ggml : build backends as libraries (#10256)
* ggml : build backends as libraries --------- Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: R0CKSTAR <xiaodong.ye@mthreads.com>
This commit is contained in:
@@ -0,0 +1,31 @@
|
||||
#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);
|
||||
}
|
||||
Reference in New Issue
Block a user