ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

vulkan: dynamic subgroup size for the remaining k quants (#10745)

Browse source 
* q5_k

q4_k

q3_k

q2_k

q6_k multi row example

* revert as multi row isnt faster for k quants
This commit is contained in:
Eve 2024-12-10 19:33:23 +00:00 • committed by GitHub
parent ae4b922614
commit dafae66cc2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 72 additions and 61 deletions
Download patch file Download diff file Expand all files Collapse all files

26
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
View file

@ -4,11 +4,12 @@
#include "mul_mat_vec_base.comp"
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE tmp[32];
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
// This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
@ -22,14 +23,17 @@ void main() {
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
const uint tid = gl_LocalInvocationID.x;
const uint itid = tid%16; // 0...16
const uint ix = tid/16;
const uint step = 8/K_QUANTS_PER_ITERATION; // 8 or 4
const uint step = 4;
const uint il = tid/step; // 0...3
const uint ir = tid - step*il; // 0...7 or 0...3
const uint n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4
const uint il = itid/step; // 0...3
const uint ir = itid - step*il; // 0...7 or 0...3
const uint n = 4;
const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const uint v_in = il % 2;
@ -40,7 +44,7 @@ void main() {
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
@ -115,7 +119,7 @@ void main() {
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}