aclnnRopeWithSinCosCache

📄 View Source Code

Applicable Products

Function

• Description: To improve performance, the inference network passes the sin and cos inputs through a cache and performs rotary position embedding computation.

• Formula:

  1. mrope mode: The shape of positions is [m, numTokens], where m is the number of elements in mropeSection, supporting 3 or 4:
  $$cosSin[i] = cosSinCache[positions[i]]$$ $$cos, sin = cosSin.chunk(2, dim=-1)$$

  • When the number of elements in mropeSection is 3:

    $$cos0 = cos[0, :, :mropeSection[0]]$$ $$cos1 = cos[1, :, mropeSection[0]:(mropeSection[0] + mropeSection[1])]$$ $$cos2 = cos[2, :, (mropeSection[0] + mropeSection[1]):(mropeSection[0] + mropeSection[1] + mropeSection[2])]$$ $$cos = torch.cat((cos0, cos1, cos2), dim=-1)$$ $$sin0 = sin[0, :, :mropeSection[0]]$$ $$sin1 = sin[1, :, mropeSection[0]:(mropeSection[0] + mropeSection[1])]$$ $$sin2 = sin[2, :, (mropeSection[0] + mropeSection[1]):(mropeSection[0] + mropeSection[1] + mropeSection[2])]$$ $$sin= torch.cat((sin0, sin1, sin2), dim=-1)$$ $$queryRot = query[..., :rotaryDim]$$ $$queryPass = query[..., rotaryDim:]$$

  • When mropeSection has 4 elements:

    $$cos = torch.cat([m[i]\ for\ i, m\ in\ enumerate(cos.split(mropeSection, dim=-1))], dim=-1)$$ $$sin = torch.cat([m[i]\ for\ i, m\ in\ enumerate(sin.split(mropeSection, dim=-1))], dim=-1)$$ $$queryRot = query[..., :rotaryDim]$$ $$queryPass = query[..., rotaryDim:]$$

  (1) rotate_half (GPT-NeoX style) calculation mode:

  $$x1, x2 = torch.chunk(queryRot, 2, dim=-1)$$ $$o1[i] = x1[i] * cos[i] - x2[i] * sin[i]$$ $$o2[i] = x2[i] * cos[i] + x1[i] * sin[i]$$ $$queryRot = torch.cat((o1, o2), dim=-1)$$ $$query = torch.cat((queryRot, queryPass), dim=-1)$$

  (2) rotate_interleaved (GPT-J style) calculation mode:

  $$x1 = queryRot[..., ::2]$$ $$x2 = queryRot[..., 1::2]$$ $$o1[i] = x1[i] * cos[i] - x2[i] * sin[i]$$ $$o2[i] = x2[i] * cos[i] + x1[i] * sin[i]$$ $$queryRot = torch.stack((o1, o2), dim=-1)$$ $$query = torch.cat((queryRot, queryPass), dim=-1)$$
  2. rope mode: The shape of positions input is [numTokens]:
  $$cosSin[i] = cosSinCache[positions[i]]$$ $$cos, sin = cosSin.chunk(2, dim=-1)$$ $$queryRot = query[..., :rotaryDim]$$ $$queryPass = query[..., rotaryDim:]$$

  (1) rotate_half (GPT-NeoX style) calculation mode:

  $$x1, x2 = torch.chunk(queryRot, 2, dim=-1)$$ $$o1[i] = x1[i] * cos[i] - x2[i] * sin[i]$$ $$o2[i] = x2[i] * cos[i] + x1[i] * sin[i]$$ $$queryRot = torch.cat((o1, o2), dim=-1)$$ $$query = torch.cat((queryRot, queryPass), dim=-1)$$

