Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml webgpu: faster matrix multiplication/matrix-vector multiplication (#17031)

Browse Source 
* Faster tensors (#8)

Add fast matrix and matrix/vector multiplication.

* Use map for shader replacements instead of pair of strings
This commit is contained in:
Reese Levine
2025-11-07 19:27:20 -08:00
committed by GitHub
parent 299f5d782c
commit 647b960bd8
8 changed files with 1247 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl
+5 -5
View File
@@ -864,8 +864,8 @@ struct MulMatParams {
broadcast3: u32
};
@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // N rows, K columns
@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // M rows, K columns (transposed)
@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
@group(0) @binding(3) var<uniform> params: MulMatParams;
@@ -891,8 +891,8 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let dst2_rem = dst3_rem % dst2_stride;
let row = dst2_rem / params.n; // output row
let col = dst2_rem % params.n; // output column
let row = dst2_rem / params.m; // output row
let col = dst2_rem % params.m; // output column
let src0_idx_base = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01;
let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11;
@@ -901,7 +901,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
for (var i: u32 = 0u; i < params.k/{{BLOCK_SIZE}}; i = i + 1u) {
sum += multiply_add(src0_idx_base, src1_idx_base, i);
}
dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.m + col] = sum;
}
#end(SHADER)