mirror of
https://github.com/RYDE-WORK/llama.cpp.git
synced 2026-01-25 09:19:08 +08:00
af148c9386
Reuse the index calculations across all of src0/src1/dst. Add a shader variant for when src0/src1 are the same dimensions and additional modulus for src1 aren't needed. Div/mod are slow, so add "fast" div/mod that have a fast path when the calculation isn't needed or can be done more cheaply.
28 lines
741 B
Plaintext
28 lines
741 B
Plaintext
#version 450
|
|
|
|
#include "types.comp"
|
|
#include "generic_binary_head.comp"
|
|
|
|
const uint num_threads = 256;
|
|
|
|
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
|
|
|
|
void main() {
|
|
uint idx = get_idx();
|
|
|
|
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
|
|
const uint num_iter = 2;
|
|
|
|
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
|
|
if (idx >= p.ne) {
|
|
continue;
|
|
}
|
|
uint i00, i01, i02, i03;
|
|
get_indices(idx, i00, i01, i02, i03);
|
|
|
|
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
|
|
|
|
idx += num_threads;
|
|
}
|
|
}
|