new API for cpu-accel

Author: Djip007

Package.swift

@@ -87,5 +87,5 @@ let package = Package(
             linkerSettings: linkerSettings
         )
     ],
-    cxxLanguageStandard: .cxx11
+    cxxLanguageStandard: .cxx17
 )

ggml/include/ggml-cpu.h

@@ -111,7 +111,6 @@ extern "C" {
         ggml_vec_dot_t           vec_dot;
         enum ggml_type           vec_dot_type;
         int64_t                  nrows; // number of rows to process simultaneously
-        int64_t                  ncols; // number of columns to process simultaneously
     };
 
     GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);

ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

@@ -9,9 +9,7 @@
 extern "C" {
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
     GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
-#endif
 
 #ifdef __cplusplus
 }

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -3,16 +3,45 @@
 #include "ggml-backend-impl.h"
 
 namespace ggml::cpu {
-    tensor_traits::~tensor_traits(){};
+    tensor_traits::~tensor_traits() {}
     extra_buffer_type::~extra_buffer_type() {}
 }
 
-const struct ggml_cpu_tensor_traits* ggml_cpu_get_tensor_traits(
-        const struct ggml_tensor * tensor)
-{
-    // host buffer can't be repacked!
-    if (tensor->buffer && !ggml_backend_buft_is_host(tensor->buffer->buft)) {
-        return (struct ggml_cpu_tensor_traits*)tensor->extra;
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
+                return true;
+            }
+        }
     }
-    return nullptr;
+/*
+    auto tensor_traits = ggml::cpu::get_tensor_traits(op);
+    if (tensor_traits) {
+        return tensor_traits->compute_forward(params, op);
+    }
+    */
+    return false;
+}
+
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
+                return true;
+            }
+        }
+    }
+/*
+    auto tensor_traits = ggml::cpu::get_tensor_traits(op);
+    if (tensor_traits) {
+        *size = tensor_traits->work_size(n_threads, op);
+        return true;
+    }
+    */
+    return false;
 }

ggml/src/ggml-cpu/ggml-cpu-traits.cpp

@@ -3,50 +3,48 @@
 #include "ggml-backend-impl.h"
 
 #ifdef  __cplusplus
+#include <vector>
+extern "C" {
+#endif
+struct ggml_compute_params {
+    // ith = thread index, nth = number of threads
+    int ith, nth;
+
+    // work buffer for all threads
+    size_t wsize;
+    void * wdata;
+
+    struct ggml_threadpool * threadpool;
+};
+
+// implementé dans ggml-cpu.c pas forcement a sa place ici.
+void ggml_barrier(struct ggml_threadpool * tp);
+
+// return true if op part of extra "accelerator"
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
+
+#ifdef  __cplusplus
+}
 namespace ggml::cpu {
     // enregistré dans tensor->extra
     class tensor_traits {
     public:
         ~tensor_traits();
-        virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) = 0;
+        virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size) = 0;
+        virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
     };
 
-    // ou mettre ca?
     class extra_buffer_type {
     public:
         ~extra_buffer_type();
         virtual bool supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
+        virtual tensor_traits* get_tensor_traits(const struct ggml_tensor * op) = 0;
     };
+
 }
-extern "C" {
-// #else
-#endif
-    // a t'on besoin d'un mapping C?
-    //bool ggml_cpu_extra_compute_forward(void* extra, struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-    // @ transferer en methode privée pour cpu_aarch64
-    typedef int (*ggml_repack_t) (struct ggml_tensor *t, int interleave_block, const void * GGML_RESTRICT data,
-                                    size_t data_size);
-    typedef void (*ggml_from_float_to_mat_t)
-                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
-    typedef void (*ggml_gemv_t)   (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                     const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)   (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                     const void * GGML_RESTRICT y, int nr, int nc);
-
-    struct ggml_cpu_tensor_traits {
-        ggml_repack_t            repack;
-        int64_t                  blck_size_interleave; // + interleave elements in blocks
-        ggml_from_float_to_mat_t from_float_to_mat;    // + mis sur le vec_dot_type ... quantize_mat_q8_0
-        enum ggml_type           vec_dot_type;         // +
-        int64_t                  nrows;                // ? number of rows to process simultaneously
-        int64_t                  ncols;                // ? number of columns to process simultaneously
-        ggml_gemv_t              gemv;                 // +
-        ggml_gemm_t              gemm;                 // +
-    };
 
-    const struct ggml_cpu_tensor_traits* ggml_cpu_get_tensor_traits(const struct ggml_tensor * tensor);
+// implementé dans ggml-cpu.cpp.
+std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type();
 
-#ifdef  __cplusplus
-}
 #endif