whisper.rn 0.4.0-rc.9 → 0.4.1

package/cpp/ggml-cpu.h

@@ -0,0 +1,143 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    // the compute plan that needs to be prepared for wsp_ggml_graph_compute()
+    // since https://github.com/ggml-org/ggml/issues/287
+    struct wsp_ggml_cplan {
+        size_t    work_size; // size of work buffer, calculated by `wsp_ggml_graph_plan()`
+        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `wsp_ggml_graph_compute()`
+
+        int n_threads;
+        struct wsp_ggml_threadpool * threadpool;
+
+        // abort wsp_ggml_graph_compute when true
+        wsp_ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
+    };
+
+    // numa strategies
+    enum wsp_ggml_numa_strategy {
+        WSP_GGML_NUMA_STRATEGY_DISABLED   = 0,
+        WSP_GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        WSP_GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        WSP_GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        WSP_GGML_NUMA_STRATEGY_MIRROR     = 4,
+        WSP_GGML_NUMA_STRATEGY_COUNT
+    };
+
+    WSP_GGML_BACKEND_API void    wsp_ggml_numa_init(enum wsp_ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    WSP_GGML_BACKEND_API bool    wsp_ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+
+    WSP_GGML_BACKEND_API struct wsp_ggml_tensor * wsp_ggml_new_i32(struct wsp_ggml_context * ctx, int32_t value);
+    WSP_GGML_BACKEND_API struct wsp_ggml_tensor * wsp_ggml_new_f32(struct wsp_ggml_context * ctx, float value);
+
+    WSP_GGML_BACKEND_API struct wsp_ggml_tensor * wsp_ggml_set_i32 (struct wsp_ggml_tensor * tensor, int32_t value);
+    WSP_GGML_BACKEND_API struct wsp_ggml_tensor * wsp_ggml_set_f32 (struct wsp_ggml_tensor * tensor, float value);
+
+    WSP_GGML_BACKEND_API int32_t wsp_ggml_get_i32_1d(const struct wsp_ggml_tensor * tensor, int i);
+    WSP_GGML_BACKEND_API void    wsp_ggml_set_i32_1d(const struct wsp_ggml_tensor * tensor, int i, int32_t value);
+
+    WSP_GGML_BACKEND_API int32_t wsp_ggml_get_i32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    WSP_GGML_BACKEND_API void    wsp_ggml_set_i32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
+    WSP_GGML_BACKEND_API float   wsp_ggml_get_f32_1d(const struct wsp_ggml_tensor * tensor, int i);
+    WSP_GGML_BACKEND_API void    wsp_ggml_set_f32_1d(const struct wsp_ggml_tensor * tensor, int i, float value);
+
+    WSP_GGML_BACKEND_API float   wsp_ggml_get_f32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    WSP_GGML_BACKEND_API void    wsp_ggml_set_f32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
+    WSP_GGML_BACKEND_API struct wsp_ggml_threadpool *      wsp_ggml_threadpool_new           (struct wsp_ggml_threadpool_params  * params);
+    WSP_GGML_BACKEND_API void                          wsp_ggml_threadpool_free          (struct wsp_ggml_threadpool * threadpool);
+    WSP_GGML_BACKEND_API int                           wsp_ggml_threadpool_get_n_threads (struct wsp_ggml_threadpool * threadpool);
+    WSP_GGML_BACKEND_API void                          wsp_ggml_threadpool_pause         (struct wsp_ggml_threadpool * threadpool);
+    WSP_GGML_BACKEND_API void                          wsp_ggml_threadpool_resume        (struct wsp_ggml_threadpool * threadpool);
+
+    // wsp_ggml_graph_plan() has to be called before wsp_ggml_graph_compute()
+    // when plan.work_size > 0, caller must allocate memory for plan.work_data
+    WSP_GGML_BACKEND_API struct wsp_ggml_cplan wsp_ggml_graph_plan(
+                  const struct wsp_ggml_cgraph * cgraph,
+                                       int   n_threads, /* = WSP_GGML_DEFAULT_N_THREADS */
+                    struct wsp_ggml_threadpool * threadpool /* = NULL */ );
+    WSP_GGML_BACKEND_API enum wsp_ggml_status  wsp_ggml_graph_compute(struct wsp_ggml_cgraph * cgraph, struct wsp_ggml_cplan * cplan);
