whisper.rn 0.4.0-rc.10 → 0.4.0-rc.11

package/cpp/ggml-cpu-aarch64.h

@@ -1,30 +1,8 @@
 #pragma once
 
+#include "ggml-cpu-traits.h"
 #include "ggml.h"
 
 // GGML internal header
 
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void wsp_quantize_mat_q8_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
-// GEMV
-void wsp_ggml_gemv_q4_0_4x4_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, const void * WSP_GGML_RESTRICT vy, int nr, int nc);
-void wsp_ggml_gemv_q4_0_4x8_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, const void * WSP_GGML_RESTRICT vy, int nr, int nc);
-void wsp_ggml_gemv_q4_0_8x8_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, const void * WSP_GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void wsp_ggml_gemm_q4_0_4x4_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, const void * WSP_GGML_RESTRICT vy, int nr, int nc);
-void wsp_ggml_gemm_q4_0_4x8_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, const void * WSP_GGML_RESTRICT vy, int nr, int nc);
-void wsp_ggml_gemm_q4_0_8x8_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, const void * WSP_GGML_RESTRICT vy, int nr, int nc);
-
-void           wsp_ggml_aarch64_repack_tensor(struct wsp_ggml_tensor * cur, enum wsp_ggml_type repack_type, const void * data, size_t data_size);
-enum wsp_ggml_type wsp_ggml_aarch64_get_optimal_repack_type(const struct wsp_ggml_tensor * cur);
-
-#ifdef __cplusplus
-}
-#endif
-
+wsp_ggml_backend_buffer_type_t wsp_ggml_backend_cpu_aarch64_buffer_type(void);

package/cpp/ggml-cpu-impl.h

@@ -15,6 +15,18 @@
 extern "C" {
 #endif
 
+struct wsp_ggml_compute_params {
+    // ith = thread index, nth = number of threads
+    int ith, nth;
+
+    // work buffer for all threads
+    size_t wsize;
+    void * wdata;
+
+    struct wsp_ggml_threadpool * threadpool;
+};
+
+
 #if defined(_MSC_VER)
 
 #define m512bh(p) p
@@ -47,6 +59,15 @@ extern "C" {
 #endif
 #endif
 
+#if defined(__s390x__) && defined(__VEC__)
+#ifndef __VXE__
+#define __VXE__
+#endif
+#ifndef __VXE2__
+#define __VXE2__
+#endif
+#endif
+
 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
 #include <sys/prctl.h>
@@ -347,25 +368,164 @@ inline static int32x4_t wsp_ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t
 #endif
 #endif
 
-#if defined(__loongarch_asx)
+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+
+#define vec_neg(a)    (-(a))                // Vector Negate
+#define vec_add(a, b) ((a) + (b))           // Vector Add
+#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
+#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
+#define vec_div(a, b) ((a) / (b))           // Vector Divide
+#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
+#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b)  ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef signed char char8x16_t __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t  int8x16_t __attribute__((vector_size(16)));
+typedef int16_t int16x8_t __attribute__((vector_size(16)));
+typedef int32_t int32x4_t __attribute__((vector_size(16)));
+
+typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute((vector_size(16)));
+
+typedef signed long long long64x2_t __attribute((vector_size(16)));
+typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
+
+typedef struct wsp_ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} wsp_ggml_uint8x16x2_t;
+
+inline static wsp_ggml_uint8x16x2_t wsp_ggml_vec_xl_u8x2(const uint8_t * ptr) {
+    wsp_ggml_uint8x16x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+typedef struct wsp_ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} wsp_ggml_uint8x16x4_t;
+
+inline static wsp_ggml_uint8x16x4_t wsp_ggml_vec_xl_u8x4(const uint8_t * ptr) {
+    wsp_ggml_uint8x16x4_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+    res.val[2] = vec_xl(32, ptr);
+    res.val[3] = vec_xl(48, ptr);
+
+    return res;
+}
+
+typedef struct wsp_ggml_int8x16x4_t {
+    int8x16_t val[4];
+} wsp_ggml_int8x16x4_t;
+
+inline static wsp_ggml_int8x16x4_t wsp_ggml_vec_xl_s8x4(const int8_t * ptr) {
+    wsp_ggml_int8x16x4_t res;
 
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+    res.val[2] = vec_xl(32, ptr);
+    res.val[3] = vec_xl(48, ptr);
 
+    return res;
+}
+
+typedef struct wsp_ggml_int16x8x2_t {
+    int16x8_t val[2];
+} wsp_ggml_int16x8x2_t;
+
+inline static wsp_ggml_int16x8x2_t wsp_ggml_vec_xl_s16x2(const int16_t * ptr) {
+    wsp_ggml_int16x8x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+/*
+    ! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
+    !          or iq4_nl for example implementation.
+*/
+inline static int8x16_t wsp_ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
+    const uchar8x16_t v_maske = {  0,  1,  4,  5,  8,  9, 12, 13,
+                                  16, 17, 20, 21, 24, 25, 28, 29 };
+
+    const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
+    const int16x8_t v_abe = vec_perm(a, b, v_maske);
+    return v_abo + v_abe;
+}
+
+inline static int32x4_t wsp_ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
+    return acc + (vec_unpackh(p) + vec_unpackl(p));
+}
+
+#endif
+
+#if defined(__loongarch_asx)
 /* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+static __m128 __lsx_vreplfr2vr_s(const float val) {
+    v4f32 res = {val, val, val, val};
+    return (__m128)res;
 }
 
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+static __m256 __lasx_xvreplfr2vr_s(const float val) {
+    v8f32 res = {val, val, val, val, val, val, val, val};
+    return (__m256)res;
 }
 #endif
 
+// TODO: move to ggml-threading
+void wsp_ggml_barrier(struct wsp_ggml_threadpool * tp);
+
 #ifdef __cplusplus
 }
 #endif