root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 03:14:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

Merge 4ff0831ce6 into 9ba399dfa7

Browse Source
This commit is contained in:
Georgi Gerganov 2024-12-24 18:02:55 -06:00 committed by GitHub
commit 2ffcbb19ff
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 20 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/src/ggml-metal/ggml-metal.m
View File

@ -2317,6 +2317,7 @@ static void ggml_metal_encode_node(
[encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
//[encoder setThreadgroupMemoryLength:4096 + 2048 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;

23
ggml/src/ggml-metal/ggml-metal.metal
View File

@ -6062,7 +6062,7 @@ kernel void kernel_mul_mm(
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
threadgroup T * sa = (threadgroup T *)(shmem);
threadgroup float * sb = (threadgroup float *)(shmem + 4096);
threadgroup half * sb = (threadgroup half *)(shmem + 4096);
const int r0 = tgpig.y;
const int r1 = tgpig.x;
@ -6077,11 +6077,11 @@ kernel void kernel_mul_mm(
const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
simdgroup_T8x8 ma[4];
simdgroup_float8x8 mb[2];
simdgroup_float8x8 mc[8];
simdgroup_half8x8 mb[2];
simdgroup_half8x8 mc[8];
for (short i = 0; i < 8; i++){
mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
mc[i] = make_filled_simdgroup_matrix<half, 8>(0.h);
}
short il = (tiitg % THREAD_PER_ROW);
@ -6115,7 +6115,7 @@ kernel void kernel_mul_mm(
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
}
*(threadgroup float2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
*(threadgroup half2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (half2x4)(*((device float2x4 *)y));
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2 + nl - 1)/nl : x;
@ -6125,7 +6125,7 @@ kernel void kernel_mul_mm(
// load matrices from threadgroup memory and conduct outer products
threadgroup const T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
threadgroup const float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
threadgroup const half * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
#pragma unroll(4)
for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
@ -6157,7 +6157,11 @@ kernel void kernel_mul_mm(
(BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
for (short i = 0; i < 8; i++) {
simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
// cast to f32
simdgroup_float8x8 mc_f32(1.0f);
simdgroup_multiply(mc_f32, mc[i], mc_f32);
simdgroup_store(mc_f32, C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
//simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
}
} else {
// block is smaller than 64x32, we should avoid writing data outside of the matrix
@ -6165,7 +6169,10 @@ kernel void kernel_mul_mm(
threadgroup float * temp_str = ((threadgroup float *) shmem) \
+ 32*(sgitg&1) + (16*(sgitg >> 1))*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);
simdgroup_float8x8 mc_f32(1.0f);
simdgroup_multiply(mc_f32, mc[i], mc_f32);
simdgroup_store(mc_f32, temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
//simdgroup_store(mc[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);