whisper.rn 0.4.0-rc.8 → 0.4.0

package/cpp/ggml-quants.h

@@ -1,262 +1,100 @@
 #pragma once
 
-#include "ggml-impl.h"
+#define WSP_GGML_COMMON_DECL_C
+#include "ggml-common.h"
 
-// GGML internal header
-
-#include <stdint.h>
-#include <stddef.h>
-
-#define QK4_0 32
-typedef struct {
-    wsp_ggml_fp16_t d;          // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(wsp_ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    wsp_ggml_fp16_t d;          // delta
-    wsp_ggml_fp16_t m;          // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(wsp_ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    wsp_ggml_fp16_t d;         // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(wsp_ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    wsp_ggml_fp16_t d;         // delta
-    wsp_ggml_fp16_t m;         // min
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(wsp_ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    wsp_ggml_fp16_t d;         // delta
-    int8_t  qs[QK8_0];     // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(wsp_ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    float d;               // delta
-    float s;               // d * sum(qs[i])
-    int8_t  qs[QK8_1];     // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// Super-block size
-#ifdef WSP_GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    wsp_ggml_fp16_t d;           // super-block scale for quantized scales
-    wsp_ggml_fp16_t dmin;        // super-block scale for quantized mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(wsp_ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef WSP_GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[2];
-    wsp_ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(wsp_ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[12];        // scales, quantized with 6 bits
-    wsp_ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(wsp_ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
+#include "ggml.h"
 
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef WSP_GGML_QKK_64
-typedef struct {
-    wsp_ggml_fp16_t d[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(wsp_ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    wsp_ggml_fp16_t d;             // super-block scale for quantized scales
-    wsp_ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(wsp_ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
+// GGML internal header
 
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef WSP_GGML_QKK_64
-typedef struct {
-    wsp_ggml_fp16_t d;               // super-block scale
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(wsp_ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    wsp_ggml_fp16_t d;               // super-block scale for quantized scales
-    wsp_ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(wsp_ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#ifdef __cplusplus
+extern "C" {
 #endif
 
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    wsp_ggml_fp16_t d;           // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(wsp_ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-// (Almost) "true" 2-bit quantization.
-// Due to the need to use blocks as per ggml dsign, it ends up using
-// 2.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    wsp_ggml_fp16_t d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(wsp_ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-// 2.3125 bpw quants
-typedef struct {
-    wsp_ggml_fp16_t d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(wsp_ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
+// NOTE: these functions are defined as WSP_GGML_API because they used by the CPU backend
 
 // Quantization
-void wsp_quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
-void wsp_quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
-void wsp_quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
-void wsp_quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
-void wsp_quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
-void wsp_quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
-
-void wsp_quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
-void wsp_quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
-void wsp_quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
-void wsp_quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
-void wsp_quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
-void wsp_quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
-
-void wsp_quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
-
-void wsp_quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
-void wsp_quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+WSP_GGML_API void wsp_quantize_row_q4_0_ref(const float * WSP_GGML_RESTRICT x, block_q4_0 * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q4_1_ref(const float * WSP_GGML_RESTRICT x, block_q4_1 * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q5_0_ref(const float * WSP_GGML_RESTRICT x, block_q5_0 * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q5_1_ref(const float * WSP_GGML_RESTRICT x, block_q5_1 * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q8_0_ref(const float * WSP_GGML_RESTRICT x, block_q8_0 * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q8_1_ref(const float * WSP_GGML_RESTRICT x, block_q8_1 * WSP_GGML_RESTRICT y, int64_t k);
+
+WSP_GGML_API void wsp_quantize_row_q2_K_ref(const float * WSP_GGML_RESTRICT x, block_q2_K * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q3_K_ref(const float * WSP_GGML_RESTRICT x, block_q3_K * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q4_K_ref(const float * WSP_GGML_RESTRICT x, block_q4_K * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q5_K_ref(const float * WSP_GGML_RESTRICT x, block_q5_K * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q6_K_ref(const float * WSP_GGML_RESTRICT x, block_q6_K * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_q8_K_ref(const float * WSP_GGML_RESTRICT x, block_q8_K * WSP_GGML_RESTRICT y, int64_t k);
+
+WSP_GGML_API void wsp_quantize_row_tq1_0_ref(const float * WSP_GGML_RESTRICT x, block_tq1_0 * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_tq2_0_ref(const float * WSP_GGML_RESTRICT x, block_tq2_0 * WSP_GGML_RESTRICT y, int64_t k);
+
+WSP_GGML_API void wsp_quantize_row_iq3_xxs_ref(const float * WSP_GGML_RESTRICT x, block_iq3_xxs * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_iq4_nl_ref (const float * WSP_GGML_RESTRICT x, block_iq4_nl  * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_iq4_xs_ref (const float * WSP_GGML_RESTRICT x, block_iq4_xs  * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_iq3_s_ref  (const float * WSP_GGML_RESTRICT x, block_iq3_s   * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_quantize_row_iq2_s_ref  (const float * WSP_GGML_RESTRICT x, block_iq2_s   * WSP_GGML_RESTRICT y, int64_t k);
 
