root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-09-22 21:16:20 +00:00
Code Issues Actions 16 Packages Projects Releases Wiki Activity

vulkan : fix Qantized Mat-Vec Mul on AMD GPUs for ncols < 64 (#8855)

Browse Source 
* Fix Vulkan mul mat vec invalid results when ncols < warp size

* Only run backend ops mul mat vec block size test if block size not already covered
This commit is contained in:
0cc4m 2024-08-05 07:52:55 +02:00 committed by GitHub
parent 5587e57a76
commit 064cdc265f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 14 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/vulkan-shaders/mul_mat_vec.comp
View File

@ -16,6 +16,13 @@ void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
const uint tid = gl_LocalInvocationID.x; const uint tid = gl_LocalInvocationID.x;
// There are not enough cols to use all threads
if (tid >= p.ncols) {
return;
}
const uint block_size = min(p.ncols, BLOCK_SIZE);
uint a_offset, b_offset, d_offset; uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset); get_offsets(a_offset, b_offset, d_offset);
@ -23,8 +30,8 @@ void main() {
tmp[tid] = FLOAT_TYPE(0.0f); tmp[tid] = FLOAT_TYPE(0.0f);
[[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) { [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
const uint col = i*BLOCK_SIZE + 2*tid; const uint col = i*block_size + 2*tid;
const uint ib = (row*p.ncols + col)/QUANT_K; // block index const uint ib = (row*p.ncols + col)/QUANT_K; // block index
const uint iqs = (col%QUANT_K)/QUANT_R; // quant index const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
const uint iybs = col - col%QUANT_K; // y block start index const uint iybs = col - col%QUANT_K; // y block start index
@ -38,7 +45,7 @@ void main() {
// sum up partial sums and write back result // sum up partial sums and write back result
barrier(); barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) { [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
if (tid < s) { if (tid < s) {
tmp[tid] += tmp[tid + s]; tmp[tid] += tmp[tid + s];
} }

7
tests/test-backend-ops.cpp
View File

@ -2271,9 +2271,10 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
for (ggml_type type_a : other_types) { for (ggml_type type_a : other_types) {
for (ggml_type type_b : {GGML_TYPE_F32}) { for (ggml_type type_b : {GGML_TYPE_F32}) {
if (ggml_blck_size(type_a) != 256) {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), { 1, 1}, {1, 1})); test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), {1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1, 1}, {1, 1})); }
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1, 1}, {1, 1}));
} }
} }