@fugood/llama.node 1.0.0-beta.5 → 1.0.0-beta.7

package/src/llama.cpp/ggml/src/ggml-cpu/repack.h

@@ -0,0 +1,98 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
+
+struct block_q4_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[96];  // scales and mins, quantized with 6 bits
+    uint8_t qs[1024];    // 4--bit quants
+};
+
+static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
+struct block_q8_Kx4 {
+    float d[4];              // delta
+    int8_t qs[QK_K * 4];     // quants
+    int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
+};
+
+static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
+
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif

package/src/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h

@@ -17,7 +17,123 @@
 //   number of elements to fit in a single register
 //
 
-#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_FMA)
+
+#define GGML_SIMD
+
+// F32 SVE
+#define GGML_F32_EPR 8
+#define DEFAULT_PG svptrue_b32()
+
+#define GGML_F32xt                        svfloat32_t
+#define GGML_F32xt_ZERO                   svdup_n_f32(0.0f)
+#define GGML_F32xt_SET1(x)                svdup_n_f32(x)
+#define GGML_F32xt_LOAD_IMPL(pg, a, ...)  svld1_f32(pg, a)
+#define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
+#define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, a, b, c)
+#define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
+#define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_MUL_IMPL(pg, a, b)     svmul_f32_m(pg, a, b)
+#define GGML_F32xt_MUL(...)               GGML_F32xt_MUL_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
+#define GGML_F32xt_REDUCE_ONE(...)        GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
+{                                                      \
+    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2);        \
+    sum3 = svadd_f32_m(DEFAULT_PG, sum3, sum4);        \
+    sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum6);        \
+    sum7 = svadd_f32_m(DEFAULT_PG, sum7, sum8);        \
+    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum3);        \
+    sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum7);        \
+    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5);        \
+    (res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1);  \
+}
+#define GGML_F32xt_REDUCE(...) GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, __VA_ARGS__)
+
+#define GGML_F32_VEC        GGML_F32xt
+#define GGML_F32_VEC_ZERO   GGML_F32xt_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32xt_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32xt_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32xt_STORE
+#define GGML_F32_VEC_FMA    GGML_F32xt_FMA
+#define GGML_F32_VEC_ADD    GGML_F32xt_ADD
+#define GGML_F32_VEC_MUL    GGML_F32xt_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32xt_REDUCE
+
+// F16 NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    #define GGML_F16_STEP 32
+    #define GGML_F16_EPR  8
+
+    #define GGML_F16x8              float16x8_t
+    #define GGML_F16x8_ZERO         vdupq_n_f16(0.0f)
+    #define GGML_F16x8_SET1(x)      vdupq_n_f16(x)
+    #define GGML_F16x8_LOAD(x)      vld1q_f16((const __fp16 *)(x))
+    #define GGML_F16x8_STORE        vst1q_f16
+    #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
+    #define GGML_F16x8_ADD          vaddq_f16
+    #define GGML_F16x8_MUL          vmulq_f16
+    #define GGML_F16x8_REDUCE(res, x)                               \
+    do {                                                            \
+        int offset = GGML_F16_ARR >> 1;                             \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
+        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
+        (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
+    } while (0)
+
+    #define GGML_F16_VEC                GGML_F16x8
+    #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
+    #define GGML_F16_VEC_SET1           GGML_F16x8_SET1
+    #define GGML_F16_VEC_LOAD(p, i)     GGML_F16x8_LOAD(p)
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
+    #define GGML_F16_VEC_FMA            GGML_F16x8_FMA
+    #define GGML_F16_VEC_ADD            GGML_F16x8_ADD
+    #define GGML_F16_VEC_MUL            GGML_F16x8_MUL
+    #define GGML_F16_VEC_REDUCE         GGML_F16x8_REDUCE
+#else
+    // if FP16 vector arithmetic is not supported, we use FP32 instead
+    // and take advantage of the vcvt_ functions to convert to/from FP16
+
+    #define GGML_F16_STEP 16
+    #define GGML_F16_EPR  4
+
+    #define GGML_F32Cx4              float32x4_t
+    #define GGML_F32Cx4_ZERO         vdupq_n_f32(0.0f)
+    #define GGML_F32Cx4_SET1(x)      vdupq_n_f32(x)
+    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
+    #define GGML_F32Cx4_STORE(x, y)  vst1_f16(x, vcvt_f16_f32(y))
+    #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
+    #define GGML_F32Cx4_ADD          vaddq_f32
+    #define GGML_F32Cx4_MUL          vmulq_f32
+    #define GGML_F32Cx4_REDUCE       GGML_F32x4_REDUCE
+
+    #define GGML_F16_VEC                GGML_F32Cx4
+    #define GGML_F16_VEC_ZERO           GGML_F32Cx4_ZERO
+    #define GGML_F16_VEC_SET1           GGML_F32Cx4_SET1
+    #define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx4_LOAD(p)
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
+    #define GGML_F16_VEC_FMA            GGML_F32Cx4_FMA
+    #define GGML_F16_VEC_ADD            GGML_F32Cx4_ADD
+    #define GGML_F16_VEC_MUL            GGML_F32Cx4_MUL
+    #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
+#endif
+
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
 
