Jeff Bolz
6a6780a232
This adds nvfp4 support for get_rows, dequant, and mul_mat(_id). For mul_mat, it does not add support for the dp4/q8_1 path, it's all via fp16/fp32.
33 lines
962 B
Plaintext
33 lines
962 B
Plaintext
#version 450
|
|
|
|
#include "dequant_head.glsl"
|
|
|
|
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer A {block_nvfp4 data_a[];};
|
|
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
|
|
|
|
void main() {
|
|
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
|
|
|
|
init_iq_shmem(gl_WorkGroupSize);
|
|
|
|
const uint tid = gl_LocalInvocationID.x % 64;
|
|
const uint sub = tid / 16;
|
|
const uint ir = tid % 16;
|
|
const uint ib = 16 * i + ir;
|
|
if (ib >= p.nel / 64) {
|
|
return;
|
|
}
|
|
|
|
const uint q_idx = 8 * sub;
|
|
const uint b_idx = 1024 * i + 64 * ir + 16 * sub;
|
|
|
|
const float d = ue4m3_to_fp32(data_a[ib].d[sub]);
|
|
|
|
[[unroll]] for (uint l = 0; l < 8; ++l) {
|
|
data_b[b_idx + l + 0] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]));
|
|
data_b[b_idx + l + 8] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] >> 4]));
|
|
}
|
|
}
|