From 42c76d1358021ccdbe8ba89109c143dd7ae166df Mon Sep 17 00:00:00 2001 From: Faisal Zaghloul Date: Thu, 29 Aug 2024 19:20:53 -0400 Subject: [PATCH] Threadpool: take 2 (#8672) * Introduce ggml_compute_threadpool - OpenMP functional: check - Vanilla ggml functional: Check - ggml w/threadpool functional: Check - OpenMP no regression: No glaring problems - Vanilla ggml no regression: No glaring problems - ggml w/threadpool no regression: No glaring problems * Minor fixes * fixed use after release bug * fixed a harmless race condition * Fix Android bulid issue * fix more race conditions * fix deadlock for cases where cgraph.n_nodes == 1 and fix --poll case * threadpool: use cpu_get_num_math to set the default number of threadpool threads This way we avoid using E-Cores and Hyperthreaded siblings. * bench: create fresh threadpool for each test For benchmarking it's better to start a fresh pool for each test with the exact number of threads needed for that test. Having larger pools is suboptimal (causes more load, etc). * atomics: always use stdatomics with clang and use relaxed memory order when polling in ggml_barrier This also removes sched_yield() calls from ggml_barrier() to match OpenMP behavior. * threadpool: make polling the default to match openmp behavior All command line args now allow for setting poll to 0 (false). * threadpool: do not wakeup threads in already paused threadpool * fix potential race condition in check_for_work * threadpool: do not create two threadpools if their params are identical * threadpool: reduce pause/resume/wakeup overhead in common cases We now start threadpool in paused state only if we have two. The resume is now implicit (ie new work) which allows for reduced locking and context-switch overhead. * threadpool: add support for hybrid polling poll params (--poll, ...) now specify "polling level", i.e. how aggresively we poll before waiting on cond.var. poll=0 means no polling, 1 means poll for 128K rounds then wait, 2 for 256K rounds, ... The default value of 50 (ie 50x128K rounds) seems like a decent default across modern platforms. We can tune this further as things evolve. * threadpool: reduce the number of barrier required New work is now indicated with an atomic counter that is incremented for each new graph that needs to be computed. This removes the need for extra barrier for clearing the "new_work" and removes the special case for trivial graphs. * threadpool: remove special-casing for disposable threadpools With the efficient hybrid polling there is no need to make disposable pools any different. This simplifies the overall logic and reduces branching. Include n_threads in debug print for disposable threadpool. Declare pause and stop flags as atomic_bool This doesn't actually generate any memory barriers and simply informs the thread sanitizer that these flags can be written & read by different threads without locking. * threadpool: do not clear barrier counters between graphs computes (fixes race with small graphs) This fixes the race condition with very small graphs where the main thread happens to start a new graph while the workers are just about to exit from barriers. * threadpool: use relaxed order for chunk sync Full memory barrier is an overkill for this since each thread works on different chunk * threadpool: remove abort_callback from threadpool state * threadpool: better naming for thread/cpumask releated functions * threadpool: consistent use of int type for n_threads params * threadpool: add support for ggml_threadpool_params_default/init Also removes the need for explicit mask_specified param. all-zero cpumask means use default (usually inherited) cpu affinity mask. * threadpool: move typedef into ggml.h * threadpool: fix apply_priority() function name * threadpool: fix swift wrapper errors due to n_threads int type cleanup * threadpool: enable --cpu-mask and other threadpool related options only if threadpool is enabled * threadpool: replace checks for compute_thread ret code with proper status check * threadpool: simplify threadpool init logic and fix main thread affinity application Most of the init code is now exactly the same between threadpool and openmp. * threadpool: update threadpool resume/pause function names * threadpool: enable openmp by default for now * threadpool: don't forget to free workers state when omp is enabled * threadpool: avoid updating process priority on the platforms that do not require it On Windows we need to change overall process priority class in order to set thread priorities, but on Linux, Mac, etc we do not need to touch the overall process settings. * threadpool: update calling thread prio and affinity only at start/resume This avoids extra syscalls for each graph_compute() * llama-bench: turn threadpool params into vectors, add output headers, etc * llama-bench: add support for cool off between tests --delay This helps for long running tests on platforms that are thermally limited (phones, laptops, etc). --delay (disabled by default) introduces the sleep for N seconds before starting each test. * threadpool: move process priority setting into the apps (bench and cli) This avoids changing the overall process priority on Windows for the apps that use ggml/llama.cpp directy. * threadpool: move all pause/resume logic into ggml * threadpool: futher api cleanup and prep for future refactoring All threadpool related functions and structs use ggml_threadpool prefix. * threadpool: minor indent fixes * threadpool: improve setprioty error message * Update examples/llama-bench/llama-bench.cpp Co-authored-by: slaren * threadpool: fix indent in set_threadpool call * use int32_t for n_thread type in public llama.cpp API * threadpool: use _new and _free instead of _create and _release * fix two more public APIs to use int32_t for n_threads * build: set _GNU_SOURCE for Adroid --------- Co-authored-by: Max Krasnyansky Co-authored-by: fmz Co-authored-by: Max Krasnyansky Co-authored-by: slaren --- common/common.cpp | 350 +++++++- common/common.h | 30 +- examples/baby-llama/baby-llama.cpp | 2 +- examples/benchmark/benchmark-matmult.cpp | 4 +- .../cvector-generator/cvector-generator.cpp | 4 +- examples/export-lora/export-lora.cpp | 2 +- examples/llama-bench/llama-bench.cpp | 125 ++- .../llama.cpp.swift/LibLlama.swift | 4 +- examples/llava/llava-cli.cpp | 4 +- examples/llava/minicpmv-cli.cpp | 2 +- examples/main/main.cpp | 37 + examples/server/server.cpp | 4 +- examples/speculative/speculative.cpp | 7 +- ggml/include/ggml-alloc.h | 4 +- ggml/include/ggml-backend.h | 1 + ggml/include/ggml.h | 43 +- ggml/src/CMakeLists.txt | 2 +- ggml/src/ggml-backend.c | 25 +- ggml/src/ggml.c | 850 ++++++++++++++---- include/llama.h | 17 +- src/llama.cpp | 48 +- tests/test-rope.cpp | 2 +- 22 files changed, 1310 insertions(+), 257 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 715adf946..9fa184725 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -251,6 +251,57 @@ int32_t cpu_get_num_math() { return cpu_get_num_physical_cores(); } +// Helper for setting process priority + +#if defined(_WIN32) + +bool set_process_priority(enum ggml_sched_priority prio) { + if (prio == GGML_SCHED_PRIO_NORMAL) { + return true; + } + + DWORD p = NORMAL_PRIORITY_CLASS; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = NORMAL_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_MEDIUM: p = ABOVE_NORMAL_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_HIGH: p = HIGH_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS; break; + } + + if (!SetPriorityClass(GetCurrentProcess(), p)) { + fprintf(stderr, "warn: failed to set process priority class %d : (%d)\n", prio, (int) GetLastError()); + return false; + } + + return true; +} + +#else // MacOS and POSIX +#include +#include + +bool set_process_priority(enum ggml_sched_priority prio) { + if (prio == GGML_SCHED_PRIO_NORMAL) { + return true; + } + + int p = 0; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = 0; break; + case GGML_SCHED_PRIO_MEDIUM: p = -5; break; + case GGML_SCHED_PRIO_HIGH: p = -10; break; + case GGML_SCHED_PRIO_REALTIME: p = -20; break; + } + + if (!setpriority(PRIO_PROCESS, 0, p)) { + fprintf(stderr, "warn: failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno); + return false; + } + return true; +} + +#endif + // // CLI argument parsing // @@ -277,6 +328,30 @@ void gpt_params_handle_model_default(gpt_params & params) { } } +void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) { + int32_t n_set = 0; + + if (cpuparams.n_threads < 0) { + // Assuming everything about cpuparams is invalid + if (role_model != nullptr) { + cpuparams = *role_model; + } else { + cpuparams.n_threads = cpu_get_num_math(); + } + } + + for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + if (cpuparams.cpumask[i]) { + n_set++; + } + } + + if (n_set && n_set < cpuparams.n_threads) { + // Not enough set bits, may experience performance issues. + fprintf(stderr, "warn: Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads); + } +} + bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; std::string arg; @@ -296,6 +371,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { } } + postprocess_cpu_params(params.cpuparams, nullptr); + postprocess_cpu_params(params.cpuparams_batch, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams_batch, ¶ms.cpuparams_batch); + if (params.prompt_cache_all && (params.interactive || params.interactive_first)) { throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } @@ -331,7 +411,7 @@ void gpt_params_parse_from_env(gpt_params & params) { get_env("LLAMA_ARG_MODEL_ALIAS", params.model_alias); get_env("LLAMA_ARG_HF_REPO", params.hf_repo); get_env("LLAMA_ARG_HF_FILE", params.hf_file); - get_env("LLAMA_ARG_THREADS", params.n_threads); + get_env("LLAMA_ARG_THREADS", params.cpuparams.n_threads); get_env("LLAMA_ARG_CTX_SIZE", params.n_ctx); get_env("LLAMA_ARG_N_PARALLEL", params.n_parallel); get_env("LLAMA_ARG_BATCH", params.n_batch); @@ -368,6 +448,79 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { return true; } +bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) { + size_t dash_loc = range.find('-'); + if (dash_loc == std::string::npos) { + fprintf(stderr, "Format of CPU range is invalid! Expected []-[].\n"); + return false; + } + + size_t start_i; + size_t end_i; + + if (dash_loc == 0) { + start_i = 0; + } else { + start_i = std::stoull(range.substr(0, dash_loc)); + if (start_i >= GGML_MAX_N_THREADS) { + fprintf(stderr, "Start index out of bounds!\n"); + return false; + } + } + + if (dash_loc == range.length() - 1) { + end_i = GGML_MAX_N_THREADS - 1; + } else { + end_i = std::stoull(range.substr(dash_loc + 1)); + if (end_i >= GGML_MAX_N_THREADS) { + fprintf(stderr, "End index out of bounds!