aclnnQuantGroupedMatmulDequant

Applicable Products

Product	Supported
Ascend 950PR/Ascend 950DT	×
Atlas A3 training products/Atlas A3 inference products	×
Atlas A2 training products/Atlas A2 inference products	×
Atlas 200I/500 A2 inference products	×
Atlas inference products	√
Atlas training products	×

Function

Description: Quantizes the input x, performs grouped matrix multiplication, and then dequantizes the result.
Formula:
1. If smoothScaleOptional is input, then
$x = x\cdot scale_{smooth}$
1. If xScaleOptional is not input, dynamic quantization is used, and the x quantization coefficient needs to be calculated.
$scale_{x}=row\_max(abs(x))/max_{quantDataType}$
1. Quantization
$x_{quantized}=round(x/scale_{x})$
1. Grouped matrix multiplication + dequantization
- 4.1 If the data type of the input $scale_{weight}$ is FLOAT32:
$\begin{aligned} x^{*}_{quantized} &= x_{quantized}[group[i-1]:group[i]]\\ out^{*}_{quantized} &= out_{quantized}[group[i-1]:group[i]]\\ scale^{*}_{x} &= \begin{cases} scale_{x}[group[i-1]:group[i]] & pertoken \\ scale_{x} & pertensor \\ \end{cases} \\ out^{*}_{quantized} &= (x^{*}_{quantized}@weight_{quantized}[i] + bias) * scale_{weight}[i] * scale^{*}_{x} \end{aligned}$
- 4.2 If the data type of the input $scale_{weight}$ is INT64: $x^{*}_{quantized} = x_{quantized}[group[i-1]:group[i]] \\ out^{*}_{quantized} = out_{quantized}[group[i-1]:group[i]] \\ scale_{weight} = torch.tensor(np.frombuffer(scale_{weight}.numpy().astype(np.int32).\\tobytes(), dtype=np.float32)).reshape(scale_{weight}) \\ out^{*}_{quantized} = (x^{*}_{quantized}@weight_{quantized}[i] + bias) * scale_{weight}[i]$ Special note: In scenario 4.2 above, $scale_{weight}$ has already undergone a matrix multiplication operation with $scale_{x}$ before being input, so the operator omits this step during internal computation. This requires that the quantization mode must be per-tensor static quantization. That is, before input, $scale_{weight}$ must be processed as follows to obtain data of the INT64 type: $scale_{weight} = scale_{weight} * scale_{x} \\ scale_{weight} = torch.tensor(np.frombuffer(scale_{weight}.numpy().astype(np.float32). \\tobytes(), dtype=np.int32).astype(np.int64)).reshape(scale_{weight})$

Function Prototype

Each operator is divided into a two-phase API. You must first call the "aclnnQuantGroupedMatmulDequantGetWorkspaceSize" API to obtain the input parameters and calculate the required workspace size based on the process, and then call the "aclnnQuantGroupedMatmulDequant" API to perform the computation.

aclnnStatus aclnnQuantGroupedMatmulDequantGetWorkspaceSize(
  const aclTensor *x,
  const aclTensor *weight,
  const aclTensor *weightScale,
  const aclTensor *groupList,
  const aclTensor *biasOptional,
  const aclTensor *xScaleOptional,
  const aclTensor *xOffsetOptional,
  const aclTensor *smoothScaleOptional,
  char            *xQuantMode,
  bool             transposeWeight,
  const aclTensor *out,
  uint64_t        *workspaceSize,
  aclOpExecutor  **executor)

aclnnStatus aclnnQuantGroupedMatmulDequant(
  void          *workspace,
  uint64_t       workspaceSize,
  aclOpExecutor *executor,
  aclrtStream    stream)

