mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2025-01-04 15:54:36 +00:00
26 lines
970 B
Plaintext
26 lines
970 B
Plaintext
/**
|
|
* Copyright (c) 2023 Nomic, Inc. All rights reserved.
|
|
*
|
|
* This software is licensed under the terms of the Software for Open Models License (SOM),
|
|
* version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
|
|
* this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
|
|
*/
|
|
|
|
void main() {
|
|
const uint i = gl_WorkGroupID.x;
|
|
const int r = inB[i + pcs.inBOff];
|
|
|
|
int z = 0;
|
|
for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
|
|
const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
|
|
const mat4 result = dequantize_block(inIndex, ind%NL);
|
|
for (uint j = 0; j < 4; ++j) {
|
|
for (uint k = 0; k < 4; ++k) {
|
|
const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
|
|
out_[outIndex] = result[j][k];
|
|
++z;
|
|
}
|
|
}
|
|
}
|
|
}
|