Skip to content

Improving 4bit quant mat mul performance by shifting position of -8 operation. #15436

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 29, 2025
Merged

Improving 4bit quant mat mul performance by shifting position of -8 operation. #15436

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
28 changes: 22 additions & 6 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 @@ -64,6 +64,12 @@ void main() {

T sums[TILE_ROWS * TILE_TXCOLS * 4];

$if QUANT_NBITS == 4:
// accumulate mat1 elements sum so -8 bias can be applied using it later.
T mat1_accum[TILE_ROWS];
$for r in range(TILE_ROWS):
mat1_accum[${r}] = T(0.0);

for (int r = 0; r < TILE_ROWS; ++r) {
$for c in range(TILE_TXCOLS):
$for j in range(4):
Expand All @@ -86,6 +92,10 @@ void main() {
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 QUANT_NBITS == 4:
// Accumulate mat1 element sum, this will be multiplied with -8 later for converting 4 bit data to a signed number.
mat1_accum[i] += mat1[i * 4 + 0] + mat1[i * 4 + 1] + mat1[i * 4 + 2] + mat1[i * 4 + 3];
}

$if WEIGHT_STORAGE == "buffer":
Expand All @@ -109,13 +119,13 @@ void main() {
packed_weight_tex = texelFetch(
t_weight, ivec2(weight_txcol + ${c}, pos + r), 0);

qmat2_vec4 = (VEC4_T(packed_weight_tex >> 4) - 8.0);
qmat2_vec4 = VEC4_T(packed_weight_tex >> 4);
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_vec4 = VEC4_T(packed_weight_tex & 0x0F);
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;
Expand Down Expand Up @@ -156,10 +166,16 @@ void main() {
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 QUANT_NBITS == 4:
scaled_sums.x = (sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 0] + mat1_accum[r] * -8.0) * scales[${c}].x;
scaled_sums.y = (sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 1] + mat1_accum[r] * -8.0) * scales[${c}].y;
scaled_sums.z = (sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 2] + mat1_accum[r] * -8.0) * scales[${c}].z;
scaled_sums.w = (sums[r * TILE_TXCOLS * 4 + ${c} * 4 + 3] + mat1_accum[r] * -8.0) * scales[${c}].w;
$else:
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) {
Expand Down
Loading