vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

add dnnl stream

Browse source
This commit is contained in:
luoyu-intel 2024-08-19 07:08:04 +00:00
parent 3d0a64f092
commit 4dc55156ee
3 changed files with 69 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml/src/ggml-sycl.cpp
View file

@ -2494,7 +2494,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
#else #else
DnnlGemmWrapper::row_gemm(*stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), auto dnnl_stream = ctx.stream_dnnl(stream);
DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>()); src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>());
const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
@ -2529,7 +2530,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
dst_dd_i, ldc))); dst_dd_i, ldc)));
#else #else
DnnlGemmWrapper::row_gemm(*stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(), auto dnnl_stream = ctx.stream_dnnl(stream);
DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>()); src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
#endif #endif
} }

93
ggml/src/ggml-sycl/common.hpp
View file

@ -19,6 +19,10 @@
#include "dpct/helper.hpp" #include "dpct/helper.hpp"
#include "ggml-sycl.h" #include "ggml-sycl.h"
#include "presets.hpp" #include "presets.hpp"
#if GGML_SYCL_DNNL
#include "dnnl.hpp"
#include "dnnl_sycl.hpp"
#endif
#define GGML_COMMON_DECL_SYCL #define GGML_COMMON_DECL_SYCL
#define GGML_COMMON_IMPL_SYCL #define GGML_COMMON_IMPL_SYCL
@ -59,7 +63,7 @@ static int g_ggml_sycl_debug = 0;
// define for XMX in Intel GPU // define for XMX in Intel GPU
// TODO: currently, it's not used for XMX really. // TODO: currently, it's not used for XMX really.
#if !defined(GGML_SYCL_FORCE_MMQ) #if !defined(GGML_SYCL_FORCE_MMQ)
#define SYCL_USE_XMX #define SYCL_USE_XMX
#endif #endif
// max batch size to use MMQ kernels when tensor cores are available // max batch size to use MMQ kernels when tensor cores are available
@ -80,16 +84,16 @@ static int g_ggml_sycl_debug = 0;
typedef sycl::queue *queue_ptr; typedef sycl::queue *queue_ptr;
enum ggml_sycl_backend_gpu_mode { enum ggml_sycl_backend_gpu_mode {
SYCL_UNSET_GPU_MODE = -1, SYCL_UNSET_GPU_MODE = -1,
SYCL_SINGLE_GPU_MODE = 0, SYCL_SINGLE_GPU_MODE = 0,
SYCL_MUL_GPU_MODE SYCL_MUL_GPU_MODE
}; };
static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size"); static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");
static void crash() { static void crash() {
int* ptr = NULL; int* ptr = NULL;
*ptr = 0; *ptr = 0;
} }
[[noreturn]] static void ggml_sycl_error( [[noreturn]] static void ggml_sycl_error(
@ -98,9 +102,9 @@ static void crash() {
const char* file, const char* file,
const int line, const int line,
const char* msg) { const char* msg) {
fprintf(stderr, "SYCL error: %s: %s\n", stmt, msg); fprintf(stderr, "SYCL error: %s: %s\n", stmt, msg);
fprintf(stderr, " in function %s at %s:%d\n", func, file, line); fprintf(stderr, " in function %s at %s:%d\n", func, file, line);
GGML_ABORT("SYCL error"); GGML_ABORT("SYCL error");
} }
#define SYCL_CHECK(err) \ #define SYCL_CHECK(err) \
@ -138,40 +142,40 @@ static int g_all_sycl_device_count = -1;
static bool g_ggml_backend_sycl_buffer_type_initialized = false; static bool g_ggml_backend_sycl_buffer_type_initialized = false;
static ggml_sycl_backend_gpu_mode g_ggml_sycl_backend_gpu_mode = static ggml_sycl_backend_gpu_mode g_ggml_sycl_backend_gpu_mode =
SYCL_UNSET_GPU_MODE; SYCL_UNSET_GPU_MODE;
static void* g_scratch_buffer = nullptr; static void* g_scratch_buffer = nullptr;
static size_t g_scratch_size = 0; // disabled by default static size_t g_scratch_size = 0; // disabled by default
static size_t g_scratch_offset = 0; static size_t g_scratch_offset = 0;