\n"); + return false; + } + } + + for (size_t i = start_i; i <= end_i; i++) { + boolmask[i] = true; + } + + return true; +} + +bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) { + // Discard potential 0x prefix + size_t start_i = 0; + if (mask.length() >= 2 && mask.substr(0, 2) == "0x") { + start_i = 2; + } + + size_t num_digits = mask.length() - start_i; + if (num_digits > 128) num_digits = 128; + + size_t end_i = num_digits + start_i; + + for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) { + char c = mask.at(i); + int8_t id = c; + + if ((c >= '0' && c <= '9')) { + id -= '0'; + } else if (c >= 'a' && c <= 'f') { + id -= 'a' - 10; + } else if (c >= 'A' && c <= 'F') { + id -= 'A' - 10; + } else { + fprintf(stderr, "Invalid hex character '%c' at position %d\n", c, int32_t(i)); + return false; + } + + boolmask[ n ] = boolmask[ n ] || ((id & 8) != 0); + boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0); + boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0); + boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0); + } + + return true; +} + #define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; } bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) { @@ -384,36 +537,142 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa } if (arg == "-t" || arg == "--threads") { CHECK_ARG - params.n_threads = std::stoi(argv[i]); - if (params.n_threads <= 0) { - params.n_threads = std::thread::hardware_concurrency(); + params.cpuparams.n_threads = std::stoi(argv[i]); + if (params.cpuparams.n_threads <= 0) { + params.cpuparams.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-C" || arg == "--cpu-mask") { + CHECK_ARG + std::string mask = argv[i]; + params.cpuparams.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask); + return true; + } + if (arg == "-Cr" || arg == "--cpu-range") { + CHECK_ARG + std::string range = argv[i]; + params.cpuparams.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.cpuparams.cpumask); + return true; + } + if (arg == "--prio") { + CHECK_ARG + params.cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict") { + CHECK_ARG + params.cpuparams.strict_cpu = std::stoul(argv[i]); + return true; + } + if (arg == "--poll") { + CHECK_ARG + params.cpuparams.poll = std::stoul(argv[i]); + return true; + } if (arg == "-tb" || arg == "--threads-batch") { CHECK_ARG - params.n_threads_batch = std::stoi(argv[i]); - if (params.n_threads_batch <= 0) { - params.n_threads_batch = std::thread::hardware_concurrency(); + params.cpuparams_batch.n_threads = std::stoi(argv[i]); + if (params.cpuparams_batch.n_threads <= 0) { + params.cpuparams_batch.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-Cb" || arg == "--cpu-mask-batch") { + CHECK_ARG + std::string mask = argv[i]; + params.cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.cpuparams_batch.cpumask); + return true; + } + if (arg == "-Crb" || arg == "--cpu-range_batch") { + CHECK_ARG + std::string range = argv[i]; + params.cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.cpuparams_batch.cpumask); + return true; + } + if (arg == "--prio-batch") { + CHECK_ARG + params.cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-batch") { + params.cpuparams_batch.strict_cpu = true; + return true; + } + if (arg == "--poll-batch") { + CHECK_ARG + params.cpuparams_batch.poll = std::stoul(argv[i]); + return true; + } if (arg == "-td" || arg == "--threads-draft") { CHECK_ARG - params.n_threads_draft = std::stoi(argv[i]); - if (params.n_threads_draft <= 0) { - params.n_threads_draft = std::thread::hardware_concurrency(); + params.draft_cpuparams.n_threads = std::stoi(argv[i]); + if (params.draft_cpuparams.n_threads <= 0) { + params.draft_cpuparams.n_threads = std::thread::hardware_concurrency(); } return true; + } + if (arg == "-Cd" || arg == "--cpu-mask-draft") { + CHECK_ARG + std::string mask = argv[i]; + params.draft_cpuparams.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.draft_cpuparams.cpumask); + return true; + } + if (arg == "-Crd" || arg == "--cpu-range-draft") { + CHECK_ARG + std::string range = argv[i]; + params.draft_cpuparams.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.draft_cpuparams.cpumask); + return true; + } + if (arg == "--prio-draft") { + CHECK_ARG + params.draft_cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-draft") { + params.draft_cpuparams.strict_cpu = true; + return true; + } + if (arg == "--poll-draft") { + CHECK_ARG + params.draft_cpuparams.poll = std::stoul(argv[i]); + return true; } if (arg == "-tbd" || arg == "--threads-batch-draft") { CHECK_ARG - params.n_threads_batch_draft = std::stoi(argv[i]); - if (params.n_threads_batch_draft <= 0) { - params.n_threads_batch_draft = std::thread::hardware_concurrency(); + params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]); + if (params.draft_cpuparams_batch.n_threads <= 0) { + params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-Crbd" || arg == "--cpu-range-batch-draft") { + CHECK_ARG + std::string range = argv[i]; + params.draft_cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.draft_cpuparams_batch.cpumask); + return true; + } + if (arg == "--prio-batch-draft") { + CHECK_ARG + params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-batch-draft") { + params.draft_cpuparams_batch.strict_cpu = true; + return true; + } + if (arg == "--poll-batch-draft") { + CHECK_ARG + params.draft_cpuparams_batch.poll = std::stoul(argv[i]); + return true; + } if (arg == "-p" || arg == "--prompt") { CHECK_ARG params.prompt = argv[i]; @@ -1498,11 +1757,40 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --no-display-prompt", "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" }); options.push_back({ "*", "-co, --color", "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" }); options.push_back({ "*", "-s, --seed SEED", "RNG seed (default: %d, use random seed for < 0)", params.seed }); - options.push_back({ "*", "-t, --threads N", "number of threads to use during generation (default: %d)", params.n_threads }); + options.push_back({ "*", "-t, --threads N", "number of threads to use during generation (default: %d)", params.cpuparams.n_threads }); options.push_back({ "*", "-tb, --threads-batch N", "number of threads to use during batch and prompt processing (default: same as --threads)" }); options.push_back({ "speculative", "-td, --threads-draft N", "number of threads to use during generation (default: same as --threads)" }); - options.push_back({ "speculative", "-tbd, --threads-batch-draft N", - "number of threads to use during batch and prompt processing (default: same as --threads-draft)" }); + options.push_back({ "speculative", "-tbd, --threads-batch-draft N","number of threads to use during batch and prompt processing (default: same as --threads-draft)" }); + +#ifndef GGML_USE_OPENMP + // these options are available only with the internal threadpool + options.push_back({ "*", "-C, --cpu-mask M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")"}); + options.push_back({ "*", "-Cr, --cpu-range lo-hi", "range of CPUs for affinity. Complements --cpu-mask"}); + options.push_back({ "*", " --cpu-strict <0|1>", "use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu}); + options.push_back({ "*", " --priority N", "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority}); + options.push_back({ "*", " --poll <0...100>", "use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll}); + + options.push_back({ "*", "-Cb, --cpu-mask-batch M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)"}); + options.push_back({ "*", "-Crb, --cpu-range-batch lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch"}); + options.push_back({ "*", " --cpu-strict-batch <0|1>","use strict CPU placement (default: same as --cpu-strict)"}); + options.push_back({ "*", " --priority-batch N", "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority)"}); + options.push_back({ "*", " --poll-batch <0|1>", "use polling to wait for work (default: same as --poll"}); + + options.push_back({ "speculative", "-Cd, --cpu-mask-draft M", "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)"}); + options.push_back({ "speculative", "-Crd, --cpu-range-draft lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft"}); + options.push_back({ "speculative", " --cpu-strict-draft <0|1>","Use strict CPU placement for draft model (default: same as --cpu-strict)"}); + options.push_back({ "speculative", " --priority-draft N", "Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: same as --priority)"}); + options.push_back({ "speculative", " --poll-draft <0|1>", "Use polling to wait for draft model work (default: same as --poll])"}); + + options.push_back({ "speculative", "-Cbd, --cpu-mask-batch-draft M","Draft model CPU affinity mask. Complements cpu-range-draft-batch (default: same as --cpu-mask-draft)"}); + options.push_back({ "speculative", "-Crbd, --cpu-range-batch-draft lo-hi", + "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)"}); + options.push_back({ "speculative", " --cpu-strict-batch-draft <0|1>", + "Use strict CPU placement for draft model (default: --cpu-strict-draft)"}); + options.push_back({ "speculative", " --priority-batch-draft N","Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority-draft)"}); + options.push_back({ "speculative", " --poll-batch-draft <0|1>","Use polling to wait for draft model work (default: --poll-draft)"}); +#endif // GGML_USE_OPENMP + options.push_back({ "speculative", " --draft N", "number of tokens to draft for speculative decoding (default: %d)", params.n_draft }); options.push_back({ "speculative", "-ps, --p-split N", "speculative decoding split probability (default: %.1f)", (double)params.p_split }); options.push_back({ "*", "-lcs, --lookup-cache-static FNAME", @@ -1774,7 +2062,6 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "export-lora", "-m, --model", "model path from which to load base model (default '%s')", params.model.c_str() }); options.push_back({ "export-lora", " --lora FNAME", "path to LoRA adapter (can be repeated to use multiple adapters)" }); options.push_back({ "export-lora", " --lora-scaled FNAME S", "path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" }); - options.push_back({ "*", "-t, --threads N", "number of threads to use during computation (default: %d)", params.n_threads }); options.push_back({ "export-lora", "-o, --output FNAME", "output file (default: '%s')", params.lora_outfile.c_str() }); printf("usage: %s [options]\n", argv[0]); @@ -1806,9 +2093,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param std::string gpt_params_get_system_info(const gpt_params & params) { std::ostringstream os; - os << "system_info: n_threads = " << params.n_threads; - if (params.n_threads_batch != -1) { - os << " (n_threads_batch = " << params.n_threads_batch << ")"; + os << "system_info: n_threads = " << params.cpuparams.n_threads; + if (params.cpuparams_batch.n_threads != -1) { + os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")"; } #if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later // TODO: windows + arm64 + mingw64 @@ -2332,8 +2619,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.n_seq_max = params.n_parallel; cparams.n_batch = params.n_batch; cparams.n_ubatch = params.n_ubatch; - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; + cparams.n_threads = params.cpuparams.n_threads; + cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ? + params.cpuparams.n_threads : params.cpuparams_batch.n_threads; cparams.seed = params.seed; cparams.logits_all = params.logits_all; cparams.embeddings = params.embedding; @@ -2359,6 +2647,22 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param return cparams; } +struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) { + struct ggml_threadpool_params tpp; + + ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults + + if (params.mask_valid) { + std::memcpy(&tpp.cpumask, ¶ms.cpumask, GGML_MAX_N_THREADS); + } + + tpp.prio = params.priority; + tpp.poll = params.poll; + tpp.strict_cpu = params.strict_cpu; + + return tpp; +} + #ifdef LLAMA_USE_CURL static bool starts_with(const std::string & str, const std::string & prefix) { @@ -3348,7 +3652,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector); fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z); - fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency()); + fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency()); fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k); fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p); fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p); diff --git a/common/common.h b/common/common.h index f603ba2be..cb5e7f6df 100644 --- a/common/common.h +++ b/common/common.h @@ -67,13 +67,18 @@ enum dimre_method { DIMRE_METHOD_MEAN, }; +struct cpu_params { + int n_threads = -1; + bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask. + bool mask_valid = false; // Default: any CPU + enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime) + bool strict_cpu = false; // Use strict CPU placement + uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling) +}; + struct gpt_params { uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed - int32_t n_threads = cpu_get_num_math(); - int32_t n_threads_draft = -1; - int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads) - int32_t n_threads_batch_draft = -1; int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 0; // context size int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS) @@ -100,6 +105,11 @@ struct gpt_params { int32_t yarn_orig_ctx = 0; // YaRN original context length float defrag_thold = -1.0f; // KV cache defragmentation threshold + struct cpu_params cpuparams; + struct cpu_params cpuparams_batch; + struct cpu_params draft_cpuparams; + struct cpu_params draft_cpuparams_batch; + ggml_backend_sched_eval_callback cb_eval = nullptr; void * cb_eval_user_data = nullptr; @@ -204,7 +214,7 @@ struct gpt_params { int32_t port = 8080; // server listens on this network port int32_t timeout_read = 600; // http read timeout in seconds int32_t timeout_write = timeout_read; // http write timeout in seconds - int32_t n_threads_http = -1; // number of threads to process HTTP requests + int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool) std::string hostname = "127.0.0.1"; std::string public_path = ""; @@ -277,6 +287,11 @@ void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params); std::string gpt_params_get_system_info(const gpt_params & params); +bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]); +bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]); +void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr); +bool set_process_priority(enum ggml_sched_priority prio); + // // String utils // @@ -327,8 +342,9 @@ struct llama_init_result { struct llama_init_result llama_init_from_gpt_params(gpt_params & params); -struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params); -struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params); +struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params); +struct llama_context_params llama_context_params_from_gpt_params (const gpt_params & params); +struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params); struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params); struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params); diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp index aca332e94..3ce91070b 100644 --- a/examples/baby-llama/baby-llama.cpp +++ b/examples/baby-llama/baby-llama.cpp @@ -18,7 +18,7 @@ constexpr float rms_norm_eps = 5e-6f; #endif static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); + struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr); if (plan.work_size > 0) { buf.resize(plan.work_size); diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index 47cb16c69..97622f4f4 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -21,7 +21,7 @@ #endif static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); + struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr); if (plan.work_size > 0) { buf.resize(plan.work_size); @@ -54,7 +54,7 @@ static void tensor_dump(const ggml_tensor * tensor, const char * name) { #define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor) struct benchmark_params_struct { - int32_t n_threads = 1; + int n_threads = 1; int32_t n_iterations = 10; }; diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp index 8fa492571..a68268388 100644 --- a/examples/cvector-generator/cvector-generator.cpp +++ b/examples/cvector-generator/cvector-generator.cpp @@ -486,8 +486,8 @@ int main(int argc, char ** argv) { if (use_pca) { // run PCA PCA::pca_params pca_params; - pca_params.n_threads = params.n_threads; - pca_params.n_batch = params.n_pca_batch; + pca_params.n_threads = params.cpuparams.n_threads; + pca_params.n_batch = params.n_pca_batch; pca_params.n_iterations = params.n_pca_iterations; PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final); } else { diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index c7e5ca788..8df457e21 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -410,7 +410,7 @@ int main(int argc, char ** argv) { g_verbose = (params.verbosity == 1); try { - lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads); + lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads); ctx.run_merge(); } catch (const std::exception & err) { fprintf(stderr, "%s\n", err.what()); diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 42918bfc7..8edadef90 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -16,6 +16,7 @@ #include #include #include +#include #include "ggml.h" #include "llama.h" @@ -225,6 +226,9 @@ struct cmd_params { std::vector type_k; std::vector type_v; std::vector n_threads; + std::vector cpu_mask; + std::vector cpu_strict; + std::vector poll; std::vector n_gpu_layers; std::vector rpc_servers; std::vector split_mode; @@ -236,6 +240,8 @@ struct cmd_params { std::vector embeddings; ggml_numa_strategy numa; int reps; + ggml_sched_priority prio; + int delay; bool verbose; output_formats output_format; output_formats output_format_stderr; @@ -251,6 +257,9 @@ static const cmd_params cmd_params_defaults = { /* type_k */ {GGML_TYPE_F16}, /* type_v */ {GGML_TYPE_F16}, /* n_threads */ {cpu_get_num_math()}, + /* cpu_mask */ {"0x0"}, + /* cpu_strict */ {false}, + /* poll */ {50}, /* n_gpu_layers */ {99}, /* rpc_servers */ {""}, /* split_mode */ {LLAMA_SPLIT_MODE_LAYER}, @@ -262,6 +271,8 @@ static const cmd_params cmd_params_defaults = { /* embeddings */ {false}, /* numa */ GGML_NUMA_STRATEGY_DISABLED, /* reps */ 5, + /* prio */ GGML_SCHED_PRIO_NORMAL, + /* delay */ 0, /* verbose */ false, /* output_format */ MARKDOWN, /* output_format_stderr */ NONE, @@ -281,6 +292,9 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -ctk, --cache-type-k (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); printf(" -ctv, --cache-type-v (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); printf(" -t, --threads (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str()); + printf(" -C, --cpu-mask (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str()); + printf(" --cpu-strict <0|1> (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str()); + printf(" --poll <0...100> (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str()); printf(" -ngl, --n-gpu-layers (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str()); printf(" -rpc, --rpc (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str()); printf(" -sm, --split-mode (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); @@ -292,6 +306,8 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str()); printf(" -ts, --tensor-split (default: 0)\n"); printf(" -r, --repetitions (default: %d)\n", cmd_params_defaults.reps); + printf(" --prio <0|1|2|3> (default: %d)\n", cmd_params_defaults.prio); + printf(" --delay <0...N> (seconds) (default: %d)\n", cmd_params_defaults.delay); printf(" -o, --output (default: %s)\n", output_format_str(cmd_params_defaults.output_format)); printf(" -oe, --output-err (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr)); printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0"); @@ -338,6 +354,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { params.output_format_stderr = cmd_params_defaults.output_format_stderr; params.reps = cmd_params_defaults.reps; params.numa = cmd_params_defaults.numa; + params.prio = cmd_params_defaults.prio; + params.delay = cmd_params_defaults.delay; for (int i = 1; i < argc; i++) { arg = argv[i]; @@ -433,6 +451,27 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { } auto p = string_split(argv[i], split_delim); params.n_threads.insert(params.n_threads.end(), p.begin(), p.end()); + } else if (arg == "-C" || arg == "--cpu-mask") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end()); + } else if (arg == "--cpu-strict") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end()); + } else if (arg == "--poll") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.poll.insert(params.poll.end(), p.begin(), p.end()); } else if (arg == "-ngl" || arg == "--n-gpu-layers") { if (++i >= argc) { invalid_param = true; @@ -541,6 +580,18 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { break; } params.reps = std::stoi(argv[i]); + } else if (arg == "--prio") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.prio = (enum ggml_sched_priority) std::stoi(argv[i]); + } else if (arg == "--delay") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.delay = std::stoi(argv[i]); } else if (arg == "-o" || arg == "--output") { if (++i >= argc) { invalid_param = true; @@ -585,6 +636,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { if (params.use_mmap.empty()) { params.use_mmap = cmd_params_defaults.use_mmap; } if (params.embeddings.empty()) { params.embeddings = cmd_params_defaults.embeddings; } if (params.n_threads.empty()) { params.n_threads = cmd_params_defaults.n_threads; } + if (params.cpu_mask.empty()) { params.cpu_mask = cmd_params_defaults.cpu_mask; } + if (params.cpu_strict.empty()) { params.cpu_strict = cmd_params_defaults.cpu_strict; } + if (params.poll.empty()) { params.poll = cmd_params_defaults.poll; } return params; } @@ -598,6 +652,9 @@ struct cmd_params_instance { ggml_type type_k; ggml_type type_v; int n_threads; + std::string cpu_mask; + bool cpu_strict; + int poll; int n_gpu_layers; std::string rpc_servers; llama_split_mode split_mode; @@ -667,7 +724,10 @@ static