File size: 1,427 Bytes
c9c3262 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | #version 450
#include "types.glsl"
#include "generic_unary_head.glsl"
#include "dequant_funcs.glsl"
#if defined(DATA_A_IQ4_NL) || defined(DATA_A_MXFP4) || defined(DATA_A_NVFP4)
// 16 invocations needed for init_iq_shmem
layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
#else
layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
#endif
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
if (gl_LocalInvocationIndex.x != 0) {
return;
}
#endif
const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
if (idx >= p.ne) {
return;
}
uint dst_idx = get_doffset() + dst_idx(idx);
uint src_idx = src0_idx_quant(idx, QUANT_K);
const uint a_offset = 0;
const uint ib = src_idx;
const vec2 dm = get_dm(ib, a_offset);
[[unroll]] for (int j = 0; j < QUANT_K; j += 4) {
vec4 v = dequantize4(ib, j / QUANT_R, a_offset);
v = v * dm.x + vec4(dm.y);
#if QUANT_R == 2
data_d[dst_idx + j/2 + 0] = v[0];
data_d[dst_idx + j/2 + QUANT_K/2 + 0] = v[1];
data_d[dst_idx + j/2 + 1] = v[2];
data_d[dst_idx + j/2 + QUANT_K/2 + 1] = v[3];
#else
data_d[dst_idx + j + 0] = v[0];
data_d[dst_idx + j + 1] = v[1];
data_d[dst_idx + j + 2] = v[2];
data_d[dst_idx + j + 3] = v[3];
#endif
}
}
|