Skip to content

issue/1278 - fix neox rope on ascend #1279

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wooway777 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

issue/1278 - fix neox rope on ascend #1279

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
94 changes: 55 additions & 39 deletions src/infiniop/ops/rope/ascend/rope_ascend_kernel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -362,6 +362,7 @@ class RoPEKernelNeox {
size_t _tile_len;
size_t _copy_len;
size_t _half_len;
size_t _half_copy_len;
size_t _batch;
size_t _nhead;

Expand Down Expand Up @@ -399,6 +400,7 @@ __aicore__ inline void RoPEKernelNeox<T, U>::init(GM_ADDR y,
this->_st_xnh = st_xnh;
this->_st_xbatch = st_xbatch;
_copy_len = alignTileLen<T>(dh, BYTE_ALIGN);
_half_copy_len = alignTileLen<T>(_half_len, BYTE_ALIGN);

_block_idx = GetBlockIdx();

Expand All @@ -410,27 +412,28 @@ __aicore__ inline void RoPEKernelNeox<T, U>::init(GM_ADDR y,

pipe.InitBuffer(_in_que, BUFFER_NUM, _copy_len * sizeof(T));
pipe.InitBuffer(_out_que, BUFFER_NUM, _copy_len * sizeof(T));
pipe.InitBuffer(_sin_que, BUFFER_NUM, _half_len * sizeof(T));
pipe.InitBuffer(_cos_que, BUFFER_NUM, _half_len * sizeof(T));
pipe.InitBuffer(_sin_que, BUFFER_NUM, _half_copy_len * sizeof(T));
pipe.InitBuffer(_cos_que, BUFFER_NUM, _half_copy_len * sizeof(T));

if constexpr (std::is_same<T, bfloat16_t>::value) {
size_t half_float_copy_len = alignTileLen<float>(_half_len, BYTE_ALIGN);
pipe.InitBuffer(_tmp_float_input, _copy_len * sizeof(float));
pipe.InitBuffer(_tmp_float_sin, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_float_cos, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_float_sin, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_float_cos, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_float_output, _tile_len * sizeof(float));
pipe.InitBuffer(_tmp_first_half, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_second_half, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_result1, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_result2, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_result3, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_result4, _half_len * sizeof(float));
pipe.InitBuffer(_tmp_first_half, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_second_half, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_result1, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_result2, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_result3, half_float_copy_len * sizeof(float));
pipe.InitBuffer(_tmp_result4, half_float_copy_len * sizeof(float));
} else {
pipe.InitBuffer(_tmp_first_half, _half_len * sizeof(T));
pipe.InitBuffer(_tmp_second_half, _half_len * sizeof(T));
pipe.InitBuffer(_tmp_result1, _half_len * sizeof(T));
pipe.InitBuffer(_tmp_result2, _half_len * sizeof(T));
pipe.InitBuffer(_tmp_result3, _half_len * sizeof(T));
pipe.InitBuffer(_tmp_result4, _half_len * sizeof(T));
pipe.InitBuffer(_tmp_first_half, _half_copy_len * sizeof(T));
pipe.InitBuffer(_tmp_second_half, _half_copy_len * sizeof(T));
pipe.InitBuffer(_tmp_result1, _half_copy_len * sizeof(T));
pipe.InitBuffer(_tmp_result2, _half_copy_len * sizeof(T));
pipe.InitBuffer(_tmp_result3, _half_copy_len * sizeof(T));
pipe.InitBuffer(_tmp_result4, _half_copy_len * sizeof(T));
}
}

Expand All @@ -446,8 +449,10 @@ __aicore__ inline void RoPEKernelNeox<T, U>::copyIn(size_t i) {
auto idx = batch_idx * _st_xbatch + i * _st_xnt + head_idx * _st_xnh;
DataCopy(input_ub, _x_gm[idx], _copy_len);
auto pos_idx = _p_gm(i);
DataCopy(sin_ub, _sin_gm[pos_idx * _half_len], _half_len);
DataCopy(cos_ub, _cos_gm[pos_idx * _half_len], _half_len);
DataCopyExtParams halfCopyParams = {1, static_cast<uint32_t>(_half_len * sizeof(T)), 0, 0, 0};
DataCopyPadExtParams<T> halfPadParams{true, 0, 0, 0};
DataCopyPad(sin_ub, _sin_gm[pos_idx * _half_len], halfCopyParams, halfPadParams);
DataCopyPad(cos_ub, _cos_gm[pos_idx * _half_len], halfCopyParams, halfPadParams);
_in_que.EnQue(input_ub);
_sin_que.EnQue(sin_ub);
_cos_que.EnQue(cos_ub);
Expand All @@ -470,26 +475,32 @@ __aicore__ inline void RoPEKernelNeox<T, U>::compute(size_t i) {
LocalTensor<float> result2_f = _tmp_result2.Get<float>();
LocalTensor<float> result3_f = _tmp_result3.Get<float>();
LocalTensor<float> result4_f = _tmp_result4.Get<float>();
size_t half_float_copy_len = alignTileLen<float>(_half_len, BYTE_ALIGN);

