aclnnFusedQuantMatmul

Applicable Products

Function

• Description: Fused quantized matrix multiplication and Gelu computation.

• Formula:

  • x1Scale, bias INT32 (no offset in this scenario):

    $$qbmmout = (x1@x2 + bias) * x2Scale * x1Scale$$

  • x1Scale, bias BFLOAT16/FLOAT16/FLOAT32 (this scenario has no offset):

    $$qbmmout = x1@x2 * x2scale * x1Scale + bias$$

  • x1Scale without bias:

    $$qbmmout = x1@x2 * x2Scale * x1Scale$$

  • The OP type is defined by the fusedOpType input and supports the following:

    • gelu_tanh operation:

      $$out = gelu\_tanh(qbmmout)$$

    • gelu_erf operation:

      $$out = gelu\_erf(qbmmout)$$

Function Prototype

Each operator is divided into a two-phase API. The "aclnnFusedQuantMatmulGetWorkspaceSize" API must be called first to obtain the workspace size required for computation and the executor that encapsulates the operator computation flow. Then, the "aclnnFusedQuantMatmul" API is called to perform the computation.

aclnnStatus aclnnFusedQuantMatmulGetWorkspaceSize(
  const aclTensor *x1,
  const aclTensor *x2,
  const aclTensor *x1Scale,
  const aclTensor *x2Scale,
  const aclTensor *yScaleOptional,
  const aclTensor *x1OffsetOptional,
  const aclTensor *x2OffsetOptional,
  const aclTensor *yOffsetOptional,
  const aclTensor *biasOptional,
  const aclTensor *x3Optional,
  const char      *fusedOpType,
  int64_t          groupSizeOptional,
  aclTensor       *out,
  uint64_t        *workspaceSize,
  aclOpExecutor   **executor)

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

aclnnFusedQuantMatmulGetWorkspaceSize

• Parameters:

  Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-Continuous Tensor
  x1	Input	Input x1 in the formula.	Only when the last m and k axes are transposed, non-contiguous tensor is supported. Non-contiguous tensors are not supported in other axis directions. The size of the last dimension cannot exceed 65535.	INT4, INT8, INT32	ND	2-6	√
  x2	Input	Input x2 in the formula.	In ND format, non-contiguous tensors with the last two axes transposed are supported. Non-contiguous tensors in other scenarios are not supported. The size of the last dimension cannot exceed 65535.	INT4, INT8, INT32	ND	2-6	√
  x1Scale	Input	Represents the quantization parameter, x1Scale in the formula.	-	FLOAT32	ND	1	-
  x2Scale	Input	Represents the quantization parameter, which is the input x2Scale in the formula.	-	FLOAT32, BFLOAT16	ND	1	-
  yScaleOptional	Input	Quantization scale parameter for output y. Used in static quantization.	Reserved parameter. Not supported in the current version. Pass nullptr.	-	-	-	-
  x1OffsetOptional	Input	Input x1Offset in the formula.	Reserved parameter. Not supported in the current version. Pass nullptr.	-	-	-	-
  x2OffsetOptional	Input	In Formula, input x2Offset.	Reserved Parameter. Not supported in the current version. Required input: nullptr.	-	-	-	-
  yOffsetOptional	Input	Input yOffset in the formula.	Reserved Parameter. Not supported in the Current Version. nullptr is the Required Input.	-	-	-	-
  biasOptional	Input	Bias in the formula.	Pass nullptr when there is no bias.	INT32, FLOAT32, BFLOAT16, FLOAT16	ND	1, 3	-
  x3Optional	Input	Input of the fused binary operation.	Currently, only fused unary operations are supported. This is a reserved parameter and is not supported in the current version. nullptr is required.	-	-	-	-
  fusedOpType	Input	In the formula, the input fusedOpType specifies the fusion mode supported by the QuantBatchMatmul operator.	The fusion mode value must be one of "gelu_erf" or "gelu_tanh".	STRING	-	-	-
  groupSizeOptional	Input	Specifies the quantization group size in the m, n, and k directions.	Reserved Parameter. Not supported in the current version. Required input nullptr.	-	-	-	-
  out	Output	Output out in the formula.	-	FLOAT16, BFLOAT16	ND	2-6	✓
  workspaceSize	Output	Returns the workspace size to be applied for on the Device side.	-	-	-	-	-
  executor	Output	Returns the op executor, which contains the operator computation flow.	-	-	-	-	-

• 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 input x1, x2, x1Scale, x2Scale, or out is a null pointer.
  ACLNN_ERR_PARAM_INVALID	161002	The shape of x1, x2, biasOptional, x1Scale, x2Scale, or out does not meet the validation conditions.
  		The data type or data format of x1, x2, biasOptional, x1Scale, x2Scale, or out is not within the supported range.
  		x1, x2, biasOptional, x1Scale, x2Scale, or out is an empty tensor.
  		The input fusedOpType is neither "gelu_tanh" nor "gelu_erf".
  		The input parameters yScaleOptional, x1OffsetOptional, x2OffsetOptional, yOffsetOptional, x3Optional, and groupSizeOptional are not supported in the current version.

