@fugood/llama.node 0.3.3 → 0.3.4

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -1,30 +1,8 @@
 #pragma once
 
+#include "ggml-cpu-traits.h"
 #include "ggml.h"
 
 // GGML internal header
 
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
-// GEMV
-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);
-
-// GEMM
-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_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
-enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
-
-#ifdef __cplusplus
-}
-#endif
-
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

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

@@ -0,0 +1,55 @@
+#ifdef GGML_USE_CPU_HBM
+
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-impl.h"
+
+#include "ggml-cpu-hbm.h"
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                                                           size_t                     size) {
+    void * ptr;
+    int    result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft                 = buft;
+    buffer->iface.free_buffer    = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+                           /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+                           /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+                           /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+                           /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+                           },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif

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

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+// GGML CPU internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);

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

@@ -15,6 +15,18 @@
 extern "C" {
 #endif
 
+struct ggml_compute_params {
+    // ith = thread index, nth = number of threads
+    int ith, nth;
+
+    // work buffer for all threads
+    size_t wsize;
+    void * wdata;
+
+    struct ggml_threadpool * threadpool;
+};
+
+
 #if defined(_MSC_VER)
 
 #define m512bh(p) p
@@ -366,6 +378,9 @@ static __m256 __lasx_xvreplfr2vr_s(float val) {
 }
 #endif
 
+// TODO: move to ggml-threading
+void ggml_barrier(struct ggml_threadpool * tp);
+
 #ifdef __cplusplus
 }
 #endif

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -1791,11 +1791,12 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             const int8x16_t y1_l = vld1q_s8(b_y1->qs);
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
-            float32_t _scale[4] = { GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                    GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
-
+            float32_t _scale[4] = {
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+            };
             float32x4_t scale = vld1q_f32(_scale);
 
             int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
@@ -1811,13 +1812,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
 
             sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
         }
-        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
         float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
 
-        vst1_f32(s,      vget_low_f32(sumv2));
+        vst1_f32(s,      vget_low_f32 (sumv2));
         vst1_f32(s + bs, vget_high_f32(sumv2));
+
         return;
     }
 #endif
@@ -2345,10 +2348,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             const block_q8_1 * restrict b_y0 = &vy0[i];
             const block_q8_1 * restrict b_y1 = &vy1[i];
 
-            float32_t summs_t[4] = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
-                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                                    GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
+            float32_t summs_t[4] = {
+                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
+                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
+                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
+                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)
+            };
             summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
 
             const uint8x16_t m4b = vdupq_n_u8(0x0F);
@@ -2369,10 +2374,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
             // mmla into int32x4_t
-            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d,
-                                   GGML_FP16_TO_FP32(b_x0->d)*b_y1->d,
-                                   GGML_FP16_TO_FP32(b_x1->d)*b_y0->d,
-                                   GGML_FP16_TO_FP32(b_x1->d)*b_y1->d};
+            float32_t _scale[4] = {
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+            };
             float32x4_t scale = vld1q_f32(_scale);
 
             int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
@@ -2387,15 +2394,17 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
             int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
             sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
         }
 
-        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
         float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
         sumv2 = vaddq_f32(sumv2, summs0);
 
         vst1_f32(s,      vget_low_f32 (sumv2));
         vst1_f32(s + bs, vget_high_f32(sumv2));
+
         return;
     }
 #endif
@@ -3372,10 +3381,12 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             const int8x16_t y1_l = vld1q_s8(b_y1->qs);
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
-            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                   GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+            float32_t _scale[4] = {
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+            };
             float32x4_t scale = vld1q_f32(_scale);
 
             int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
@@ -3391,13 +3402,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
 
             sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                                                       l1, r1)), l2, r2)), l3, r3))), scale);
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
         }
-        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
         float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
 
-        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s,      vget_low_f32 (sumv2));
         vst1_f32(s + bs, vget_high_f32(sumv2));
+
         return;
     }
 #endif

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

@@ -0,0 +1,36 @@
+#include "ggml-cpu-traits.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+
+namespace ggml::cpu {
+tensor_traits::~tensor_traits() {}
+
+extra_buffer_type::~extra_buffer_type() {}
+}  // namespace ggml::cpu
+
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}

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

@@ -0,0 +1,38 @@
+#pragma once
+#include "ggml-backend-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+#    include <vector>
+extern "C" {
+#endif
+
+// return true if op part of extra "accelerator"
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
+
+#ifdef __cplusplus
+}
+
+namespace ggml::cpu {
+// register in tensor->extra
+class tensor_traits {
+  public:
+    virtual ~tensor_traits();
+    virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size)        = 0;
+    virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
+};
+
+class extra_buffer_type {
+  public:
+    virtual ~extra_buffer_type();
+    virtual bool            supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
+    virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op)                   = 0;
+};
+}  // namespace ggml::cpu
+
+// implemented in ggml-cpu.cpp.
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
+
+#endif