[luci/pass] introduce quantize weight with GPTQ pass

compiler/luci/pass/include/luci/Pass/QuantizeWeightsWithGPTQPass.h

@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __LUCI_QUANTIZE_WEIGHTS_WITH_GPTQ_PASS_H__
+#define __LUCI_QUANTIZE_WEIGHTS_WITH_GPTQ_PASS_H__
+
+#include <luci/Pass/QuantizationParameters.h>
+#include <luci/IR/CircleNode.h>
+
+#include <logo/Pass.h>
+#include <loco.h>
+
+#include <unordered_map>
+
+namespace luci
+{
+
+using HessianMap = std::unordered_map<const luci::CircleNode *, std::vector<float>>;
+
+/**
+ * @brief Pass to quantize weights
+ */
+class QuantizeWeightsWithGPTQPass : public logo::Pass
+{
+public:
+  struct Context
+  {
+    loco::DataType input_model_dtype = loco::DataType::Unknown;
+    loco::DataType output_model_dtype = loco::DataType::Unknown;
+    QuantizationGranularity granularity = QuantizationGranularity::ChannelWise;
+    std::vector<LayerInfo> layers_info;
+  };
+
+public:
+  QuantizeWeightsWithGPTQPass(std::unique_ptr<Context> &&ctx) : _ctx{std::move(ctx)}
+  {
+    // DO NOTHING
+  }
+
+  QuantizeWeightsWithGPTQPass(std::unique_ptr<Context> &&ctx, HessianMap *hessian_map)
+    : _ctx{std::move(ctx)}, _hessian_map{hessian_map}
+  {
+    // DO NOTHING
+  }
+
+public:
+  QuantizeWeightsWithGPTQPass(loco::DataType input_model_dtype, loco::DataType output_model_dtype,
+                              QuantizationGranularity granularity)
+  {
+    _ctx = std::make_unique<Context>();
+    {
+      _ctx->input_model_dtype = input_model_dtype;
+      _ctx->output_model_dtype = output_model_dtype;
+      _ctx->granularity = granularity;
+    }
+  }
+  virtual const char *name(void) const { return "luci::QuantizeWeightsWithGPTQPass"; }
+
+public:
+  bool run(loco::Graph *graph);
+
+private:
+  std::unique_ptr<Context> _ctx;
+  HessianMap *_hessian_map;
+};
+
+} // namespace luci
+
+#endif //__LUCI_QUANTIZE_WEIGHTS_WITH_GPTQ_PASS_H__

compiler/luci/pass/src/QuantizationUtils.cpp

@@ -292,6 +292,11 @@ uint32_t cal_offset(loco::TensorShape &dimension, uint32_t *indices)
          indices[2] * dimension.dim(3).value() + indices[3];
 }
 
+uint32_t cal_offset_2d(loco::TensorShape &dimension, uint32_t *indices)
+{
+  return indices[0] * dimension.dim(1).value() + indices[1];
+}
+
 // Activation (ofm) qtype is determined in different ways.
 // 1. Pre-defined values: Some Ops have pre-defined qparams (ex: LOGISTIC, TANH)
 // 2. Integer scale: Output of some Ops should be integers (ex: FLOOR, CEIL)

compiler/luci/pass/src/QuantizationUtils.h

@@ -53,6 +53,7 @@ bool get_channel_dim_index(CircleConst *node, loco::TensorShape &dimension,
 
 // Calculate offset of the given indices in dimension
 uint32_t cal_offset(loco::TensorShape &dimension, uint32_t *indices);
+uint32_t cal_offset_2d(loco::TensorShape &dimension, uint32_t *indices);
 
 // Backward propagation of concatenation qparam
 void propagate_concat_quantparam(luci::CircleConcatenation *concat);
@@ -63,6 +64,8 @@ void propagate_pad_v2_quantparam(luci::CirclePadV2 *pad_v2);
 // Return true if the node is quantized
 bool is_quantized(const CircleNode *node);
 
+uint8_t fp32_to_uint8_cast(float f);
+
 // Return true if the node is fp32
 bool is_fp32(const CircleNode *node);