mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

opencl: add sqr, sqrt, mean and ssm_conv (#17476)

Browse Source 
* opencl: add sqr

* opencl: add sqrt

* opencl: add mean

* opencl: add ssm_conv

* opencl: add missing cl_khr_fp16

* opencl: do sqrt in f32 then convert to f16 for better precision
This commit is contained in:
lhez
2025-11-26 13:29:58 -08:00
committed by GitHub
parent 5449367b21
commit 7cba58bbea
6 changed files with 557 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/kernels/mean.cl
+39
View File
@@ -0,0 +1,39 @@
kernel void kernel_mean_f32(
global float * src0,
ulong offset0,
global float * dst,
ulong offsetd,
int ne00,
int ne01,
int ne02,
int ne03,
ulong nb01,
ulong nb02,
ulong nb03,
ulong nb1,
ulong nb2,
ulong nb3
) {
src0 = (global float *)((global char *)src0 + offset0);
dst = (global float *)((global char *)dst + offsetd);
int i3 = get_global_id(2);
int i2 = get_global_id(1);
int i1 = get_global_id(0);
if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
return;
}
global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
global float * dst_row = (global float *) ((global char *) dst + i1*nb1 + i2*nb2 + i3*nb3);
float row_sum = 0;
for (int i0 = 0; i0 < ne00; i0++) {
row_sum += src_row[i0];
}
dst_row[0] = row_sum / ne00;
}
ggml/src/ggml-opencl/kernels/sqr.cl
+53
View File
@@ -0,0 +1,53 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
kernel void kernel_sqr_cont_f32(
global float * src0,
ulong offset0,
global float * dst,
ulong offsetd
) {
src0 = (global float*)((global char*)src0 + offset0);
dst = (global float*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = src0[gid] * src0[gid];
}
kernel void kernel_sqr_cont_f32_4(
global float4 * src0,
ulong offset0,
global float4 * dst,
ulong offsetd
) {
src0 = (global float4*)((global char*)src0 + offset0);
dst = (global float4*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = src0[gid] * src0[gid];
}
kernel void kernel_sqr_cont_f16(
global half * src0,
ulong offset0,
global half * dst,
ulong offsetd
) {
src0 = (global half*)((global char*)src0 + offset0);
dst = (global half*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = src0[gid] * src0[gid];
}
kernel void kernel_sqr_cont_f16_4(
global half4 * src0,
ulong offset0,
global half4 * dst,
ulong offsetd
) {
src0 = (global half4*)((global char*)src0 + offset0);
dst = (global half4*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = src0[gid] * src0[gid];
}
ggml/src/ggml-opencl/kernels/sqrt.cl
+53
View File
@@ -0,0 +1,53 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
kernel void kernel_sqrt_cont_f32(
global float * src0,
ulong offset0,
global float * dst,
ulong offsetd
) {
src0 = (global float*)((global char*)src0 + offset0);
dst = (global float*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = sqrt(src0[gid]);
}
kernel void kernel_sqrt_cont_f32_4(
global float4 * src0,
ulong offset0,
global float4 * dst,
ulong offsetd
) {
src0 = (global float4*)((global char*)src0 + offset0);
dst = (global float4*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = sqrt(src0[gid]);
}
kernel void kernel_sqrt_cont_f16(
global half * src0,
ulong offset0,
global half * dst,
ulong offsetd
) {
src0 = (global half*)((global char*)src0 + offset0);
dst = (global half*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = convert_half(sqrt(convert_float(src0[gid])));
}
kernel void kernel_sqrt_cont_f16_4(
global half4 * src0,
ulong offset0,
global half4 * dst,
ulong offsetd
) {
src0 = (global half4*)((global char*)src0 + offset0);
dst = (global half4*)((global char*)dst + offsetd);
uint gid = get_global_id(0);
dst[gid] = convert_half4(sqrt(convert_float4(src0[gid])));
}
ggml/src/ggml-opencl/kernels/ssm_conv.cl
+77
View File
@@ -0,0 +1,77 @@
kernel void kernel_ssm_conv_f32_f32(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb00,
ulong nb01,
ulong nb02,
int ne10,
ulong nb11,
ulong nb0,
ulong nb1,
ulong nb2
){
src0 = src0 + offset0;
src1 = src1 + offset1;
dst = dst + offsetd;
int ir = get_global_id(0);
int i2 = get_global_id(1);
int i3 = get_global_id(2);
int nc = ne10;
global float * s = (global float *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
global float * c = (global float *) (src1 + ir*nb11);
global float * d = (global float *) (dst + ir*nb0 + i2*nb1 + i3*nb2);
float sumf = 0.0f;
for (int i0 = 0; i0 < nc; ++i0) {
sumf += s[i0] * c[i0];
}
d[0] = sumf;
}
kernel void kernel_ssm_conv_f32_f32_4(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb00,
ulong nb01,
ulong nb02,
int ne10,
ulong nb11,
ulong nb0,
ulong nb1,
ulong nb2
) {
src0 = src0 + offset0;
src1 = src1 + offset1;
dst = dst + offsetd;
int ir = get_global_id(0);
int i2 = get_global_id(1);
int i3 = get_global_id(2);
int nc = ne10;
global float4 * s = (global float4 *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
global float4 * c = (global float4 *) (src1 + ir*nb11);
global float * d = (global float *) (dst + ir*nb0 + i2*nb1 + i3*nb2);
float sumf = 0.0f;
for (int i0 = 0; i0 < nc/4; ++i0) {
sumf += dot(s[i0], c[i0]);
}
d[0] = sumf;
}