Use symmetric weights for convs and int8 in the default quantizer (#8344)

Summary:

As titled. int8 should give better performance with Cadence kernels, since they're not improving uint8 anymore.
The upcoming (quantized) convolution kernel needs symmetric weights, so we make that change as well.

Reviewed By: zonglinpeng

Differential Revision: D69405797

Author: mcremon-meta

backends/cadence/aot/export_example.py

@@ -9,8 +9,6 @@
 import logging
 import tempfile
 
-import torch
-
 from executorch.backends.cadence.aot.ops_registrations import *  # noqa
 from typing import Any, Tuple
 
@@ -23,38 +21,15 @@
 from executorch.backends.cadence.aot.quantizer.quantizer import CadenceDefaultQuantizer
 from executorch.backends.cadence.runtime import runtime
 from executorch.backends.cadence.runtime.executor import BundledProgramManager
-from executorch.backends.xnnpack.quantizer.xnnpack_quantizer_utils import (
-    QuantizationConfig,
-    QuantizationSpec,
-)
 from executorch.exir import ExecutorchProgramManager
 from torch import nn
-from torch.ao.quantization.observer import HistogramObserver, MinMaxObserver
 
 from .utils import save_bpte_program, save_pte_program
 
 
 FORMAT = "[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s"
 logging.basicConfig(level=logging.INFO, format=FORMAT)
 
-act_qspec = QuantizationSpec(
-    dtype=torch.int8,
-    quant_min=-128,
-    quant_max=127,
-    qscheme=torch.per_tensor_affine,
-    is_dynamic=False,
-    observer_or_fake_quant_ctr=HistogramObserver.with_args(eps=2**-12),
-)
-
-wgt_qspec = QuantizationSpec(
-    dtype=torch.int8,
-    quant_min=-128,
-    quant_max=127,
-    qscheme=torch.per_tensor_affine,
-    is_dynamic=False,
-    observer_or_fake_quant_ctr=MinMaxObserver,
-)
-
 
 def export_model(
     model: nn.Module,
@@ -66,15 +41,8 @@ def export_model(
     working_dir = tempfile.mkdtemp(dir="/tmp")
     logging.debug(f"Created work directory {working_dir}")
 
-    qconfig = QuantizationConfig(
-        act_qspec,
-        act_qspec,
-        wgt_qspec,
-        None,
-    )
-
     # Instantiate the quantizer
-    quantizer = CadenceDefaultQuantizer(qconfig)
+    quantizer = CadenceDefaultQuantizer()
 
     # Convert the model
     converted_model = convert_pt2(model, example_inputs, quantizer)

backends/cadence/aot/ops_registrations.py

@@ -576,7 +576,7 @@ def quantized_relu_per_tensor_meta(
     out_multiplier: int,
     out_shift: int,
 ) -> torch.Tensor:
-    return input.new_empty(input.size(), dtype=torch.uint8)
+    return input.new_empty(input.size(), dtype=input.dtype)
 
 
 @register_fake("cadence::fully_connected")

backends/cadence/aot/quantizer/quantizer.py

@@ -40,30 +40,46 @@
 from torch.ao.quantization.quantizer.composable_quantizer import ComposableQuantizer
 
 
-act_qspec = QuantizationSpec(
-    dtype=torch.uint8,
-    quant_min=0,
-    quant_max=255,
+act_qspec_asym8u = QuantizationSpec(
+    dtype=torch.int8,
+    quant_min=-128,
+    quant_max=127,
     qscheme=torch.per_tensor_affine,
     is_dynamic=False,
     observer_or_fake_quant_ctr=HistogramObserver.with_args(eps=2**-12),
 )
 
-wgt_qspec = QuantizationSpec(
-    dtype=torch.uint8,
-    quant_min=0,
-    quant_max=255,
+wgt_qspec_asym8u = QuantizationSpec(
+    dtype=torch.int8,
+    quant_min=-128,
+    quant_max=127,
     qscheme=torch.per_tensor_affine,
     is_dynamic=False,
     observer_or_fake_quant_ctr=MinMaxObserver,
 )
 
+wgt_qspec_asym8s = QuantizationSpec(
+    dtype=torch.int8,
+    quant_min=-128,
+    quant_max=127,
+    qscheme=torch.per_tensor_symmetric,
+    is_dynamic=False,
+    observer_or_fake_quant_ctr=MinMaxObserver,
+)
+
 bias_qspec: Optional[QuantizationSpec] = None
 
-_default_qconfig = QuantizationConfig(
-    act_qspec,
-    act_qspec,
-    wgt_qspec,
+qconfig_A8uW8u = QuantizationConfig(
+    act_qspec_asym8u,
+    act_qspec_asym8u,
+    wgt_qspec_asym8u,
+    None,
+)
+
+qconfig_A8uW8s = QuantizationConfig(
+    act_qspec_asym8u,
+    act_qspec_asym8u,
+    wgt_qspec_asym8s,
     None,
 )
 
@@ -147,19 +163,17 @@ def get_supported_operators(cls) -> List[OperatorConfig]:
         return []
 