 #define GGML_SIMD
 
@@ -828,10 +944,8 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     for (int i = 0; i < offset; ++i) {              \
         x[i] = vec_add(x[i], x[offset + i]);        \
     }                                               \
-    res = vec_extract(x[0], 0) +                    \
-          vec_extract(x[0], 1) +                    \
-          vec_extract(x[0], 2) +                    \
-          vec_extract(x[0], 3);                     \
+    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
+    res = tmp[0] + tmp[1];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4

package/src/llama.cpp/ggml/src/ggml-cpu/{ggml-cpu-traits.cpp → traits.cpp}

@@ -1,4 +1,4 @@
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"

package/src/llama.cpp/ggml/src/ggml-cpu/vec.cpp

@@ -17,29 +17,98 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
 
 #if defined(GGML_SIMD)
     float sumf = 0.0f;
-    const int np = (n & ~(GGML_F32_STEP - 1));
 
-    GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+        const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
+
+        const int np = (n & ~(ggml_f32_step - 1));
+        svfloat32_t sum1 = svdup_n_f32(0.0f);
+        svfloat32_t sum2 = svdup_n_f32(0.0f);
+        svfloat32_t sum3 = svdup_n_f32(0.0f);
+        svfloat32_t sum4 = svdup_n_f32(0.0f);
+        svfloat32_t sum5 = svdup_n_f32(0.0f);
+        svfloat32_t sum6 = svdup_n_f32(0.0f);
+        svfloat32_t sum7 = svdup_n_f32(0.0f);
+        svfloat32_t sum8 = svdup_n_f32(0.0f);
+        svfloat32_t ax1,ax2,ax3,ax4,ax5,ax6,ax7,ax8;
+        svfloat32_t ay1,ay2,ay3,ay4,ay5,ay6,ay7,ay8;
+        for (int i = 0; i < np; i += ggml_f32_step) {
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+
+            ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+            sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
+
+            ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
+            ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
+            sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
+
+            ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
+            ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
+            sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
+
+            ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
+            ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
+            sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
+
+            ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
+            ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
+            sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
+
+            ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
+            ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
+            sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
+
+            ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
+            ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
+            sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
+        }
+        // leftovers
+        // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
+        const int np2 = (n & ~(ggml_f32_epr - 1));
+        for (int i = np; i < np2; i += ggml_f32_epr) {
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+        }
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        if (np2 < n) {
+            svbool_t pg = svwhilelt_b32(np2, n);
+            ax1 = svld1_f32(pg, x + np2);
+            ay1 = svld1_f32(pg, y + np2);
+            sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
+        }
+        // reduce sum1,sum2 to sum1
+        GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    GGML_F32_VEC ax[GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
+        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+        GGML_F32_VEC ax[GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
 
-            sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+                sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
+            }
         }
-    }
 
-    // reduce sum0..sum3 to sum0
-    GGML_F32_VEC_REDUCE(sumf, sum);
+        // reduce sum0..sum3 to sum0
+        GGML_F32_VEC_REDUCE(sumf, sum);
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        sumf += x[i]*y[i];
-    }
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            sumf += x[i]*y[i];
+        }
+    #endif
 #else
     // scalar
     ggml_float sumf = 0.0;

package/src/llama.cpp/ggml/src/ggml-cpu/vec.h

@@ -5,6 +5,7 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 #include "ggml.h"
+#include "ggml-cpu.h"
 
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
@@ -148,27 +149,108 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 
 inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F32_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
 
-    GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+        const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
 
-    GGML_F32_VEC ax[GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
+        const int np = (n & ~(ggml_f32_step - 1));
+        svfloat32_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+        svfloat32_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+        for (int i = 0; i < np; i += ggml_f32_step) {
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx);
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
 
-            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            GGML_F32_VEC_STORE(y + i, ay1);
+
+            ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
+
+            GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
+
+            ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
+            ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
+            ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
+
+            GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
+
+            ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
+            ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
+            ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
+
+            GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
+
+            ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
+            ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
+            ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
+
+            GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
+
+            ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
+            ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
+            ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
+
+            GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
+
+            ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
+            ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
+            ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
+
+            GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
+
+            ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
+            ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
+            ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
+
+            GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
         }
-    }
+        // leftovers
+        // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
+        const int np2 = (n & ~(ggml_f32_epr - 1));
+        for (int i = np; i < np2; i += ggml_f32_epr) {
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+
+            GGML_F32_VEC_STORE(y + i, ay1);
+        }
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        if (np2 < n) {
+            svbool_t pg =svwhilelt_b32(np2, n);
+            ax1 = svld1_f32(pg, x + np2);
+            ay1 = svld1_f32(pg, y + np2);
+            ay1 = svmad_f32_m(pg, ax1, vx, ay1);
+
+            svst1_f32(pg, y + np2, ay1);
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] += x[i]*v;
-    }
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+
+        GGML_F32_VEC ax[GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx);
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] += x[i]*v;
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -220,36 +302,45 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
     }
 
