mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-11-03 09:22:01 +00:00
9e405b6e2e
op_getrows_f32 is required since https://github.com/ggerganov/llama.cpp/pull/6122 for the Vulkan w/ Kompute backend to be functional. As such, implement this op to make this backend functional again.
32 lines
762 B
Plaintext
32 lines
762 B
Plaintext
#version 450
|
|
|
|
#include "common.comp"
|
|
|
|
layout(local_size_x = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer tensorInA { float 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_f32(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_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
|
|
}
|