sync up with latest wasi-nn spec (#3530)

Author: lum1n0us

core/iwasm/libraries/wasi-nn/README.md

@@ -21,9 +21,21 @@ $ cmake -DWAMR_BUILD_WASI_NN=1 <other options> ...
 
 ### Wasm
 
-The definition of functions provided by WASI-NN (Wasm imports) is in the header file _core/iwasm/libraries/wasi-nn/wasi_nn.h_.
+The definition of functions provided by WASI-NN (Wasm imports) is in the header file [wasi_nn.h](_core/iwasm/libraries/wasi-nn/wasi_nn.h_). By only including this file in a WASM application you will bind WASI-NN into your module.
 
-By only including this file in a WASM application you will bind WASI-NN into your module.
+For some historical reasons, there are two sets of functions in the header file. The first set is the original one, and the second set is the new one. The new set is recommended to use. In code, `WASM_ENABLE_WASI_EPHEMERAL_NN` is used to control which set of functions to use. If `WASM_ENABLE_WASI_EPHEMERAL_NN` is defined, the new set of functions will be used. Otherwise, the original set of functions will be used.
+
+There is a big difference between the two sets of functions, `tensor_type`.
+
+``` c
+#if WASM_ENABLE_WASI_EPHEMERAL_NN != 0
+typedef enum { fp16 = 0, fp32, fp64, bf16, u8, i32, i64 } tensor_type;
+#else
+typedef enum { fp16 = 0, fp32, up8, ip32 } tensor_type;
+#endif /* WASM_ENABLE_WASI_EPHEMERAL_NN != 0 */
+```
+
+It is required to recompile the Wasm application if you want to switch between the two sets of functions.
 
 ## Tests
 
@@ -41,7 +53,10 @@ Build the runtime image for your execution target type.
 - `tpu`
 
 ```bash
-EXECUTION_TYPE=cpu docker build -t wasi-nn-${EXECUTION_TYPE} -f core/iwasm/libraries/wasi-nn/test/Dockerfile.${EXECUTION_TYPE} .
+$ pwd
+<somewhere>/wasm-micro-runtime
+
+$ EXECUTION_TYPE=cpu docker build -t wasi-nn-${EXECUTION_TYPE} -f core/iwasm/libraries/wasi-nn/test/Dockerfile.${EXECUTION_TYPE} .
 ```
 
 ### Build wasm app

core/iwasm/libraries/wasi-nn/include/wasi_nn.h

@@ -29,6 +29,10 @@ load(graph_builder_array *builder, graph_encoding encoding,
      execution_target target, graph *g)
     __attribute__((import_module("wasi_nn")));
 
+wasi_nn_error
+load_by_name(const char *name, graph *g)
+    __attribute__((import_module("wasi_nn")));
+
 /**
  * INFERENCE
  *

core/iwasm/libraries/wasi-nn/include/wasi_nn_types.h

@@ -18,20 +18,31 @@ extern "C" {
  *
  */
 
-// Error codes returned by functions in this API.
+// sync up with
+// https://github.com/WebAssembly/wasi-nn/blob/main/wit/wasi-nn.wit#L136 Error
+// codes returned by functions in this API.
 typedef enum {
     // No error occurred.
     success = 0,
     // Caller module passed an invalid argument.
     invalid_argument,
     // Invalid encoding.
     invalid_encoding,
-    // Caller module is missing a memory export.
-    missing_memory,
-    // Device or resource busy.
-    busy,
+    // The operation timed out.
+    timeout,
     // Runtime Error.
     runtime_error,
+    // Unsupported operation.
+    unsupported_operation,
+    // Graph is too large.
+    too_large,
+    // Graph not found.
+    not_found,
+    // The operation is insecure or has insufficient privilege to be performed.
+    // e.g., cannot access a hardware feature requested
+    security,
+    // The operation failed for an unspecified reason.
+    unknown,
 } wasi_nn_error;
 
