Migrations
/
whisper.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge 5567eaaa1c into 1a91c19af9

Browse Source
pull/387/merge
Abitofevrything 2 years ago committed by GitHub
parent 1a91c19af9 5567eaaa1c
commit 13204489d7
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 245 additions and 165 deletions
Show Stats Download Patch File Download Diff File

4
examples/talk.wasm/gpt-2.cpp
Unescape View File

@ -837,14 +837,14 @@ struct gpt2_context * gpt2_init(const char * path_model) {
// load the model
{
const int64_t t_start_us = ggml_time_us();
const int64_t t_start_us = ggml_real_time_us();
if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
return nullptr;
}
const int64_t t_load_us = ggml_time_us() - t_start_us;
const int64_t t_load_us = ggml_real_time_us() - t_start_us;
printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
}

4
examples/talk/gpt-2.cpp
Unescape View File

@ -837,7 +837,7 @@ struct gpt2_context * gpt2_init(const char * path_model) {
// load the model
{
const int64_t t_start_us = ggml_time_us();
const int64_t t_start_us = ggml_time_real_us();
if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, path_model);
@ -845,7 +845,7 @@ struct gpt2_context * gpt2_init(const char * path_model) {
return nullptr;
}
const int64_t t_load_us = ggml_time_us() - t_start_us;
const int64_t t_load_us = ggml_time_real_us() - t_start_us;
printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
}

60
extra/bench-all.sh
Unescape View File

@ -29,42 +29,60 @@ printf "Running benchmark for all models\n"
printf "This can take a while!\n"
printf "\n"
printf "| CPU | OS | Config | Model | Th | Load | Enc. | Commit |\n"
printf "| --- | -- | ------ | ----- | -- | ---- | ---- | ------ |\n"
cat >&2 << EOF
How to interpret these results:
- CPU is your CPU model
- OS is your current operating system
- Model is the GGML model being benchmarked
- Threads is the number of threads used
- Load is the time your computer took to load the model
- Encode is the time it took to run the Whisper encoder
- Time is reported as (real / process):
real: This is the wall-clock time in ms
process: This the CPU time. If you're using multiple threads, the time spent in each thread will be added together.
The proces time should be approximately (Proc Enc. / Threads).
If it isn't, you likely have another program making use of the CPU
- Commit is the current git commit.
EOF
printf "| CPU | OS | Config | Model | Th | Load | Encode | Commit |\n"
printf "| ------- | ------ | ---------------- | ------- | -- | ----- | ------------- | -------- |\n"
for model in "${models[@]}"; do
# run once to heat-up the cache
./bench -m ./models/ggml-$model.bin -t $n_threads 2>/dev/null 1>/dev/null
./bench -m "./models/ggml-$model.bin" -t "$n_threads" 2>/dev/null 1>/dev/null
# actual run
# store stderr output in a variable in order to parse it later
output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)
output=$(./bench -m "./models/ggml-$model.bin" -t "$n_threads" 2>&1)
# parse the output:
load_time=$(echo "$output" | grep "load time" | awk '{print $5}')
encode_time=$(echo "$output" | grep "encode time" | awk '{print $5}')
load_proc=$(echo "$output" | grep "load time" | awk '{print $8}')
load_real=$(echo "$output" | grep "load time" | awk '{print $5}')
encode_proc=$(echo "$output" | grep "encode time" | awk '{print $8}')
encode_real=$(echo "$output" | grep "encode time" | awk '{print $5}')
system_info=$(echo "$output" | grep "system_info")
n_threads=$(echo "$output" | grep "system_info" | awk '{print $4}')
# floor to milliseconds
load_time=${load_time%.*}
encode_time=${encode_time%.*}
load_proc=${load_proc%.*}
load_real=${load_real%.*}
encode_proc=${encode_proc%.*}
encode_real=${encode_real%.*}
config=""
load_str="$load_real"
encode_str="$encode_real / $encode_proc"
if [[ $system_info == *"AVX2 = 1"* ]]; then
config="$config AVX2"
fi
if [[ $system_info == *"NEON = 1"* ]]; then
config="$config NEON"
fi
if [[ $system_info == *"BLAS = 1"* ]]; then
config="$config BLAS"
fi
config=$(echo "$system_info" | sed 's/ | /\n/g' | tail -n +2 | awk '/ = 1/{print $1}' | tr '\n' ' ')
commit=$(git rev-parse --short HEAD)
printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
printf "| <todo> | <todo> | %-16s | %-7s | %-2s | %-5s | %-13s | %-8s |\n" \
"$config" \
"$model" \
"$n_threads" \
"$load_str" \
"$encode_str" \
"$commit"
done

