CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-19 11:57:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Support ne[3]>1 in noncontig matrix-vector multiply (#15015)

Browse Source
This commit is contained in:
Jeff Bolz
2025-08-02 03:48:30 -05:00
committed by GitHub
parent 339bd0268c
commit ec0b18802c
3 changed files with 30 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
View File

@@ -26,6 +26,9 @@ layout (push_constant) uniform parameter
uint ne12;
uint b_offset;
uint d_offset;
uint nb03;
uint nb13;
uint nb23;
} p;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -34,6 +37,7 @@ void main() {
const uint tid = gl_LocalInvocationID.x;
const uint row_x = gl_GlobalInvocationID.y;
const uint channel = gl_GlobalInvocationID.z;
const uint i3 = gl_WorkGroupID.x;
const uint channel_x = channel / p.channel_x_divisor;
const uint channel_y = channel % p.ne12;
@@ -41,7 +45,7 @@ void main() {
const uint nrows_dst = p.nrows_x;
const uint row_dst = row_x;
const uint idst = channel*nrows_dst + row_dst;
const uint idst = i3*p.nb23 + channel*nrows_dst + row_dst;
FLOAT_TYPE temp = 0.0f;
@@ -58,8 +62,8 @@ void main() {
const uint row_y = col_x;
const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel_y*p.channel_stride_y + row_y;
const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
const vec4 av4 = vec4(data_a_v4[ix / 4]);
const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -74,8 +78,8 @@ void main() {
const uint row_y = col_x;
const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel_y*p.channel_stride_y + row_y;
const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
const vec4 av4 = vec4(data_a_v4[ix / 4]);
const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -91,8 +95,8 @@ void main() {
const uint row_y = col_x;
const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel_y*p.channel_stride_y + row_y;
const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);