+
+    // same as wsp_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
+    WSP_GGML_BACKEND_API enum wsp_ggml_status  wsp_ggml_graph_compute_with_ctx(struct wsp_ggml_context * ctx, struct wsp_ggml_cgraph * cgraph, int n_threads);
+
+    //
+    // system info
+    //
+
+    // x86
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_sse3       (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_ssse3      (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx        (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx_vnni   (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx2       (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_bmi2       (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_f16c       (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_fma        (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx512     (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx512_vbmi(void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx512_vnni(void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_avx512_bf16(void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_amx_int8   (void);
+    // ARM
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_neon       (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_arm_fma    (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_fp16_va    (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_dotprod    (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_matmul_int8(void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_sve        (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_sme        (void);
+    // other
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_riscv_v    (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_vsx        (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_vxe        (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_wasm_simd  (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_llamafile  (void);
+
+    // Internal types and functions exposed for tests and benchmarks
+
+    typedef void (*wsp_ggml_vec_dot_t)  (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT x, size_t bx,
+                                       const void * WSP_GGML_RESTRICT y, size_t by, int nrc);
+
+    struct wsp_ggml_type_traits_cpu {
+        wsp_ggml_from_float_t        from_float;
+        wsp_ggml_vec_dot_t           vec_dot;
+        enum wsp_ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+    };
+
+    WSP_GGML_BACKEND_API const struct wsp_ggml_type_traits_cpu * wsp_ggml_get_type_traits_cpu(enum wsp_ggml_type type);
+
+    WSP_GGML_BACKEND_API void wsp_ggml_cpu_init(void);
+
+    //
+    // CPU backend
+    //
+
+    WSP_GGML_BACKEND_API wsp_ggml_backend_t wsp_ggml_backend_cpu_init(void);
+
+    WSP_GGML_BACKEND_API bool wsp_ggml_backend_is_cpu                (wsp_ggml_backend_t backend);
+    WSP_GGML_BACKEND_API void wsp_ggml_backend_cpu_set_n_threads     (wsp_ggml_backend_t backend_cpu, int n_threads);
+    WSP_GGML_BACKEND_API void wsp_ggml_backend_cpu_set_threadpool    (wsp_ggml_backend_t backend_cpu, wsp_ggml_threadpool_t threadpool);
+    WSP_GGML_BACKEND_API void wsp_ggml_backend_cpu_set_abort_callback(wsp_ggml_backend_t backend_cpu, wsp_ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    WSP_GGML_BACKEND_API wsp_ggml_backend_reg_t wsp_ggml_backend_cpu_reg(void);
+
+    WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp32_to_fp16(const float *, wsp_ggml_fp16_t *, int64_t);
+    WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp16_to_fp32(const wsp_ggml_fp16_t *, float *, int64_t);
+    WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp32_to_bf16(const float *, wsp_ggml_bf16_t *, int64_t);
+    WSP_GGML_BACKEND_API void wsp_ggml_cpu_bf16_to_fp32(const wsp_ggml_bf16_t *, float *, int64_t);
+
+#ifdef __cplusplus
+}
+#endif

