happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge d058fc4373 into 58062860af

This commit is contained in:
Alfred 2025-12-17 12:25:40 +08:00 committed by GitHub
parent 58062860af d058fc4373
commit 6919ebc677
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-hexagon/htp/htp-ops.h
View File

@ -50,6 +50,8 @@ struct htp_ops_context {
struct fastdiv_values src1_div21; // fastdiv values for ne2 * ne1 struct fastdiv_values src1_div21; // fastdiv values for ne2 * ne1
uint32_t flags; uint32_t flags;
atomic_uint shared_atomic_lock;
}; };
int op_matmul(struct htp_ops_context * octx); int op_matmul(struct htp_ops_context * octx);

5
ggml/src/ggml-hexagon/htp/matmul-ops.c
View File

@ -1712,6 +1712,7 @@ static void quantize_fp32_q8x4x2(const struct htp_tensor * src,
static void htp_quantize_fp32_q8x4x2(unsigned int n, unsigned int i, void * data) { static void htp_quantize_fp32_q8x4x2(unsigned int n, unsigned int i, void * data) {
struct htp_ops_context * octx = data; struct htp_ops_context * octx = data;
quantize_fp32_q8x4x2(&octx->src1, octx->src1_spad.data, &octx->src0_spad, n, i, octx->src1_nrows_per_thread); quantize_fp32_q8x4x2(&octx->src1, octx->src1_spad.data, &octx->src0_spad, n, i, octx->src1_nrows_per_thread);
atomic_fetch_add(&octx->shared_atomic_lock, 1);
} }
// ** matmul callbacks for worker_pool // ** matmul callbacks for worker_pool
@ -2027,6 +2028,8 @@ int op_matmul(struct htp_ops_context * octx) {
octx->src0_nrows_per_thread = (src0_nrows + octx->n_threads - 1) / octx->n_threads; octx->src0_nrows_per_thread = (src0_nrows + octx->n_threads - 1) / octx->n_threads;
octx->src0_nrows_per_thread += (octx->src0_nrows_per_thread & 1); // round up to even octx->src0_nrows_per_thread += (octx->src0_nrows_per_thread & 1); // round up to even
atomic_store(&octx->shared_atomic_lock, 0);
if (need_quant) { if (need_quant) {
// Run quant jobs // Run quant jobs
const uint32_t n_quant_jobs = MIN(src1_nrows, octx->n_threads); const uint32_t n_quant_jobs = MIN(src1_nrows, octx->n_threads);
@ -2034,6 +2037,8 @@ int op_matmul(struct htp_ops_context * octx) {
worker_pool_run_func(octx->ctx->worker_pool, quant_job_func, octx, n_quant_jobs); worker_pool_run_func(octx->ctx->worker_pool, quant_job_func, octx, n_quant_jobs);
} }
FARF(HIGH, "matmul-%s : quant jobs finished! Atomic lock: %u\n", op_type, atomic_load(&octx->shared_atomic_lock));
if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) { if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
// Run matmul jobs // Run matmul jobs
const uint32_t n_matmul_jobs = octx->n_threads; const uint32_t n_matmul_jobs = octx->n_threads;