vulkan: Support Q1_0 (#21539)

* vulkan: Support Q1_0

* use get_dm
This commit is contained in:
Jeff Bolz
2026-04-10 08:35:27 +02:00
committed by GitHub
parent e095a482a0
commit 7b69125331
9 changed files with 161 additions and 4 deletions
@@ -184,6 +184,31 @@ void quantize(uint dst_idx, uint src_idx)
}
#endif
#if defined(DATA_A_Q1_0)
void quantize(uint dst_idx, uint src_idx)
{
float sum_abs = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_Q1_0; j++) {
sum_abs += abs(data_s[src_idx + j]);
}
const float d = sum_abs / QUANT_K_Q1_0;
data_q[dst_idx].d = float16_t(d);
[[unroll]] for (int j = 0; j < QUANT_K_Q1_0 / 8; ++j) {
data_q[dst_idx].qs[j] = uint8_t(0);
}
[[unroll]] for (int j = 0; j < QUANT_K_Q1_0; ++j) {
if (data_s[src_idx + j] >= 0.0) {
data_q[dst_idx].qs[j / 8] |= uint8_t(1 << (j % 8));
}
}
}
#endif
#if defined(DATA_A_IQ4_NL)
uint best_index(float x) {
if (x <= kvalues_iq4nl[0]) return 0;