package/cpp/ggml-impl.h

@@ -3,21 +3,44 @@
 // GGML internal header
 
 #include "ggml.h"
+#include "gguf.h"
 
 #include <assert.h>
+#include <math.h>
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stdbool.h>
 #include <stdint.h>
+#include <string.h>
+
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-#undef MIN
-#undef MAX
+void wsp_ggml_print_backtrace(void);
+
+#ifndef MIN
+#    define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
 
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#ifndef MAX
+#    define MAX(a, b) ((a) > (b) ? (a) : (b))
+#endif
 
 // required for mmap as gguf only guarantees 32-byte alignment
 #define TENSOR_ALIGNMENT 32
@@ -27,22 +50,36 @@ extern "C" {
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
 #ifndef __cplusplus
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
+    #ifndef static_assert
+        #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+            #define static_assert(cond, msg) _Static_assert(cond, msg)
+        #else
+            #define static_assert(cond, msg) struct global_scope_noop_trick
+        #endif
+    #endif
 #endif
 
+static inline int wsp_ggml_up32(int n) {
+    return (n + 31) & ~31;
+}
+
+//static inline int wsp_ggml_up64(int n) {
+//    return (n + 63) & ~63;
+//}
+
+static inline int wsp_ggml_up(int n, int m) {
+    // assert m is a power of 2
+    WSP_GGML_ASSERT((m & (m - 1)) == 0);
+    return (n + m - 1) & ~(m - 1);
+}
+
 //
 // logging
 //
 
 WSP_GGML_ATTRIBUTE_FORMAT(2, 3)
-void wsp_ggml_log_internal        (enum wsp_ggml_log_level level, const char * format, ...);
-void wsp_ggml_log_callback_default(enum wsp_ggml_log_level level, const char * text, void * user_data);
+WSP_GGML_API void wsp_ggml_log_internal        (enum wsp_ggml_log_level level, const char * format, ...);
+WSP_GGML_API void wsp_ggml_log_callback_default(enum wsp_ggml_log_level level, const char * text, void * user_data);
 
 #define WSP_GGML_LOG(...)       wsp_ggml_log_internal(WSP_GGML_LOG_LEVEL_NONE , __VA_ARGS__)
 #define WSP_GGML_LOG_INFO(...)  wsp_ggml_log_internal(WSP_GGML_LOG_LEVEL_INFO , __VA_ARGS__)
@@ -51,6 +88,78 @@ void wsp_ggml_log_callback_default(enum wsp_ggml_log_level level, const char * t
 #define WSP_GGML_LOG_DEBUG(...) wsp_ggml_log_internal(WSP_GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
 #define WSP_GGML_LOG_CONT(...)  wsp_ggml_log_internal(WSP_GGML_LOG_LEVEL_CONT , __VA_ARGS__)
 
+#define WSP_GGML_DEBUG 0
+
+#if (WSP_GGML_DEBUG >= 1)
+#define WSP_GGML_PRINT_DEBUG(...) WSP_GGML_LOG_DEBUG(__VA_ARGS__)
+#else
+#define WSP_GGML_PRINT_DEBUG(...)
+#endif
+
+#if (WSP_GGML_DEBUG >= 5)
+#define WSP_GGML_PRINT_DEBUG_5(...) WSP_GGML_LOG_DEBUG(__VA_ARGS__)
+#else
+#define WSP_GGML_PRINT_DEBUG_5(...)
+#endif
+
+#if (WSP_GGML_DEBUG >= 10)
+#define WSP_GGML_PRINT_DEBUG_10(...) WSP_GGML_LOG_DEBUG(__VA_ARGS__)
+#else
+#define WSP_GGML_PRINT_DEBUG_10(...)
+#endif
+
+// tensor params
+
+static void wsp_ggml_set_op_params(struct wsp_ggml_tensor * tensor, const void * params, size_t params_size) {
+    WSP_GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
+    assert(params_size <= WSP_GGML_MAX_OP_PARAMS);
+    memcpy(tensor->op_params, params, params_size);
+}
+
+static int32_t wsp_ggml_get_op_params_i32(const struct wsp_ggml_tensor * tensor, uint32_t i) {
+    assert(i < WSP_GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    return ((const int32_t *)(tensor->op_params))[i];
+}
+
+static float wsp_ggml_get_op_params_f32(const struct wsp_ggml_tensor * tensor, uint32_t i) {
+    assert(i < WSP_GGML_MAX_OP_PARAMS / sizeof(float));
+    return ((const float *)(tensor->op_params))[i];
+}
+
+static void wsp_ggml_set_op_params_i32(struct wsp_ggml_tensor * tensor, uint32_t i, int32_t value) {
+    assert(i < WSP_GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    ((int32_t *)(tensor->op_params))[i] = value;
+}
+
+static void wsp_ggml_set_op_params_f32(struct wsp_ggml_tensor * tensor, uint32_t i, float value) {
+    assert(i < WSP_GGML_MAX_OP_PARAMS / sizeof(float));
+    ((float *)(tensor->op_params))[i] = value;
+}
+
+struct wsp_ggml_map_custom1_op_params {
+    wsp_ggml_custom1_op_t  fun;
+    int                n_tasks;
+    void             * userdata;
+};
+
+struct wsp_ggml_map_custom2_op_params {
+    wsp_ggml_custom2_op_t   fun;
+    int                 n_tasks;
+    void              * userdata;
+};
+
+struct wsp_ggml_map_custom3_op_params {
+    wsp_ggml_custom3_op_t fun;
+    int               n_tasks;
+    void            * userdata;
+};
+
+struct wsp_ggml_custom_op_params {
+    wsp_ggml_custom_op_t fun;
+    int              n_tasks;
+    void           * userdata;
+};
+
 // bitset
 
