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
35 lines
1.4 KiB
Plaintext
35 lines
1.4 KiB
Plaintext
#version 450
|
|
|
|
#include "dequant_head.glsl"
|
|
|
|
layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
|
|
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
|
|
|
|
void main() {
|
|
[[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
|
|
const uint i = gl_WorkGroupID.x * 256 + wgy;
|
|
if (i >= p.nel / QUANT_K) {
|
|
return;
|
|
}
|
|
|
|
const uint tid = gl_LocalInvocationID.x;
|
|
const uint ip = tid / 32;
|
|
const uint il = tid - 32 * ip;
|
|
const uint is = 8 * ip + il / 16;
|
|
|
|
const uint y_idx = i * QUANT_K + 128 * ip + il;
|
|
|
|
const uint ql_idx = 32 * ip + il;
|
|
const uint8_t qs = data_a[i].qs[32 * ip + il];
|
|
|
|
FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].dm.x);
|
|
FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].dm.y);
|
|
data_b[y_idx + 0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
|
|
data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
|
|
data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
|
|
data_b[y_idx + 96] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+6] & 0xF) * ((qs >> 6) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+6] >> 4));
|
|
}
|
|
}
|