167
ggml.c
Unescape View File

@ -282,54 +282,83 @@ ggml_fp16_t ggml_fp32_to_fp16(float x) {
#if defined(_MSC_VER) || defined(__MINGW32__)
static int64_t timer_freq;
static HANDLE current_process_handle;
void ggml_time_init(void) {
LARGE_INTEGER frequency;
QueryPerformanceFrequency(&frequency);
timer_freq = frequency.QuadPart;
current_process_handle = GetCurrentProcess();
}
int64_t ggml_time_ms(void) {
int64_t ggml_time_real_ms(void) {
LARGE_INTEGER t;
QueryPerformanceCounter(&t);
return (t.QuadPart * 1000) / timer_freq;
}
int64_t ggml_time_us(void) {
int64_t ggml_time_real_us(void) {
LARGE_INTEGER t;
QueryPerformanceCounter(&t);
return (t.QuadPart * 1000000) / timer_freq;
}
// Query only user time for process times as CLOCK_PROCESS_CPUTIME_ID does not include kernel time on Unix systems.
int64_t ggml_time_proc_ms(void) {
FILETIME user_time;
GetProcessTimes(current_process_handle, NULL, NULL, NULL, &user_time);
ULARGE_INTEGER t;
t.u.LowPart = user_time.dwLowDateTime;
t.u.HighPart = user_time.dwHighDateTime;
return t.QuadPart / 10000;
}
int64_t ggml_time_proc_us(void) {
FILETIME user_time;
GetProcessTimes(current_process_handle, NULL, NULL, NULL, &user_time);
ULARGE_INTEGER t;
t.u.LowPart = user_time.dwLowDateTime;
t.u.HighPart = user_time.dwHighDateTime;
return t.QuadPart / 10;
}
#else
void ggml_time_init(void) {}
int64_t ggml_time_ms(void) {
int64_t ggml_time_real_ms(void) {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return (int64_t)ts.tv_sec*1000 + (int64_t)ts.tv_nsec/1000000;
}
int64_t ggml_time_us(void) {
int64_t ggml_time_real_us(void) {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return (int64_t)ts.tv_sec*1000000 + (int64_t)ts.tv_nsec/1000;
}
#endif
int64_t ggml_cycles(void) {
return clock();
int64_t ggml_time_proc_ms(void) {
struct timespec ts;
clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
return (int64_t)ts.tv_sec*1000 + (int64_t)ts.tv_nsec/1000000;
}
int64_t ggml_cycles_per_ms(void) {
return CLOCKS_PER_SEC/1000;
int64_t ggml_time_proc_us(void) {
struct timespec ts;
clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
return (int64_t)ts.tv_sec*1000000 + (int64_t)ts.tv_nsec/1000;
}
#endif
#ifdef GGML_PERF
#define ggml_perf_time_ms() ggml_time_ms()
#define ggml_perf_time_us() ggml_time_us()
#define ggml_perf_cycles() ggml_cycles()
#define ggml_perf_cycles_per_ms() ggml_cycles_per_ms()
#define GGML_PERF_TIME_REAL_US() ggml_time_real_us()
#define GGML_PERF_TIME_PROC_US() ggml_time_proc_us()
#else
#define ggml_perf_time_ms() 0
#define ggml_perf_time_us() 0
#define ggml_perf_cycles() 0
#define ggml_perf_cycles_per_ms() 0
#define GGML_PERF_TIME_REAL_US() 0
#define GGML_PERF_TIME_PROC_US() 0
#endif
//
@ -1477,7 +1506,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
if (is_first_call) {
// initialize GELU, EXP and F32 tables
{
const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
const uint64_t t_start = GGML_PERF_TIME_REAL_US(); UNUSED(t_start);
ggml_fp16_t ii;
for (int i = 0; i < (1 << 16); ++i) {
@ -1488,14 +1517,14 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
table_exp_f16[i] = GGML_FP32_TO_FP16(exp(f));
}
const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
const uint64_t t_end = GGML_PERF_TIME_REAL_US(); UNUSED(t_end);
GGML_PRINT_DEBUG("%s: GELU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
}
// initialize g_state
{
const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
const uint64_t t_start = GGML_PERF_TIME_REAL_US(); UNUSED(t_start);
g_state = (struct ggml_state) {
/*.contexts =*/ { { 0 } },
@ -1505,7 +1534,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
g_state.contexts[i].used = false;
}
const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
const uint64_t t_end = GGML_PERF_TIME_REAL_US(); UNUSED(t_end);
GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
}
@ -1657,8 +1686,8 @@ struct ggml_tensor * ggml_new_tensor_impl(
/*.opt =*/ { NULL },
/*.n_tasks =*/ 0,
/*.perf_runs =*/ 0,