aclnnQuantGroupedMatmulDequantGetWorkspaceSize

Parameters

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-Contiguous Tensor
x	Input	Represents the left matrix of the input. Required parameter, x in the formula.	Empty tensors are not supported. The shape supports 2 dimensions, with each dimension representing: (m, k).	FLOAT16	ND	2	√
weight	Input	Represents the right matrix of the input. Required parameter. weight_{quantized} in the formula.	Empty tensors are not supported.	INT8	FRACTAL_NZ, ND	3, 5	Yes
weightScale	Input	Represents the quantization coefficient of weight. Required parameter. Corresponds to scale_{weight} in the formula.	The shape is 2D (g, n), where g and n are consistent with g and n of weight. When the data type is INT64, xScaleOptional must be of data type FLOAT16, and xQuantMode must be pertensor.	FLOAT32, INT64	ND	2	√
groupList	Input	Group in the formula. Represents the prefix sum result of input tensor x grouped in the MoE scenario, with a 1-dimensional shape [g].	-	INT64	ND	1	√
biasOptional	Input	Represents the offset for computation. An optional parameter, the bias in the formula.	Currently, only a null pointer is supported.	-	-	-	-
xScaleOptional	Input	Represents the quantization coefficient of x, an optional parameter, scale_{x} in the formula.	When xQuantMode is pertensor, the shape is 1-dimensional (1,); when xQuantMode is pertoken, the shape is 1-dimensional (m,), where m is consistent with m of the input x. If empty, dynamic quantization is used. When the data type is FLOAT16, it must require that the weightScale data type be INT64 and the xQuantMode value be pertensor. Supports empty tensors.	FLOAT32, FLOAT16	ND	1	√
xOffsetOptional	Input	Represents the offset of x, an optional parameter.	Currently, only a null pointer is supported.	-	-	-	-
smoothScaleOptional	Input	Represents the smoothing coefficient of x. An optional parameter, which is scale_{smooth} in the formula.	The shape is 1D (k,), where k is consistent with k of x. An empty tensor is supported.	FLOAT16	ND	1	√
xQuantMode	Input	Specifies the quantization mode of input x.	Supports pertoken and pertensor. For dynamic quantization, only pertoken is supported. pertoken indicates that each token (a row) has its own quantization parameter; pertensor indicates that the entire tensor uses a unified quantization parameter.	STRING	-	-	-
transposeWeight	Input	Indicates whether to transpose the input weight.	Currently only supports true.	BOOL	-	-	-
out	Output	Computation result. Required parameter. out in the formula.	Shape supports 2 dimensions, where each dimension represents: (m, n). Here, m is consistent with m of x, and n is consistent with n of weight.	FLOAT16	ND	2	×
workspaceSize	Output	Returns the workspace size to be applied on the Device side.	-	-	-	-	-
executor	Output	Returns the operator executor, which contains the operator computation process.	-	-	-	-	-

In ND format, the shape of weight supports 3 dimensions.
- When transposeWeight is true, each dimension represents: (g, n, k).
- When transposeWeight is false, each dimension represents: (g, k, n).
In FRACTAL_NZ format, the shape of weight supports 5 dimensions.
- When transposeWeight is true, each dimension represents: (g, k1, n1, n0, k0), where k0 = 32, n0 = 16, and k1 and k of x must satisfy the following relationship: ceilDiv(k, 32) = k1.
- When transposeWeight is false, each dimension represents: (g, n1, k1, k0, n0), where k0 = 16, n0 = 32, and k1 and k of x must satisfy the following relationship: ceilDiv(k, 16) = k1.
- The aclnnCalculateMatmulWeightSizeV2 API and aclnnTransMatmulWeight API can be used to convert the input format from ND to the FRACTAL_NZ format.

Return Value

aclnnStatus: return code. For details, see aclnn Return Code.

The first-phase API performs input parameter validation and returns an error in the following scenarios:

Return Value	Error Code	Description
ACLNN_ERR_PARAM_NULLPTR	161001	The passed-in parameter is a required input, output, or required attribute and is a null pointer.
ACLNN_ERR_PARAM_INVALID	161002	If the input parameter type is aclTensor and its data type is not within the supported range. n or k in the shape of weight is not divisible by 16.
aclnnAdvanceStepGetWorkspaceSize failed	561002	If the passed-in parameter type is aclTensor and its shape does not match the above parameter description.

aclnnFatreluMul

Parameters

Parameter	Input/Output	Description
workspace	Input	Device-side workspace memory address.
workspaceSize	Input	Device-side workspace size, obtained from the first-phase API aclnnQuantGroupedMatmulDequantGetWorkspaceSize.
executor	Input	Operator executor that contains the operator computation process.
stream	Input	Specifies the stream that executes the task.

Return Value

aclnnStatus: return code. For details, see aclnn Return Code.

Constraints

Deterministic computation:
- aclnnQuantGroupedMatmulDequant defaults to a deterministic implementation.
Both n and k must be integer multiples of 16.
When the weightScale data type is INT64, the mandatory requirement is that the xScaleOptional data type is FLOAT16 and the xQuantMode value is pertensor; when the xScaleOptional data type is FLOAT16, the mandatory requirement is that the weightScale data type is INT64 and the xQuantMode value is pertensor.

Example

The following provides example code for reference only. For details about compilation and running, see Compile and Run Samples.

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_quant_grouped_matmul_dequant.h"

#define CHECK_RET(cond, return_expr) \
  do {                               \
    if (!(cond)) {                   \
      return_expr;                   \
    }                                \
  } while (0)

#define LOG_PRINT(message, ...)     \
  do {                              \
    printf(message, ##__VA_ARGS__); \
  } while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) {
  int64_t shapeSize = 1;
  for (auto i : shape) {
    shapeSize *= i;
  }
  return shapeSize;
}

void PrintOutResult(std::vector<int64_t> &shape, void** deviceAddr) {
  auto size = GetShapeSize(shape);
  std::vector<float> resultData(size, 0);
  auto ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]),
                         *deviceAddr, size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return);
  for (int64_t i = 0; i < size; i++) {
    LOG_PRINT("mean result[%ld] is: %f\n", i, resultData[i]);
  }
}

int Init(int32_t deviceId, aclrtStream* stream) {
  // Boilerplate: resource initialization.
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
}