Cast(input_f, input_ub, AscendC::RoundMode::CAST_NONE, _copy_len);
Cast(sin_f, sin_ub, AscendC::RoundMode::CAST_NONE, _half_len);
Cast(cos_f, cos_ub, AscendC::RoundMode::CAST_NONE, _half_len);

for (size_t j = 0; j < _half_len; j++) {
first_half_f(j) = input_f(j);
second_half_f(j) = input_f(_half_len + j);
Cast(sin_f, sin_ub, AscendC::RoundMode::CAST_NONE, half_float_copy_len);
Cast(cos_f, cos_ub, AscendC::RoundMode::CAST_NONE, half_float_copy_len);

for (size_t j = 0; j < half_float_copy_len; j++) {
if (j < _half_len) {
first_half_f(j) = input_f(j);
second_half_f(j) = input_f(_half_len + j);
} else {
first_half_f(j) = 0.0f;
second_half_f(j) = 0.0f;
}
}
PipeBarrier<PIPE_V>();

Mul<float>(result1_f, first_half_f, cos_f, _half_len);
Mul<float>(result2_f, second_half_f, sin_f, _half_len);
Mul<float>(result1_f, first_half_f, cos_f, half_float_copy_len);
Mul<float>(result2_f, second_half_f, sin_f, half_float_copy_len);
PipeBarrier<PIPE_V>();
Sub<float>(result3_f, result1_f, result2_f, _half_len);
Sub<float>(result3_f, result1_f, result2_f, half_float_copy_len);

Mul<float>(result1_f, first_half_f, sin_f, _half_len);
Mul<float>(result2_f, second_half_f, cos_f, _half_len);
Mul<float>(result1_f, first_half_f, sin_f, half_float_copy_len);
Mul<float>(result2_f, second_half_f, cos_f, half_float_copy_len);
PipeBarrier<PIPE_V>();
Add<float>(result4_f, result1_f, result2_f, _half_len);
Add<float>(result4_f, result1_f, result2_f, half_float_copy_len);

LocalTensor<float> output_f = _tmp_float_output.Get<float>();
for (size_t j = 0; j < _half_len; j++) {
Expand All @@ -507,21 +518,26 @@ __aicore__ inline void RoPEKernelNeox<T, U>::compute(size_t i) {
LocalTensor<T> result3 = _tmp_result3.Get<T>();
LocalTensor<T> result4 = _tmp_result4.Get<T>();

for (size_t j = 0; j < _half_len; j++) {
first_half(j) = input_ub(j);
second_half(j) = input_ub(_half_len + j);
for (size_t j = 0; j < _half_copy_len; j++) {
if (j < _half_len) {
first_half(j) = input_ub(j);
second_half(j) = input_ub(_half_len + j);
} else {
first_half(j) = static_cast<T>(0);
second_half(j) = static_cast<T>(0);
}
}
PipeBarrier<PIPE_V>();

Mul<T>(result1, first_half, cos_ub, _half_len);
Mul<T>(result2, second_half, sin_ub, _half_len);
Mul<T>(result1, first_half, cos_ub, _half_copy_len);
Mul<T>(result2, second_half, sin_ub, _half_copy_len);
PipeBarrier<PIPE_V>();
Sub<T>(result3, result1, result2, _half_len);
Sub<T>(result3, result1, result2, _half_copy_len);

Mul<T>(result1, first_half, sin_ub, _half_len);
Mul<T>(result2, second_half, cos_ub, _half_len);
Mul<T>(result1, first_half, sin_ub, _half_copy_len);
Mul<T>(result2, second_half, cos_ub, _half_copy_len);
PipeBarrier<PIPE_V>();
Add<T>(result4, result1, result2, _half_len);
Add<T>(result4, result1, result2, _half_copy_len);

for (size_t j = 0; j < _half_len; j++) {
output_ub(j) = result3(j);
Expand Down
Loading