llama-cpp-capacitor 0.0.6 → 0.0.8

package/cpp/ggml-cpu/ggml-cpu-impl.h

@@ -0,0 +1,525 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct lm_ggml_compute_params {
+    // ith = thread index, nth = number of threads
+    int ith, nth;
+
+    // work buffer for all threads
+    size_t wsize;
+    void * wdata;
+
+    struct lm_ggml_threadpool * threadpool;
+};
+
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__s390x__) && defined(__VEC__)
+#ifndef __VXE__
+#define __VXE__
+#endif  // __VXE__
+#ifndef __VXE2__
+#define __VXE2__
+#endif  // __VXE2__
+#endif  // __s390x__ && __VEC__
+
+#if defined(__s390x__) && defined(LM_GGML_NNPA)
+#ifndef __NNPA__
+#define __NNPA__
+#endif  // __NNPA__
+#endif  // __s390x__ && LM_GGML_NNPA
+
+#if defined(__ARM_FEATURE_SVE)
+#include <sys/prctl.h>
+#endif
+
+#if defined(__ARM_NEON)
+
+// ref: https://github.com/ggml-org/llama.cpp/pull/5404
+#ifdef _MSC_VER
+#define lm_ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+#else
+#define lm_ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct lm_ggml_int16x8x2_t {
+    int16x8_t val[2];
+} lm_ggml_int16x8x2_t;
+
+inline static lm_ggml_int16x8x2_t lm_ggml_vld1q_s16_x2(const int16_t * ptr) {
+    lm_ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct lm_ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} lm_ggml_uint8x16x2_t;
+
+inline static lm_ggml_uint8x16x2_t lm_ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    lm_ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct lm_ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} lm_ggml_uint8x16x4_t;
+
+inline static lm_ggml_uint8x16x4_t lm_ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    lm_ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct lm_ggml_int8x16x2_t {
+    int8x16_t val[2];
+} lm_ggml_int8x16x2_t;
+
+inline static lm_ggml_int8x16x2_t lm_ggml_vld1q_s8_x2(const int8_t * ptr) {
+    lm_ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct lm_ggml_int8x16x4_t {
+    int8x16_t val[4];
+} lm_ggml_int8x16x4_t;
+
+inline static lm_ggml_int8x16x4_t lm_ggml_vld1q_s8_x4(const int8_t * ptr) {
+    lm_ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t lm_ggml_vqtbl1q_s8(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;
+}
+
+// NOTE: not tested
+inline static uint8x16_t lm_ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_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;
+}
+
+#else
+
+#define lm_ggml_int16x8x2_t  int16x8x2_t
+#define lm_ggml_uint8x16x2_t uint8x16x2_t
+#define lm_ggml_uint8x16x4_t uint8x16x4_t
+#define lm_ggml_int8x16x2_t  int8x16x2_t
+#define lm_ggml_int8x16x4_t  int8x16x4_t
+
+#define lm_ggml_vld1q_s16_x2 vld1q_s16_x2
+#define lm_ggml_vld1q_u8_x2  vld1q_u8_x2
+#define lm_ggml_vld1q_u8_x4  vld1q_u8_x4
+#define lm_ggml_vld1q_s8_x2  vld1q_s8_x2
+#define lm_ggml_vld1q_s8_x4  vld1q_s8_x4
+#define lm_ggml_vqtbl1q_s8   vqtbl1q_s8
+#define lm_ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t lm_ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define lm_ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#endif
+
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#endif
+
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#elif defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#include <immintrin.h>
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#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 lm_ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} lm_ggml_uint8x16x2_t;
+
+inline static lm_ggml_uint8x16x2_t lm_ggml_vec_xl_u8x2(const uint8_t * ptr) {
+    lm_ggml_uint8x16x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+typedef struct lm_ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} lm_ggml_uint8x16x4_t;
+
+inline static lm_ggml_uint8x16x4_t lm_ggml_vec_xl_u8x4(const uint8_t * ptr) {
+    lm_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 lm_ggml_int8x16x4_t {
+    int8x16_t val[4];
+} lm_ggml_int8x16x4_t;
+
+inline static lm_ggml_int8x16x4_t lm_ggml_vec_xl_s8x4(const int8_t * ptr) {
+    lm_ggml_int8x16x4_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 lm_ggml_int16x8x2_t {
+    int16x8_t val[2];
+} lm_ggml_int16x8x2_t;
+
+inline static lm_ggml_int16x8x2_t lm_ggml_vec_xl_s16x2(const int16_t * ptr) {
+    lm_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 lm_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;
+}
+
+/**
+ * @see https://github.com/ggml-org/llama.cpp/pull/14037
+ */
+inline float vec_hsum(float32x4_t v) {
+    float32x4_t v_temp = v + vec_reve(v);
+    return v_temp[0] + v_temp[1];
+}
+
+inline static int32x4_t lm_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(const float val) {
+    v4f32 res = {val, val, val, val};
+    return (__m128)res;
+}
+
+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 lm_ggml_barrier(struct lm_ggml_threadpool * tp);
+
+void lm_ggml_threadpool_chunk_set(struct lm_ggml_threadpool * tp, int value);
+int  lm_ggml_threadpool_chunk_add(struct lm_ggml_threadpool * tp, int value);
+
+#ifdef __cplusplus
+}
+#endif