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

Merge 0b0003b491 into 2973a65ecb

This commit is contained in:
Ethan 2025-12-16 22:20:44 -05:00 committed by GitHub
parent 2973a65ecb 0b0003b491
commit 0bf1be5052
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 88 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-hexagon/CMakeLists.txt
View File

@ -11,6 +11,7 @@ target_include_directories(htp_iface PUBLIC
${HEXAGON_SDK_ROOT}/incs
${HEXAGON_SDK_ROOT}/incs/stddef
${HEXAGON_SDK_ROOT}/utils/examples
${HEXAGON_SDK_ROOT}/libs/itrace/inc
${CMAKE_CURRENT_SOURCE_DIR}/htp
${CMAKE_CURRENT_BINARY_DIR})
@ -32,6 +33,7 @@ ggml_add_backend_library(${TARGET_NAME}
ggml-hexagon.cpp htp-utils.c htp-utils.h ../../include/ggml-hexagon.h)
target_link_libraries(${TARGET_NAME} PRIVATE htp_iface)
target_link_libraries(${TARGET_NAME} PRIVATE ${HEXAGON_SDK_ROOT}/libs/itrace/prebuilt/android_aarch64/libitrace.so)
target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/htp ${CMAKE_CURRENT_BINARY_DIR})
# Build HTP bits