 typedef uint32_t wsp_ggml_bitset_t;
@@ -99,7 +208,7 @@ void wsp_ggml_hash_set_reset(struct wsp_ggml_hash_set * hash_set);
 static bool wsp_ggml_hash_contains(const struct wsp_ggml_hash_set * hash_set, struct wsp_ggml_tensor * key);
 
 // returns WSP_GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
-static size_t wsp_ggml_hash_find(const struct wsp_ggml_hash_set * hash_set, struct wsp_ggml_tensor * key);
+static size_t wsp_ggml_hash_find(const struct wsp_ggml_hash_set * hash_set, const struct wsp_ggml_tensor * key);
 
 // returns WSP_GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 static size_t wsp_ggml_hash_insert(struct wsp_ggml_hash_set * hash_set, struct wsp_ggml_tensor * key);
@@ -113,7 +222,7 @@ static inline size_t wsp_ggml_hash(const struct wsp_ggml_tensor * p) {
     return (size_t)(uintptr_t)p >> 4;
 }
 
-static size_t wsp_ggml_hash_find(const struct wsp_ggml_hash_set * hash_set, struct wsp_ggml_tensor * key) {
+static size_t wsp_ggml_hash_find(const struct wsp_ggml_hash_set * hash_set, const struct wsp_ggml_tensor * key) {
     size_t h = wsp_ggml_hash(key) % hash_set->size;
 
     // linear probing
@@ -184,26 +293,311 @@ enum wsp_ggml_cgraph_eval_order {
 };
 
 struct wsp_ggml_cgraph {
-    int size;
-    int n_nodes;
-    int n_leafs;
+    int size;    // maximum number of nodes/leafs/grads/grad_accs
+    int n_nodes; // number of nodes currently in use
+    int n_leafs; // number of leafs currently in use
 
-    struct wsp_ggml_tensor ** nodes;
-    struct wsp_ggml_tensor ** grads;
-    struct wsp_ggml_tensor ** leafs;
+    struct wsp_ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
+    struct wsp_ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
+    struct wsp_ggml_tensor ** grad_accs; // accumulators for node gradients
+    struct wsp_ggml_tensor ** leafs;     // tensors with constant data
 
     struct wsp_ggml_hash_set visited_hash_set;
 
     enum wsp_ggml_cgraph_eval_order order;
 };
 
+// returns a slice of cgraph with nodes [i0, i1)
+// the slice does not have leafs or gradients
+// if you need the gradients, get them from the original graph
 struct wsp_ggml_cgraph wsp_ggml_graph_view(struct wsp_ggml_cgraph * cgraph, int i0, int i1);
 
