whispercpp 1.3.2 → 1.3.3

data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c

@@ -3,11 +3,11 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-threading.h"
 #include "unary-ops.h"
 #include "binary-ops.h"
@@ -72,15 +72,13 @@
 #define UNUSED GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0)
 
+// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
+float ggml_table_f32_f16[1 << 16];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
-    int has_neon;
-    int has_dotprod;
-    int has_i8mm;
-    int has_sve;
     int sve_cnt;
-    int has_sme;
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
+} ggml_arm_arch_features = { 0 };
 #endif
 
 
@@ -197,6 +195,7 @@ typedef pthread_t ggml_thread_t;
 
 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32] = {
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp32,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
         .vec_dot_type             = GGML_TYPE_F32,
         .nrows                    = 1,
@@ -270,7 +269,11 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q4_K,
         .vec_dot                  = ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
         .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_Q5_K] = {
         .from_float               = quantize_row_q5_K,
@@ -555,6 +558,14 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value) {
+    atomic_store_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
+int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value) {
+    return atomic_fetch_add_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
     cpu_set_t cpuset;
@@ -666,87 +677,15 @@ bool ggml_is_numa(void) {
 
 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM (1 << 13)
-#endif
-
-#if !defined(HWCAP2_SME)
-#define HWCAP2_SME (1 << 23)
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__)
-    uint32_t hwcap = getauxval(AT_HWCAP);
-    uint32_t hwcap2 = getauxval(AT_HWCAP2);
-
-    ggml_arm_arch_features.has_neon    = !!(hwcap & HWCAP_ASIMD);
-    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-    ggml_arm_arch_features.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
-    ggml_arm_arch_features.has_sve     = !!(hwcap & HWCAP_SVE);
-    ggml_arm_arch_features.has_sme     = !!(hwcap2 & HWCAP2_SME);
-
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
 #endif
-#elif defined(__APPLE__)
-    int oldp = 0;
-    size_t size = sizeof(oldp);
-    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_neon = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_dotprod = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_i8mm = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_sme = oldp;
-
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#else
-// Run-time CPU feature detection not implemented for this platform, fallback to compile time
-#if defined(__ARM_NEON)
-    ggml_arm_arch_features.has_neon = 1;
-#else
-    ggml_arm_arch_features.has_neon = 0;
-#endif
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    ggml_arm_arch_features.has_i8mm = 1;
-#else
-    ggml_arm_arch_features.has_i8mm = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-    ggml_arm_arch_features.has_sve = 1;
-    ggml_arm_arch_features.sve_cnt = 16;
-#else
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
-    ggml_arm_arch_features.has_sme = 1;
-#else
-    ggml_arm_arch_features.has_sme = 0;
-#endif
-#endif
 }
-#endif
+
+#endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -801,7 +740,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
             {
                 assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                 for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -860,7 +799,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
             {
                 assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                 for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -911,7 +850,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
@@ -956,7 +895,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -985,7 +924,7 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
         case GGML_TYPE_I32:
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
         case GGML_TYPE_BF16:
             return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
@@ -1012,7 +951,7 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1050,7 +989,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
             }
         case GGML_TYPE_F16:
             {
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
@@ -1089,7 +1028,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1116,7 +1055,7 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
         case GGML_TYPE_I32:
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
         case GGML_TYPE_BF16:
             return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
@@ -1143,7 +1082,7 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1254,7 +1193,7 @@ static void ggml_compute_forward_mul_mat_one_chunk(
     }
 }
 
-static void ggml_compute_forward_mul_mat(
+void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
 
