Skip to content

ggml-cpu: vectorize the fused RMS-norm kernel (ggml_compute_forward_rms_norm_f32) #32

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
codspeed-hq wants to merge 1 commit into
base: master
Choose a base branch
from
Open

ggml-cpu: vectorize the fused RMS-norm kernel (ggml_compute_forward_rms_norm_f32) #32

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
17 changes: 9 additions & 8 deletions ggml/src/ggml-cpu/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3825,11 +3825,11 @@ static void ggml_compute_forward_rms_norm_f32(
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);

ggml_float sum = 0.0;
// worth switching to explicit SIMD?
for (int64_t i00 = 0; i00 < ne00; i00++) {
sum += (ggml_float)(x[i00] * x[i00]);
}
// sum of squares over the row, computed with the vectorized
// dot-product kernel (sum(x*x)) instead of a scalar loop.
float sumf = 0.0f;
ggml_vec_dot_f32(ne00, &sumf, 0, x, 0, x, 0, 1);
const ggml_float sum = (ggml_float) sumf;

const float mean = sum/ne00;
const float scale = 1.0f/sqrtf(mean + eps);
Expand All @@ -3845,9 +3845,10 @@ static void ggml_compute_forward_rms_norm_f32(
const int64_t i13 = i03 % ne13;
const float * w = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13);

for (int64_t i00 = 0; i00 < ne00; i00++) {
y[i00] = x[i00] * scale * w[i00];
}
// y[i] = x[i]*scale*w[i], vectorized. Kept as two separate
// multiplies so the result is bit-identical to the scalar
// expression above.
ggml_vec_scale_mul_f32(ne00, y, x, w, scale);
} else {
memcpy(y, x, ne00 * sizeof(float));
ggml_vec_scale_f32(ne00, y, scale);
Expand Down
75 changes: 75 additions & 0 deletions ggml/src/ggml-cpu/vec.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -652,6 +652,81 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
#endif
}

// y[i] = (x[i]*s)*w[i]
// The two multiplies are kept separate (no FMA) so the result matches the
// scalar expression `x[i]*s*w[i]` bit-for-bit.
inline static void ggml_vec_scale_mul_f32(const int n, float * y, const float * x, const float * w, const float s) {
#if defined(GGML_SIMD)
#if defined(__ARM_FEATURE_SVE)
const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
const int ggml_f32_epr = sve_register_length / 32;
const int ggml_f32_step = 2 * ggml_f32_epr;

GGML_F32_VEC vs = GGML_F32_VEC_SET1(s);
const int np = (n & ~(ggml_f32_step - 1));
svfloat32_t ax1, ax2, aw1, aw2;
for (int i = 0; i < np; i += ggml_f32_step) {
ax1 = GGML_F32_VEC_LOAD(x + i);
aw1 = GGML_F32_VEC_LOAD(w + i);
ax1 = GGML_F32_VEC_MUL(ax1, vs);
ax1 = GGML_F32_VEC_MUL(ax1, aw1);
GGML_F32_VEC_STORE(y + i, ax1);

ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
aw2 = GGML_F32_VEC_LOAD(w + i + 1*ggml_f32_epr);
ax2 = GGML_F32_VEC_MUL(ax2, vs);
ax2 = GGML_F32_VEC_MUL(ax2, aw2);
GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ax2);
}
// leftovers: predicated tail
for (int i = np; i < n; i += ggml_f32_epr) {
svbool_t pg = svwhilelt_b32(i, n);
svfloat32_t vx = svld1_f32(pg, x + i);
svfloat32_t vw = svld1_f32(pg, w + i);
vx = svmul_f32_m(pg, vx, vs);
vx = svmul_f32_m(pg, vx, vw);
svst1_f32(pg, y + i, vx);
}
#elif defined(__riscv_v_intrinsic)
for (int i = 0, avl; i < n; i += avl) {
avl = __riscv_vsetvl_e32m8(n - i);
vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
vfloat32m8_t aw = __riscv_vle32_v_f32m8(&w[i], avl);
ax = __riscv_vfmul_vf_f32m8(ax, s, avl);
ax = __riscv_vfmul_vv_f32m8(ax, aw, avl);
__riscv_vse32_v_f32m8(&y[i], ax, avl);
}
#else
const int np = (n & ~(GGML_F32_STEP - 1));

GGML_F32_VEC vs = GGML_F32_VEC_SET1(s);

GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC aw[GGML_F32_ARR];

for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
aw[j] = GGML_F32_VEC_LOAD(w + i + j*GGML_F32_EPR);
ax[j] = GGML_F32_VEC_MUL(ax[j], vs);
ax[j] = GGML_F32_VEC_MUL(ax[j], aw[j]);
GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ax[j]);
}
}

// leftovers
for (int i = np; i < n; ++i) {
y[i] = x[i]*s*w[i];
}
#endif
#else
// scalar
for (int i = 0; i < n; ++i) {
y[i] = x[i]*s*w[i];
}
#endif
}

inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, const float s, const float b) {
#if defined(GGML_USE_ACCELERATE)
vDSP_vsmsa(x, 1, &s, &b, y, 1, n);
Expand Down