 // Memory allocation
 
-void * wsp_ggml_aligned_malloc(size_t size);
-void wsp_ggml_aligned_free(void * ptr, size_t size);
+WSP_GGML_API void * wsp_ggml_aligned_malloc(size_t size);
+WSP_GGML_API void wsp_ggml_aligned_free(void * ptr, size_t size);
+
+// FP16 to FP32 conversion
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+//
+// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
+// for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
+//
+#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+    #define WSP_GGML_COMPUTE_FP16_TO_FP32(x) wsp_ggml_compute_fp16_to_fp32(x)
+    #define WSP_GGML_COMPUTE_FP32_TO_FP16(x) wsp_ggml_compute_fp32_to_fp16(x)
+
+    #define WSP_GGML_FP16_TO_FP32(x) wsp_ggml_compute_fp16_to_fp32(x)
+
+    static inline float wsp_ggml_compute_fp16_to_fp32(wsp_ggml_fp16_t h) {
+        __fp16 tmp;
+        memcpy(&tmp, &h, sizeof(wsp_ggml_fp16_t));
+        return (float)tmp;
+    }
+
+    static inline wsp_ggml_fp16_t wsp_ggml_compute_fp32_to_fp16(float f) {
+        wsp_ggml_fp16_t res;
+        __fp16 tmp = f;
+        memcpy(&res, &tmp, sizeof(wsp_ggml_fp16_t));
+        return res;
+    }
+
+#elif defined(__F16C__)
+
+    #ifdef _MSC_VER
+        #define WSP_GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+        #define WSP_GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+    #else
+        #define WSP_GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+        #define WSP_GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+    #endif
+
+#elif defined(__POWER9_VECTOR__)
+
+    #define WSP_GGML_COMPUTE_FP16_TO_FP32(x) wsp_ggml_compute_fp16_to_fp32(x)
+    #define WSP_GGML_COMPUTE_FP32_TO_FP16(x) wsp_ggml_compute_fp32_to_fp16(x)
+    /* the inline asm below is about 12% faster than the lookup method */
+    #define WSP_GGML_FP16_TO_FP32(x) WSP_GGML_COMPUTE_FP16_TO_FP32(x)
+    #define WSP_GGML_FP32_TO_FP16(x) WSP_GGML_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float wsp_ggml_compute_fp16_to_fp32(wsp_ggml_fp16_t h) {
+        float f;
+        double d;
+        __asm__(
+            "mtfprd %0,%2\n"
+            "xscvhpdp %0,%0\n"
+            "frsp %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=f"(f):
+            /* in */   "r"(h));
+        return f;
+    }
+
+    static inline wsp_ggml_fp16_t wsp_ggml_compute_fp32_to_fp16(float f) {
+        double d;
+        wsp_ggml_fp16_t r;
+        __asm__( /* xscvdphp can work on double or single precision */
+            "xscvdphp %0,%2\n"
+            "mffprd %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=r"(r):
+            /* in */   "f"(f));
+        return r;
+    }
+
+#elif defined(__riscv) && defined(__riscv_zfhmin)
+
+    static inline float wsp_ggml_compute_fp16_to_fp32(wsp_ggml_fp16_t h) {
+        float f;
+        __asm__(
+            "fmv.h.x %[f], %[h]\n\t"
+            "fcvt.s.h %[f], %[f]"
+            : [f] "=&f" (f)
+            : [h] "r" (h)
+        );
+        return f;
+    }
+
+    static inline wsp_ggml_fp16_t wsp_ggml_compute_fp32_to_fp16(float f) {
+        wsp_ggml_fp16_t res;
+        __asm__(
+            "fcvt.h.s %[f], %[f]\n\t"
+            "fmv.x.h %[h], %[f]"
+            : [h] "=&r" (res)
+            : [f] "f" (f)
+        );
+        return res;
+    }
+
+    #define WSP_GGML_COMPUTE_FP16_TO_FP32(x) wsp_ggml_compute_fp16_to_fp32(x)
+    #define WSP_GGML_COMPUTE_FP32_TO_FP16(x) wsp_ggml_compute_fp32_to_fp16(x)
+    #define WSP_GGML_FP16_TO_FP32(x) WSP_GGML_COMPUTE_FP16_TO_FP32(x)
+    #define WSP_GGML_FP32_TO_FP16(x) WSP_GGML_COMPUTE_FP32_TO_FP16(x)
+
+#else
+
+    // FP16 <-> FP32
+    // ref: https://github.com/Maratyszcza/FP16
+
+    static inline float fp32_from_bits(uint32_t w) {
+        union {
+            uint32_t as_bits;
+            float as_value;
+        } fp32;
+        fp32.as_bits = w;
+        return fp32.as_value;
