vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

kompute : fix ggml_add kernel

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-19 00:09:16 +02:00
parent 610394fff8
commit 1453215165
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 9 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

16
kompute-shaders/op_add.comp
View file

@ -30,6 +30,7 @@ layout(push_constant) uniform PushConstants {
int nb1;
int nb2;
int nb3;
//int offs; // TODO: needed for GGML_OP_ACC, see metal code
} pcs;
// general-purpose kernel for addition of two tensors
@ -44,15 +45,14 @@ void main() {
const uint i12 = i02 % pcs.ne12;
const uint i11 = i01 % pcs.ne11;
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + gl_SubgroupInvocationID.x*pcs.nb00) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11 + gl_SubgroupInvocationID.x*pcs.nb10) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1 + gl_SubgroupInvocationID.x*pcs.nb0 ) / 4);
int offs = 0; // TMP (see above)
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11 ) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1 + offs) / 4);
for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
out_[pcs.outOff + dst_off] = inA[pcs.inAOff + src0_off] + inB[pcs.inBOff + src1_off];
src0_off += gl_WorkGroupSize.x*pcs.ne00;
src1_off += gl_WorkGroupSize.x*pcs.ne10;
dst_off += gl_WorkGroupSize.x*pcs.ne0;
const uint i10 = i0 % pcs.ne10;
out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
}
}