  (2) rotate_interleaved (GPT-J style) calculation mode:

  $$x1 = queryRot[..., ::2]$$ $$x2 = queryRot[..., 1::2]$$ $$o1[i] = x1[i] * cos[i] - x2[i] * sin[i]$$ $$o2[i] = x2[i] * cos[i] + x1[i] * sin[i]$$ $$queryRot = torch.stack((o1, o2), dim=-1)$$ $$query = torch.cat((queryRot, queryPass), dim=-1)$$

Function Prototype

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

aclnnStatus aclnnRopeWithSinCosCacheGetWorkspaceSize(
    const aclTensor   *positions,
    const aclTensor   *queryIn,
    const aclTensor   *keyIn,
    const aclTensor   *cosSinCache,
    const aclIntArray *mropeSection,
    int64_t            headSize,
    bool               isNeoxStyle,
    aclTensor         *queryOut,
    aclTensor         *keyOut,
    uint64_t          *workspaceSize,
    aclOpExecutor     **executor)

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

aclnnRopeWithSinCosCacheGetWorkspaceSize

• Parameters:

  Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
  positions	Input	positions in the formula, used to select the positional encoding tensor.	Empty tensors are not supported. In rope mode, the shape is (numTokens). In mrope mode, the shape is (3, numTokens) or (4, numTokens).	INT64	ND	1-2	√
  queryIn	Input	The query in the formula, which is the first tensor to perform rotary position embedding.	Empty tensors are not supported. Must be a 2D tensor with shape (numTokens, numQHeads*headSize).	BFLOAT16, FLOAT16, FLOAT32	ND	2	√
  keyIn	Input	The second tensor on which rotary position embedding is to be performed.	Empty tensors are not supported. It must be a 2D tensor with the shape (numTokens, numKHeads*headSize).	BFLOAT16, FLOAT16, FLOAT32	ND	2	√
  cosSinCache	Input	Represents the positional encoding tensor involved in the computation.	Empty tensors are not supported. Must be a 2D tensor with shape (maxSeqLen, rotaryDim), where maxSeqLen represents the maximum sequence length processed by the model, and rotaryDim represents the dimension size of the rotary position embedding.	BFLOAT16, FLOAT16, FLOAT32	ND	2	√
  mropeSection	Input	mropeSection in the formula, used to integrate the input positional encoding tensor information in mrope mode.	The input mropeSection attribute indicates that mrope mode is enabled. When mrope mode is not enabled (i.e., rope mode), the input is nullptr.	aclIntArray	-	-	-
  headSize	Input	Size of each attention head dimension.	-	INT64	-	-	-
  isNeoxStyle	Input	Indicates whether to use the GPT-NeoX calculation mode.	true indicates the GPT-NeoX style calculation mode. false indicates the GPT-J style calculation mode.	BOOL	-	-	-
  queryOut	Output	Result of query after performing rotary position embedding.	The data type is the same as query. Must be a 2D tensor with the shape (numTokens, numQHeads*headSize).	FLOAT32, FLOAT16, BFLOAT16	ND	2	×
  keyOut	Output	Result of performing rotary position embedding on key.	The data type is the same as key. Must be a 2D tensor with the shape (numTokens, numKHeads*headSize).	FLOAT32, FLOAT16, BFLOAT16	ND	2	×
  workspaceSize	Output	Returns the workspace size that the user needs to apply 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	Return Code	Description
  ACLNN_ERR_PARAM_NULLPTR	161001	A required input, required output, or required attribute passed in is a null pointer.
  ACLNN_ERR_PARAM_INVALID	161002	The data type and data format of the input and output are not within the supported range.
  ACLNN_ERR_INNER_TILING_ERROR	561002	The shape information of multiple input tensors does not match.
  		The shape information between the input attribute and the input tensor does not match.
  ACLNN_ERR_INNER_TILING_ERROR	361001	The query or key is not 64-byte aligned.
  		rotaryDim > headSize.

aclnnRopeWithSinCosCache

• Parameters:

  Parameter	Input/Output	Description
  workspace	Input	Workspace memory address applied for on the Device side.
  workspaceSize	Input	Workspace size applied for on the Device side, obtained from the first-phase API aclnnRopeWithSinCosCacheGetWorkspaceSize.
  executor	Input	Operator executor, which contains the operator computation flow.
  stream	Input	Specifies the stream for task execution.

