Skip to content

Doubling tile texel col count for mat mul op to improve performance. #15192

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 19, 2025
Merged

Doubling tile texel col count for mat mul op to improve performance. #15192

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 7 additions & 2 deletions backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_coop.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,12 +98,17 @@ void main() {
// Preload weight tensor
[[unroll]] for (int r = 0; r < 4; r++) {
$if QUANT_NBITS == 4:
$if WEIGHT_STORAGE == "buffer":
u8vec4 packed_weight_tex;
$else:
uvec4 packed_weight_tex;

$for c in range(0, TILE_TXCOLS, 2):
$if WEIGHT_STORAGE == "buffer":
qmat2_bufi = (pos + r) * weight_row_txstride + weight_txcol;
const u8vec4 packed_weight_tex = t_weight[qmat2_bufi + ${c}]
packed_weight_tex = t_weight[qmat2_bufi + ${c}]
$else:
const uvec4 packed_weight_tex = texelFetch(
packed_weight_tex = texelFetch(
t_weight, ivec2(weight_txcol + ${c}, pos + r), 0);

qmat2[r][${c}] = (VEC4_T((packed_weight_tex & 0xF0) >> 4) - 8.0);
Expand Down
2 changes: 1 addition & 1 deletion backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_coop.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ linear_qcsnw_coop:
WEIGHT_STORAGE: texture2d
SCALES_STORAGE: texture2d
TILE_ROWS: 4
TILE_TXCOLS: 1
TILE_TXCOLS: 2
QUANT_NBITS: 8
generate_variant_forall:
TILE_ROWS:
Expand Down
9 changes: 7 additions & 2 deletions backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_tiled.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,12 +106,17 @@ void main() {
for (int r = 0; r < 4; r++) {
VEC4_T qmat2[TILE_TXCOLS];
$if QUANT_NBITS == 4:
$if WEIGHT_STORAGE == "buffer":
u8vec4 packed_weight_tex;
$else:
uvec4 packed_weight_tex;

$for c in range(0, TILE_TXCOLS, 2):
$if WEIGHT_STORAGE == "buffer":
qmat2_bufi = (pos + r) * weight_row_txstride + weight_txcol;
const u8vec4 packed_weight_tex = t_weight[qmat2_bufi + ${c}]
packed_weight_tex = t_weight[qmat2_bufi + ${c}]
$else:
const uvec4 packed_weight_tex = texelFetch(
packed_weight_tex = texelFetch(
t_weight, u16vec2(weight_txcol + ${c}, pos + r), 0);

qmat2[${c}] = (VEC4_T(packed_weight_tex >> 4) - 8.0);
Expand Down
2 changes: 1 addition & 1 deletion backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_tiled.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ linear_qcsnw_tiled:
WEIGHT_STORAGE: texture2d
SCALES_STORAGE: texture2d
TILE_ROWS: 4
TILE_TXCOLS: 1
TILE_TXCOLS: 2
QUANT_NBITS: 8
generate_variant_forall:
TILE_ROWS:
Expand Down
4 changes: 2 additions & 2 deletions backends/vulkan/runtime/graph/ops/impl/QuantizedLinearQCSNW.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,7 +73,7 @@ utils::uvec3 linear_qcsnw_tiled_global_wg_size(
}

// Number of output texels in the output tile
uint32_t out_tile_ntxcols = 1;
uint32_t out_tile_ntxcols = 2;
if (quant_nbits == 4) {
out_tile_ntxcols = 2;
}
Expand Down Expand Up @@ -324,7 +324,7 @@ void add_linear_qcsnw_tiled_node(
}

// Number of output texels in the output tile
uint32_t out_tile_ntxcols = 1;
uint32_t out_tile_ntxcols = 2;
if (quant_nbits == 4) {
out_tile_ntxcols = 2;
}
Expand Down
Loading