 /**
@@ -48,8 +59,14 @@ typedef struct {
     uint32_t size;
 } tensor_dimensions;
 
+#if WASM_ENABLE_WASI_EPHEMERAL_NN != 0
+// sync up with
+// https://github.com/WebAssembly/wasi-nn/blob/main/wit/wasi-nn.wit#L27
 // The type of the elements in a tensor.
+typedef enum { fp16 = 0, fp32, fp64, bf16, u8, i32, i64 } tensor_type;
+#else
 typedef enum { fp16 = 0, fp32, up8, ip32 } tensor_type;
+#endif /* WASM_ENABLE_WASI_EPHEMERAL_NN != 0 */
 
 // The tensor data.
 //
@@ -96,6 +113,8 @@ typedef struct {
 // An execution graph for performing inference (i.e., a model).
 typedef uint32_t graph;
 
+// sync up with
+// https://github.com/WebAssembly/wasi-nn/blob/main/wit/wasi-nn.wit#L75
 // Describes the encoding of the graph. This allows the API to be implemented by
 // various backends that encode (i.e., serialize) their graph IR with different
 // formats.
@@ -105,7 +124,8 @@ typedef enum {
     tensorflow,
     pytorch,
     tensorflowlite,
-    backend_amount
+    ggml,
+    autodetect,
 } graph_encoding;
 
 // Define where the graph should be executed.
@@ -118,6 +138,7 @@ typedef uint32_t graph_execution_context;
 
 typedef wasi_nn_error (*LOAD)(void *, graph_builder_array *, graph_encoding,
                               execution_target, graph *);
+typedef wasi_nn_error (*LOAD_BY_NAME)(void *, const char *, uint32_t, graph *);
 typedef wasi_nn_error (*INIT_EXECUTION_CONTEXT)(void *, graph,
                                                 graph_execution_context *);
 typedef wasi_nn_error (*SET_INPUT)(void *, graph_execution_context, uint32_t,
@@ -126,11 +147,12 @@ typedef wasi_nn_error (*COMPUTE)(void *, graph_execution_context);
 typedef wasi_nn_error (*GET_OUTPUT)(void *, graph_execution_context, uint32_t,
                                     tensor_data, uint32_t *);
 /* wasi-nn general APIs */
-typedef void (*BACKEND_INITIALIZE)(void **);
-typedef void (*BACKEND_DEINITIALIZE)(void *);
+typedef wasi_nn_error (*BACKEND_INITIALIZE)(void **);
+typedef wasi_nn_error (*BACKEND_DEINITIALIZE)(void *);
 
 typedef struct {
     LOAD load;
+    LOAD_BY_NAME load_by_name;
     INIT_EXECUTION_CONTEXT init_execution_context;
     SET_INPUT set_input;
     COMPUTE compute;
@@ -140,7 +162,7 @@ typedef struct {
 } api_function;
 
 bool
-wasi_nn_register_backend(graph_encoding backend_code, api_function apis);
+wasi_nn_register_backend(api_function apis);
 
 #ifdef __cplusplus
 }

core/iwasm/libraries/wasi-nn/src/utils/wasi_nn_app_native.c

@@ -76,7 +76,7 @@ graph_builder_array_app_native(wasm_module_inst_t instance,
     graph_builder *builder = (graph_builder *)wasm_runtime_malloc(
         array_size * sizeof(graph_builder));
     if (builder == NULL)
-        return missing_memory;
+        return too_large;
 
     for (uint32_t i = 0; i < array_size; ++i) {
         wasi_nn_error res;
@@ -149,7 +149,7 @@ tensor_dimensions_app_native(wasm_module_inst_t instance,
     *dimensions =
         (tensor_dimensions *)wasm_runtime_malloc(sizeof(tensor_dimensions));
     if (dimensions == NULL)
-        return missing_memory;
+        return too_large;
 
     (*dimensions)->size = dimensions_wasm->size;
     (*dimensions)->buf = (uint32_t *)wasm_runtime_addr_app_to_native(