• Return value

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

Constraints

• Deterministic computation:

  • aclnnRopeWithSinCosCache defaults to a deterministic implementation.

• queryIn, keyIn, and cosSinCache only support 2D shape input.

• The data types of queryIn, keyIn, and cosSinCache inputs must be consistent.

• headSize: a multiple of 32 when the data type is BFLOAT16 or FLOAT16, and a multiple of 16 when the data type is FLOAT32.

• rotaryDim: always less than or equal to headSize; a multiple of 32 when the data type is BFLOAT16 or FLOAT16, and a multiple of 16 when the data type is FLOAT32; in mrope mode, the sum of all elements in mropeSection must equal half of the rotaryDim value.

• The value of the input tensor positions must be less than maxSeqLen of dimension 0 of cosSinCache.

• In mrope mode, mropeSection currently supports only [16, 24, 24], [24, 20, 20], [8, 12, 12], and [16, 16, 16, 16].

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/level2/aclnn_rope_with_sin_cos_cache.h"
#include <iostream>

#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 for 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 aclCreateTensor 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;
}

int main() {
  // 1. (Boilerplate) Initialize the device/stream. Refer to the acl API manual.
  // Fill in the deviceId based on your 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.
  std::vector<int64_t> positionsShape = {2};
  std::vector<int64_t> queryInShape = {2, 64};
  std::vector<int64_t> keyInShape = {2, 64};
  std::vector<int64_t> cosSinCacheShape = {2, 32};
  std::vector<int64_t> queryOutShape = {2, 64};
  std::vector<int64_t> keyOutShape = {2, 64};
  void* positionsDeviceAddr = nullptr;
  void* queryInDeviceAddr = nullptr;
  void* keyInDeviceAddr = nullptr;
  void* cosSinCacheDeviceAddr = nullptr;
  void* queryOutDeviceAddr = nullptr;
  void* keyOutDeviceAddr = nullptr;

  aclTensor* positions = nullptr;
  aclTensor* queryIn = nullptr;
  aclTensor* keyIn = nullptr;
  aclTensor* cosSinCache = nullptr;
  int64_t headSize = 32;
  bool isNeoxStyle = true;
  aclTensor *queryOut = nullptr;
  aclTensor *keyOut = nullptr;

