Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: support all backend tests for SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU/NORM (#24582)

Browse Source 
* vulkan: make SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU use unary.comp

* vulkan: make NORM support noncontig

* add noncontiguous row test cases for norm/l2_norm, handle this in the CPU backend and l2_norm.comp

* fix supports_op for cuda and webgpu
This commit is contained in:
Jeff Bolz
2026-06-23 09:48:24 -05:00
committed by GitHub
parent 92e854ab83
commit 72a9269172
15 changed files with 171 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ops.cpp
+50 -23
View File
@@ -3688,8 +3688,6 @@ static void ggml_compute_forward_norm_f32(
GGML_ASSERT(ggml_are_same_shape(src0, dst));
GGML_ASSERT(src0->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;
@@ -3703,25 +3701,49 @@ static void ggml_compute_forward_norm_f32(
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
float sum = 0.0;
ggml_vec_sum_f32(ne00, &sum, x);
float mean = sum/ne00;
if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
const float * xf = (const float *) x;
float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
float variance = 0;
float sum = 0.0;
ggml_vec_sum_f32(ne00, &sum, xf);
float mean = sum/ne00;
float * yf = (float *) y;
float variance = 0;
#ifdef GGML_USE_ACCELERATE
mean = -mean;
vDSP_vsadd(x, 1, &mean, y, 1, ne00);
vDSP_measqv(y, 1, &variance, ne00);
mean = -mean;
vDSP_vsadd(xf, 1, &mean, yf, 1, ne00);
vDSP_measqv(yf, 1, &variance, ne00);
#else
variance = ggml_vec_cvar_f32(ne00, y, x, mean);
variance = ggml_vec_cvar_f32(ne00, yf, xf, mean);
#endif //GGML_USE_ACCELERATE
const float scale = 1.0f/sqrtf(variance + eps);
ggml_vec_scale_f32(ne00, y, scale);
const float scale = 1.0f/sqrtf(variance + eps);
ggml_vec_scale_f32(ne00, yf, scale);
} else {
float sum = 0.0;
for (int64_t i00 = 0; i00 < ne00; i00++) {
sum += *(const float *) (x + i00*nb00);
}
const float mean = sum/ne00;
float variance = 0.0f;
for (int64_t i00 = 0; i00 < ne00; i00++) {
const float v = *(const float *) (x + i00*nb00) - mean;
*(float *) (y + i00*nb0) = v;
variance += v * v;
}
variance /= ne00;
const float scale = 1.0f/sqrtf(variance + eps);
for (int64_t i00 = 0; i00 < ne00; i00++) {
*(float *) (y + i00*nb0) *= scale;
}
}
}
}
}
@@ -4142,8 +4164,6 @@ static void ggml_compute_forward_l2_norm_f32(
GGML_ASSERT(ggml_are_same_shape(src0, dst));
GGML_ASSERT(src0->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;
@@ -4158,20 +4178,27 @@ static void ggml_compute_forward_l2_norm_f32(
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
ggml_float sum = 0.0;
for (int64_t i00 = 0; i00 < ne00; i00++) {
sum += (ggml_float)(x[i00] * x[i00]);
const float xi = *(const float *) (x + i00*nb00);
sum += (ggml_float)(xi * xi);
}
float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
memcpy(y, x, ne00 * sizeof(float));
const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
ggml_vec_scale_f32(ne00, y, scale);
char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
memcpy(y, x, ne00 * sizeof(float));
ggml_vec_scale_f32(ne00, (float *) y, scale);
} else {
for (int64_t i00 = 0; i00 < ne00; i00++) {
const float xi = *(const float *) (x + i00*nb00);
*(float *) (y + i00*nb0) = xi * scale;
}
}
}
}
}