* Copyright (c) 2025 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.
*/
#include <gtest/gtest.h>
#define private public
#define protected public
#include "kernel_operator.h"
#include "kernel_utils.h"
#include <math.h>
#include <iostream>
#include <fstream>
using namespace std;
using namespace AscendC;
namespace AscendC {
template <typename srcType, typename dstType, typename scaleType, bool has_offset, bool has_workspace, bool tensor_offset>
class KernelQuantizePerToken {
public:
__aicore__ inline KernelQuantizePerToken() {}
__aicore__ inline void Init(GM_ADDR src_gm, GM_ADDR dst_gm, GM_ADDR scale_gm, GM_ADDR offset_gm,
uint32_t m_num, uint32_t n_num, uint32_t calCount_num)
{
m = m_num;
n = n_num;
calCount = calCount_num;
output_m = m;
uint32_t align_size = 32;
src_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src_gm), m*n);
dst_global.SetGlobalBuffer(reinterpret_cast<__gm__ dstType *>(dst_gm), m*n);
scale_global.SetGlobalBuffer(reinterpret_cast<__gm__ scaleType *>(scale_gm), align_size);
offset_global.SetGlobalBuffer(reinterpret_cast<__gm__ scaleType *>(offset_gm), align_size);
pipe.InitBuffer(inQueue, 1, m * n * sizeof(srcType));
pipe.InitBuffer(inQueueScale, 1, align_size * sizeof(scaleType));
pipe.InitBuffer(inQueueOffset, 1, align_size * sizeof(scaleType));
pipe.InitBuffer(outQueue, 1, m * n * sizeof(dstType));
}
__aicore__ inline void Process()
{
CopyIn();
Compute();
CopyOut();
}
private:
__aicore__ inline void CopyIn()
{
LocalTensor<srcType> srcLocal = inQueue.AllocTensor<srcType>();
LocalTensor<scaleType> scaleLocal = inQueueScale.AllocTensor<scaleType>();
LocalTensor<scaleType> offsetLocal = inQueueOffset.AllocTensor<scaleType>();
DataCopy(srcLocal, src_global, m*n);
DataCopy(scaleLocal, scale_global, 16);
DataCopy(offsetLocal, offset_global, 16);
inQueue.EnQue(srcLocal);
inQueueScale.EnQue(scaleLocal);
inQueueOffset.EnQue(offsetLocal);
}
__aicore__ inline void Compute()
{
LocalTensor<dstType> dstLocal = outQueue.AllocTensor<dstType>();
LocalTensor<srcType> srcLocal = inQueue.DeQue<srcType>();
LocalTensor<scaleType> scaleLocal = inQueueScale.DeQue<scaleType>();
LocalTensor<scaleType> offsetLocal = inQueueOffset.DeQue<scaleType>();
constexpr static QuantizeConfig config = {QuantizePolicy::PER_TOKEN, has_offset};
QuantizeParams para;
para.m = m;
para.n = n;
if constexpr (has_workspace) {
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
if constexpr (tensor_offset) {
Quantize<config>(dstLocal, srcLocal, sharedTmpBuffer, scaleLocal, offsetLocal, para);
} else {
Quantize<config>(dstLocal, srcLocal, sharedTmpBuffer, scaleLocal, offsetLocal.GetValue(0), para);
}
} else {
if constexpr (tensor_offset) {
Quantize<config>(dstLocal, srcLocal, scaleLocal, offsetLocal, para);
} else {
Quantize<config>(dstLocal, srcLocal, scaleLocal, offsetLocal.GetValue(0), para);
}
}
outQueue.EnQue<dstType>(dstLocal);
inQueue.FreeTensor(srcLocal);
inQueueScale.FreeTensor(scaleLocal);
inQueueOffset.FreeTensor(offsetLocal);
}
__aicore__ inline void CopyOut()
{
LocalTensor<dstType> dstLocal = outQueue.DeQue<dstType>();
DataCopy(dst_global, dstLocal, output_m * n);
outQueue.FreeTensor(dstLocal);
}
private:
GlobalTensor<srcType> src_global;
GlobalTensor<scaleType> scale_global;
GlobalTensor<scaleType> offset_global;
GlobalTensor<dstType> dst_global;
TPipe pipe;
TQue<TPosition::VECIN, 1> inQueue, inQueueScale, inQueueOffset;
TQue<TPosition::VECOUT, 1> outQueue;
uint32_t m;
uint32_t n;
uint32_t calCount;
uint32_t output_m;
};
}
template <typename srcType, typename dstType, typename scaleType, bool has_offset, bool has_workspace, bool tensor_offset>
__global__ __aicore__ void Entry(__gm__ uint8_t* srcGm, __gm__ uint8_t* dstGm, __gm__ uint8_t* scaleGm, __gm__ uint8_t* offsetGm,
uint32_t m, uint32_t n, uint32_t calCount)
{
AscendC::KernelQuantizePerToken<srcType, dstType, scaleType, has_offset, has_workspace, tensor_offset> op;
op.Init(srcGm, dstGm, scaleGm, offsetGm, m, n, calCount);
op.Process();
}
struct QuantizePerTokenTestParams {
uint32_t m;
uint32_t n;
uint32_t calCount;
void (*cal_func)(uint8_t*, uint8_t*, uint8_t*, uint8_t*, uint32_t, uint32_t, uint32_t);
};