/*.perf_cycles =*/ 0,
/*.perf_time_us =*/ 0,
/*.perf_time_proc_us =*/ 0,
/*.perf_time_real_us =*/ 0,
/*.data =*/ data == NULL ? (void *)(result + 1) : data,
/*.pad =*/ { 0 },
};
@ -4307,7 +4336,7 @@ static void ggml_compute_forward_mul_mat_f32(
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
struct ggml_tensor * dst) {
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int ne00 = src0->ne[0];
@ -4403,7 +4432,7 @@ static void ggml_compute_forward_mul_mat_f32(
}
}
//printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
//printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (GGML_PERF_TIME_PROC_US() - t0)/1000.0, ne0, ne1, ne2, ne3);
return;
}
@ -4533,7 +4562,7 @@ static void ggml_compute_forward_mul_mat_f32(
}
}
//int64_t t1 = ggml_perf_time_us();
//int64_t t1 = GGML_PERF_TIME_PROC_US();
//static int64_t acc = 0;
//acc += t1 - t0;
//if (t1 - t0 > 10) {
@ -4552,7 +4581,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
struct ggml_tensor * dst) {
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int ne00 = src0->ne[0];
@ -4686,7 +4715,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
}
}
//printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
//printf("CBLAS = %f ms, %d x %d x %d x %d\n", (GGML_PERF_TIME_PROC_US() - t0)/1000.0, ne0, ne1, ne2, ne3);
return;
}
@ -4840,7 +4869,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
}
}
//int64_t t1 = ggml_time_us();
//int64_t t1 = GGML_PERF_TIME_REAL_US();
//static int64_t acc = 0;
//acc += t1 - t0;
//if (t1 - t0 > 10) {
@ -5308,7 +5337,7 @@ static void ggml_compute_forward_conv_1d_1s_f16_f32(
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int ne00 = src0->ne[0];
@ -5428,7 +5457,7 @@ static void ggml_compute_forward_conv_1d_1s_f32(
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int ne00 = src0->ne[0];
@ -5574,7 +5603,7 @@ static void ggml_compute_forward_conv_1d_2s_f16_f32(
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int ne00 = src0->ne[0];
@ -5694,7 +5723,7 @@ static void ggml_compute_forward_conv_1d_2s_f32(
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int ne00 = src0->ne[0];
@ -5838,7 +5867,7 @@ static void ggml_compute_forward_flash_attn_f32(
const struct ggml_tensor * v,
const bool masked,
struct ggml_tensor * dst) {
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int neq0 = q->ne[0];
@ -6047,7 +6076,7 @@ static void ggml_compute_forward_flash_attn_f16(
const struct ggml_tensor * v,
const bool masked,
struct ggml_tensor * dst) {
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int neq0 = q->ne[0];
@ -6322,7 +6351,7 @@ static void ggml_compute_forward_flash_ff_f16(
const struct ggml_tensor * c0, // F16 proj_w
const struct ggml_tensor * c1, // F32 proj_b
struct ggml_tensor * dst) {
int64_t t0 = ggml_perf_time_us();
int64_t t0 = GGML_PERF_TIME_PROC_US();
UNUSED(t0);
const int nea0 = a->ne[0];
@ -7001,8 +7030,8 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) {
/*.grads =*/ { NULL },
/*.leafs =*/ { NULL },
/*.perf_runs =*/ 0,
/*.perf_cycles =*/ 0,
/*.perf_time_us =*/ 0,
/*.perf_time_proc_us =*/ 0,
/*.perf_time_real_us =*/ 0,
};
ggml_build_forward_impl(&result, tensor, false);
@ -7411,8 +7440,8 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
}
}
const int64_t perf_start_cycles = ggml_perf_cycles();
const int64_t perf_start_time_us = ggml_perf_time_us();
const int64_t perf_time_proc_start_us = GGML_PERF_TIME_PROC_US();
const int64_t perf_time_real_start_us = GGML_PERF_TIME_REAL_US();
for (int i = 0; i < cgraph->n_nodes; i++) {
GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, i, cgraph->n_nodes);
@ -7424,8 +7453,8 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
// continue;
//}
const int64_t perf_node_start_cycles = ggml_perf_cycles();
const int64_t perf_node_start_time_us = ggml_perf_time_us();
const int64_t perf_node_time_proc_start_us = GGML_PERF_TIME_PROC_US();
const int64_t perf_node_time_real_start_us = GGML_PERF_TIME_REAL_US();
// INIT