aclnnFusedQuantMatmul

• Parameters:

  Parameter	Input/Output	Description
  workspace	Input	Memory address of the workspace applied on the Device side.
  workspaceSize	Input	Size of the workspace applied on the Device side, obtained from the first-phase API aclnnFusedQuantMatmulGetWorkspaceSize.
  executor	Input	Operator executor that contains the operator computation flow.
  stream	Input	Specifies the Stream on which the task is executed.

• Return value

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

Constraints

• Deterministic description:

  • Atlas A3 training products/Atlas A3 inference products, Atlas A2 training products/Atlas A2 inference products: aclnnFusedQuantMatmul defaults to a deterministic implementation.

• The input and output support the following data type combinations:

  x1	x2	x1Scale	x2Scale	x2OffsetOptional	yScaleOptional	biasOptional	yOffsetOptional	out
  INT8	INT8	FLOAT32	FLOAT32/BFLOAT16	null	null	null/INT32/BFLOAT16/FLOAT32	null	BFLOAT16
  INT8	INT8	FLOAT32	FLOAT32	null	null	null/INT32/FLOAT16/FLOAT32	null	FLOAT16
  INT4/INT32	INT4/INT32	FLOAT32	FLOAT32/BFLOAT16	null	null	null/INT32/BFLOAT16/FLOAT32	null	BFLOAT16
  INT4/INT32	INT4/INT32	FLOAT32	FLOAT32	null	null	null/INT32/FLOAT16/FLOAT32	null	FLOAT16

• This API supports x1 pertoken quantization and x2 perchannel/pertensor quantization. The constraints on the input dtype combinations of x1, x2, x1Scale, and x2Scale supported by different quantization modes are as follows:

  • The x1 data type supports INT8, INT32, and INT4.

    • When the data type is INT32 or INT4, it is an INT4 quantization scenario:

      • Currently, only ND input is supported.

      • Currently, only non-transposed input is supported.

      • The inner axis of x1 must be an even number.

    • When the data type is INT32, each INT32 data stores eight INT4 data values, with the corresponding dimension representation: (batch, m, k // 8). k must be a multiple of 8.

  • x2 supports INT8, INT32, and INT4 data types.

    • When the data type is INT32 or INT4, it is an INT4 quantization scenario:

      • This API supports only the 2D ND format.

      • The inner axis of x2 must be an even number.

    • When the data type is INT32, each INT32 value stores eight INT4 values:

    • You can use the aclnnConvertWeightToINT4Pack API to convert the data format of x2 from INT32 (one int32 stores one int4 in bits 0-3) to INT32 (one int32 stores eight int4 values) or INT4 (one int4 represents one int4). See aclnnConvertWeightToINT4Pack.

  • x1Scale constraints are as follows:

    • The shape supports 1D with a shape of (m,). The data type supports FLOAT32.

  • x2Scale constraints are as follows:

    • The shape supports 1D with a shape of (n,) or (1,), where n is consistent with n of x2. The data type supports FLOAT32 and BFLOAT16.

  • Constraints for biasOptional are as follows:

    • The shape supports 1D and 3D. In the INT4 quantization scenario, only 1D biasOptional is supported, with the shape (n). For 3D, the biasOptional shape is (batch, 1, n).

    • The data type supports int32, float32, bfloat16, or float16.

  • Constraints for out are as follows:

    • The shape supports 2D to 6D, (batch, m, n). The data type supports FLOAT16 and BFLOAT16.

Example

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

• Atlas A3 training products/Atlas A3 inference products, Atlas A2 training products/Atlas A2 inference products:

x1 is INT8, x2 is INT8, x1Scale is FLOAT32, and x2Scale is FLOAT32.

#include <iostream>
#include <memory>
#include <vector>

#include "acl/acl.h"
#include "aclnnop/aclnn_fused_quant_matmul.h"

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

#define CHECK_FREE_RET(cond, return_expr) \
    do {                                  \
        if (!(cond)) {                    \
            Finalize(deviceId, stream);   \
            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;
}

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);
    // Call aclrtMemcpy to copy host-side data to device-side memory.
    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];
    }

