales, zero_points, axis=1, dtype=torch.quint4x2)
+        ref_res = _quantize_per_channel_sub_byte_ref(r, scales, zero_points, 1, 4)
+        self.assertTrue(np.allclose(qr.int_repr(), ref_res))
+
     def test_qtensor_permute(self):
         scale = 0.02
         zero_point = 1
@@ -516,7 +562,9 @@ def test_qtensor_per_channel_load_save(self):
         scales = torch.rand(10, dtype=torch.double) * 0.02 + 0.01
         zero_points = torch.round(torch.rand(10) * 20 + 1).to(torch.long)
         # quint32, cuda is not supported yet
-        for dtype in [torch.quint8, torch.qint8]:
+        for dtype in [torch.quint8, torch.qint8, torch.quint4x2]:
+            if dtype == torch.quint4x2:
+                zero_points = torch.ones(10, dtype=torch.float)
             qr = torch.quantize_per_channel(r, scales, zero_points, 1, dtype)
             with tempfile.NamedTemporaryFile() as f:
                 # Serializing and Deserializing Tensor