mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2025-01-04 15:54:36 +00:00
40 lines
1.1 KiB
Plaintext
40 lines
1.1 KiB
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.
|
|
*/
|
|
|
|
#version 450
|
|
|
|
#include "common.comp"
|
|
|
|
layout(local_size_x = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
|
|
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
|
|
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
|
|
|
|
layout (push_constant) uniform parameter {
|
|
uint inAOff;
|
|
uint inBOff;
|
|
uint outOff;
|
|
int ne00;
|
|
int nb01;
|
|
int nb1;
|
|
} pcs;
|
|
|
|
void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
|
|
for (int j = 0; j < k; j++) {
|
|
out_[y + j] = inA[x + j];
|
|
}
|
|
}
|
|
|
|
void main() {
|
|
const uint i = gl_WorkGroupID.x;
|
|
const int r = inB[i + pcs.inBOff];
|
|
|
|
dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1 + pcs.outOff, pcs.ne00);
|
|
}
|