Skip to content

[ET-VK] Improve q8 matmul by increasing TILE_N4 #14597

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
Sep 25, 2025
Merged

[ET-VK] Improve q8 matmul by increasing TILE_N4 #14597

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
2 changes: 1 addition & 1 deletion backends/vulkan/runtime/graph/ops/glsl/linear_fp_output_tile_int8_int8_compute.glslh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,7 @@ void accumulate_out_tile_with_int_accum(
input_zp_vec * weight_sums.data[n4] + accum.data[m][n4];
out_tile.data[m][n4] =
fma(VEC4_T(accum_adjusted),
VEC4_T(input_q_scale * weight_scales.data[0]),
VEC4_T(input_q_scale * weight_scales.data[n4]),
out_tile.data[m][n4]);
}
}
Expand Down
2 changes: 1 addition & 1 deletion backends/vulkan/runtime/graph/ops/glsl/linear_q8ta_q8csw_tiled.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ linear_q8ta_q8csw_tiled:
PACKED_INT8_INPUT_STORAGE: buffer
WEIGHT_STORAGE: texture2d
TILE_M4: 1
TILE_N4: 1
TILE_N4: 2
TILE_K4: 1
generate_variant_forall:
DTYPE:
Expand Down
4 changes: 4 additions & 0 deletions backends/vulkan/runtime/graph/ops/impl/QuantizedLinear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,6 +77,10 @@ utils::uvec3 quantized_linear_global_wg_size(
M_per_tile = 1;
}

if (shader.kernel_name.find("q8ta_q8csw_tiled") != std::string::npos) {
N_per_tile = 8;
}

const uint32_t num_N_tiles = utils::div_up(N, N_per_tile);
const uint32_t num_M_tiles = utils::div_up(M, M_per_tile);

Expand Down
Loading