class QuantizePerTokenTestsuite : public testing::Test, public testing::WithParamInterface<QuantizePerTokenTestParams> {
protected:
void SetUp()
{
AscendC::SetGCoreType(2);
}
void TearDown()
{
AscendC::SetGCoreType(0);
}
};
INSTANTIATE_TEST_CASE_P(TEST_OPEARATION_ASCEND_QUANT_PER_TOKEN, QuantizePerTokenTestsuite,
::testing::Values(
QuantizePerTokenTestParams{ 10, 256, 1280, Entry<half, fp8_e5m2_t, half, false, true, true> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<half, fp8_e4m3fn_t, half, true, false, true> },
QuantizePerTokenTestParams{ 7, 256, 512, Entry<bfloat16_t, fp8_e5m2_t, bfloat16_t, false, true, true> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<bfloat16_t, fp8_e4m3fn_t, bfloat16_t, true, false, true> },
QuantizePerTokenTestParams{ 10, 256, 1280, Entry<float, fp8_e5m2_t, float, false, true, true> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<float, fp8_e4m3fn_t, float, true, false, true> },
QuantizePerTokenTestParams{ 7, 256, 512, Entry<half, fp8_e5m2_t, float, false, true, true> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<half, fp8_e4m3fn_t, float, true, false, true> },
QuantizePerTokenTestParams{ 7, 256, 512, Entry<bfloat16_t, fp8_e5m2_t, float, false, true, true> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<bfloat16_t, fp8_e4m3fn_t, float, true, false, true> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<half, hifloat8_t, half, true, false, true> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<bfloat16_t, hifloat8_t, bfloat16_t, true, false, true> },
QuantizePerTokenTestParams{ 6, 288, 1440, Entry<float, hifloat8_t, float, true, true, true> },
QuantizePerTokenTestParams{ 8, 512, 3584, Entry<half, hifloat8_t, float, true, false, true> },
QuantizePerTokenTestParams{ 4, 256, 768, Entry<bfloat16_t, hifloat8_t, float, true, true, true> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<half, int8_t, half, true, false, true> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<bfloat16_t, int8_t, bfloat16_t, true, false, true> },
QuantizePerTokenTestParams{ 6, 288, 1440, Entry<float, int8_t, float, true, true, true> },
QuantizePerTokenTestParams{ 8, 512, 3584, Entry<half, int8_t, float, true, false, true> },
QuantizePerTokenTestParams{ 4, 256, 768, Entry<bfloat16_t, int8_t, float, true, true, true> },
QuantizePerTokenTestParams{ 10, 256, 1280, Entry<half, fp8_e5m2_t, half, false, true, false> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<half, fp8_e4m3fn_t, half, true, false, false> },
QuantizePerTokenTestParams{ 10, 256, 1280, Entry<float, fp8_e5m2_t, float, false, true, false> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<float, fp8_e4m3fn_t, float, true, false, false> },
QuantizePerTokenTestParams{ 7, 256, 512, Entry<half, fp8_e5m2_t, float, false, true, false> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<half, fp8_e4m3fn_t, float, true, false, false> },
QuantizePerTokenTestParams{ 7, 256, 512, Entry<bfloat16_t, fp8_e5m2_t, float, false, true, false> },
QuantizePerTokenTestParams{ 7, 288, 864, Entry<bfloat16_t, fp8_e4m3fn_t, float, true, false, false> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<half, hifloat8_t, half, true, false, false> },
QuantizePerTokenTestParams{ 6, 288, 1440, Entry<float, hifloat8_t, float, true, true, false> },
QuantizePerTokenTestParams{ 8, 512, 3584, Entry<half, hifloat8_t, float, true, false, false> },
QuantizePerTokenTestParams{ 4, 256, 768, Entry<bfloat16_t, hifloat8_t, float, true, true, false> },
QuantizePerTokenTestParams{ 10, 288, 1152, Entry<half, int8_t, half, true, false, false> },
QuantizePerTokenTestParams{ 6, 288, 1440, Entry<float, int8_t, float, true, true, false> },
QuantizePerTokenTestParams{ 8, 512, 3584, Entry<half, int8_t, float, true, false, false> },
QuantizePerTokenTestParams{ 4, 256, 768, Entry<bfloat16_t, int8_t, float, true, true, false> }
));
TEST_P(QuantizePerTokenTestsuite, QuantizePerTokenOpTestCase)
{
const int dataBytes = 4;
auto param = GetParam();
uint8_t srcGm[param.m * param.n * dataBytes];
uint8_t dstGm[param.m * param.n * dataBytes];
uint8_t scaleGm[param.m * dataBytes];
uint8_t offsetGm[param.m * dataBytes];
param.cal_func(srcGm, dstGm, scaleGm, offsetGm, param.m, param.n, param.calCount);
}
