diff options
Diffstat (limited to 'ggml/src/ggml-sycl.cpp')
| -rw-r--r-- | ggml/src/ggml-sycl.cpp | 64 |
1 files changed, 60 insertions, 4 deletions
diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index d8eb86c2..e0f8536a 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -37,6 +37,7 @@ #include "ggml-backend-impl.h" #include "ggml-sycl/backend.hpp" +#include "ggml-sycl/common.hpp" #include "ggml-sycl/presets.hpp" bool ggml_sycl_loaded(void); @@ -83,12 +84,64 @@ static __dpct_inline__ float op_add(const float a, const float b) { return a + b; } -static __dpct_inline__ float op_mul(const float a, const float b) { - return a * b; +static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, + size_t offset, size_t size) { + GGML_SYCL_DEBUG(" [SYCL] call %s\n", __func__); + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context; + SYCL_CHECK(ggml_sycl_set_device(ctx->device)); + auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue()); + if (size == 0) { + return; // Nothing to do + } + if (tensor->data == nullptr) { + GGML_ABORT("Error: Tensor data pointer is null.\n"); + } + void * target_ptr = static_cast<char *>(tensor->data) + offset; + SYCL_CHECK(CHECK_TRY_ERROR((*stream).memset(target_ptr, value, size))); + SYCL_CHECK(CHECK_TRY_ERROR((*stream).wait())); +} + +static void ggml_backend_sycl_buffer_reset(ggml_backend_buffer_t buffer) { + GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__); + if (buffer == nullptr) { + return; + } + + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context; + + if (ctx != nullptr) { + for (ggml_tensor_extra_gpu * extra : ctx->tensor_extras) { + release_extra_gpu(extra); + } + ctx->tensor_extras.clear(); // reset the tensor_extras vector + } } -static __dpct_inline__ float op_div(const float a, const float b) { - return a / b; +static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { + /* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer, + /* .get_base = */ ggml_backend_sycl_buffer_get_base, + /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, + /* .memset_tensor = */ ggml_backend_sycl_buffer_memset_tensor, + /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor, + /* .clear = */ ggml_backend_sycl_buffer_clear, + /* .reset = */ ggml_backend_sycl_buffer_reset, +}; + +// sycl buffer type +struct ggml_backend_sycl_buffer_type_context { + int device; + std::string name; + + // each buffer type has its own stream + queue_ptr stream = nullptr; +}; + +static const char * ggml_backend_sycl_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; + + return ctx->name.c_str(); } template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t> @@ -4407,7 +4460,9 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { /* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer, /* .get_base = */ ggml_backend_sycl_buffer_get_base, /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, + /* .memset_tensor = */ ggml_backend_sycl_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, + /* .memset_tensor =*/ NULL, /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor, /* .clear = */ ggml_backend_sycl_buffer_clear, @@ -4810,6 +4865,7 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = { /* .free_buffer = */ ggml_backend_sycl_split_buffer_free_buffer, /* .get_base = */ ggml_backend_sycl_split_buffer_get_base, /* .init_tensor = */ ggml_backend_sycl_split_buffer_init_tensor, + /* .memset_tensor =*/ NULL, /* .set_tensor = */ ggml_backend_sycl_split_buffer_set_tensor, /* .get_tensor = */ ggml_backend_sycl_split_buffer_get_tensor, /* .cpy_tensor = */ NULL, |