Jeff Bolz
788e07dc91
* vulkan: Support Q2_0 The backend perf tests for mat-vec-mul weren't very good at first (worse than q2_k), doubling the rows per workgroup made a big difference. * reorder * resolve merge conflict, adjust err threshold for f16->q2_0 set_rows
30 lines
938 B
Plaintext
30 lines
938 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_q2_0 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;
|
|
|
|
const uint tid = gl_LocalInvocationID.x % 64;
|
|
const uint il = tid / 4;
|
|
const uint ir = tid % 4;
|
|
const uint ib = 4*i + ir;
|
|
if (ib >= p.nel / QUANT_K_Q2_0) {
|
|
return;
|
|
}
|
|
|
|
const uint b_idx = 256*i + QUANT_K_Q2_0*ir + 4*il;
|
|
const uint bits = uint(data_a[ib].qs[il]);
|
|
const float d = float(data_a[ib].d);
|
|
|
|
data_b[b_idx ] = D_TYPE(d * (float(bits & 3u) - 1.0f));
|
|
data_b[b_idx + 1] = D_TYPE(d * (float((bits >> 2u) & 3u) - 1.0f));
|
|
data_b[b_idx + 2] = D_TYPE(d * (float((bits >> 4u) & 3u) - 1.0f));
|
|
data_b[b_idx + 3] = D_TYPE(d * (float(bits >> 6u) - 1.0f));
|
|
}
|