whisper.rn 0.4.0-rc.9 → 0.4.0

package/cpp/ggml-cpu/quants.h

@@ -0,0 +1,89 @@
+#pragma once
+
+#define WSP_GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void wsp_quantize_row_q4_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q4_1(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q5_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q5_1(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q8_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q8_1(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+
+void wsp_quantize_row_q2_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q3_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q4_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q5_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q6_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_q8_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+
+void wsp_quantize_row_tq1_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_tq2_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+
+void wsp_quantize_row_iq4_nl (const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_quantize_row_iq4_xs (const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void wsp_ggml_vec_dot_q4_0_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q4_1_q8_1(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q5_0_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q5_1_q8_1(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q8_0_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
+void wsp_ggml_vec_dot_q2_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q3_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q4_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q5_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q6_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
+void wsp_ggml_vec_dot_tq1_0_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_tq2_0_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
+void wsp_ggml_vec_dot_iq2_xxs_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq2_xs_q8_K (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq2_s_q8_K  (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq3_xxs_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq1_s_q8_K  (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq1_m_q8_K  (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq4_nl_q8_0 (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq4_xs_q8_K (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq3_s_q8_K  (int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
+// Generic implementation
+void wsp_quantize_row_q8_0_generic(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT vy, int64_t k);
+void wsp_quantize_row_q8_1_generic(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT vy, int64_t k);
+void wsp_quantize_row_q8_K_generic(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+void wsp_ggml_vec_dot_q4_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q4_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q5_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q5_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q8_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_tq1_0_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_tq2_0_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q2_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q3_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q4_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_q5_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy,  size_t by, int nrc);
+void wsp_ggml_vec_dot_q6_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq2_s_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq3_s_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq1_s_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq1_m_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+void wsp_ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
+#ifdef __cplusplus
+}
+#endif