+    }
+
+    static inline uint32_t fp32_to_bits(float f) {
+        union {
+            float as_value;
+            uint32_t as_bits;
+        } fp32;
+        fp32.as_value = f;
+        return fp32.as_bits;
+    }
+
+    static inline float wsp_ggml_compute_fp16_to_fp32(wsp_ggml_fp16_t h) {
+        const uint32_t w = (uint32_t) h << 16;
+        const uint32_t sign = w & UINT32_C(0x80000000);
+        const uint32_t two_w = w + w;
+
+        const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        const float exp_scale = 0x1.0p-112f;
+    #else
+        const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+    #endif
+        const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+        const uint32_t magic_mask = UINT32_C(126) << 23;
+        const float magic_bias = 0.5f;
+        const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+        const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+        const uint32_t result = sign |
+            (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+        return fp32_from_bits(result);
+    }
+
+    static inline wsp_ggml_fp16_t wsp_ggml_compute_fp32_to_fp16(float f) {
+    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        const float scale_to_inf = 0x1.0p+112f;
+        const float scale_to_zero = 0x1.0p-110f;
+    #else
+        const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+        const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+    #endif
+        float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+        const uint32_t w = fp32_to_bits(f);
+        const uint32_t shl1_w = w + w;
+        const uint32_t sign = w & UINT32_C(0x80000000);
+        uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+        if (bias < UINT32_C(0x71000000)) {
+            bias = UINT32_C(0x71000000);
+        }
+
+        base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+        const uint32_t bits = fp32_to_bits(base);
+        const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+        const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+        const uint32_t nonsign = exp_bits + mantissa_bits;
+        return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+    }
+
+    #define WSP_GGML_COMPUTE_FP16_TO_FP32(x) wsp_ggml_compute_fp16_to_fp32(x)
+    #define WSP_GGML_COMPUTE_FP32_TO_FP16(x) wsp_ggml_compute_fp32_to_fp16(x)
+
+#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in wsp_ggml_init()
+WSP_GGML_API float wsp_ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into wsp_ggml_lookup_fp16_to_fp32,
+// so we define WSP_GGML_FP16_TO_FP32 and WSP_GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(WSP_GGML_FP16_TO_FP32)
+inline static float wsp_ggml_lookup_fp16_to_fp32(wsp_ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return wsp_ggml_table_f32_f16[s];
+}
+
+#define WSP_GGML_FP16_TO_FP32(x) wsp_ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(WSP_GGML_FP32_TO_FP16)
+#define WSP_GGML_FP32_TO_FP16(x) WSP_GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float wsp_ggml_compute_bf16_to_fp32(wsp_ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline wsp_ggml_bf16_t wsp_ggml_compute_fp32_to_bf16(float s) {
+    wsp_ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define WSP_GGML_FP32_TO_BF16(x) wsp_ggml_compute_fp32_to_bf16(x)
+#define WSP_GGML_BF16_TO_FP32(x) wsp_ggml_compute_bf16_to_fp32(x)
 
 #ifdef __cplusplus
 }
 #endif
+
+#ifdef __cplusplus
+#include <vector>
+
+// expose GGUF internals for test code
+WSP_GGML_API size_t wsp_gguf_type_size(enum wsp_gguf_type type);
+WSP_GGML_API struct wsp_gguf_context * wsp_gguf_init_from_file_impl(FILE * file, struct wsp_gguf_init_params params);
+WSP_GGML_API void wsp_gguf_write_to_buf(const struct wsp_gguf_context * ctx, std::vector<int8_t> & buf, bool only_meta);
+#endif // __cplusplus