struct ggml_compute_params params = {
@ -7550,12 +7579,12 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
// performance stats (node)
{
int64_t perf_cycles_cur = ggml_perf_cycles() - perf_node_start_cycles;
int64_t perf_time_us_cur = ggml_perf_time_us() - perf_node_start_time_us;
int64_t perf_cur_time_proc_us = GGML_PERF_TIME_PROC_US() - perf_node_time_proc_start_us;
int64_t perf_cur_time_real_us = GGML_PERF_TIME_REAL_US() - perf_node_time_real_start_us;
node->perf_runs++;
node->perf_cycles += perf_cycles_cur;
node->perf_time_us += perf_time_us_cur;
node->perf_time_proc_us += perf_cur_time_proc_us;
node->perf_time_real_us += perf_cur_time_real_us;
}
}
@ -7575,19 +7604,19 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
// performance stats (graph)
{
int64_t perf_cycles_cur = ggml_perf_cycles() - perf_start_cycles;
int64_t perf_time_us_cur = ggml_perf_time_us() - perf_start_time_us;
int64_t perf_cur_time_proc_us = GGML_PERF_TIME_PROC_US() - perf_time_proc_start_us;
int64_t perf_cur_time_real_us = GGML_PERF_TIME_REAL_US() - perf_time_real_start_us;
cgraph->perf_runs++;
cgraph->perf_cycles += perf_cycles_cur;
cgraph->perf_time_us += perf_time_us_cur;
cgraph->perf_time_proc_us += perf_cur_time_proc_us;
cgraph->perf_time_real_us += perf_cur_time_real_us;
GGML_PRINT_DEBUG("%s: perf (%d) - cpu = %.3f / %.3f ms, wall = %.3f / %.3f ms\n",
__func__, cgraph->perf_runs,
(double) perf_cycles_cur / (double) ggml_cycles_per_ms(),
(double) cgraph->perf_cycles / (double) ggml_cycles_per_ms() / (double) cgraph->perf_runs,
(double) perf_time_us_cur / 1000.0,
(double) cgraph->perf_time_us / 1000.0 / cgraph->perf_runs);
(double) perf_cur_time_proc_us / 1000.0,
(double) cgraph->perf_time_proc_us / 1000.0 / (double) cgraph->perf_runs,
(double) perf_cur_time_real_us / 1000.0,
(double) cgraph->perf_time_real_us / 1000.0 / (double) cgraph->perf_runs);
}
}
@ -7613,16 +7642,16 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i];
perf_total_per_op_us[node->op] += node->perf_time_us;
perf_total_per_op_us[node->op] += node->perf_time_real_us;
GGML_PRINT(" - %3d: [ %6d, %6d, %6d] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
i,
node->ne[0], node->ne[1], node->ne[2],
GGML_OP_LABEL[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
(double) node->perf_cycles / (double) ggml_cycles_per_ms(),
(double) node->perf_cycles / (double) ggml_cycles_per_ms() / (double) node->perf_runs,
(double) node->perf_time_us / 1000.0,
(double) node->perf_time_us / 1000.0 / node->perf_runs);
(double) node->perf_time_proc_us / 1000.0,
(double) node->perf_time_proc_us / 1000.0 / (double) node->perf_runs,
(double) node->perf_time_real_us / 1000.0,
(double) node->perf_time_real_us / 1000.0 / node->perf_runs);
}
GGML_PRINT("n_leafs = %d\n", cgraph->n_leafs);
@ -7901,10 +7930,10 @@ static enum ggml_opt_result ggml_opt_adam(
ggml_get_f32_1d(ps[i], 0), ggml_get_f32_1d(ps[i]->grad, 0));
}
const int64_t t_start_wall = ggml_time_us();
const int64_t t_start_cpu = ggml_cycles();
UNUSED(t_start_wall);
UNUSED(t_start_cpu);
const int64_t t_real_start_us = GGML_PERF_TIME_REAL_US();
const int64_t t_process_start_us = GGML_PERF_TIME_PROC_US();
UNUSED(t_real_start_us);
UNUSED(t_process_start_us);
{
// update the gradient
@ -7984,13 +8013,13 @@ static enum ggml_opt_result ggml_opt_adam(
fx_prev = fx;
{
const int64_t t_end_cpu = ggml_cycles();
GGML_PRINT_DEBUG("time iter: %5.3f s\n", ((float)(t_end_cpu - t_start_cpu))/CLOCKS_PER_SEC);
UNUSED(t_end_cpu);
const int64_t t_process_end_us = GGML_PERF_TIME_PROC_US();
GGML_PRINT_DEBUG("time iter: %5.3f s\n", (t_process_end_us - t_process_start_us)/1e6);
UNUSED(t_process_end_us);
const int64_t t_end_wall = ggml_time_us();
GGML_PRINT_DEBUG("wall time iter: %5.3f s\n", (t_end_wall - t_start_wall)/1e6);
UNUSED(t_end_wall);
const int64_t t_real_end_us = GGML_PERF_TIME_REAL_US();
GGML_PRINT_DEBUG("wall time iter: %5.3f s\n", (t_real_end_us - t_real_start_us)/1e6);
UNUSED(t_real_end_us);
}
}