 // Dequantization
-void wsp_dewsp_quantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
-//void wsp_dewsp_quantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
-
-void wsp_dewsp_quantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k);
-void wsp_dewsp_quantize_row_iq2_xs (const block_iq2_xs  * restrict x, float * restrict y, int k);
-
-// Dot product
-void wsp_ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+WSP_GGML_API void wsp_dewsp_quantize_row_q4_0(const block_q4_0 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q4_1(const block_q4_1 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q5_0(const block_q5_0 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q5_1(const block_q5_1 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q8_0(const block_q8_0 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+//WSP_GGML_API void wsp_dewsp_quantize_row_q8_1(const block_q8_1 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+
+WSP_GGML_API void wsp_dewsp_quantize_row_q2_K(const block_q2_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q3_K(const block_q3_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q4_K(const block_q4_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q5_K(const block_q5_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q6_K(const block_q6_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_q8_K(const block_q8_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+
+WSP_GGML_API void wsp_dewsp_quantize_row_tq1_0(const block_tq1_0 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_tq2_0(const block_tq2_0 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+
+WSP_GGML_API void wsp_dewsp_quantize_row_iq2_xxs(const block_iq2_xxs * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq2_xs (const block_iq2_xs  * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq2_s  (const block_iq2_s   * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq3_xxs(const block_iq3_xxs * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq1_s  (const block_iq1_s   * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq1_m  (const block_iq1_m   * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq4_nl (const block_iq4_nl  * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq4_xs (const block_iq4_xs  * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+WSP_GGML_API void wsp_dewsp_quantize_row_iq3_s  (const block_iq3_s   * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
 
-void wsp_ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void wsp_ggml_vec_dot_iq2_xs_q8_K (int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-//
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-//
-size_t wsp_quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q3_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q4_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q5_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q6_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q4_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q4_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q5_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t wsp_quantize_q5_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-
-void iq2xs_init_impl(int grid_size);
-void iq2xs_free_impl(int grid_size);
+WSP_GGML_API size_t wsp_quantize_iq2_xxs(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq2_xs (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq2_s  (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq3_xxs(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq1_s  (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq1_m  (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq4_nl (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq4_xs (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_iq3_s  (const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+WSP_GGML_API size_t wsp_quantize_tq1_0(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_tq2_0(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+WSP_GGML_API size_t wsp_quantize_q2_K(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q3_K(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q4_K(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q5_K(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q6_K(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q4_0(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q4_1(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q5_0(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q5_1(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+WSP_GGML_API size_t wsp_quantize_q8_0(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+WSP_GGML_API void wsp_iq2xs_init_impl(enum wsp_ggml_type type);
+WSP_GGML_API void wsp_iq2xs_free_impl(enum wsp_ggml_type type);
+WSP_GGML_API void wsp_iq3xs_init_impl(int grid_size);
+WSP_GGML_API void wsp_iq3xs_free_impl(int grid_size);
+
+#ifdef __cplusplus
+}
+#endif

package/cpp/ggml-threading.cpp

@@ -0,0 +1,12 @@
+#include "ggml-threading.h"
+#include <mutex>
+
+std::mutex wsp_ggml_critical_section_mutex;
+
+void wsp_ggml_critical_section_start() {
+    wsp_ggml_critical_section_mutex.lock();
+}
+
+void wsp_ggml_critical_section_end(void) {
+    wsp_ggml_critical_section_mutex.unlock();
+}

package/cpp/ggml-threading.h

@@ -0,0 +1,14 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+WSP_GGML_API void wsp_ggml_critical_section_start(void);
+WSP_GGML_API void wsp_ggml_critical_section_end(void);
+
+#ifdef __cplusplus
+}
+#endif