@@ -1879,6 +1818,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_get_rows_back(params, tensor);
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                ggml_compute_forward_set_rows(params, tensor);
+            } break;
         case GGML_OP_DIAG:
             {
                 ggml_compute_forward_diag(params, tensor);
@@ -1923,6 +1866,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
+        case GGML_OP_CONV_2D:
+            {
+                ggml_compute_forward_conv_2d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D_DW:
             {
                 ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -1955,6 +1902,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad_reflect_1d(params, tensor);
             } break;
+        case GGML_OP_ROLL:
+            {
+                ggml_compute_forward_roll(params, tensor);
+            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -2002,6 +1953,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_unary(params, tensor);
             } break;
+        case GGML_OP_GLU:
+            {
+                ggml_compute_forward_glu(params, tensor);
+            } break;
         case GGML_OP_GET_REL_POS:
             {
                 ggml_compute_forward_get_rel_pos(params, tensor);
@@ -2212,6 +2167,18 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     GGML_ABORT("fatal error");
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(node)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                    {
+                        n_tasks = n_threads;
+                    } break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            break;
         case GGML_OP_SILU_BACK:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -2228,6 +2195,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_GET_ROWS:
+        case GGML_OP_SET_ROWS:
             {
                 // FIXME: get_rows can use additional threads, but the cost of launching additional threads
                 // decreases performance with GPU offloading
@@ -2264,6 +2232,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2279,6 +2248,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_ROLL:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -2414,12 +2384,32 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     // This is up to the applications.
     DWORD p = THREAD_PRIORITY_NORMAL;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p = THREAD_PRIORITY_BELOW_NORMAL;  break;
         case GGML_SCHED_PRIO_NORMAL:   p = THREAD_PRIORITY_NORMAL;        break;
         case GGML_SCHED_PRIO_MEDIUM:   p = THREAD_PRIORITY_ABOVE_NORMAL;  break;
         case GGML_SCHED_PRIO_HIGH:     p = THREAD_PRIORITY_HIGHEST;       break;
         case GGML_SCHED_PRIO_REALTIME: p = THREAD_PRIORITY_TIME_CRITICAL; break;
     }
 
+    if (prio != GGML_SCHED_PRIO_LOW) {
+        // Tell Windows that this thread should not be throttled (needs its own CPU core).
+        // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
+        // all our threads onto the first 4 cores which results in terrible performance with
+        // n_threads > 4
+        #if _WIN32_WINNT >= 0x0602
+        THREAD_POWER_THROTTLING_STATE t;
+        ZeroMemory(&t, sizeof(t));
+        t.Version     = THREAD_POWER_THROTTLING_CURRENT_VERSION;
+        t.ControlMask = THREAD_POWER_THROTTLING_EXECUTION_SPEED;
+        t.StateMask   = 0;
+
+        if (!SetThreadInformation(GetCurrentThread(), ThreadPowerThrottling, &t, sizeof(t))) {
+            GGML_LOG_DEBUG("failed to disable thread power throttling %d : (%d)\n", prio, (int) GetLastError());
+            return false;
+        }
+        #endif
+    }
+
     if (prio == GGML_SCHED_PRIO_NORMAL) {
         // Keep inherited policy/priority
         return true;
@@ -2447,6 +2437,8 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     struct sched_param p;
     int32_t policy = SCHED_OTHER;
     switch (prio) {
+        // TODO: there seems to be no way to set lower prio on Apple platforms
+        case GGML_SCHED_PRIO_LOW:      policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
         case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;
@@ -2503,6 +2495,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     struct sched_param p;
     int32_t policy = SCHED_OTHER;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      policy = SCHED_BATCH; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
         case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;
@@ -2758,6 +2751,10 @@ struct ggml_cplan ggml_graph_plan(
                             GGML_ABORT("fatal error");
                         }
                     } break;
+                case GGML_OP_CONV_2D:
+                    {
+                        cur = GGML_IM2COL_WORK_SIZE;
+                    } break;
                 case GGML_OP_CONV_TRANSPOSE_2D:
                     {
                         const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -3158,6 +3155,10 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+void ggml_cpu_fp32_to_fp32(const float * x, float * y, int64_t n) {
+    memcpy(y, x, n * sizeof(float));
+}
+
 void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
     int64_t i = 0;
 #if defined(__F16C__)
@@ -3178,9 +3179,24 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
         _mm_storel_epi64((__m128i *)(y + i), y_vec);
     }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
+        float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
+        uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
+    for (; i + 3 < n; i += 4) {
+        float32x4_t v_x = vec_xl(0, (const float *)(x + i));
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
 #endif
     for (; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(x[i]);
     }
 }
 
