Ruben Ortlam
bcf5bda6f5
* vulkan: add mmq q2_k integer dot support * Refactor mmq caching * Reduce mmq register use * Load 4 quant blocks into shared memory in one step * Pack q2_k blocks into caches of 32 * Use 32-bit accumulators for integer dot matmul * Add q4_k mmq * Add q3_k mmq * Add q5_k mmq * Add q6_k mmq * Add mxfp4 mmq, enable MMQ MUL_MAT_ID * Fix mmv dm loads
33 lines
941 B
Plaintext
33 lines
941 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_mxfp4 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 il = tid/32;
|
|
const uint ir = tid%32;
|
|
const uint ib = 32*i + ir;
|
|
if (ib >= p.nel / 32) {
|
|
return;
|
|
}
|
|
|
|
const uint q_idx = 8*il;
|
|
const uint b_idx = 1024*i + 32*ir + q_idx;
|
|
|
|
const float d = e8m0_to_fp32(data_a[ib].e);
|
|
|
|
[[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 + 16] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] >> 4]));
|
|
}
|
|
}
|