Skip to content

[Layer] Reduce repeated sin and cos embedding table data to optimize ROPE perf. #266

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

Merged
pujiang2018 merged 1 commit into from
Mar 16, 2024
Merged

[Layer] Reduce repeated sin and cos embedding table data to optimize ROPE perf. #266

Show file tree
Hide file tree
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
76 changes: 43 additions & 33 deletions src/layers/rotary_embedding.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,9 +16,7 @@

#include "compile_util.h"

static int max_seq_len_cached = -1;
static int inv_freq_size = -1;
static float *inv_freq;
static float *emb_cos = nullptr;
static float *emb_sin = nullptr;

Expand All @@ -29,38 +27,35 @@ LlamaRotaryEmbedding::LlamaRotaryEmbedding(const int dim, const int max_position
if (!initialized) {
initialized = true;

max_seq_len_cached = max_position_embeddings;
inv_freq_size = (dim + 1) / 2;
inv_freq = (float *)malloc(inv_freq_size * sizeof(float));
float *inv_freq = (float *)malloc(inv_freq_size * sizeof(float));
for (size_t i = 0; i < inv_freq_size; i++) {
inv_freq[i] = 1.0 / pow(base, float(i * 2) / dim);
}

llamaCalEmb();
llamaCalEmb(inv_freq, max_position_embeddings);
} else if (dim != inv_freq_size * 2) {
printf("Incorrect dim=%d, inv_freq_size=%d\n", dim, inv_freq_size);
exit(-1);
}
};

void LlamaRotaryEmbedding::llamaCalEmb() {
emb_cos = (float *)aligned_alloc(64, max_seq_len_cached * (inv_freq_size * 2) * sizeof(float));
emb_sin = (float *)aligned_alloc(64, max_seq_len_cached * (inv_freq_size * 2) * sizeof(float));
void LlamaRotaryEmbedding::llamaCalEmb(const float *inv_freq, const int max_position_embeddings) {
emb_cos = (float *)aligned_alloc(64, max_position_embeddings * inv_freq_size * sizeof(float));
emb_sin = (float *)aligned_alloc(64, max_position_embeddings * inv_freq_size * sizeof(float));

#pragma omp parallel for
for (size_t i = 0; i < max_seq_len_cached; i++) {
float *pcos = emb_cos + i * inv_freq_size * 2;
float *psin = emb_sin + i * inv_freq_size * 2;
for (size_t i = 0; i < max_position_embeddings; i++) {
float *pcos = emb_cos + i * inv_freq_size;
float *psin = emb_sin + i * inv_freq_size;

for (size_t j = 0; j < inv_freq_size; j++) {
float tmp = i * inv_freq[j];
float cos_tmp = std::cos(tmp);
float sin_tmp = std::sin(tmp);

pcos[j] = cos_tmp;
pcos[j + inv_freq_size] = cos_tmp;
psin[j] = sin_tmp;
psin[j + inv_freq_size] = sin_tmp;
}
}
}
Expand All @@ -84,14 +79,31 @@ void LlamaRotaryEmbedding::llamaCalEmb() {
// position_ids: an array in the size of seq_len
// query and key is the matrix like below:
//
// |<------------------------------ head_size * head_num --------------------------------->|
// |_head_size|___________________________________________________________________________________
// | | | | | | | | | ^
// | | | | | | | | | |
// | | | | | | | | | bs*seq_len
// | | | | | | | | | |
// | | | | | | | | | |
// |__________|__________|__________|__________|__________|__________|__________|__________|____v__
// |<------------------------------ head_num * head_size --------------------------------->|
// |_head_size|_____________________________________________________________________________ _ _ _ _
// | | | | | | | | | ^
// | | | | | | | | | |
// | | | | | | | | | bs*seq_len
// | | | | | | | | | |
// | | | | | | | | | |
// |__________|__________|__________|__________|__________|__________|__________|__________|_ _ _v_
//
// inv_freq:
// _____
// |_____| 1
// head_size/2
//
// emb_cos: emb_sin:
// _____ _____
// | | | |
// | | | |
// | | | |
// | | | | max_position_embeddings
// | | | |
// | | | |
// |_____| |_____|
// head_size/2 head_size/2

void LlamaRotaryEmbedding::forward(
float *query, float *key, int qStride, int kStride, const int *qkShape, const int *positionIds) {
int dim = inv_freq_size * 2;
Expand All @@ -115,22 +127,22 @@ void LlamaRotaryEmbedding::forward(
for (int bs = 0; bs < batchSize; ++bs) {
for (int seq = 0; seq < seqLen; ++seq) {
int pos = positionIds[seq];
float *pcos = emb_cos + pos * dim;
float *psin = emb_sin + pos * dim;
float *pcos = emb_cos + pos * half;
float *psin = emb_sin + pos * half;