  std::vector<int64_t> positionsHostData = {0, 1};
  std::vector<float> queryInHostData = {74, 54, 84, 125, 23, 78, 37, 72, 27, 98, 34, 107, 29, 23, 54, 60, 70, 49,
                                        119, 54, 29, 54, 41, 99, 27, 62, 5, 46, 108, 39, 24, 123, 33, 82, 6, 40, 88,
                                        24, 6, 116, 38, 119, 110, 5, 30, 79, 87, 18, 29, 100, 90, 24, 21, 93, 63, 68,
                                        34, 112, 119, 48, 74, 43, 85, 64, 14, 49, 128, 59, 18, 37, 123, 76, 14, 63, 10,
                                        39, 107, 124, 79, 16, 17, 76, 80, 47, 90, 41, 58, 82, 75, 80, 69, 37, 74, 36, 54,
                                        26, 32, 54, 13, 100, 105, 15, 13, 69, 122, 26, 94, 59, 29, 14, 60, 8, 24, 17, 45,
                                        33, 107, 122, 63, 111, 75, 128, 68, 31, 105, 6, 82, 99};
  std::vector<float> keyInHostData = {112, 32, 66, 114, 69, 31, 117, 122, 77, 57, 78, 119, 115, 25, 54, 27, 122, 65, 15, 85,
                                      33, 16, 36, 6, 95, 15, 43, 6, 66, 91, 14, 101, 78, 51, 110, 74, 56, 30, 127, 61, 53, 29,
                                      32, 65, 114, 77, 26, 116, 89, 38, 75, 14, 96, 91, 87, 34, 25, 42, 57, 26, 51, 43, 23, 42,
                                      40, 17, 98, 117, 53, 75, 68, 75, 38, 41, 115, 76, 67, 22, 76, 10, 24, 46, 85, 54, 61, 114,
                                      10, 59, 6, 123, 58, 10, 115, 9, 13, 58, 66, 120, 23, 30, 83, 13, 11, 76, 18, 82, 57, 4,
                                      117, 105, 8, 73, 127, 5, 91, 56, 12, 125, 20, 3, 104, 40, 46, 18, 89, 63, 99, 104};
  std::vector<float> cosSinCacheHostData = {112, 32, 66, 114, 69, 31, 117, 122, 77, 57, 78, 119, 115, 25, 54, 27, 122, 65, 15, 85,
                                      33, 16, 36, 6, 95, 15, 43, 6, 66, 91, 14, 101, 78, 51, 110, 74, 56, 30, 127, 61, 53, 29,
                                      32, 65, 114, 77, 26, 116, 89, 38, 75, 14, 96, 91, 87, 34, 25, 42, 57, 26, 51, 43, 23, 42};
  std::vector<float> queryOutHostData = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
  std::vector<float> keyOutHostData = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};

  ret = CreateAclTensor(positionsHostData, positionsShape,
                        &positionsDeviceAddr, aclDataType::ACL_INT64,
                        &positions);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(queryInHostData, queryInShape, &queryInDeviceAddr,
                      aclDataType::ACL_FLOAT, &queryIn);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(keyInHostData, keyInShape, &keyInDeviceAddr,
                      aclDataType::ACL_FLOAT, &keyIn);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(cosSinCacheHostData, cosSinCacheShape, &cosSinCacheDeviceAddr,
                      aclDataType::ACL_FLOAT, &cosSinCache);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(queryOutHostData, queryOutShape, &queryOutDeviceAddr, aclDataType::ACL_FLOAT,
                        &queryOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(keyOutHostData, keyOutShape, &keyOutDeviceAddr, aclDataType::ACL_FLOAT,
                        &keyOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

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

  // Call the first-phase API of aclnnRopeWithSinCosCache.
  ret = aclnnRopeWithSinCosCacheGetWorkspaceSize(positions, queryIn, keyIn, cosSinCache, nullptr, headSize, isNeoxStyle,
                                               queryOut, keyOut, &workspaceSize, &executor);
  CHECK_RET(
      ret == ACL_SUCCESS,
      LOG_PRINT("aclnnRopeWithSinCosCacheGetWorkspaceSize failed. ERROR: %d\n", ret);
      return ret);

  // Apply for device memory based on the workspaceSize calculated 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 aclnnRopeWithSinCosCache.
  ret = aclnnRopeWithSinCosCache(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS,
            LOG_PRINT("aclnnRopeWithSinCosCache 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 side memory to the host side. Modify based on the specific API definition.
  PrintOutResult(queryOutShape, &queryOutDeviceAddr);
  PrintOutResult(keyOutShape, &keyOutDeviceAddr);

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(positions);
  aclDestroyTensor(queryIn);
  aclDestroyTensor(keyIn);
  aclDestroyTensor(cosSinCache);
  aclDestroyTensor(queryOut);
  aclDestroyTensor(keyOut);

  // 7. Release device resources.
  aclrtFree(positionsDeviceAddr);
  aclrtFree(queryInDeviceAddr);
  aclrtFree(keyInDeviceAddr);
  aclrtFree(cosSinCacheDeviceAddr);
  aclrtFree(queryOutDeviceAddr);
  aclrtFree(keyOutDeviceAddr);

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

  return 0;
}