82
ggml/src/ggml-hexagon/ggml-hexagon.cpp
View File

@ -11,6 +11,9 @@
#include <string>
#include <stdexcept>
#include "itrace.h"
#include "itrace_types.h"
#include "itrace_cpu_events.h"
#ifdef _WIN32
# include <sal.h>
# ifndef _WINDOWS
@ -45,6 +48,9 @@ static int opt_arch = 0; // autodetect
static int opt_etm = 0;
static int opt_verbose = 0;
static int opt_profile = 0;
static int opt_trace = 0;
static itrace_logger_handle_t g_itrace_logger_handle = NULL;
static itrace_profiler_handle_t g_itrace_cpu_profiler_handle = NULL;
static int opt_hostbuf = 1;
static int opt_experimental = 0;
@ -282,6 +288,9 @@ void ggml_hexagon_session::enqueue(struct htp_general_req &req, struct dspqueue_
// Flush HTP response queue i.e wait for all outstanding requests to complete
void ggml_hexagon_session::flush() {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, "session-flush", NULL);
}
dspqueue_t q = this->queue;
// Repeatedly read packets from the queue until it's empty. We don't
@ -332,6 +341,10 @@ void ggml_hexagon_session::flush() {
this->op_pending--; // atomic dec
}
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
itrace_flush_logs(g_itrace_logger_handle);
}
}
// ** backend buffers
@ -694,6 +707,9 @@ static void init_row_q4x4x2(block_q4_0 * x, int64_t k) {
// repack q4_0 data into q4x4x2 tensor
static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, (std::string("repack-q4_0-q4x4x2-") + t->name).c_str(), NULL);
}
int64_t nrows = ggml_nrows(t);
size_t row_size = ggml_row_size(t->type, t->ne[0]);
@ -751,6 +767,9 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
ggml_aligned_free(buf_pd, row_size_pd);
ggml_aligned_free(buf_rp, row_size_rp);
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
// repack q4x4x2 tensor into q4_0 data
@ -2320,6 +2339,9 @@ static void hex_dump_dspbuf(const struct ggml_tensor * t, const dspqueue_buffer
}
static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, (std::string("mul-mat-") + op->name).c_str(), NULL);
}
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const struct ggml_tensor * dst = op;
@ -2387,9 +2409,16 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
(uint32_t) src1->ne[2], (uint32_t) src1->ne[3], dst->name, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1],
(uint32_t) dst->ne[2], (uint32_t) dst->ne[3], sess->prof_usecs, sess->prof_cycles, sess->prof_pkts,
(float) sess->prof_cycles / sess->prof_pkts, (unsigned long long) t2 - t1);
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flags) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, (std::string("mul-mat-id-") + op->name).c_str(), NULL);
}
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const struct ggml_tensor * src2 = op->src[2];
@ -2466,9 +2495,15 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
(uint32_t) src2->ne[3], dst->name, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],
(uint32_t) dst->ne[3], sess->prof_usecs, sess->prof_cycles, sess->prof_pkts,
(float) sess->prof_cycles / sess->prof_pkts, (unsigned long long) t2 - t1);
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, "binary", NULL);
}
const struct ggml_tensor * node = op;
const struct ggml_tensor * src0 = node->src[0];
const struct ggml_tensor * src1 = node->src[1];
@ -2555,9 +2590,15 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
(uint32_t) src1->ne[2], (uint32_t) src1->ne[3], dst->name, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1],
(uint32_t) dst->ne[2], (uint32_t) dst->ne[3], sess->prof_usecs, sess->prof_cycles, sess->prof_pkts,
(float) sess->prof_cycles / sess->prof_pkts, (unsigned long long) t2 - t1);
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, (std::string("add-id-") + op->name).c_str(), NULL);
}
const struct ggml_tensor * node = op;
const struct ggml_tensor * src0 = node->src[0];
const struct ggml_tensor * src1 = node->src[1];
@ -2630,9 +2671,15 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
(uint32_t) src1->ne[2], (uint32_t) src1->ne[3], dst->name, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1],
(uint32_t) dst->ne[2], (uint32_t) dst->ne[3], sess->prof_usecs, sess->prof_cycles, sess->prof_pkts,
(float) sess->prof_cycles / sess->prof_pkts, (unsigned long long) t2 - t1);
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, "unary", NULL);
}
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const struct ggml_tensor * dst = op;
@ -2763,9 +2810,15 @@ static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
(uint32_t) dst->ne[2], (uint32_t) dst->ne[3], sess->prof_usecs, sess->prof_cycles, sess->prof_pkts,
(float) sess->prof_cycles / sess->prof_pkts, (unsigned long long) t2 - t1);
}
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, "rope", NULL);
}
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const struct ggml_tensor * src2 = op->src[2];
@ -2871,6 +2924,9 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
(uint32_t) dst->ne[2], (uint32_t) dst->ne[3], sess->prof_usecs, sess->prof_cycles, sess->prof_pkts,
(float) sess->prof_cycles / sess->prof_pkts, (unsigned long long) t2 - t1);
}
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
}
static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
@ -2879,6 +2935,13 @@ static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
}
static void ggml_backend_hexagon_free(ggml_backend_t backend) {
// Flush and close itrace logger if profiling was enabled
if (opt_trace) {
itrace_flush_logs(g_itrace_logger_handle);
itrace_close_logger(g_itrace_logger_handle);
HEX_VERBOSE("ggml-hex: close itrace\n");
}
// we just need to delete the backend here
// the sessions are allocated & freed as part of the registry
delete backend;
@ -2908,6 +2971,10 @@ static inline int last_compute_op(ggml_cgraph * graph) {
static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
auto sess = static_cast<ggml_hexagon_session *>(backend->context);
if (opt_trace) {
itrace_start_section(g_itrace_cpu_profiler_handle, (std::string("hexagon-graph-compute-") + sess->name).c_str(), NULL);
}
HEX_VERBOSE("ggml-hex: %s graph-compute n_nodes %d\n", sess->name.c_str(), graph->n_nodes);
const int last = last_compute_op(graph);
@ -2980,6 +3047,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
// Wait until all pending ops complete
sess->flush();
if (opt_trace) {
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
return GGML_STATUS_SUCCESS;
}
@ -3473,6 +3544,17 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
opt_profile = getenv("GGML_HEXAGON_PROFILE") != nullptr;
opt_etm = getenv("GGML_HEXAGON_ETM") != nullptr;
opt_experimental = getenv("GGML_HEXAGON_EXPERIMENTAL") != nullptr;
opt_trace = getenv("GGML_HEXAGON_TRACE") != nullptr;
// Initialize itrace if profiling is enabled
if (opt_trace) {
HEX_VERBOSE("ggml-hex: open itrace\n");
itrace_open_logger(CPU_DOMAIN_ID, &g_itrace_logger_handle);
itrace_open_profiler(g_itrace_logger_handle, CPU_DOMAIN_ID, 4*1024*1024, &g_itrace_cpu_profiler_handle);
itrace_start_section(g_itrace_cpu_profiler_handle, "open-itrace", NULL);
itrace_end_section(g_itrace_cpu_profiler_handle, NULL);
}
const char * str_opmask = getenv("GGML_HEXAGON_OPMASK");
if (str_opmask != nullptr) {

5
scripts/snapdragon/adb/run-cli.sh
View File

@ -30,6 +30,9 @@ sched=
profile=
[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF GGML_HEXAGON_OPSYNC=1"
trace=
[ "$TRACE" != "" ] && trace="GGML_HEXAGON_TRACE=$TRACE"
opmask=
[ "$OPMASK" != "" ] && opmask="GGML_HEXAGON_OPMASK=$OPMASK"
@ -45,7 +48,7 @@ adb $adbserial shell " \
cd $basedir; ulimit -c unlimited; \
LD_LIBRARY_PATH=$basedir/$branch/lib \
ADSP_LIBRARY_PATH=$basedir/$branch/lib \
$verbose $experimental $sched $opmask $profile $nhvx $ndev \
$verbose $experimental $sched $opmask $profile $nhvx $ndev $trace \
./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model \
--poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 \
--ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on \