float *q = query + bs * seqLen * qStride + seq * qStride + head * dim;
float *k = key + bs * seqLen * kStride + seq * kStride + head * dim;
#pragma omp simd
for (int i = 0; i < half; ++i) {
if (head < qHeads) {
auto q1 = q[i];
q[i] = q[i] * pcos[i] - q[i + half] * psin[i];
q[i + half] = q[i + half] * pcos[i + half] + q1 * psin[i + half];
q[i] = q1 * pcos[i] - q[i + half] * psin[i];
q[i + half] = q[i + half] * pcos[i] + q1 * psin[i];
}
if (head < kHeads) {
auto k1 = k[i];
k[i] = k[i] * pcos[i] - k[i + half] * psin[i];
k[i + half] = k[i + half] * pcos[i + half] + k1 * psin[i + half];
k[i] = k1 * pcos[i] - k[i + half] * psin[i];
k[i + half] = k[i + half] * pcos[i] + k1 * psin[i];
}
}
}
Expand All @@ -155,8 +167,8 @@ void LlamaRotaryEmbedding::forward(
for (int bs = 0; bs < batchSize; ++bs) {
for (int seq = 0; seq < seqLen; ++seq) {
int pos = positionIds[seq];
float *pcos = emb_cos + pos * dim;
float *psin = emb_sin + pos * dim;
float *pcos = emb_cos + pos * half;
float *psin = emb_sin + pos * half;

bfloat16_t *q = query + bs * seqLen * qStride + seq * qStride + head * dim;
bfloat16_t *k = key + bs * seqLen * kStride + seq * kStride + head * dim;
Expand All @@ -167,9 +179,7 @@ void LlamaRotaryEmbedding::forward(
__mmask16 mask = (remain >= 16 ? 0xffff : (1 << remain) - 1);

__m512 pCosVec = _mm512_maskz_loadu_ps(mask, &pcos[i]);
__m512 pCosHalfVec = _mm512_maskz_loadu_ps(mask, &pcos[i + half]);
__m512 pSinVec = _mm512_maskz_loadu_ps(mask, &psin[i]);
__m512 pSinHalfVec = _mm512_maskz_loadu_ps(mask, &psin[i + half]);

// Compute something like:
// q[i] = q[i] * pcos[i] - q[i + half] * psin[i];
Expand All @@ -178,7 +188,7 @@ void LlamaRotaryEmbedding::forward(
__m512 qVec = bfloat16_t::cvt_bf16_to_fp32(_mm256_maskz_loadu_epi16(mask, &q[i]));
__m512 qHalfVec = bfloat16_t::cvt_bf16_to_fp32(_mm256_maskz_loadu_epi16(mask, &q[i + half]));
__m512 qNew = _mm512_fmsub_ps(qVec, pCosVec, _mm512_mul_ps(qHalfVec, pSinVec));
__m512 qHalfNew = _mm512_fmadd_ps(qHalfVec, pCosHalfVec, _mm512_mul_ps(qVec, pSinHalfVec));
__m512 qHalfNew = _mm512_fmadd_ps(qHalfVec, pCosVec, _mm512_mul_ps(qVec, pSinVec));
_mm256_mask_storeu_epi16(&q[i], mask, bfloat16_t::cvt_fp32_to_bf16(qNew));
_mm256_mask_storeu_epi16(&q[i + half], mask, bfloat16_t::cvt_fp32_to_bf16(qHalfNew));
}
Expand All @@ -187,7 +197,7 @@ void LlamaRotaryEmbedding::forward(
__m512 kVec = bfloat16_t::cvt_bf16_to_fp32(_mm256_maskz_loadu_epi16(mask, &k[i]));
__m512 kHalfVec = bfloat16_t::cvt_bf16_to_fp32(_mm256_maskz_loadu_epi16(mask, &k[i + half]));
__m512 kNew = _mm512_fmsub_ps(kVec, pCosVec, _mm512_mul_ps(kHalfVec, pSinVec));
__m512 kHalfNew = _mm512_fmadd_ps(kHalfVec, pCosHalfVec, _mm512_mul_ps(kVec, pSinHalfVec));
__m512 kHalfNew = _mm512_fmadd_ps(kHalfVec, pCosVec, _mm512_mul_ps(kVec, pSinVec));
_mm256_mask_storeu_epi16(&k[i], mask, bfloat16_t::cvt_fp32_to_bf16(kNew));
_mm256_mask_storeu_epi16(&k[i + half], mask, bfloat16_t::cvt_fp32_to_bf16(kHalfNew));
}
Expand Down
2 changes: 1 addition & 1 deletion src/layers/rotary_embedding.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@ class LlamaRotaryEmbedding {
bfloat16_t *query, bfloat16_t *key, int qStride, int kStride, const int *qkShape, const int *positionIds);

private:
void llamaCalEmb();
void llamaCalEmb(const float *inv_freq, const int max_position_embeddings);

private:
static bool initialized;
Expand Down