@@ -3204,9 +3220,25 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i + 0));
+        vec_xst(v_yl, 0, (float *)(y + i + 4));
+    }
+    for (; i + 3 < n; i += 4) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i));
+    }
 #endif
+
     for (; i < n; ++i) {
-        y[i] = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 }
 
@@ -3406,9 +3438,17 @@ int ggml_cpu_has_vxe(void) {
 #endif
 }
 
+int ggml_cpu_has_nnpa(void) {
+#if defined(GGML_NNPA)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
-    return ggml_arm_arch_features.has_neon;
+    return 1;
 #else
     return 0;
 #endif
@@ -3416,7 +3456,7 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return ggml_arm_arch_features.has_dotprod;
+    return 1;
 #else
     return 0;
 #endif
@@ -3424,7 +3464,7 @@ int ggml_cpu_has_dotprod(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return ggml_arm_arch_features.has_sve;
+    return 1;
 #else
     return 0;
 #endif
@@ -3432,7 +3472,7 @@ int ggml_cpu_has_sve(void) {
 
 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return ggml_arm_arch_features.has_i8mm;
+    return 1;
 #else
     return 0;
 #endif
@@ -3448,14 +3488,14 @@ int ggml_cpu_get_sve_cnt(void) {
 
 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return ggml_arm_arch_features.has_sme;
+    return 1;
 #else
     return 0;
 #endif
 }
 
 void ggml_cpu_init(void) {
-    // needed to initialize f16 tables
+    // needed to initialize ggml_time
     {
         struct ggml_init_params params = { 0, NULL, false };
         struct ggml_context * ctx = ggml_init(params);
@@ -3476,9 +3516,10 @@ void ggml_cpu_init(void) {
                     uint16_t u16;
                     ggml_fp16_t fp16;
                 } u = {i};
-                float f = GGML_FP16_TO_FP32(u.fp16);
-                ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
-                ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
+                float f = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
+                ggml_table_f32_f16[i] = f;
+                ggml_table_gelu_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_f32(f));
+                ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }
 
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);

data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -1,8 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-aarch64.h"
-#include "ggml-cpu-traits.h"
+#include "repack.h"
+#include "traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
 
@@ -11,7 +11,7 @@
 #include <vector>
 
 #ifdef GGML_USE_CPU_HBM
-#    include "ggml-cpu-hbm.h"
+#    include "hbm.h"
 #endif
 
 #ifdef GGML_USE_CPU_KLEIDIAI
@@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
-        if (ggml_backend_cpu_aarch64_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#ifdef GGML_USE_CPU_REPACK
+        if (ggml_backend_cpu_repack_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_repack_buffer_type());
         }
 #endif
 
@@ -416,6 +416,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
 
     switch (op->op) {
         case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
             return
                 op->type != GGML_TYPE_IQ3_XXS &&
                 op->type != GGML_TYPE_IQ3_S   &&
@@ -578,6 +579,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_vxe()) {
             features.push_back({ "VXE", "1" });
         }
+        if (ggml_cpu_has_nnpa()) {
+            features.push_back({ "NNPA", "1" });
+        }
         if (ggml_cpu_has_wasm_simd()) {
             features.push_back({ "WASM_SIMD", "1" });
         }
@@ -596,8 +600,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_CPU_KLEIDIAI
         features.push_back({ "KLEIDIAI", "1" });
     #endif
-    #ifdef GGML_USE_CPU_AARCH64
-        features.push_back({ "AARCH64_REPACK", "1" });
+    #ifdef GGML_USE_CPU_REPACK
+        features.push_back({ "REPACK", "1" });
     #endif
 
         features.push_back({ nullptr, nullptr });

data/ext/sources/ggml/src/ggml-cpu/{ggml-cpu-hbm.cpp → hbm.cpp}

@@ -5,7 +5,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 
-#include "ggml-cpu-hbm.h"
+#include "hbm.h"
 
 // buffer type HBM
 

data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -26,7 +26,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "kernels.h"