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) {
  auto size = GetShapeSize(shape) * sizeof(T);
  // Call aclrtMalloc to apply for device-side memory.
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
  // Copy data from the host side to the device-side memory by calling aclrtMemcpy.
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // Calculate the strides of a contiguous tensor.
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) {
    strides[i] = shape[i + 1] * strides[i + 1];
  }

  // Create an aclTensor by calling the aclCreateTensor API.
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
}

int main() {
  // 1. (Boilerplate) Initialize the device/stream. See the ACL API manual.
  // Set deviceId based on the actual device.
  int32_t deviceId = 0;
  aclrtStream stream;
  auto ret = Init(deviceId, &stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. Construct the input and output. Customize the construction based on the API interface.
  int G = 4;
  int M = 64;
  int K = 256;
  int N = 512;

  char quantMode[16] = "pertoken";
  bool transposeWeight = true;

  std::vector<int64_t> xShape = {M,K};
  std::vector<int64_t> weightShape = {G,N,K};
  std::vector<int64_t> weightScaleShape = {G,N};
  std::vector<int64_t> xScaleShape = {M};
  std::vector<int64_t> smoothScaleShape = {K};
  std::vector<int64_t> outShape = {M,N};
  std::vector<int64_t> groupListShape = {G};

  void* xDeviceAddr = nullptr;
  void* weightDeviceAddr = nullptr;
  void* weightScaleDeviceAddr = nullptr;
  void* groupListDeviceAddr = nullptr;
  void* xScaleDeviceAddr = nullptr;
  void* smoothScaleDeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;

  aclTensor* x = nullptr;
  aclTensor* weight = nullptr;
  aclTensor* weightScale = nullptr;
  aclTensor* groupList = nullptr;
  aclTensor* bias = nullptr;
  aclTensor* xScale = nullptr;
  aclTensor* xOffset = nullptr;
  aclTensor* smoothScale = nullptr;
  aclTensor* out = nullptr;

  std::vector<uint16_t> xHostData(GetShapeSize(xShape));
  std::vector<uint8_t> weightHostData(GetShapeSize(weightShape));
  std::vector<float> weightScaleHostData(GetShapeSize(weightScaleShape));
  std::vector<int64_t> groupListHostData(GetShapeSize(groupListShape));
  groupListHostData[0] = 7;
  groupListHostData[0] = 32;
  groupListHostData[0] = 40;
  groupListHostData[0] = 64;
  std::vector<float> xScaleHostData(GetShapeSize(xScaleShape));
  std::vector<uint16_t> smoothScaleHostData(GetShapeSize(smoothScaleShape));
  std::vector<uint16_t> outHostData(GetShapeSize(outShape));

  ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_FLOAT16, &x);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(weightHostData, weightShape, &weightDeviceAddr, aclDataType::ACL_INT8, &weight);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(weightScaleHostData, weightScaleShape, &weightScaleDeviceAddr, aclDataType::ACL_FLOAT, &weightScale);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(groupListHostData, groupListShape, &groupListDeviceAddr, aclDataType::ACL_INT64, &groupList);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(xScaleHostData, xScaleShape, &xScaleDeviceAddr, aclDataType::ACL_FLOAT, &xScale);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(smoothScaleHostData, smoothScaleShape, &smoothScaleDeviceAddr, aclDataType::ACL_FLOAT16, &smoothScale);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT16, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3. Call the CANN operator library API. Replace with the specific API name.
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;

  // Call the first-phase API of aclnnQuantGroupedMatmulDequant.
  ret = aclnnQuantGroupedMatmulDequantGetWorkspaceSize(x, weight, weightScale, groupList,
                                                bias, xScale, xOffset, smoothScale,
                                                quantMode, transposeWeight, out,
                                                &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQuantGroupedMatmulDequantGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);

  // Allocate device memory based on the workspaceSize computed by the first-phase API.
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) {
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
  }

  // Call the second-phase API of aclnnQuantGroupedMatmulDequant.
  ret = aclnnQuantGroupedMatmulDequant(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQuantGroupedMatmulDequant failed. ERROR: %d\n", ret); return ret);

  // 4. (Boilerplate) Synchronously wait for task execution to complete.
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. Obtain the output value and copy the result from the device memory to the host memory. Modify based on the specific API definition.
  PrintOutResult(outShape, &outDeviceAddr);

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(x);
  aclDestroyTensor(weight);
  aclDestroyTensor(weightScale);
  aclDestroyTensor(groupList);
  aclDestroyTensor(xScale);
  aclDestroyTensor(smoothScale);
  aclDestroyTensor(out);

  // 7. Release device resources.
  aclrtFree(xDeviceAddr);
  aclrtFree(weightDeviceAddr);
  aclrtFree(weightScaleDeviceAddr);
  aclrtFree(groupListDeviceAddr);
  aclrtFree(xScaleDeviceAddr);
  aclrtFree(smoothScaleDeviceAddr);
  aclrtFree(outDeviceAddr);

  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();

  return 0;
}