Skip to content

Improving quantized matmul performance by devectorizing shader. #15274

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+39 −26
Open

Improving quantized matmul performance by devectorizing shader. #15274

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b4010f2
Minor perf improvements to quantized mat mul shader. (#15261)
trivedivivek Oct 21, 2025
1a04777
Improving quantized matmul performance by devectorizing shader. (#15274)
trivedivivek Oct 21, 2025
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
65 changes: 39 additions & 26 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 @@ -62,36 +62,30 @@ void main() {
return;
}

VEC4_T sums[TILE_ROWS][TILE_TXCOLS];
T sums[TILE_ROWS * TILE_TXCOLS * 4];

for (int r = 0; r < TILE_ROWS; ++r) {
$for c in range(TILE_TXCOLS):
sums[r][${c}] = VEC4_T(0.0);
$for j in range(4):
sums[r * TILE_TXCOLS * 4 + ${c} * 4 + ${j}] = T(0.0);
}

const int in_row_txstride = div4(in_sizes.x);

for (int pos = 0, txpos = 0;
pos < in_sizes.x;
txpos < in_row_txstride;
pos += 4, txpos += 1) {

T mat1[TILE_ROWS][4];

$if IN_STORAGE == "buffer":
uint in_row_txstride = div4(in_sizes.x);
T mat1[TILE_ROWS * 4];

// Preload input tensor
for (int i = 0; i < TILE_ROWS; i++) {
$if IN_STORAGE == "buffer":
VEC4_T tmp = t_in[(out_row + i) * in_row_txstride + txpos];
mat1[i][0] = tmp.x;
mat1[i][1] = tmp.y;
mat1[i][2] = tmp.z;
mat1[i][3] = tmp.w;
VEC4_T mat1_vec4 = t_in[(out_row + i) * in_row_txstride + txpos];
$else:
VEC4_T tmp = VEC4_T(texelFetch(t_in, ivec3(txpos, out_row + i, 0), 0));
mat1[i][0] = tmp.x;
mat1[i][1] = tmp.y;
mat1[i][2] = tmp.z;
mat1[i][3] = tmp.w;
VEC4_T mat1_vec4 = VEC4_T(texelFetch(t_in, ivec3(txpos, out_row + i, 0), 0));
$for j in range(4):
mat1[i * 4 + ${j}] = mat1_vec4[${j}];
}

$if WEIGHT_STORAGE == "buffer":
Expand All @@ -100,7 +94,9 @@ void main() {

// Preload weight tensor
for (int r = 0; r < 4; r++) {
VEC4_T qmat2[TILE_TXCOLS];
T qmat2[TILE_TXCOLS * 4];
VEC4_T qmat2_vec4;

$if QUANT_NBITS == 4:
$if WEIGHT_STORAGE == "buffer":
u8vec4 packed_weight_tex;
Expand All @@ -115,20 +111,31 @@ void main() {
packed_weight_tex = texelFetch(
t_weight, ivec2(weight_txcol + ${c}, pos + r), 0);

qmat2[${c}] = (VEC4_T(packed_weight_tex >> 4) - 8.0);
qmat2[${c + 1}] = (VEC4_T(packed_weight_tex & 0x0F) - 8.0);
qmat2_vec4 = (VEC4_T(packed_weight_tex >> 4) - 8.0);
qmat2[${c} * 4 * TILE_TXCOLS + 0] = qmat2_vec4.x;
qmat2[${c} * 4 * TILE_TXCOLS + 1] = qmat2_vec4.y;
qmat2[${c} * 4 * TILE_TXCOLS + 2] = qmat2_vec4.z;
qmat2[${c} * 4 * TILE_TXCOLS + 3] = qmat2_vec4.w;

qmat2_vec4 = (VEC4_T(packed_weight_tex & 0x0F) - 8.0);
qmat2[${c} * 4 * TILE_TXCOLS + 4] = qmat2_vec4.x;
qmat2[${c} * 4 * TILE_TXCOLS + 5] = qmat2_vec4.y;
qmat2[${c} * 4 * TILE_TXCOLS + 6] = qmat2_vec4.z;
qmat2[${c} * 4 * TILE_TXCOLS + 7] = qmat2_vec4.w;
$else:
$for c in range(TILE_TXCOLS):
$if WEIGHT_STORAGE == "buffer":
qmat2_bufi = (pos + r) * weight_row_txstride + out_txcol;
qmat2[${c}] = t_weight[qmat2_bufi + ${c}];
qmat2_vec4 = t_weight[qmat2_bufi + ${c}];
$else:
qmat2[${c}] = VEC4_T(
texelFetch(t_weight, ivec2(out_txcol + ${c}, pos + r), 0));
qmat2_vec4 = VEC4_T(texelFetch(t_weight, ivec2(out_txcol + ${c}, pos + r), 0));
$for j in range(4):
qmat2[${c} * 4 + ${j}] = qmat2_vec4[${j}];

for (int tr = 0; tr < TILE_ROWS; ++tr) {
$for c in range(TILE_TXCOLS):
sums[tr][${c}] += qmat2[${c}] * mat1[tr][r];
$for j in range(4):
sums[tr * TILE_TXCOLS * 4 + ${c} * 4 + ${j}] += qmat2[${c} * 4 + ${j}] * mat1[tr * 4 + r];
}
}
}
Expand All @@ -147,16 +154,22 @@ void main() {
uint out_row_txstride = div4(out_sizes.x);

for (int r = 0; r < TILE_ROWS; ++r) {
VEC4_T scaled_sums;
$for c in range(TILE_TXCOLS):
scaled_sums.x = sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 0] * scales[${c}].x;
scaled_sums.y = sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 1] * scales[${c}].y;
scaled_sums.z = sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 2] * scales[${c}].z;
scaled_sums.w = sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 3] * scales[${c}].w;

$if OUT_STORAGE == "buffer":
if (out_row + r < out_sizes.y) {
out_bufi = (out_row + r) * out_row_txstride + out_txcol;
t_out[out_bufi + ${c}] = sums[r][${c}] * scales[${c}];
t_out[out_bufi + ${c}] = scaled_sums;
}
$else:
imageStore(
t_out,
ivec3(out_txcol + ${c}, out_row + r, 0),
sums[r][${c}] * scales[${c}]);
scaled_sums);
}
}
Loading