 
-def get_cadence_default_quantizer_list_with_config(
-    quantization_config: QuantizationConfig,
-) -> List[Quantizer]:
+def get_cadence_default_quantizers() -> List[Quantizer]:
     return [
-        CadenceAtenQuantizer(AddmmPattern(), quantization_config),
-        CadenceAtenQuantizer(BmmPattern(), quantization_config),
-        CadenceAtenQuantizer(Conv1dPattern(), quantization_config),
-        CadenceAtenQuantizer(Conv2dPattern(), quantization_config),
-        CadenceAtenQuantizer(LayerNormPattern(), quantization_config),
-        CadenceAtenQuantizer(LinearPattern(), quantization_config),
-        CadenceAtenQuantizer(MatmulPattern(), quantization_config),
-        CadenceAtenQuantizer(ReluPattern0(), quantization_config),
-        CadenceAtenQuantizer(ReluPattern1(), quantization_config),
+        CadenceAtenQuantizer(AddmmPattern(), qconfig_A8uW8u),
+        CadenceAtenQuantizer(BmmPattern(), qconfig_A8uW8u),
+        CadenceAtenQuantizer(Conv1dPattern(), qconfig_A8uW8s),
+        CadenceAtenQuantizer(Conv2dPattern(), qconfig_A8uW8s),
+        CadenceAtenQuantizer(LayerNormPattern(), qconfig_A8uW8u),
+        CadenceAtenQuantizer(LinearPattern(), qconfig_A8uW8u),
+        CadenceAtenQuantizer(MatmulPattern(), qconfig_A8uW8u),
+        CadenceAtenQuantizer(ReluPattern0(), qconfig_A8uW8u),
+        CadenceAtenQuantizer(ReluPattern1(), qconfig_A8uW8u),
     ]
 
 
@@ -178,10 +192,9 @@ class CadenceDefaultQuantizer(CadenceQuantizer):
     Default quantizer for Cadence backend.
     """
 
-    def __init__(self, qconfig: Optional[QuantizationConfig] = None) -> None:
-        if qconfig is None:
-            qconfig = _default_qconfig
-        quantizers = get_cadence_default_quantizer_list_with_config(qconfig)
+    def __init__(self, quantizers: Optional[list[Quantizer]] = None) -> None:
+        if quantizers is None:
+            quantizers = get_cadence_default_quantizers()
         super().__init__(quantizers)
 
 

backends/cadence/hifi/operators/op_quantized_relu_out.cpp

@@ -18,33 +18,6 @@ namespace impl {
 namespace HiFi {
 namespace native {
 
-template <typename T>
-void quantized_relu_(
-    const Tensor& input,
-    const Tensor& in_zero_point,
-    const int64_t out_zero_point,
-    const Tensor& out_multiplier,
-    const Tensor& out_shift,
-    Tensor& output) {
-  T q_zero_point = in_zero_point.const_data_ptr<T>()[0];
-  const T* __restrict__ in = input.const_data_ptr<T>();
-  T* __restrict__ out = output.mutable_data_ptr<T>();
-
-  const int32_t* __restrict__ out_multiplier_data =
-      out_multiplier.const_data_ptr<int32_t>();
-  const int32_t* __restrict__ out_shift_data =
-      out_shift.const_data_ptr<int32_t>();
-
-  // Compute the out_scale from out_multiplier and out_shift
-  const float out_scale =
-      -out_multiplier_data[0] * 1.0 / (1 << 31) * pow(2, out_shift_data[0]);
-
-  for (size_t i = 0, e = input.numel(); i < e; ++i) {
-    float temp = in[i] > q_zero_point ? (in[i] - q_zero_point) : 0;
-    out[i] = kernels::quantize<T>(temp, out_scale, (int32_t)out_zero_point);
-  }
-}
-
 void quantized_relu_per_tensor_out(
     KernelRuntimeContext& ctx,
     const Tensor& input,
@@ -68,7 +41,7 @@ void quantized_relu_per_tensor_out(
         _out_multiplier,
         _out_shift,
         _out_zero_point,
-        _out_zero_point,
+        0,
         255,
         input.numel());
 
@@ -85,7 +58,7 @@ void quantized_relu_per_tensor_out(
         _out_multiplier,
         _out_shift,
         _out_zero_point,
-        _out_zero_point,
+        -128,
         127,
         input.numel());
 
@@ -107,9 +80,7 @@ void quantized_relu_per_tensor_out(
     const Tensor& out_multiplier,
     const Tensor& out_shift,
     Tensor& output) {
-  const uint8_t* p_in = input.const_data_ptr<uint8_t>();
-  uint8_t* p_out = output.mutable_data_ptr<uint8_t>();
-  uint8_t _in_zero_point = in_zero_point.const_data_ptr<uint8_t>()[0];
+  int8_t _in_zero_point = in_zero_point.const_data_ptr<int8_t>()[0];
   int32_t _out_multiplier = out_multiplier.const_data_ptr<int32_t>()[0];
   int32_t _out_shift = out_shift.const_data_ptr<int32_t>()[0];