45 lines
1.2 KiB
Plaintext
45 lines
1.2 KiB
Plaintext
#version 450
|
|
|
|
#include "common.comp"
|
|
|
|
#define NL 16
|
|
#define BYTES_FOR_TYPE 4 /*bytes for float*/
|
|
#define SIZE_OF_BLOCK sizeof_block_q6_k
|
|
|
|
layout(local_size_x = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer tensorInA { uint8_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;
|
|
|
|
block_q6_k get_unaligned_block_q6_k(uint index) {
|
|
block_q6_k fres;
|
|
[[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
|
|
fres.ql[it] = inA[index + it];
|
|
}
|
|
[[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
|
|
fres.qh[it] = inA[index + QK_K/2 + it];
|
|
}
|
|
[[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
|
|
fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
|
|
}
|
|
fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
|
|
return fres;
|
|
}
|
|
|
|
mat4 dequantize_block(uint index, uint il) {
|
|
const block_q6_k block = get_unaligned_block_q6_k(index);
|
|
return dequantize_q6_k(block, il);
|
|
}
|
|
|
|
#include "op_getrows.comp"
|