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

void Finalize(int32_t deviceId, aclrtStream stream)
{
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
}

int aclnnFusedQuantMatmulTest(int32_t deviceId, 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 inputs and outputs. Custom construction based on the API is required.
    std::vector<int64_t> x1Shape = {2, 8192};   // (m,k)
    std::vector<int64_t> x2Shape = {8192, 128}; // (k,n)
    std::vector<int64_t> x1ScaleShape = {2}; // x1ScaleShape = [M]
    std::vector<int64_t> x2ScaleShape = {128}; // x2ScaleShape = [N]
    std::vector<int64_t> outShape = {2, 128};

    void* x1DeviceAddr = nullptr;
    void* x2DeviceAddr = nullptr;
    void* x2ScaleDeviceAddr = nullptr;
    void* x1ScaleDeviceAddr = nullptr;
    void* outDeviceAddr = nullptr;
    aclTensor* x1 = nullptr;
    aclTensor* x2 = nullptr;
    aclTensor* x2Scale = nullptr;
    aclTensor* x1Scale = nullptr;
    aclTensor* out = nullptr;
    std::vector<int8_t> x1HostData(GetShapeSize(x1Shape), 1);
    std::vector<int8_t> x2HostData(GetShapeSize(x2Shape), 1);
    std::vector<float> x1ScaleHostData(GetShapeSize(x1ScaleShape), 1);
    std::vector<float> x2ScaleHostData(GetShapeSize(x2ScaleShape), 1);
    std::vector<uint16_t> outHostData(GetShapeSize(outShape), 1);

    // Create the x1 aclTensor.
    ret = CreateAclTensor(x1HostData, x1Shape, &x1DeviceAddr, aclDataType::ACL_INT8, &x1);
    std::unique_ptr<aclTensor, aclnnStatus (*)(const aclTensor*)> x1TensorPtr(x1, aclDestroyTensor);
    std::unique_ptr<void, aclError (*)(void*)> x1DeviceAddrPtr(x1DeviceAddr, aclrtFree);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the x2 aclTensor.
    ret = CreateAclTensor(x2HostData, x2Shape, &x2DeviceAddr, aclDataType::ACL_INT8, &x2);
    std::unique_ptr<aclTensor, aclnnStatus (*)(const aclTensor*)> x2TensorPtr(x2, aclDestroyTensor);
    std::unique_ptr<void, aclError (*)(void*)> x2DeviceAddrPtr(x2DeviceAddr, aclrtFree);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the x1Scale aclTensor.
    ret = CreateAclTensor(x1ScaleHostData, x1ScaleShape, &x1ScaleDeviceAddr, aclDataType::ACL_FLOAT, &x1Scale);
    std::unique_ptr<aclTensor, aclnnStatus (*)(const aclTensor*)> x1ScaleTensorPtr(x1Scale, aclDestroyTensor);
    std::unique_ptr<void, aclError (*)(void*)> x1ScaleDeviceAddrPtr(x1ScaleDeviceAddr, aclrtFree);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the x2Scale aclTensor.
    ret = CreateAclTensor(x2ScaleHostData, x2ScaleShape, &x2ScaleDeviceAddr, aclDataType::ACL_FLOAT, &x2Scale);
    std::unique_ptr<aclTensor, aclnnStatus (*)(const aclTensor*)> scaleTensorPtr(x2Scale, aclDestroyTensor);
    std::unique_ptr<void, aclError (*)(void*)> x2ScaleDeviceAddrPtr(x2ScaleDeviceAddr, aclrtFree);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the out aclTensor.
    ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT16, &out);
    std::unique_ptr<aclTensor, aclnnStatus (*)(const aclTensor*)> outTensorPtr(out, aclDestroyTensor);
    std::unique_ptr<void, aclError (*)(void*)> outDeviceAddrPtr(outDeviceAddr, aclrtFree);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    int64_t groupSize = 0;
    const char fusedOpType[] = "gelu_tanh";

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

    ret = aclnnFusedQuantMatmulGetWorkspaceSize(
        x1, x2, x1Scale, x2Scale, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, fusedOpType, groupSize, out,
        &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFusedQuantMatmulGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
    // Apply for device memory based on the workspaceSize calculated by the first-phase API.
    void* workspaceAddr = nullptr;
    std::unique_ptr<void, aclError (*)(void*)> workspaceAddrPtr(nullptr, aclrtFree);
    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);
        workspaceAddrPtr.reset(workspaceAddr);
    }
    // Call the second-phase API of aclnnFusedQuantMatmul.
    ret = aclnnFusedQuantMatmul(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFusedQuantMatmul 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. Copy the result from the device-side memory to the host side. Modify this based on the specific API definition.
    auto size = GetShapeSize(outShape);
    std::vector<uint16_t> resultData(
        size, 0); // In C, fp16 data cannot be printed directly. Read it as uint16 and manually convert it to fp16 via binary conversion.
    ret = aclrtMemcpy(
        resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr, 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 ret);
    for (int64_t i = 0; i < size; i++) {
        LOG_PRINT("result[%ld] is: %u\n", i, resultData[i]);
    }
    return ACL_SUCCESS;
}

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