root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 19:04:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

metal : reorder write loop in mul mat kernel + style (#10231)
Some checks failed
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full-cuda.Dockerfile platforms:linux/amd64 tag:full-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full-musa.Dockerfile platforms:linux/amd64 tag:full-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full.Dockerfile platforms:linux/amd64,linux/arm64 tag:full]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-cuda.Dockerfile platforms:linux/amd64 tag:light-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-intel.Dockerfile platforms:linux/amd64 tag:light-intel]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-musa.Dockerfile platforms:linux/amd64 tag:light-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli.Dockerfile platforms:linux/amd64,linux/arm64 tag:light]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-cuda.Dockerfile platforms:linux/amd64 tag:server-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-intel.Dockerfile platforms:linux/amd64 tag:server-intel]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-musa.Dockerfile platforms:linux/amd64 tag:server-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server.Dockerfile platforms:linux/amd64,linux/arm64 tag:server]) (push) Waiting to run
Details
Nix CI / nix-eval (macos-latest) (push) Waiting to run
Details
Nix CI / nix-eval (ubuntu-latest) (push) Waiting to run
Details
Nix CI / nix-build (macos-latest) (push) Waiting to run
Details
Nix CI / nix-build (ubuntu-latest) (push) Waiting to run
Details
flake8 Lint / Lint (push) Waiting to run
Details
update-flake-lock / lockfile (push) Has been cancelled
Details

Browse Source 
* metal : reorder write loop

* metal : int -> short, style

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-11-09 11:53:13 +02:00 committed by GitHub
parent 39a334a9aa
commit 6423c65aa8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 44 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

44
ggml/src/ggml-metal.metal
View File

@ -6327,9 +6327,10 @@ kernel void kernel_mul_mm(device const uchar * src0,
simdgroup_T8x8 ma[4];
simdgroup_float8x8 mb[2];
simdgroup_float8x8 c_res[8];
for (int i = 0; i < 8; i++){
c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
simdgroup_float8x8 mc[8];
for (short i = 0; i < 8; i++){
mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
}
short il = (tiitg % THREAD_PER_ROW);
@ -6354,7 +6355,7 @@ kernel void kernel_mul_mm(device const uchar * src0,
threadgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll(16)
for (int i = 0; i < 16; i++) {
for (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
@ -6373,14 +6374,14 @@ kernel void kernel_mul_mm(device const uchar * src0,
threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
#pragma unroll(4)
for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
#pragma unroll(4)
for (int i = 0; i < 4; i++) {
for (short i = 0; i < 4; i++) {
simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
}
simdgroup_barrier(mem_flags::mem_none);
#pragma unroll(2)
for (int i = 0; i < 2; i++) {
for (short i = 0; i < 2; i++) {
simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
}
@ -6388,8 +6389,8 @@ kernel void kernel_mul_mm(device const uchar * src0,
lsmb += BLOCK_SIZE_N/SG_MAT_ROW * SG_MAT_SIZE;
#pragma unroll(8)
for (int i = 0; i < 8; i++){
simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
for (short i = 0; i < 8; i++){
simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
}
}
}
@ -6397,25 +6398,36 @@ kernel void kernel_mul_mm(device const uchar * src0,
if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg & 1)) \
+ (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
for (short i = 0; i < 8; i++) {
simdgroup_store(mc[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
}
} else {
// block is smaller than 64x32, we should avoid writing data outside of the matrix
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup float * temp_str = ((threadgroup float *) shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
for (short i = 0; i < 8; i++) {
simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
if (sgitg == 0) {
for (int i = 0; i < n_rows; i++) {
for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
*(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
device float * D = dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*ne0 + im*ne1*ne0;
device float4 * D4 = (device float4 *) D;
threadgroup float * C = temp_str + (j*BLOCK_SIZE_M);
threadgroup float4 * C4 = (threadgroup float4 *) C;
int i = 0;
for (; i < n_rows/4; i++) {
*(D4 + i) = *(C4 + i);
}
i *= 4;
for (; i < n_rows; i++) {
*(D + i) = *(C + i);
}
}
}