16
ggml.h
Unescape View File

@ -275,8 +275,8 @@ struct ggml_tensor {
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
int64_t perf_time_proc_us;
int64_t perf_time_real_us;
void * data;
char padding[8];
@ -297,8 +297,8 @@ struct ggml_cgraph {
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
int64_t perf_time_proc_us;
int64_t perf_time_real_us;
};
struct ggml_init_params {
@ -308,10 +308,10 @@ struct ggml_init_params {
};
void ggml_time_init(void); // call this once at the beginning of the program
int64_t ggml_time_ms(void);
int64_t ggml_time_us(void);
int64_t ggml_cycles(void);
int64_t ggml_cycles_per_ms(void);
int64_t ggml_time_real_ms(void);
int64_t ggml_time_real_us(void);
int64_t ggml_time_proc_ms(void);
int64_t ggml_time_proc_us(void);
void ggml_print_object (const struct ggml_object * obj);
void ggml_print_objects(const struct ggml_context * ctx);

111
whisper.cpp
Unescape View File

@ -467,12 +467,18 @@ struct whisper_decoder {
};
struct whisper_context {
int64_t t_load_us = 0;
int64_t t_mel_us = 0;
int64_t t_sample_us = 0;
int64_t t_encode_us = 0;
int64_t t_decode_us = 0;
int64_t t_start_us = 0;
int64_t t_load_real_us = 0;
int64_t t_load_proc_us = 0;
int64_t t_mel_real_us = 0;
int64_t t_mel_proc_us = 0;
int64_t t_sample_real_us = 0;
int64_t t_sample_proc_us = 0;
int64_t t_encode_real_us = 0;
int64_t t_encode_proc_us = 0;
int64_t t_decode_real_us = 0;
int64_t t_decode_proc_us = 0;
int64_t t_start_real_us = 0;
int64_t t_start_proc_us = 0;
ggml_type wtype; // weight type (FP32 or FP16)
@ -597,9 +603,11 @@ static void kv_cache_free(struct whisper_kv_cache & cache) {
static bool whisper_model_load(struct whisper_model_loader * loader, whisper_context & wctx) {
fprintf(stderr, "%s: loading model\n", __func__);
const int64_t t_start_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
wctx.t_start_us = t_start_us;
wctx.t_start_real_us = t_start_real_us;
wctx.t_start_proc_us = t_start_proc_us;
auto & model = wctx.model;
auto & vocab = wctx.vocab;
@ -1208,7 +1216,8 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
wctx.rng = std::mt19937(0);
wctx.t_load_us = ggml_time_us() - t_start_us;
wctx.t_load_real_us = ggml_time_real_us() - t_start_real_us;
wctx.t_load_proc_us = ggml_time_proc_us() - t_start_proc_us;
return true;
}
@ -1226,7 +1235,8 @@ static bool whisper_encode(
whisper_context & wctx,
const int mel_offset,
const int n_threads) {
const int64_t t_start_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
const auto & model = wctx.model;
const auto & mel_inp = wctx.mel;
@ -1619,7 +1629,8 @@ static bool whisper_encode(
ggml_free(ctx0);
wctx.t_encode_us += ggml_time_us() - t_start_us;
wctx.t_encode_real_us += ggml_time_real_us() - t_start_real_us;
wctx.t_encode_proc_us += ggml_time_proc_us() - t_start_proc_us;
return true;
}
@ -1641,7 +1652,8 @@ static bool whisper_decode(
const int n_tokens,
const int n_past,
const int n_threads) {
const int64_t t_start_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
