Skip to content

feat(ggml): vectorize row conversion functions #16587

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

Closed
wants to merge 4 commits into from
+118 −6
Closed

feat(ggml): vectorize row conversion functions #16587

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3308759
Leverage the existing GGML_F32_VEC helpers to broadcast the fill valu…
Oct 11, 2025
dff1173
Vectorize additional f32 helper loops
Oct 11, 2025
e4189f7
Normalize f32 helper tails for ggml vec ops
Oct 13, 2025
bd95063
feat(ggml): vectorize row conversion functions
Oct 15, 2025
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
96 changes: 91 additions & 5 deletions ggml/src/ggml-cpu/vec.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,16 +77,85 @@ inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp
z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
}
}
inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; }
inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; }
inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; }
inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) {
int i = 0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));

GGML_F32_VEC vv = GGML_F32_VEC_SET1(v);

for (; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; ++j) {
GGML_F32_VEC ax = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
GGML_F32_VEC az = GGML_F32_VEC_ADD(ax, vv);
GGML_F32_VEC_STORE(z + i + j*GGML_F32_EPR, az);
}
}
#endif
for (; i < n; ++i) {
z[i] = x[i] + v;
}
}
inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) {
int i = 0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));

for (; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; ++j) {
GGML_F32_VEC ay = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
GGML_F32_VEC ax = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
ay = GGML_F32_VEC_ADD(ay, ax);
GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay);
}
}
#endif
for (; i < n; ++i) {
y[i] += x[i];
}
}
inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) {
int i = 0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));

GGML_F32_VEC vv = GGML_F32_VEC_SET1(v);

for (; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; ++j) {
GGML_F32_VEC ay = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
ay = GGML_F32_VEC_ADD(ay, vv);
GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay);
}
}
#endif
for (; i < n; ++i) {
y[i] += v;
}
}
inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; }
inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_FP16_TO_FP32(y[i]));
}
}
inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_set_f32 (const int n, float * x, const float v) {
int i = 0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));

GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);

for (; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; ++j) {
GGML_F32_VEC_STORE(x + i + j*GGML_F32_EPR, vx);
}
}
#endif
for (; i < n; ++i) {
x[i] = v;
}
}
inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; }
inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
Expand All @@ -95,7 +164,24 @@ inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp
}
}

inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; }
inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) {
int i = 0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));

for (; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; ++j) {
GGML_F32_VEC ax = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
GGML_F32_VEC ay = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
GGML_F32_VEC az = GGML_F32_VEC_MUL(ax, ay);
GGML_F32_VEC_STORE(z + i + j*GGML_F32_EPR, az);
}
}
#endif
for (; i < n; ++i) {
z[i] = x[i]*y[i];
}
}
inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_FP16_TO_FP32(y[i]));
Expand Down
28 changes: 27 additions & 1 deletion ggml/src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -428,20 +428,46 @@ ggml_bf16_t ggml_fp32_to_bf16(float x) {
}

void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
for (int64_t i = 0; i < n; i++) {
int i = 0;
#if defined(__F16C__)
for (; i + 7 < n; i += 8) {
__m128i x_i = _mm_loadu_si128((__m128i *)(x + i));
__m256 y_v = _mm256_cvtph_ps(x_i);
_mm256_storeu_ps(y + i, y_v);
}
#endif
for (; i < n; i++) {
y[i] = GGML_FP16_TO_FP32(x[i]);
}
}

void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
int i = 0;
#if defined(__F16C__)
for (; i + 7 < n; i += 8) {
__m256 x_v = _mm256_loadu_ps(x + i);
__m128i y_v = _mm256_cvtps_ph(x_v, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
_mm_storeu_si128((__m128i *)(y + i), y_v);
}
#endif
for (; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(x[i]);
}
}

void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
int i = 0;
#if defined(__AVX512F__)
for (; i + 15 < n; i += 16) {
__m512 y_v = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((__m256i *)(x + i))), 16));
_mm512_storeu_ps(y + i, y_v);
}
#elif defined(__AVX2__)
for (; i + 7 < n; i += 8) {
__m256 y_v = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(x + i))), 16));
_mm256_storeu_ps(y + i, y_v);
}
#endif
for (; i < n; ++i) {
y[i] = GGML_BF16_TO_FP32(x[i]);
}
Expand Down