* Copyright (c) 2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file uniform.cpp
* \brief Uniform random number generator implementation
*/
#include "interface/utils/operator_tracer.h"
#include "passes/pass_utils/graph_utils.h"
#include "interface/function/function.h"
#include "interface/program/program.h"
#include "tilefwk/error_code.h"
#include "interface/operation/operation_common.h"
namespace npu::tile_fwk {
namespace {
const std::vector<NPUArch> UNIFORM_SUPPORTED_ARCHITECTURES = {NPUArch::DAV_3510};
}
LogicalTensorPtr TensorUniform(
Function& function, LogicalTensorPtr& result, const Element& key, const SymbolicScalar& counter0,
const Element& counter1, const Element& rounds, const std::vector<int64_t>& shape, DataType dtype)
{
auto& op = function.AddOperation(Opcode::OP_UNIFORM, {}, {result});
std::vector<Element> scalars = {key, counter1, rounds, Element(DT_INT32, static_cast<int32_t>(dtype))};
op.SetAttribute(OpAttributeKey::vectorScalar, scalars);
op.SetAttribute(OpAttributeKey::dynScalar, counter0);
op.SetAttribute(OP_ATTR_PREFIX + "SHAPE", shape);
return result;
}
static void TiledUniformBuildIn(
Function& function, const TileShape& tileShape, const LogicalTensorPtr& result, TileInfo& resultTileInfo,
uint64_t key, const SymbolicScalar& counter0, uint64_t counter1, uint16_t rounds, DataType dtype)
{
auto& vecTile = tileShape.GetVecTile();
const size_t ALIGN_SIZE = 32;
int64_t uint32BufferSize = vecTile[0];
int64_t uint32BufferBytes = ((uint32BufferSize * BytesOf(DT_UINT32) + ALIGN_SIZE - 1) / ALIGN_SIZE) * ALIGN_SIZE;
int64_t tempByteSize = uint32BufferBytes;
if (dtype == DT_FP16 || dtype == DT_BF16) {
int64_t uint32BufferLowBytes =
((uint32BufferSize * BytesOf(DT_UINT32) + ALIGN_SIZE - 1) / ALIGN_SIZE) * ALIGN_SIZE;
tempByteSize += uint32BufferLowBytes;
}
for (int64_t i = 0; i < result->shape[0]; i += vecTile[0]) {
resultTileInfo.offset[0] = i;
resultTileInfo.shape[0] = std::min(result->shape[0] - resultTileInfo.offset[0], vecTile[0]);
auto resultTile = result->View(function, resultTileInfo.shape, resultTileInfo.offset);
std::vector<int64_t> tempShape({static_cast<int64_t>(tempByteSize)});
auto tempTensor = std::make_shared<LogicalTensor>(function, DT_UINT8, tempShape);
SymbolicScalar tileCounter0 = counter0 + i / 4;
uint64_t tileCounter1 = counter1;
auto& op = function.AddOperation(Opcode::OP_UNIFORM, {}, {resultTile, tempTensor});
std::vector<Element> scalars = {
Element(DT_UINT64, key), Element(DT_UINT64, tileCounter1), Element(DT_UINT16, rounds),
Element(DT_INT32, static_cast<int32_t>(dtype))};
op.SetAttribute(OpAttributeKey::vectorScalar, scalars);
op.SetAttribute(OpAttributeKey::dynScalar, tileCounter0);
op.SetAttribute(OP_ATTR_PREFIX + "SHAPE", resultTileInfo.shape);
}
}
void UniformOperationTileFunc(
Function& function, const TileShape& tileShape, [[maybe_unused]] const std::vector<LogicalTensorPtr>& iOperand,
const std::vector<LogicalTensorPtr>& oOperand, const Operation& op)
{
auto& vecTile = tileShape.GetVecTile();
for (size_t i = 0; i < vecTile.size(); ++i) {
ASSERT(VectorErrorCode::ERR_PARAM_INVALID, vecTile[i] % 4 == 0)
<< "Uniform: tileShape[" << i << "] must be a multiple of 4, but got " << vecTile[i];
}
auto scalars = op.GetVectorElementAttribute(OpAttributeKey::vectorScalar);
uint64_t key = scalars[0].Cast<uint64_t>();
uint64_t counter1 = scalars[1].Cast<uint64_t>();
uint16_t rounds = scalars[2].Cast<uint16_t>();
DataType dtype = static_cast<DataType>(scalars[3].Cast<int32_t>());
SymbolicScalar counter0 = op.GetSymbolicScalarAttribute(OpAttributeKey::dynScalar);
auto shapeAttr = op.GetVectorIntAttribute(OP_ATTR_PREFIX + "SHAPE");
std::vector<int64_t> shape;
for (auto dim : shapeAttr) {
shape.push_back(static_cast<int64_t>(dim));
}
TileInfo resultTileInfo(shape.size(), shape.size());
TiledUniformBuildIn(function, tileShape, oOperand[0], resultTileInfo, key, counter0, counter1, rounds, dtype);
}
static Tensor RealUniform(
uint64_t key, const SymbolicScalar& counter0, uint64_t counter1, const std::vector<int64_t>& shape, uint16_t rounds,
DataType dtype)
{
DECLARE_TRACER();
auto resTensor = Tensor(dtype, shape);
RETURN_CALL(
Uniform, *Program::GetInstance().GetCurrentFunction(), resTensor.GetStorage(), Element(DT_UINT64, key),
counter0, Element(DT_UINT64, counter1), Element(DT_UINT16, rounds), shape, dtype);
}
Tensor Uniform(
const Element& key, const SymbolicScalar& counter0, const Element& counter1, const std::vector<int64_t>& shape,
const Element& rounds, DataType dtype)
{
DECLARE_TRACER();
CheckSupportedNPUArch(UNIFORM_SUPPORTED_ARCHITECTURES, "Uniform");
uint16_t roundsVal = rounds.Cast<uint16_t>();
ASSERT(VectorErrorCode::ERR_PARAM_INVALID, shape.size() == 1) << "Uniform: shape must be 1-dimensional";
ASSERT(VectorErrorCode::ERR_PARAM_INVALID, roundsVal == 7 || roundsVal == 10) << "Uniform: rounds must be 7 or 10";
ASSERT(VectorErrorCode::ERR_PARAM_INVALID, dtype == DT_FP32 || dtype == DT_FP16 || dtype == DT_BF16)
<< "Uniform: dtype must be DT_FP32, DT_FP16 or DT_BF16";
CheckDstShapeSize(shape, "UNIFORM");
return RealUniform(key.Cast<uint64_t>(), counter0, counter1.Cast<uint64_t>(), shape, roundsVal, dtype);
}
REGISTER_OPERATION_TILED_FUNC(OP_UNIFORM, Opcode::OP_UNIFORM, UniformOperationTileFunc);
}