const auto & model = wctx.model;
const auto & hparams = model.hparams;
@ -1992,7 +2004,8 @@ static bool whisper_decode(
ggml_free(ctx0);
wctx.t_decode_us += ggml_time_us() - t_start_us;
wctx.t_decode_real_us += ggml_time_real_us() - t_start_real_us;
wctx.t_decode_proc_us += ggml_time_proc_us() - t_start_proc_us;
return true;
}
@ -2107,7 +2120,8 @@ static bool log_mel_spectrogram(
const whisper_filters & filters,
const bool speed_up,
whisper_mel & mel) {
const int64_t t_start_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
// Hanning window
std::vector<float> hann;
@ -2216,7 +2230,8 @@ static bool log_mel_spectrogram(
mel.data[i] = (mel.data[i] + 4.0)/4.0;
}
wctx.t_mel_us += ggml_time_us() - t_start_us;
wctx.t_mel_real_us = ggml_time_real_us() - t_start_real_us;
wctx.t_mel_proc_us = ggml_time_proc_us() - t_start_proc_us;
return true;
}
@ -2642,21 +2657,25 @@ whisper_token whisper_token_transcribe(void) {
}
void whisper_print_timings(struct whisper_context * ctx) {
const int64_t t_end_us = ggml_time_us();
const int64_t t_real_end_us = ggml_time_real_us();
const int64_t t_proc_end_us = ggml_time_proc_us();
fprintf(stderr, "\n");
fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, ctx->t_load_us/1000.0f);
fprintf(stderr, "%s: mel time = %8.2f ms\n", __func__, ctx->t_mel_us/1000.0f);
fprintf(stderr, "%s: sample time = %8.2f ms\n", __func__, ctx->t_sample_us/1000.0f);
fprintf(stderr, "%s: encode time = %8.2f ms / %.2f ms per layer\n", __func__, ctx->t_encode_us/1000.0f, ctx->t_encode_us/1000.0f/ctx->model.hparams.n_audio_layer);
fprintf(stderr, "%s: decode time = %8.2f ms / %.2f ms per layer\n", __func__, ctx->t_decode_us/1000.0f, ctx->t_decode_us/1000.0f/ctx->model.hparams.n_text_layer);
fprintf(stderr, "%s: total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f);
fprintf(stderr, "%s: load time = %8.2f ms (proc %8.2f ms)\n", __func__, ctx->t_load_real_us/1000.0f, ctx->t_load_proc_us/1000.0f);
fprintf(stderr, "%s: mel time = %8.2f ms (proc %8.2f ms)\n", __func__, ctx->t_mel_real_us/1000.0f, ctx->t_mel_proc_us/1000.0f);
fprintf(stderr, "%s: sample time = %8.2f ms (proc %8.2f ms)\n", __func__, ctx->t_sample_real_us/1000.0f, ctx->t_sample_proc_us/1000.0f);
fprintf(stderr, "%s: encode time = %8.2f ms (proc %8.2f ms)\n", __func__, ctx->t_encode_real_us/1000.0f, ctx->t_encode_proc_us/1000.0f);
fprintf(stderr, "%s: decode time = %8.2f ms (proc %8.2f ms)\n", __func__, ctx->t_decode_real_us/1000.0f, ctx->t_decode_proc_us/1000.0f);
fprintf(stderr, "%s: total time = %8.2f ms (proc %8.2f ms)\n", __func__, (t_real_end_us - ctx->t_start_real_us)/1000.f, (t_proc_end_us - ctx->t_start_proc_us)/1000.0f);
}
void whisper_reset_timings(struct whisper_context * ctx) {
ctx->t_sample_us = 0;
ctx->t_encode_us = 0;
ctx->t_decode_us = 0;
ctx->t_sample_real_us = 0;
ctx->t_sample_proc_us = 0;
ctx->t_encode_real_us = 0;
ctx->t_encode_proc_us = 0;
ctx->t_decode_real_us = 0;