std::vector get_cmd_params_instances(const cmd_param for (const auto & tv : params.type_v) for (const auto & nkvo : params.no_kv_offload) for (const auto & fa : params.flash_attn) - for (const auto & nt : params.n_threads) { + for (const auto & nt : params.n_threads) + for (const auto & cm : params.cpu_mask) + for (const auto & cs : params.cpu_strict) + for (const auto & pl : params.poll) { for (const auto & n_prompt : params.n_prompt) { if (n_prompt == 0) { continue; @@ -681,6 +741,9 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_k = */ tk, /* .type_v = */ tv, /* .n_threads = */ nt, + /* .cpu_mask = */ cm, + /* .cpu_strict = */ cs, + /* .poll = */ pl, /* .n_gpu_layers = */ nl, /* .rpc_servers = */ rpc, /* .split_mode = */ sm, @@ -707,6 +770,9 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_k = */ tk, /* .type_v = */ tv, /* .n_threads = */ nt, + /* .cpu_mask = */ cm, + /* .cpu_strict = */ cs, + /* .poll = */ pl, /* .n_gpu_layers = */ nl, /* .rpc_servers = */ rpc, /* .split_mode = */ sm, @@ -733,6 +799,9 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_k = */ tk, /* .type_v = */ tv, /* .n_threads = */ nt, + /* .cpu_mask = */ cm, + /* .cpu_strict = */ cs, + /* .poll = */ pl, /* .n_gpu_layers = */ nl, /* .rpc_servers = */ rpc, /* .split_mode = */ sm, @@ -769,6 +838,9 @@ struct test { int n_batch; int n_ubatch; int n_threads; + std::string cpu_mask; + bool cpu_strict; + int poll; bool has_rpc; ggml_type type_k; ggml_type type_v; @@ -795,6 +867,9 @@ struct test { n_batch = inst.n_batch; n_ubatch = inst.n_ubatch; n_threads = inst.n_threads; + cpu_mask = inst.cpu_mask; + cpu_strict = inst.cpu_strict; + poll = inst.poll; has_rpc = !inst.rpc_servers.empty(); type_k = inst.type_k; type_v = inst.type_v; @@ -872,13 +947,14 @@ struct test { "cpu_info", "gpu_info", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", - "n_threads", "type_k", "type_v", + "n_threads", "cpu_mask", "cpu_strict", "poll", + "type_k", "type_v", "n_gpu_layers", "split_mode", "main_gpu", "no_kv_offload", "flash_attn", "tensor_split", "use_mmap", "embeddings", "n_prompt", "n_gen", "test_time", "avg_ns", "stddev_ns", - "avg_ts", "stddev_ts" + "avg_ts", "stddev_ts", }; return fields; } @@ -887,7 +963,7 @@ struct test { static field_type get_field_type(const std::string & field) { if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || - field == "n_threads" || + field == "n_threads" || field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" || field == "main_gpu" || field == "n_prompt" || field == "n_gen" || @@ -896,6 +972,7 @@ struct test { } if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" || field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" || + field == "cpu_strict" || field == "flash_attn" || field == "use_mmap" || field == "embeddings") { return BOOL; } @@ -928,7 +1005,8 @@ struct test { cpu_info, gpu_info, model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params), std::to_string(n_batch), std::to_string(n_ubatch), - std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v), + std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll), + ggml_type_name(type_k), ggml_type_name(type_v), std::to_string(n_gpu_layers), split_mode_str(split_mode), std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn), tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings), @@ -1067,7 +1145,7 @@ struct markdown_printer : public printer { return -30; } if (field == "t/s") { - return 16; + return 20; } if (field == "size" || field == "params") { return 10; @@ -1149,6 +1227,15 @@ struct markdown_printer : public printer { if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) { fields.emplace_back("n_threads"); } + if (params.cpu_mask.size() > 1 || params.cpu_mask != cmd_params_defaults.cpu_mask) { + fields.emplace_back("cpu_mask"); + } + if (params.cpu_strict.size() > 1 || params.cpu_strict != cmd_params_defaults.cpu_strict) { + fields.emplace_back("cpu_strict"); + } + if (params.poll.size() > 1 || params.poll != cmd_params_defaults.poll) { + fields.emplace_back("poll"); + } if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) { fields.emplace_back("n_batch"); } @@ -1383,6 +1470,8 @@ int main(int argc, char ** argv) { llama_backend_init(); llama_numa_init(params.numa); + set_process_priority(params.prio); + // initialize printer std::unique_ptr p = create_printer(params.output_format); std::unique_ptr p_err = create_printer(params.output_format_stderr); @@ -1428,6 +1517,28 @@ int main(int argc, char ** argv) { llama_kv_cache_clear(ctx); + // cool off before the test + if (params.delay) { + std::this_thread::sleep_for(std::chrono::seconds(params.delay)); + } + + struct ggml_threadpool_params tpp = ggml_threadpool_params_default(t.n_threads); + if (!parse_cpu_mask(t.cpu_mask, tpp.cpumask)) { + LOG_TEE("%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str()); + exit(1); + } + tpp.strict_cpu = t.cpu_strict; + tpp.poll = t.poll; + tpp.prio = params.prio; + + struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp); + if (!threadpool) { + LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads); + exit(1); + } + + llama_attach_threadpool(ctx, threadpool, NULL); + // warmup run if (t.n_prompt > 0) { //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads); @@ -1466,6 +1577,8 @@ int main(int argc, char ** argv) { llama_print_timings(ctx); llama_free(ctx); + + ggml_threadpool_free(threadpool); } llama_free_model(lmodel); diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index 58c32ca53..48b7840ae 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -71,8 +71,8 @@ actor LlamaContext { var ctx_params = llama_context_default_params() ctx_params.seed = 1234 ctx_params.n_ctx = 2048 - ctx_params.n_threads = UInt32(n_threads) - ctx_params.n_threads_batch = UInt32(n_threads) + ctx_params.n_threads = Int32(n_threads) + ctx_params.n_threads_batch = Int32(n_threads) let context = llama_new_context_with_model(model, ctx_params) guard let context else { diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index 8c7dd2ae3..86b39f20e 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -129,14 +129,14 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para if (!params->image.empty()) { LOG_TEE("using base64 encoded image instead of command line image path\n"); } - embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt); + embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt); if (!embed) { LOG_TEE("%s: can't load image from prompt\n", __func__); return NULL; } params->prompt = remove_image_from_prompt(prompt); } else { - embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, fname.c_str()); + embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str()); if (!embed) { fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str()); return NULL; diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp index 379fc295f..f500ea5b9 100644 --- a/examples/llava/minicpmv-cli.cpp +++ b/examples/llava/minicpmv-cli.cpp @@ -180,7 +180,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling, static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){ auto ctx_clip = clip_init_context(params); - auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str()); + auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str()); if (!embeds) { std::cerr << "error: failed to load image " << fname << ". Terminating\n\n"; return NULL; diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 4a342ad03..2c05afb04 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -221,6 +221,40 @@ int main(int argc, char ** argv) { return 1; } + LOG("%s: llama threadpool init = n_threads = %d\n", + __func__, + (int) params.cpuparams.n_threads + ); + struct ggml_threadpool_params tpp_batch = + ggml_threadpool_params_from_cpu_params(params.cpuparams_batch); + struct ggml_threadpool_params tpp = + ggml_threadpool_params_from_cpu_params(params.cpuparams); + + set_process_priority(params.cpuparams.priority); + + struct ggml_threadpool * threadpool_batch = NULL; + if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) { + threadpool_batch = ggml_threadpool_new(&tpp_batch); + if (!threadpool_batch) { + LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads); + exit(1); + } + + // Start the non-batch threadpool in the paused state + tpp.paused = true; + } + + struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp); + if (!threadpool) { + LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads); + exit(1); + } + + llama_attach_threadpool(ctx, threadpool, threadpool_batch); + if (ctx_guidance) { + llama_attach_threadpool(ctx_guidance, threadpool, threadpool_batch); + } + const int n_ctx_train = llama_n_ctx_train(model); const int n_ctx = llama_n_ctx(ctx); LOG("n_ctx: %d\n", n_ctx); @@ -989,6 +1023,9 @@ int main(int argc, char ** argv) { llama_sampling_free(ctx_sampling); llama_backend_free(); + ggml_threadpool_free(threadpool); + ggml_threadpool_free(threadpool_batch); + #ifndef LOG_DISABLE_LOGS LOG_TEE("Log end\n"); #endif // LOG_DISABLE_LOGS diff --git a/examples/server/server.cpp b/examples/server/server.cpp index c37182fe4..cc938e80d 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2534,8 +2534,8 @@ int main(int argc, char ** argv) { }); LOG_INFO("system info", { - {"n_threads", params.n_threads}, - {"n_threads_batch", params.n_threads_batch}, + {"n_threads", params.cpuparams.n_threads}, + {"n_threads_batch", params.cpuparams_batch.n_threads}, {"total_threads", std::thread::hardware_concurrency()}, {"system_info", llama_print_system_info()}, }); diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index b051a18f1..1616edecb 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -73,10 +73,11 @@ int main(int argc, char ** argv) { // load the draft model params.model = params.model_draft; params.n_gpu_layers = params.n_gpu_layers_draft; - if (params.n_threads_draft > 0) { - params.n_threads = params.n_threads_draft; + if (params.draft_cpuparams.n_threads > 0) { + params.cpuparams.n_threads = params.draft_cpuparams.n_threads; } - params.n_threads_batch = params.n_threads_batch_draft; + + params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads; llama_init_result llama_init_dft = llama_init_from_gpt_params(params); model_dft = llama_init_dft.model; ctx_dft = llama_init_dft.context; diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h index 434c13b34..0dff47d65 100644 --- a/ggml/include/ggml-alloc.h +++ b/ggml/include/ggml-alloc.h @@ -7,8 +7,8 @@ extern "C" { #endif typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t; -typedef struct ggml_backend_buffer * ggml_backend_buffer_t; -typedef struct ggml_backend * ggml_backend_t; +typedef struct ggml_backend_buffer * ggml_backend_buffer_t; +typedef struct ggml_backend * ggml_backend_t; // Tensor allocator struct ggml_tallocr { diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h index e73b9a745..e497b6d02 100644 --- a/ggml/include/ggml-backend.h +++ b/ggml/include/ggml-backend.h @@ -103,6 +103,7 @@ extern "C" { GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend); GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads); + GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool); GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data); // Create a backend buffer from an existing pointer diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index b11d047ae..5233a9995 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -231,6 +231,8 @@ #define GGML_MAX_SRC 10 #ifndef GGML_MAX_NAME #define GGML_MAX_NAME 64 +#define GGML_MAX_N_THREADS 512 + #endif #define GGML_MAX_OP_PARAMS 64 #define GGML_DEFAULT_N_THREADS 4 @@ -628,6 +630,29 @@ extern "C" { // If it returns true, the computation is aborted typedef bool (*ggml_abort_callback)(void * data); + // Scheduling priorities + enum ggml_sched_priority { + GGML_SCHED_PRIO_NORMAL, + GGML_SCHED_PRIO_MEDIUM, + GGML_SCHED_PRIO_HIGH, + GGML_SCHED_PRIO_REALTIME + }; + + // Threadpool params + // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults + struct ggml_threadpool_params { + bool cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings) + int n_threads; // number of threads + enum ggml_sched_priority prio; // thread priority + uint32_t poll; // polling level (0 - no polling, 100 - aggressive polling) + bool strict_cpu; // strict cpu placement + bool paused; // start in paused state + }; + + struct ggml_threadpool; // forward declaration, see ggml.c + + typedef struct ggml_threadpool * ggml_threadpool_t; + // the compute plan that needs to be prepared for ggml_graph_compute() // since https://github.com/ggerganov/ggml/issues/287 struct ggml_cplan { @@ -635,6 +660,7 @@ extern "C" { uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()` int n_threads; + struct ggml_threadpool * threadpool; // abort ggml_graph_compute when true ggml_abort_callback abort_callback; @@ -2057,10 +2083,23 @@ extern "C" { GGML_API size_t ggml_graph_overhead(void); GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads); + GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads); + GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params *p, int n_threads); + GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params *p0, const struct ggml_threadpool_params *p1); + GGML_API struct ggml_threadpool* ggml_threadpool_new (struct ggml_threadpool_params * params); + GGML_API void ggml_threadpool_free (struct ggml_threadpool * threadpool); + GGML_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool); + GGML_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool); + GGML_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool); + // ggml_graph_plan() has to be called before ggml_graph_compute() // when plan.work_size > 0, caller must allocate memory for plan.work_data - GGML_API struct ggml_cplan ggml_graph_plan (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); - GGML_API enum ggml_status ggml_graph_compute( struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); + GGML_API struct ggml_cplan ggml_graph_plan( + const struct ggml_cgraph * cgraph, + int n_threads, /* = GGML_DEFAULT_N_THREADS */ + struct ggml_threadpool * threadpool /* = NULL */ ); + GGML_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); + // same as ggml_graph_compute() but the work data is allocated as a part of the context // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index ff84b9bb5..ec7d30825 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -1247,7 +1247,7 @@ endif() # Data types, macros and functions related to controlling CPU affinity and # some memory allocation are available on Linux through GNU extensions in libc -if (CMAKE_SYSTEM_NAME MATCHES "Linux") +if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android") add_compile_definitions(_GNU_SOURCE) endif() diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.c index 8856967c9..5b877db35 100644 --- a/ggml/src/ggml-backend.c +++ b/ggml/src/ggml-backend.c @@ -722,9 +722,11 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) { #endif struct ggml_backend_cpu_context { - int n_threads; - void * work_data; - size_t work_size; + int n_threads; + ggml_threadpool_t threadpool; + + void * work_data; + size_t work_size; ggml_abort_callback abort_callback; void * abort_callback_data; @@ -759,7 +761,7 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu)); - cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); + cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool); cpu_plan->cgraph = *cgraph; // FIXME: deep copy if (cpu_plan->cplan.work_size > 0) { @@ -796,7 +798,7 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backe GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; - struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); + struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool); if (cpu_ctx->work_size < cplan.work_size) { free(cpu_ctx->work_data); @@ -873,6 +875,7 @@ ggml_backend_t ggml_backend_cpu_init(void) { } ctx->n_threads = GGML_DEFAULT_N_THREADS; + ctx->threadpool = NULL; ctx->work_data = NULL; ctx->work_size = 0; ctx->abort_callback = NULL; @@ -903,6 +906,18 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) { ctx->n_threads = n_threads; } +void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) { + GGML_ASSERT(ggml_backend_is_cpu(backend_cpu)); + + struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context; + + if (ctx->threadpool && ctx->threadpool != threadpool) { + // already had a different threadpool, pause/suspend it before switching + ggml_threadpool_pause(ctx->threadpool); + } + ctx->threadpool = threadpool; +} + void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) { GGML_ASSERT(ggml_backend_is_cpu(backend_cpu)); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 9c105fd35..dc6cdca0b 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -69,23 +69,42 @@ int ggml_sve_cnt_b = 0; #endif #include +#if !defined(__clang__) typedef volatile LONG atomic_int; typedef atomic_int atomic_bool; typedef atomic_int atomic_flag; #define ATOMIC_FLAG_INIT 0 +typedef enum { + memory_order_relaxed, + memory_order_consume, + memory_order_acquire, + memory_order_release, + memory_order_acq_rel, + memory_order_seq_cst +} memory_order; + static void atomic_store(atomic_int * ptr, LONG val) { InterlockedExchange(ptr, val); } +static void atomic_store_explicit(atomic_int * ptr, LONG val, memory_order mo) { + // TODO: add support for explicit memory order + InterlockedExchange(ptr, val); +} static LONG atomic_load(atomic_int * ptr) { return InterlockedCompareExchange(ptr, 0, 0); } +static LONG atomic_load_explicit(atomic_int * ptr, memory_order mo) { + // TODO: add support for explicit memory order + return InterlockedCompareExchange(ptr, 0, 0); +} static LONG atomic_fetch_add(atomic_int * ptr, LONG inc) { return InterlockedExchangeAdd(ptr, inc); } -static LONG atomic_fetch_sub(atomic_int * ptr, LONG dec) { - return atomic_fetch_add(ptr, -(dec)); +static LONG atomic_fetch_add_explicit(atomic_int * ptr, LONG inc, memory_order mo) { + // TODO: add support for explicit memory order + return InterlockedExchangeAdd(ptr, inc); } static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) { return InterlockedExchange(ptr, 1); @@ -93,6 +112,9 @@ static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) { static void atomic_flag_clear(atomic_flag * ptr) { InterlockedExchange(ptr, 0); } +#else // clang +#include +#endif typedef HANDLE pthread_t; @@ -121,8 +143,10 @@ static int sched_yield (void) { return 0; } #else + #include #include +#include typedef void * thread_ret_t; @@ -1868,28 +1892,102 @@ struct ggml_context_container { struct ggml_context context; }; -struct ggml_compute_state_shared { - const struct ggml_cgraph * cgraph; - const struct ggml_cplan * cplan; +// +// Threading defs +// - int n_threads; +typedef pthread_t ggml_thread_t; + +#if defined(_WIN32) + +typedef CONDITION_VARIABLE ggml_cond_t; +typedef SRWLOCK ggml_mutex_t; + +#define ggml_mutex_init(m) InitializeSRWLock(m) +#define ggml_mutex_destroy(m) +#define ggml_mutex_lock(m) AcquireSRWLockExclusive(m) +#define ggml_mutex_unlock(m) ReleaseSRWLockExclusive(m) +#define ggml_mutex_lock_shared(m) AcquireSRWLockShared(m) +#define ggml_mutex_unlock_shared(m) ReleaseSRWLockShared(m) + +#define ggml_cond_init(c) InitializeConditionVariable(c) +#define ggml_cond_destroy(c) +#define ggml_cond_wait(c, m) SleepConditionVariableSRW(c, m, INFINITE, CONDITION_VARIABLE_LOCKMODE_SHARED) +#define ggml_cond_broadcast(c) WakeAllConditionVariable(c) + +#define ggml_thread_create pthread_create +#define ggml_thread_join pthread_join + +#else + +typedef pthread_cond_t ggml_cond_t; +typedef pthread_mutex_t ggml_mutex_t; + +#define ggml_mutex_init(m) pthread_mutex_init(m, NULL) +#define ggml_mutex_destroy(m) pthread_mutex_destroy(m) +#define ggml_mutex_lock(m) pthread_mutex_lock(m) +#define ggml_mutex_unlock(m) pthread_mutex_unlock(m) +#define ggml_mutex_lock_shared(m) pthread_mutex_lock(m) +#define ggml_mutex_unlock_shared(m) pthread_mutex_unlock(m) + +#define ggml_lock_init(x) UNUSED(x) +#define ggml_lock_destroy(x) UNUSED(x) +#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) +#define ggml_lock_lock(x) _mm_pause() +#else +#define ggml_lock_lock(x) UNUSED(x) +#endif +#define ggml_lock_unlock(x) UNUSED(x) + +#define GGML_LOCK_INITIALIZER 0 +#define ggml_cond_init(c) pthread_cond_init(c, NULL) +#define ggml_cond_destroy(c) pthread_cond_destroy(c) +#define ggml_cond_wait(c, m) pthread_cond_wait(c, m) +#define ggml_cond_broadcast(c) pthread_cond_broadcast(c) + +#define ggml_thread_create pthread_create +#define ggml_thread_join pthread_join + +#endif + +// Threadpool def +struct ggml_threadpool { + ggml_mutex_t mutex; // mutex for cond.var + ggml_cond_t cond; // cond.var for waiting for new work + + struct ggml_cgraph * cgraph; + struct ggml_cplan * cplan; // synchronization primitives + atomic_int n_graph; // incremented when there is work to be done (i.e each graph) atomic_int n_barrier; atomic_int n_barrier_passed; + atomic_int current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads. - ggml_abort_callback abort_callback; // abort ggml_graph_compute when true - void * abort_callback_data; + // these are atomic as an annotation for thread-sanitizer + atomic_bool stop; // Used for stopping the threadpool altogether + atomic_bool pause; // Used for pausing the threadpool or individual threads - atomic_int current_chunk; // currently processing chunk during mul_mat, shared between all the threads + struct ggml_compute_state * workers; // per thread state + int n_threads_max; // number of threads in the pool + int n_threads_cur; // number of threads used in the current graph + + int32_t prio; // Scheduling priority + uint32_t poll; // Polling level (0 - no polling) enum ggml_status ec; }; +// Per-thread state struct ggml_compute_state { +#ifndef GGML_USE_OPENMP ggml_thread_t thrd; + bool cpumask[GGML_MAX_N_THREADS]; + int last_graph; + bool pending; +#endif + struct ggml_threadpool * threadpool; int ith; - struct ggml_compute_state_shared * shared; }; struct ggml_compute_params { @@ -1900,7 +1998,7 @@ struct ggml_compute_params { size_t wsize; void * wdata; - struct ggml_compute_state_shared * shared; + struct ggml_threadpool * threadpool; }; // @@ -2971,6 +3069,19 @@ static_assert(GGML_UNARY_OP_COUNT == 13, "GGML_UNARY_OP_COUNT != 13"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); +// Helpers for polling loops +#if defined(__aarch64__) && ( defined(__clang__) || defined(__GNUC__) ) +static inline void ggml_thread_cpu_relax(void) { + __asm__ volatile("yield" ::: "memory"); +} +#elif defined(__x86_64__) +static inline void ggml_thread_cpu_relax(void) { + _mm_pause(); +} +#else +static inline void ggml_thread_cpu_relax(void) {;} +#endif + // // NUMA support // @@ -3018,42 +3129,36 @@ inline static void ggml_critical_section_start(void) { } #ifdef GGML_USE_OPENMP -static void ggml_barrier(struct ggml_compute_state_shared * shared) { - if (shared->n_threads == 1) { +static void ggml_barrier(struct ggml_threadpool * threadpool) { + if (threadpool->n_threads_cur == 1) { return; } #pragma omp barrier } #else -static void ggml_barrier(struct ggml_compute_state_shared * shared) { - if (shared->n_threads == 1) { +static void ggml_barrier(struct ggml_threadpool * threadpool) { + if (threadpool->n_threads_cur == 1) { return; } - atomic_int * n_barrier = &shared->n_barrier; - atomic_int * n_barrier_passed = &shared->n_barrier_passed; + atomic_int * n_barrier = &threadpool->n_barrier; + atomic_int * n_barrier_passed = &threadpool->n_barrier_passed; - int n_threads = shared->n_threads; - int passed_old = atomic_load(n_barrier_passed); + int n_threads = threadpool->n_threads_cur; + int passed_old = atomic_load_explicit(n_barrier_passed, memory_order_relaxed); if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) { // last thread atomic_store(n_barrier, 0); - atomic_fetch_add(n_barrier_passed, 1); + atomic_fetch_add_explicit(n_barrier_passed, 1, memory_order_relaxed); } else { // wait for other threads - const int n_spin_before_sleep = 100000; while (true) { - for (int i = 0; i < n_spin_before_sleep; i++) { - if (atomic_load(n_barrier_passed) != passed_old) { - return; - } - #if defined(__SSE3__) - _mm_pause(); - #endif + if (atomic_load_explicit(n_barrier_passed, memory_order_relaxed) != passed_old) { + return; } - sched_yield(); + ggml_thread_cpu_relax(); } } } @@ -10148,7 +10253,7 @@ static void ggml_compute_forward_acc_f32( ((char *) src0->data), ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } const int ith = params->ith; @@ -12622,10 +12727,10 @@ UseGgmlGemm1:; if (ith == 0) { // Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start. - atomic_store(¶ms->shared->current_chunk, nth); + atomic_store_explicit(¶ms->threadpool->current_chunk, nth, memory_order_relaxed); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); #if GGML_USE_LLAMAFILE if (src1->type != vec_dot_type) { @@ -12733,7 +12838,7 @@ UseGgmlGemm2:; break; } - current_chunk = atomic_fetch_add(¶ms->shared->current_chunk, 1); + current_chunk = atomic_fetch_add_explicit(¶ms->threadpool->current_chunk, 1, memory_order_relaxed); } } @@ -12828,7 +12933,7 @@ static void ggml_compute_forward_mul_mat_id( } } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // compute each matrix multiplication in sequence for (int cur_a = 0; cur_a < n_as; ++cur_a) { @@ -12982,7 +13087,7 @@ static void ggml_compute_forward_out_prod_f32( if (ith == 0) { ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // dst[:,:,:,:] = 0 // for i2,i3: @@ -13100,7 +13205,7 @@ static void ggml_compute_forward_out_prod_q_f32( if (ith == 0) { ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // parallelize by last three dimensions @@ -13286,7 +13391,7 @@ static void ggml_compute_forward_set_f32( ((char *) src0->data), ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } const int ith = params->ith; @@ -13865,7 +13970,7 @@ static void ggml_compute_forward_diag_mask_f32( ((char *) src0->data), ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } // TODO: handle transposed/permuted matrices @@ -14641,7 +14746,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32( // need to zero dst since we are accumulating into it memset(dst->data, 0, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; @@ -14729,7 +14834,7 @@ static void ggml_compute_forward_conv_transpose_1d_f32( // need to zero dst since we are accumulating into it memset(dst->data, 0, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; @@ -15109,7 +15214,7 @@ static void ggml_compute_forward_conv_transpose_2d( memset(dst->data, 0, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int32_t stride = ggml_get_op_params_i32(dst, 0); @@ -15977,7 +16082,7 @@ static void ggml_compute_forward_flash_attn_back_f32( if (ith == 0) { memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int64_t elem_q = ggml_nelements(q); const int64_t elem_k = ggml_nelements(k); @@ -16668,7 +16773,7 @@ static void ggml_compute_forward_add_rel_pos_f32( if (params->ith == 0) { memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L357-L359 @@ -16953,7 +17058,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( if (ith == 0) { memset(sums, 0, sizeof(float) * (nth + nth * nc)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // rows per thread const int dr = (nr + nth - 1)/nth; @@ -16994,7 +17099,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( } #endif } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); if (ith == 0) { float * dp = (float *) dst->data; @@ -18810,65 +18915,6 @@ void ggml_graph_clear(struct ggml_cgraph * cgraph) { ggml_hash_set_reset(&cgraph->visited_hash_set); } -// -// thread data -// -// synchronization is done via busy loops -// I tried using spin locks, but not sure how to use them correctly - the things I tried were slower than busy loops -// - -#ifdef __APPLE__ - -//#include -// -//typedef os_unfair_lock ggml_lock_t; -// -//#define ggml_lock_init(x) UNUSED(x) -//#define ggml_lock_destroy(x) UNUSED(x) -//#define ggml_lock_lock os_unfair_lock_lock -//#define ggml_lock_unlock os_unfair_lock_unlock -// -//#define GGML_LOCK_INITIALIZER OS_UNFAIR_LOCK_INIT - -typedef int ggml_lock_t; - -#define ggml_lock_init(x) UNUSED(x) -#define ggml_lock_destroy(x) UNUSED(x) -#define ggml_lock_lock(x) UNUSED(x) -#define ggml_lock_unlock(x) UNUSED(x) - -#define GGML_LOCK_INITIALIZER 0 - -#define ggml_thread_create pthread_create -#define ggml_thread_join pthread_join - -#else - -//typedef pthread_spinlock_t ggml_lock_t; - -//#define ggml_lock_init(x) pthread_spin_init(x, PTHREAD_PROCESS_PRIVATE) -//#define ggml_lock_destroy pthread_spin_destroy -//#define ggml_lock_lock pthread_spin_lock -//#define ggml_lock_unlock pthread_spin_unlock - -typedef int ggml_lock_t; - -#define ggml_lock_init(x) UNUSED(x) -#define ggml_lock_destroy(x) UNUSED(x) -#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) -#define ggml_lock_lock(x) _mm_pause() -#else -#define ggml_lock_lock(x) UNUSED(x) -#endif -#define ggml_lock_unlock(x) UNUSED(x) - -#define GGML_LOCK_INITIALIZER 0 - -#define ggml_thread_create pthread_create -#define ggml_thread_join pthread_join - -#endif - // Android's libc implementation "bionic" does not support setting affinity #if defined(__gnu_linux__) static void set_numa_thread_affinity(int thread_n) { @@ -19149,9 +19195,268 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { return n_tasks; } -struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threads) { +static thread_ret_t ggml_graph_compute_secondary_thread(void* data); + +#if defined(_WIN32) +#include "windows.h" + +// TODO: support > 64 CPUs +bool ggml_thread_apply_affinity(bool * mask) { + HANDLE h = GetCurrentThread(); + uint64_t bitmask = 0ULL; + + assert(GGML_MAX_N_THREADS >= 64); + + for (int32_t i = 0; i < 8; i++) { + int32_t idx = i * 8; + uint8_t val = 0; + val |= mask[idx + 0] << 0; + val |= mask[idx + 1] << 1; + val |= mask[idx + 2] << 2; + val |= mask[idx + 3] << 3; + val |= mask[idx + 4] << 4; + val |= mask[idx + 5] << 5; + val |= mask[idx + 6] << 6; + val |= mask[idx + 7] << 7; + bitmask |= (uint64_t)val << idx; + } + + for (int32_t i = 64; i < GGML_MAX_N_THREADS; i++) { + if (mask[i]) { + fprintf(stderr, "warn: setting thread-affinity for > 64 CPUs isn't supported on windows!\n"); + break; + } + } + + DWORD_PTR m = (DWORD_PTR)bitmask; + + m = SetThreadAffinityMask(h, m); + + return m != 0; +} + +static bool ggml_thread_apply_priority(int32_t prio) { + // Note that on Windows the Process Priority Class must be updated in order to set Thread priority. + // This is up to the applications. + DWORD p = THREAD_PRIORITY_NORMAL; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = THREAD_PRIORITY_NORMAL; break; + case GGML_SCHED_PRIO_MEDIUM: p = THREAD_PRIORITY_ABOVE_NORMAL; break; + case GGML_SCHED_PRIO_HIGH: p = THREAD_PRIORITY_HIGHEST; break; + case GGML_SCHED_PRIO_REALTIME: p = THREAD_PRIORITY_TIME_CRITICAL; break; + } + + if (prio == GGML_SCHED_PRIO_NORMAL) { + // Keep inherited policy/priority + return true; + } + + if (!SetThreadPriority(GetCurrentThread(), p)) { + fprintf(stderr, "warn: failed to set thread priority %d : (%d)\n", prio, (int) GetLastError()); + return false; + } + + return true; +} + +#elif defined(__APPLE__) +#include +#include + +static bool ggml_thread_apply_affinity(const bool * mask) { + // Not supported on Apple platforms + UNUSED(mask); + return true; +} + +static bool ggml_thread_apply_priority(int32_t prio) { + struct sched_param p; + int32_t policy = SCHED_OTHER; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: policy = SCHED_OTHER; p.sched_priority = 0; break; + case GGML_SCHED_PRIO_MEDIUM: policy = SCHED_FIFO; p.sched_priority = 40; break; + case GGML_SCHED_PRIO_HIGH: policy = SCHED_FIFO; p.sched_priority = 80; break; + case GGML_SCHED_PRIO_REALTIME: policy = SCHED_FIFO; p.sched_priority = 90; break; + } + + if (prio == GGML_SCHED_PRIO_NORMAL) { + // Keep inherited policy/priority + return true; + } + + int32_t err = pthread_setschedparam(pthread_self(), policy, &p); + if (err != 0) { + fprintf(stderr, "warn: failed to set thread priority %d : %s (%d)\n", prio, strerror(err), err); + return false; + } + + return true; +} + +#else // posix? + +static bool ggml_thread_apply_affinity(const bool * mask) { + cpu_set_t cpuset; + int err; + + CPU_ZERO(&cpuset); + + for (uint32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + if (mask[i]) { + GGML_PRINT_DEBUG("Thread %lx: adding %d to cpuset\n", pthread_self(), i); + CPU_SET(i, &cpuset); + } + } + +#ifdef __ANDROID__ + err = sched_setaffinity(0, sizeof(cpuset), &cpuset); + if (err < 0) { + err = errno; + } +#else + err = pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset); +#endif + if (err != 0) { + fprintf(stderr, "warn: failed to set affinity mask 0x%llx : %s (%d)\n", (unsigned long long)mask, strerror(err), err); + return false; + } + + return true; +} + +static bool ggml_thread_apply_priority(int32_t prio) { + struct sched_param p; + int32_t policy = SCHED_OTHER; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: policy = SCHED_OTHER; p.sched_priority = 0; break; + case GGML_SCHED_PRIO_MEDIUM: policy = SCHED_FIFO; p.sched_priority = 40; break; + case GGML_SCHED_PRIO_HIGH: policy = SCHED_FIFO; p.sched_priority = 80; break; + case GGML_SCHED_PRIO_REALTIME: policy = SCHED_FIFO; p.sched_priority = 90; break; + } + + if (prio == GGML_SCHED_PRIO_NORMAL) { + // Keep inherited policy/priority + return true; + } + + int32_t err = pthread_setschedparam(pthread_self(), policy, &p); + if (err != 0) { + fprintf(stderr, "warn: failed to set thread priority %d : %s (%d)\n", prio, strerror(err), err); + return false; + } + + return true; +} + +#endif + +static bool ggml_thread_cpumask_is_valid(const bool * mask) { + for (int i = 0; i < GGML_MAX_N_THREADS; i++) { + if (mask[i]) { return true; } + } + return false; +} + +static void ggml_thread_cpumask_next(const bool * global_mask, bool * local_mask, bool strict, int32_t* iter) { + if (!strict) { + memcpy(local_mask, global_mask, GGML_MAX_N_THREADS); + return; + } else { + memset(local_mask, 0, GGML_MAX_N_THREADS); + int32_t base_idx = *iter; + for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + int32_t idx = base_idx + i; + if (idx >= GGML_MAX_N_THREADS) { + // Just a cheaper modulo + idx -= GGML_MAX_N_THREADS; + } + if (global_mask[idx]) { + local_mask[idx] = 1; + *iter = idx + 1; + return; + } + } + } +} + +void ggml_threadpool_free(struct ggml_threadpool* threadpool) { + if (!threadpool) return; + +#ifndef GGML_USE_OPENMP + struct ggml_compute_state* workers = threadpool->workers; + const int n_threads = threadpool->n_threads_max; + + ggml_mutex_lock(&threadpool->mutex); + + threadpool->stop = true; + threadpool->pause = false; + + ggml_cond_broadcast(&threadpool->cond); + ggml_mutex_unlock(&threadpool->mutex); + + for (int j = 1; j < n_threads; j++) { + int32_t rc = ggml_thread_join(workers[j].thrd, NULL); + GGML_ASSERT(rc == GGML_EXIT_SUCCESS || rc == GGML_EXIT_ABORTED); + UNUSED(rc); + } + + ggml_mutex_destroy(&threadpool->mutex); + ggml_cond_destroy(&threadpool->cond); +#endif // GGML_USE_OPENMP + + GGML_ALIGNED_FREE(threadpool->workers); + GGML_ALIGNED_FREE(threadpool); +} + +#ifndef GGML_USE_OPENMP +// pause/resume must be called under mutex +static void ggml_threadpool_pause_locked(struct ggml_threadpool * threadpool) { + GGML_PRINT_DEBUG("Pausing threadpool\n"); + threadpool->pause = true; + ggml_cond_broadcast(&threadpool->cond); +} + +static void ggml_threadpool_resume_locked(struct ggml_threadpool * threadpool) { + GGML_PRINT_DEBUG("Resuming threadpool\n"); + threadpool->pause = false; + ggml_cond_broadcast(&threadpool->cond); +} +#endif + +void ggml_threadpool_pause(struct ggml_threadpool * threadpool) { +#ifndef GGML_USE_OPENMP + ggml_mutex_lock(&threadpool->mutex); + if (!threadpool->pause) { + ggml_threadpool_pause_locked(threadpool); + } + ggml_mutex_unlock(&threadpool->mutex); +#else + UNUSED(threadpool); +#endif +} + +void ggml_threadpool_resume(struct ggml_threadpool * threadpool) { +#ifndef GGML_USE_OPENMP + ggml_mutex_lock(&threadpool->mutex); + if (threadpool->pause) { + ggml_threadpool_resume_locked(threadpool); + } + ggml_mutex_unlock(&threadpool->mutex); +#else + UNUSED(threadpool); +#endif +} + +struct ggml_cplan ggml_graph_plan( + const struct ggml_cgraph * cgraph, + int n_threads, + struct ggml_threadpool * threadpool) { + + if (threadpool == NULL) { + GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads); + } if (n_threads <= 0) { - n_threads = GGML_DEFAULT_N_THREADS; + n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS; } size_t work_size = 0; @@ -19307,12 +19612,13 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa } if (work_size > 0) { - work_size += CACHE_LINE_SIZE*(n_threads - 1); + work_size += CACHE_LINE_SIZE*(n_threads); } - cplan.n_threads = MIN(max_tasks, n_threads); - cplan.work_size = work_size; - cplan.work_data = NULL; + cplan.threadpool = threadpool; + cplan.n_threads = MIN(max_tasks, n_threads); + cplan.work_size = work_size; + cplan.work_data = NULL; return cplan; } @@ -19320,17 +19626,17 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; - const struct ggml_cgraph * cgraph = state->shared->cgraph; - const struct ggml_cplan * cplan = state->shared->cplan; + const struct ggml_cgraph * cgraph = state->threadpool->cgraph; + const struct ggml_cplan * cplan = state->threadpool->cplan; set_numa_thread_affinity(state->ith); struct ggml_compute_params params = { - /*.ith =*/ state->ith, - /*.nth =*/ state->shared->n_threads, - /*.wsize =*/ cplan->work_size, - /*.wdata =*/ cplan->work_data, - /*.shared=*/ state->shared, + /*.ith =*/ state->ith, + /*.nth =*/ state->threadpool->n_threads_cur, + /*.wsize =*/ cplan->work_size, + /*.wdata =*/ cplan->work_data, + /*.threadpool=*/ state->threadpool, }; for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) { @@ -19339,12 +19645,12 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { ggml_compute_forward(¶ms, node); if (state->ith == 0 && cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { - state->shared->ec = GGML_STATUS_ABORTED; + state->threadpool->ec = GGML_STATUS_ABORTED; } - ggml_barrier(state->shared); + ggml_barrier(state->threadpool); - if (state->shared->ec != GGML_STATUS_SUCCESS) { + if (state->threadpool->ec != GGML_STATUS_SUCCESS) { break; } } @@ -19352,24 +19658,243 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { return 0; } +#ifndef GGML_USE_OPENMP + +static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) { + struct ggml_threadpool * threadpool = state->threadpool; + + if (state->pending || threadpool->stop || threadpool->pause) { return true; } + + // check for new graph/work + int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed); + if (new_graph != state->last_graph) { + state->pending = (state->ith < threadpool->n_threads_cur); + state->last_graph = new_graph; + } + + return state->pending; +} + +static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) { + struct ggml_threadpool * threadpool = state->threadpool; + + // This seems to make 0 ... 100 a decent range for polling level across modern processors. + // Perhaps, we can adjust it dynamically based on load and things. + const uint64_t n_rounds = 1024UL * 128 * threadpool->poll; + + for (uint64_t i=0; !ggml_graph_compute_ready(state) && ipending; +} + +static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state * state) { + struct ggml_threadpool * threadpool = state->threadpool; + + if (ggml_graph_compute_poll_for_work(state)) { + return state->pending; + } + + ggml_mutex_lock_shared(&threadpool->mutex); + while (!ggml_graph_compute_ready(state)) { + // No new work. Wait for the signal. + GGML_PRINT_DEBUG("thread #%d waiting for work\n", state->ith); + ggml_cond_wait(&threadpool->cond, &threadpool->mutex); + } + ggml_mutex_unlock_shared(&threadpool->mutex); + + return state->pending; +} + +static thread_ret_t ggml_graph_compute_secondary_thread(void* data) { + struct ggml_compute_state * state = (struct ggml_compute_state *) data; + struct ggml_threadpool * threadpool = state->threadpool; + + ggml_thread_apply_priority(threadpool->prio); + if (ggml_thread_cpumask_is_valid(state->cpumask)) { + ggml_thread_apply_affinity(state->cpumask); + } + + while (true) { + // Check if we need to sleep + while (threadpool->pause) { + GGML_PRINT_DEBUG("thread #%d inside pause loop\n", state->ith); + ggml_mutex_lock_shared(&threadpool->mutex); + if (threadpool->pause) { + ggml_cond_wait(&threadpool->cond, &threadpool->mutex); + } + GGML_PRINT_DEBUG("thread #%d resuming after wait\n", state->ith); + ggml_mutex_unlock_shared(&threadpool->mutex); + } + + // This needs to be checked for after the cond_wait + if (threadpool->stop) break; + + // Check if there is new work + // The main thread is the only one that can dispatch new work + + ggml_graph_compute_check_for_work(state); + if (state->pending) { + state->pending = false; + + ggml_graph_compute_thread(state); + } + } + + return (thread_ret_t) 0; +} + +// Start processing new graph +static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool) +{ + // always take the mutex here because the worker threads are doing hybrid poll/wait + + ggml_mutex_lock(&threadpool->mutex); + + atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_relaxed); + + if (threadpool->pause) { + // Update main thread prio and affinity to match the threadpool settings + ggml_thread_apply_priority(threadpool->prio); + if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) { + ggml_thread_apply_affinity(threadpool->workers[0].cpumask); + } + + // resume does cond broadcast + ggml_threadpool_resume_locked(threadpool); + } else { + ggml_cond_broadcast(&threadpool->cond); + } + + ggml_mutex_unlock(&threadpool->mutex); +} + +#endif // GGML_USE_OPENMP + +void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) { + p->n_threads = n_threads; + p->prio = 0; // default priority (usually means normal or inherited) + p->poll = 50; // hybrid-polling enabled + p->strict_cpu = false; // no strict placement (all threads share same cpumask) + p->paused = false; // threads are ready to go + memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited) +} + +struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) { + struct ggml_threadpool_params p; + ggml_threadpool_params_init(&p, n_threads); + return p; +} + +bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) { + if (p0->n_threads != p1->n_threads ) return false; + if (p0->prio != p1->prio ) return false; + if (p0->poll != p1->poll ) return false; + if (p0->strict_cpu != p1->strict_cpu ) return false; + return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0; +} + +static struct ggml_threadpool * ggml_threadpool_new_impl( + struct ggml_threadpool_params * tpp, + struct ggml_cgraph * cgraph, + struct ggml_cplan * cplan) { + + struct ggml_threadpool * threadpool = + GGML_ALIGNED_MALLOC(sizeof(struct ggml_threadpool)); + { + threadpool->cgraph = cgraph; + threadpool->cplan = cplan; + threadpool->n_graph = 0; + threadpool->n_barrier = 0; + threadpool->n_barrier_passed = 0; + threadpool->current_chunk = 0; + threadpool->stop = false; + threadpool->pause = tpp->paused; + threadpool->workers = NULL; + threadpool->n_threads_max = tpp->n_threads; + threadpool->n_threads_cur = tpp->n_threads; + threadpool->poll = tpp->poll; + threadpool->prio = tpp->prio; + threadpool->ec = GGML_STATUS_SUCCESS; + } + + // Allocate and init workers state + const size_t workers_size = sizeof(struct ggml_compute_state) * tpp->n_threads; + struct ggml_compute_state * workers = GGML_ALIGNED_MALLOC(workers_size); + + memset(workers, 0, workers_size); + for (int j = 0; j < tpp->n_threads; j++) { + workers[j].threadpool = threadpool; + workers[j].ith = j; + } + + threadpool->workers = workers; + +#ifndef GGML_USE_OPENMP + ggml_mutex_init(&threadpool->mutex); + ggml_cond_init(&threadpool->cond); + + // Spin the threads for all workers, and update CPU placements. + // Place the main thread last (towards the higher numbered CPU cores). + + int32_t cpumask_iter = 0; + + for (int j = 1; j < tpp->n_threads; j++) { + ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter); + + int32_t rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_secondary_thread, &workers[j]); + GGML_ASSERT(rc == 0); + } + + ggml_thread_cpumask_next(tpp->cpumask, workers[0].cpumask, tpp->strict_cpu, &cpumask_iter); + + if (!threadpool->pause) { + // Update main thread prio and affinity at the start, otherwise we'll do it in resume + ggml_thread_apply_priority(threadpool->prio); + if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) { + ggml_thread_apply_affinity(threadpool->workers[0].cpumask); + } + } +#endif // GGML_USE_OPENMP + + return threadpool; +} + +struct ggml_threadpool * ggml_threadpool_new(struct ggml_threadpool_params * tpp) { + return ggml_threadpool_new_impl(tpp, NULL, NULL); +} + enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { GGML_ASSERT(cplan); GGML_ASSERT(cplan->n_threads > 0); GGML_ASSERT(cplan->work_size == 0 || cplan->work_data != NULL); - int n_threads = cplan->n_threads; + int n_threads = cplan->n_threads; + struct ggml_threadpool * threadpool = cplan->threadpool; - struct ggml_compute_state_shared state_shared = { - /*.cgraph =*/ cgraph, - /*.cgraph_plan =*/ cplan, - /*.n_threads =*/ n_threads, - /*.n_barrier =*/ 0, - /*.n_barrier_passed =*/ 0, - /*.abort_callback =*/ NULL, - /*.abort_callback_data =*/ NULL, - /*.current_chunk =*/ 0, - /*.ec =*/ GGML_STATUS_SUCCESS, - }; + bool disposable_threadpool = false; + + if (threadpool == NULL) { + GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads); + disposable_threadpool = true; + + struct ggml_threadpool_params ttp = ggml_threadpool_params_default(n_threads); + threadpool = ggml_threadpool_new_impl(&ttp, cgraph, cplan); + } else { + // Reset some of the parameters that need resetting + // No worker threads should be accessing the parameters below at this stage + threadpool->cgraph = cgraph; + threadpool->cplan = cplan; + threadpool->n_threads_cur = n_threads; + threadpool->current_chunk = 0; + threadpool->ec = GGML_STATUS_SUCCESS; + } + + if (n_threads > threadpool->n_threads_max) { + GGML_PRINT("WARNING: cplan is requesting more threads than the threadpool contains. Expect a bad time!\n"); + } #ifdef GGML_USE_OPENMP if (n_threads > 1) { @@ -19379,63 +19904,36 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl { // update the number of threads from the actual number of threads that we got from OpenMP n_threads = omp_get_num_threads(); - state_shared.n_threads = n_threads; + threadpool->n_threads_cur = n_threads; } - struct ggml_compute_state worker = { - .thrd = 0, - .ith = omp_get_thread_num(), - .shared = &state_shared, - }; - ggml_graph_compute_thread(&worker); + ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]); } } else { - struct ggml_compute_state worker = { - .thrd = 0, - .ith = 0, - .shared = &state_shared, - }; - ggml_graph_compute_thread(&worker); + ggml_graph_compute_thread(&threadpool->workers[0]); } #else - struct ggml_compute_state * workers = alloca(sizeof(struct ggml_compute_state)*n_threads); + // Kick all threads to start the new graph + ggml_graph_compute_kickoff(threadpool); - for (int j = 0; j < n_threads; ++j) { - workers[j] = (struct ggml_compute_state) { - .thrd = 0, - .ith = j, - .shared = &state_shared, - }; - } - - // create thread pool - for (int j = 1; j < n_threads; ++j) { - const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]); - GGML_ASSERT(rc == 0); - UNUSED(rc); - } - - // this is a work thread too - ggml_graph_compute_thread(&workers[0]); - - // join or kill thread pool - if (n_threads > 1) { - for (int j = 1; j < n_threads; j++) { - const int rc = ggml_thread_join(workers[j].thrd, NULL); - GGML_ASSERT(rc == 0); - UNUSED(rc); - } - } + // This is a work thread too + ggml_graph_compute_thread(&threadpool->workers[0]); #endif // don't leave affinity set on the main thread clear_numa_thread_affinity(); - return state_shared.ec; + enum ggml_status ret = threadpool->ec; + + if (disposable_threadpool) { + ggml_threadpool_free(threadpool); + } + + return ret; } enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { - struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads); + struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads, NULL); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); @@ -20251,7 +20749,7 @@ static enum ggml_opt_result ggml_opt_adam( float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values - struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads); + struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads, NULL); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; @@ -20598,7 +21096,7 @@ static enum ggml_opt_result ggml_opt_lbfgs( opt->iter = iter; } - struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads); + struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads, NULL); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; diff --git a/include/llama.h b/include/llama.h index 6cca6320b..c3bda9e02 100644 --- a/include/llama.h +++ b/include/llama.h @@ -304,8 +304,8 @@ extern "C" { uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode uint32_t n_ubatch; // physical maximum batch size uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models) - uint32_t n_threads; // number of threads to use for generation - uint32_t n_threads_batch; // number of threads to use for batch processing + int32_t n_threads; // number of threads to use for generation + int32_t n_threads_batch; // number of threads to use for batch processing enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id @@ -428,6 +428,13 @@ extern "C" { //optional: LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); + // Optional: an auto threadpool gets created in ggml if not passed explicitly + LLAMA_API void llama_attach_threadpool( + struct llama_context * ctx, + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch); + LLAMA_API void llama_detach_threadpool(struct llama_context * ctx); + // Call once at the end of the program - currently only used for MPI LLAMA_API void llama_backend_free(void); @@ -837,13 +844,13 @@ extern "C" { // Set the number of threads used for decoding // n_threads is the number of threads used for generation (single token) // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) - LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); + LLAMA_API void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch); // Get the number of threads used for generation of a single token. - LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx); + LLAMA_API int32_t llama_n_threads(struct llama_context * ctx); // Get the number of threads used for prompt and batch processing (multiple token). - LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx); + LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx); // Set whether the model is in embeddings mode or not // If true, embeddings will be returned but logits will not diff --git a/src/llama.cpp b/src/llama.cpp index 8d5f24783..2274296b4 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -2373,8 +2373,8 @@ struct llama_cparams { uint32_t n_batch; uint32_t n_ubatch; uint32_t n_seq_max; - uint32_t n_threads; // number of threads to use for generation - uint32_t n_threads_batch; // number of threads to use for batch processing + int n_threads; // number of threads to use for generation + int n_threads_batch; // number of threads to use for batch processing float rope_freq_base; float rope_freq_scale; @@ -3091,6 +3091,9 @@ struct llama_context { #endif ggml_backend_t backend_cpu = nullptr; + ggml_threadpool_t threadpool = nullptr; + ggml_threadpool_t threadpool_batch = nullptr; + bool has_evaluated_once = false; int64_t t_start_us; @@ -15494,9 +15497,10 @@ static void llama_output_reorder(struct llama_context * ctx) { } static void llama_graph_compute( - llama_context & lctx, - ggml_cgraph * gf, - int n_threads) { + llama_context & lctx, + ggml_cgraph * gf, + int n_threads, + ggml_threadpool * threadpool) { #ifdef GGML_USE_METAL if (ggml_backend_is_metal(lctx.backend_metal)) { ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads); @@ -15505,6 +15509,7 @@ static void llama_graph_compute( if (lctx.backend_cpu != nullptr) { ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads); + ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool); ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data); } #ifdef GGML_USE_BLAS @@ -15625,6 +15630,8 @@ static int llama_decode_internal( } int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch; + GGML_ASSERT(n_threads > 0); // non-causal masks do not use the KV cache @@ -15686,7 +15693,7 @@ static int llama_decode_internal( llama_set_inputs(lctx, ubatch); - llama_graph_compute(lctx, gf, n_threads); + llama_graph_compute(lctx, gf, n_threads, threadpool); // update the kv ring buffer { @@ -15863,7 +15870,9 @@ static int llama_encode_internal( lctx.inp_embd_enc = NULL; lctx.n_outputs = n_tokens; - const int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch; + GGML_ASSERT(n_threads > 0); ggml_backend_sched_reset(lctx.sched); @@ -15895,7 +15904,7 @@ static int llama_encode_internal( llama_set_inputs(lctx, ubatch); - llama_graph_compute(lctx, gf, n_threads); + llama_graph_compute(lctx, gf, n_threads, threadpool); // extract embeddings if (embd) { @@ -16177,7 +16186,7 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids); - llama_graph_compute(lctx, gf, lctx.cparams.n_threads); + llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool); #endif //const int64_t t_end = ggml_time_us(); @@ -16203,7 +16212,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { llama_set_k_shift(lctx); - llama_graph_compute(lctx, gf, lctx.cparams.n_threads); + llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool); need_reserve = true; } @@ -17451,6 +17460,19 @@ void llama_numa_init(enum ggml_numa_strategy numa) { } } +void llama_attach_threadpool( + struct llama_context * ctx, + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch) { + ctx->threadpool = threadpool; + ctx->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool; +} + +void llama_detach_threadpool(struct llama_context * ctx) { + ctx->threadpool = nullptr; + ctx->threadpool_batch = nullptr; +} + void llama_backend_free(void) { ggml_quantize_free(); } @@ -19367,16 +19389,16 @@ size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepa } } -void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) { +void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) { ctx->cparams.n_threads = n_threads; ctx->cparams.n_threads_batch = n_threads_batch; } -uint32_t llama_n_threads(struct llama_context * ctx) { +int32_t llama_n_threads(struct llama_context * ctx) { return ctx->cparams.n_threads; } -uint32_t llama_n_threads_batch(struct llama_context * ctx) { +int32_t llama_n_threads_batch(struct llama_context * ctx) { return ctx->cparams.n_threads_batch; } diff --git a/tests/test-rope.cpp b/tests/test-rope.cpp index 8159e276a..246bb227d 100644 --- a/tests/test-rope.cpp +++ b/tests/test-rope.cpp @@ -113,7 +113,7 @@ static struct ggml_tensor * get_random_tensor_f32( } static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); + struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr); if (plan.work_size > 0) { buf.resize(plan.work_size);