CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-10 10:27:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

opencl: add f16 for add, sub, mul, div (#14984)

Browse Source
This commit is contained in:
lhez
2025-08-01 04:15:44 -07:00
committed by GitHub
parent baad94885d
commit 1c872f71fb
5 changed files with 414 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
ggml/src/ggml-opencl/kernels/add.cl
View File

@@ -81,3 +81,76 @@ kernel void kernel_add_row(
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] + src1[idx1];
}
kernel void kernel_add_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
int ne00,
int ne01,
int ne02,
int ne03,
ulong nb00,
ulong nb01,
ulong nb02,
ulong nb03,
int ne10,
int ne11,
int ne12,
int ne13,
ulong nb10,
ulong nb11,
ulong nb12,
ulong nb13,
int ne0,
int ne1,
int ne2,
int ne3,
ulong nb0,
ulong nb1,
ulong nb2,
ulong nb3
) {
src0 = src0 + offset0;
src1 = src1 + offset1;
dst = dst + offsetd;
int i03 = get_group_id(2);
int i02 = get_group_id(1);
int i01 = get_group_id(0);
int i13 = i03 % ne13;
int i12 = i02 % ne12;
int i11 = i01 % ne11;
global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const int i10 = i0 % ne10;
*((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) + *((global half *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_add_row_f16(
global half4 * src0,
ulong offset0,
global half4 * src1,
ulong offset1,
global half4 * dst,
ulong offsetd,
int ne
) {
src0 = (global half4*)((global char*)src0 + offset0);
src1 = (global half4*)((global char*)src1 + offset1);
dst = (global half4*)((global char*)dst + offsetd);
// This performs better than using %.
uint gid = get_global_id(0);
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] + src1[idx1];
}

66
ggml/src/ggml-opencl/kernels/div.cl
View File

@@ -70,3 +70,69 @@ kernel void kernel_div_row(
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] / src1[idx1];
}
kernel void kernel_div_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb00,
ulong nb01,
ulong nb02,
ulong nb03,
int ne10,
int ne11,
int ne12,
int ne13,
ulong nb10,
ulong nb11,
ulong nb12,
ulong nb13,
int ne0,
ulong nb0,
ulong nb1,
ulong nb2,
ulong nb3
) {
src0 = src0 + offset0;
src1 = src1 + offset1;
dst = dst + offsetd;
int i03 = get_group_id(2);
int i02 = get_group_id(1);
int i01 = get_group_id(0);
int i13 = i03 % ne13;
int i12 = i02 % ne12;
int i11 = i01 % ne11;
global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const int i10 = i0 % ne10;
*((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) / *((global half *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_div_row_f16(
global half4 * src0,
ulong offset0,
global half4 * src1,
ulong offset1,
global half4 * dst,
ulong offsetd,
int ne
) {
src0 = (global half4*)((global char*)src0 + offset0);
src1 = (global half4*)((global char*)src1 + offset1);
dst = (global half4*)((global char*)dst + offsetd);
// This performs better than using %.
uint gid = get_global_id(0);
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] / src1[idx1];
}

73
ggml/src/ggml-opencl/kernels/mul.cl
View File

@@ -77,3 +77,76 @@ kernel void kernel_mul_row(
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] * src1[idx1];
}
kernel void kernel_mul_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
int ne00,
int ne01,
int ne02,
int ne03,
ulong nb00,
ulong nb01,
ulong nb02,
ulong nb03,
int ne10,
int ne11,
int ne12,
int ne13,
ulong nb10,
ulong nb11,
ulong nb12,
ulong nb13,
int ne0,
int ne1,
int ne2,
int ne3,
ulong nb0,
ulong nb1,
ulong nb2,
ulong nb3
) {
src0 = src0 + offset0;
src1 = src1 + offset1;
dst = dst + offsetd;
int i03 = get_group_id(2);
int i02 = get_group_id(1);
int i01 = get_group_id(0);
int i13 = i03 % ne13;
int i12 = i02 % ne12;
int i11 = i01 % ne11;
global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const int i10 = i0 % ne10;
*((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) * *((global half *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_mul_row_f16(
global half4 * src0,
ulong offset0,
global half4 * src1,
ulong offset1,
global half4 * dst,
ulong offsetd,
int ne
) {
src0 = (global half4*)((global char*)src0 + offset0);
src1 = (global half4*)((global char*)src1 + offset1);
dst = (global half4*)((global char*)dst + offsetd);
// This performs better than using %.
uint gid = get_global_id(0);
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] * src1[idx1];
}

66
ggml/src/ggml-opencl/kernels/sub.cl
View File

@@ -70,3 +70,69 @@ kernel void kernel_sub_row(
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] - src1[idx1];
}
kernel void kernel_sub_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb00,
ulong nb01,
ulong nb02,
ulong nb03,
int ne10,
int ne11,
int ne12,
int ne13,
ulong nb10,
ulong nb11,
ulong nb12,
ulong nb13,
int ne0,
ulong nb0,
ulong nb1,
ulong nb2,
ulong nb3
) {
src0 = src0 + offset0;
src1 = src1 + offset1;
dst = dst + offsetd;
int i03 = get_group_id(2);
int i02 = get_group_id(1);
int i01 = get_group_id(0);
int i13 = i03 % ne13;
int i12 = i02 % ne12;
int i11 = i01 % ne11;
global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const int i10 = i0 % ne10;
*((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) - *((global half *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_sub_row_f16(
global half4 * src0,
ulong offset0,
global half4 * src1,
ulong offset1,
global half4 * dst,
ulong offsetd,
int ne
) {
src0 = (global half4*)((global char*)src0 + offset0);
src1 = (global half4*)((global char*)src1 + offset1);
dst = (global half4*)((global char*)dst + offsetd);
// This performs better than using %.
uint gid = get_global_id(0);
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
dst[gid] = src0[gid] - src1[idx1];
}