ctx->t_decode_proc_us = 0;
}
const char * whisper_print_system_info(void) {
@ -3455,7 +3474,8 @@ int whisper_full(
}
{
const int64_t t_start_sample_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
whisper_process_logits(*ctx, params, ctx->decoders[0], t_cur);
@ -3474,12 +3494,14 @@ int whisper_full(
memcpy(decoder.logprobs.data(), ctx->decoders[0].logprobs.data(), decoder.logprobs.size()*sizeof(decoder.logprobs[0]));
}
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
ctx->t_sample_real_us += ggml_time_real_us() - t_start_real_us;
ctx->t_sample_proc_us += ggml_time_proc_us() - t_start_proc_us;
}
}
for (int i = 0, n_max = whisper_n_text_ctx(ctx)/2 - 4; i < n_max; ++i) {
const int64_t t_start_sample_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
// store the KV caches of all decoders when doing beam-search
if (params.strategy == whisper_sampling_strategy::WHISPER_SAMPLING_BEAM_SEARCH) {
@ -3672,7 +3694,8 @@ int whisper_full(
}
}
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
ctx->t_sample_real_us += ggml_time_real_us() - t_start_real_us;
ctx->t_sample_proc_us += ggml_time_proc_us() - t_start_proc_us;
// obtain logits for the next token
for (int j = 0; j < n_decoders_cur; ++j) {
@ -3693,13 +3716,15 @@ int whisper_full(
}
{
const int64_t t_start_sample_us = ggml_time_us();
const int64_t t_start_real_us = ggml_time_real_us();
const int64_t t_start_proc_us = ggml_time_proc_us();
whisper_process_logits(*ctx, params, decoder, t_cur);
++decoder.kv_self.n;
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
ctx->t_sample_real_us += ggml_time_real_us() - t_start_real_us;
ctx->t_sample_proc_us += ggml_time_proc_us() - t_start_proc_us;
}
}
}
@ -3993,10 +4018,14 @@ int whisper_full_parallel(
}
}
ctx->t_mel_us += ctxs[i].t_mel_us;
ctx->t_sample_us += ctxs[i].t_sample_us;
ctx->t_encode_us += ctxs[i].t_encode_us;
ctx->t_decode_us += ctxs[i].t_decode_us;
ctx->t_mel_real_us += ctxs[i].t_mel_real_us;
ctx->t_mel_proc_us += ctxs[i].t_mel_proc_us;
ctx->t_sample_real_us += ctxs[i].t_sample_real_us;
ctx->t_sample_proc_us += ctxs[i].t_sample_proc_us;
ctx->t_encode_real_us += ctxs[i].t_encode_real_us;
ctx->t_encode_proc_us += ctxs[i].t_encode_proc_us;
ctx->t_decode_real_us += ctxs[i].t_decode_real_us;
ctx->t_decode_proc_us += ctxs[i].t_decode_proc_us;
kv_cache_free(ctx->kv_cross);
@ -4006,10 +4035,14 @@ int whisper_full_parallel(
}
// average the timings
ctx->t_mel_us /= n_processors;
ctx->t_sample_us /= n_processors;
ctx->t_encode_us /= n_processors;
ctx->t_decode_us /= n_processors;
ctx->t_mel_real_us /= n_processors;
ctx->t_mel_proc_us /= n_processors;
ctx->t_sample_real_us /= n_processors;
ctx->t_sample_proc_us /= n_processors;
ctx->t_encode_real_us /= n_processors;
ctx->t_encode_proc_us /= n_processors;
ctx->t_decode_real_us /= n_processors;
ctx->t_decode_proc_us /= n_processors;
// print information about the audio boundaries
fprintf(stderr, "\n");

Write Preview
Loading…
Cancel
Save