[[noreturn]] static inline void bad_arch(const sycl::stream& stream_ct1) { [[noreturn]] static inline void bad_arch(const sycl::stream& stream_ct1) {
stream_ct1 << "ERROR: ggml-sycl was compiled without support for the " stream_ct1 << "ERROR: ggml-sycl was compiled without support for the "
"current GPU architecture.\n"; "current GPU architecture.\n";
// __trap(); // __trap();
std::exit(1); std::exit(1);
(void)bad_arch; // suppress unused function warning (void)bad_arch; // suppress unused function warning
} }
int get_current_device_id(); int get_current_device_id();
inline dpct::err0 ggml_sycl_set_device(const int device) try { inline dpct::err0 ggml_sycl_set_device(const int device) try {
int current_device_id; int current_device_id;
SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id())); SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id()));
// GGML_SYCL_DEBUG("ggml_sycl_set_device device_id=%d, // GGML_SYCL_DEBUG("ggml_sycl_set_device device_id=%d,
// current_device_id=%d\n", device, current_device); // current_device_id=%d\n", device, current_device);
if (device == current_device_id) { if (device == current_device_id) {
return 0; return 0;
} }
return CHECK_TRY_ERROR(dpct::select_device(device)); return CHECK_TRY_ERROR(dpct::select_device(device));
} catch (sycl::exception const& exc) { } catch (sycl::exception const& exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__ std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl; << ", line:" << __LINE__ << std::endl;
crash(); crash();
std::exit(1); std::exit(1);
} }
////////////////////// //////////////////////
@ -249,10 +253,10 @@ struct ggml_sycl_pool_alloc {
// backend interface // backend interface
struct ggml_tensor_extra_gpu { struct ggml_tensor_extra_gpu {
void* data_device[GGML_SYCL_MAX_DEVICES]; // 1 pointer for each device for split void* data_device[GGML_SYCL_MAX_DEVICES]; // 1 pointer for each device for split
// tensors // tensors
dpct::event_ptr events[GGML_SYCL_MAX_DEVICES] dpct::event_ptr events[GGML_SYCL_MAX_DEVICES]
[GGML_SYCL_MAX_STREAMS]; // events for synchronizing multiple GPUs [GGML_SYCL_MAX_STREAMS]; // events for synchronizing multiple GPUs
}; };
struct ggml_backend_sycl_context { struct ggml_backend_sycl_context {
@ -277,6 +281,33 @@ struct ggml_backend_sycl_context {
return stream(device, 0); return stream(device, 0);
} }
#if GGML_SYCL_DNNL
dnnl::stream make_stream(sycl::queue& q) {
// Get the device associated with the queue
sycl::device dev = q.get_device();
// Get the context associated with the queue
sycl::context ctx = q.get_context();
const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx);
dnnl::stream stream = dnnl::sycl_interop::make_stream(eng, q);
return stream;
}
std::unordered_map<sycl::queue*, dnnl::stream> stream_map;
dnnl::stream stream_dnnl(int device, int _stream) {
auto q = stream(device, _stream);
return stream_dnnl(q);
}
dnnl::stream stream_dnnl(sycl::queue* qptr) {
auto it = stream_map.find(qptr);
if (it == stream_map.end()) {
stream_map[qptr] = make_stream(*qptr);
}
return it->second;
}
dnnl::stream stream_dnnl() {
return stream_dnnl(device, 0);
}
#endif
// pool // pool
std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES]; std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];

5
ggml/src/ggml-sycl/gemm.hpp
View file

@ -17,11 +17,12 @@
#include <iostream> #include <iostream>
#include "ggml-sycl.h" #include "ggml-sycl.h"
#include "dnnl.hpp"
#include "dnnl_sycl.hpp"
#if GGML_SYCL_DNNL #if GGML_SYCL_DNNL
#include "dnnl.hpp"
#include "dnnl_sycl.hpp"
class DnnlGemmWrapper { class DnnlGemmWrapper {
public: public:
using dt = dnnl::memory::data_type; using dt = dnnl::memory::data_type;