 #if defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F32_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
+        // scalar Route to scalar implementation       //TODO: Write SVE code
+        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+            for (int i = 0; i < n; ++i) {
+                y[i] += x[k][i]*v[k][0];
+            }
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
+        GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
 
-    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-        vx[k] = GGML_F32_VEC_SET1(v[k][0]);
-    }
+        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+            vx[k] = GGML_F32_VEC_SET1(v[k][0]);
+        }
 
-    GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
+        GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
 
-            for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-                ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
-                ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
-            }
+                for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+                    ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
+                    ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
+                }
 
-            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
         }
-    }
 
-    // leftovers
-    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-        for (int i = np; i < n; ++i) {
-            y[i] += x[k][i]*v[k][0];
+        // leftovers
+        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+            for (int i = np; i < n; ++i) {
+                y[i] += x[k][i]*v[k][0];
+            }
         }
-    }
+    #endif
 #else
     // scalar
     for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
@@ -265,25 +356,53 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)
     vDSP_vsmul(y, 1, &v, y, 1, n);
 #elif defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F32_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+        const int ggml_f32_step = 2 * ggml_f32_epr;
+
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+        const int np = (n & ~(ggml_f32_step - 1));
+        svfloat32_t ay1;
+        svfloat32_t ay2;
+        for (int i = 0; i < np; i += ggml_f32_step) {
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            ay1 = GGML_F32_VEC_MUL(ay1, vx);
+            GGML_F32_VEC_STORE(y + i, ay1);
+
+            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_MUL(ay2, vx);
+            GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
+        }
+        // leftovers
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        if (np < n) {
+            svbool_t pg = svwhilelt_b32(np, n);
+            ay1 = svld1_f32(pg, y + np);
+            ay1 = svmul_f32_m(pg, ay1, vx);
+            svst1_f32(pg, y + np, ay1);
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
 
-    GGML_F32_VEC ay[GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_MUL(ay[j], vx);
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_MUL(ay[j], vx);
 
-            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
         }
-    }
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] *= v;
-    }
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] *= v;
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -528,6 +647,42 @@ inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
 #error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461"
 #endif
 
+/* Below function was borrowed from the GitHub repository:
+https://github.com/openvinotoolkit/openvino/blob/master/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/common.hpp */
+#if defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    inline static svfloat32_t exp_ps_sve(svbool_t pg, svfloat32_t src) {
+        // Constants
+        const svfloat32_t log2_e = svdup_n_f32(1.4426950409f);
+        const svfloat32_t ln2 = svdup_n_f32(0.6931473921f);
+        const svfloat32_t half_ln2_sq = svdup_n_f32(0.2413862043f);
+        const svuint32_t not_mask17 = svdup_n_u32(~((1u << 17) - 1));
+        const svfloat32_t one = svdup_n_f32(1.0f);
+        const svfloat32_t inactive1 = svdup_n_f32(0.0f);
+        const svint32_t inactive2 = svdup_n_s32(0);
+
+        // Algorithm starts here
+        svfloat32_t t0 = svmul_f32_m(pg, src, log2_e);  // y = x * log2(e)
+        svfloat32_t t1 = svrintm_f32_m(inactive1, pg, t0);         // rount to int (float)
+        svint32_t t2 = svcvt_s32_f32_m(inactive2, pg, t1);         // n
+
+        t1 = svsub_f32_m(pg, t0, t1);   // a = y - floor(y)
+        t1 = svadd_f32_m(pg, t1, one);  // b = a + 1
+
+        svuint32_t t3 = svlsr_n_u32_m(pg, svreinterpret_u32_f32(t1), 17);  // v = b >> 17 (u32)
+        svfloat32_t t4 = svexpa_f32(t3);                                   // c = fexpa(v)
+        t4 = svscale_f32_m(pg, t4, t2);                                    // fexpa(v) * 2^(n)
+
+        // and_(t2.d, t1.d, not_mask17.d)
+        svfloat32_t t5 = svreinterpret_f32_u32(svand_u32_m(pg, svreinterpret_u32_f32(t1), not_mask17));
+        t5 = svsub_f32_m(pg, t1, t5);                // z
+        t0 = svmla_f32_m(pg, ln2, t5, half_ln2_sq);  // ln2 + half_ln2_sq * z
+        t0 = svmla_f32_m(pg, one, t5, t0);           // 1 + (ln2 * z) + (half_ln2_sq * z * z)
+        t0 = svmul_f32_m(pg, t0, t4);                // Final result
+
+        return t0;
+    }
+#endif
+
 #if defined(__ARM_NEON) && defined(__aarch64__)
 
 // adapted from arm limited optimized routine