root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-13 12:10:18 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

metal : fix soft_max kernels

Browse Source 
ref: 1914017863
This commit is contained in:
Georgi Gerganov 2023-12-13 10:54:17 +02:00
parent 82e4f64578
commit ab558ac2b3
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 13 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml-metal.m
View File

@ -1269,6 +1269,8 @@ void ggml_metal_graph_compute(
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
if (id_src1) { if (id_src1) {
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
} }
[encoder setBuffer:id_dst offset:offs_dst atIndex:2]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
@ -1520,7 +1522,7 @@ void ggml_metal_graph_compute(
else if (src0t == GGML_TYPE_Q6_K) { else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else { } else {
int64_t ny = (ne11 + nrows - 1)/nrows; const int64_t ny = (ne11 + nrows - 1)/nrows;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} }
} }

16
ggml-metal.metal
View File

@ -347,9 +347,9 @@ kernel void kernel_soft_max(
const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
device const float * pmask = src1 ? src1 + i01*ne00 : nullptr; device const float * pmask = src1 != src0 ? src1 + i01*ne00 : nullptr;
device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
// parallel max // parallel max
float lmax = -INFINITY; float lmax = -INFINITY;
@ -386,6 +386,8 @@ kernel void kernel_soft_max(
} }
float sum = simd_sum(lsum); float sum = simd_sum(lsum);
threadgroup_barrier(mem_flags::mem_threadgroup);
if (ntg > N_SIMDWIDTH) { if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) { if (sgitg == 0) {
buf[tiisg] = 0.0f; buf[tiisg] = 0.0f;
@ -428,9 +430,9 @@ kernel void kernel_soft_max_4(
const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
device const float4 * pmask = src1 ? (device const float4 *)(src1 + i01*ne00) : nullptr; device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 + i01*ne00) : nullptr;
device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
// parallel max // parallel max
float4 lmax4 = -INFINITY; float4 lmax4 = -INFINITY;
@ -468,6 +470,8 @@ kernel void kernel_soft_max_4(
} }
const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3]; const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
threadgroup_barrier(mem_flags::mem_threadgroup);
float sum = simd_sum(lsum); float sum = simd_sum(lsum);
if (ntg > N_SIMDWIDTH) { if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) { if (sgitg == 0) {