ggml : backport llama.cpp updates (close #709)

- About x2 overall performance improvement on Apple Silicon
- Results should now be the same for different number of threads (not
tested)

ggml.h

@@ -236,6 +236,7 @@ enum ggml_op {
 
     GGML_OP_SCALE,
     GGML_OP_CPY,
+    GGML_OP_CONT,
     GGML_OP_RESHAPE,
     GGML_OP_VIEW,
     GGML_OP_PERMUTE,
@@ -253,16 +254,29 @@ enum ggml_op {
     GGML_OP_COUNT,
 };
 
+
+// ggml object
+struct ggml_object {
+    size_t offs;
+    size_t size;
+
+    struct ggml_object * next;
+
+    char padding[8];
+};
+
+static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
+
 // n-dimensional tensor
 struct ggml_tensor {
     enum ggml_type type;
 
     int    n_dims;
-    int    ne[GGML_MAX_DIMS]; // number of elements
-    size_t nb[GGML_MAX_DIMS]; // stride in bytes:
-                              // nb[0] = sizeof(type)
-                              // nb[1] = nb[0]   * ne[0] + padding
-                              // nb[i] = nb[i-1] * ne[i-1]
+    int64_t ne[GGML_MAX_DIMS]; // number of elements
+    size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
+                               // nb[0] = sizeof(type)
+                               // nb[1] = nb[0]   * ne[0] + padding
+                               // nb[i] = nb[i-1] * ne[i-1]
 
     // compute data
     enum ggml_op op;
@@ -316,6 +330,7 @@ struct ggml_init_params {
     // memory pool
     size_t mem_size;   // bytes
     void * mem_buffer; // if NULL, memory will be allocated internally
+    bool   no_alloc;   // don't allocate memory for the tensor data
 };
 
 void    ggml_time_init(void); // call this once at the beginning of the program
@@ -327,8 +342,8 @@ int64_t ggml_cycles_per_ms(void);
 void ggml_print_object (const struct ggml_object * obj);
 void ggml_print_objects(const struct ggml_context * ctx);
 
-int    ggml_nelements(const struct ggml_tensor * tensor);
-size_t ggml_nbytes   (const struct ggml_tensor * tensor);
+int64_t ggml_nelements(const struct ggml_tensor * tensor);
+size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
 
 int    ggml_blck_size (enum ggml_type type);
 size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
@@ -343,40 +358,37 @@ size_t ggml_used_mem(const struct ggml_context * ctx);
 
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
 
-bool ggml_mlock_supported(void);
-bool ggml_mlock(struct ggml_context * ctx, char ** err_p);
-
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int *ne);
+        const int64_t *ne);
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0);
+        int64_t ne0);
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1);
+        int64_t ne0,
+        int64_t ne1);
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2);
 
 struct ggml_tensor * ggml_new_tensor_4d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2,
-        int    ne3);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2,
+        int64_t ne3);
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
 struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
@@ -514,6 +526,11 @@ struct ggml_tensor * ggml_cpy(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b);
 
+// make contiguous
+struct ggml_tensor * ggml_cont(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // return view(a), b specifies the new shape
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape(
@@ -526,33 +543,43 @@ struct ggml_tensor * ggml_reshape(
 struct ggml_tensor * ggml_reshape_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1);
+        int64_t               ne0,
+        int64_t               ne1);
 
 // return view(a)
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape_3d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2);
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2);
 
 // offset in bytes
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
+        int64_t               ne0,
         size_t                offset);
 
 struct ggml_tensor * ggml_view_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
+        int64_t               ne0,
+        int64_t               ne1,
         size_t                nb1, // row stride in bytes
         size_t                offset);
 
+struct ggml_tensor * ggml_view_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        size_t                nb1, // row   stride in bytes
+        size_t                nb2, // slice stride in bytes
+        size_t                offset);
+
 struct ggml_tensor * ggml_permute(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -748,8 +775,8 @@ enum ggml_opt_result ggml_opt(
 // quantization
 //
 
-size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
-size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
 
 //
 // system info
@@ -768,6 +795,30 @@ int ggml_cpu_has_blas(void);
 int ggml_cpu_has_sse3(void);
 int ggml_cpu_has_vsx(void);
 
+
+//
+// Internal types and functions exposed for tests and benchmarks
+//
+
+#ifdef  __cplusplus
+// restrict not standard in C++
+#define GGML_RESTRICT
+#else
+#define GGML_RESTRICT restrict
+#endif
+typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+typedef void (*quantize_row_q_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+typedef void (*vec_dot_q_t)(const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+
+typedef struct {
+    dequantize_row_q_t dequantize_row_q;
+    quantize_row_q_t   quantize_row_q;
+    quantize_row_q_t   quantize_row_q_reference;
+    vec_dot_q_t        vec_dot_q;
+} quantize_fns_t;
+
+quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
+
 #ifdef  __cplusplus
 }
 #endif

whisper.cpp

@@ -654,9 +654,11 @@ static bool kv_cache_init(
                                  int   n_ctx) {
     cache.buf.resize(mem_bytes);
 
-    struct ggml_init_params params;
-    params.mem_size   = cache.buf.size();
-    params.mem_buffer = cache.buf.data();
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ cache.buf.size(),
+        /*.mem_buffer =*/ cache.buf.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     cache.ctx = ggml_init(params);
 
@@ -688,9 +690,11 @@ static bool kv_cache_reinit(struct whisper_kv_cache & cache) {
 
     WHISPER_ASSERT(cache.buf.size() >= 2*n_elements*ggml_type_size(wtype));
 
-    struct ggml_init_params params;
-    params.mem_size   = cache.buf.size();
-    params.mem_buffer = cache.buf.data();
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ cache.buf.size(),
+        /*.mem_buffer =*/ cache.buf.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     cache.ctx = ggml_init(params);
 
@@ -1028,9 +1032,11 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
     // create the ggml context
     {
-        struct ggml_init_params params;
-        params.mem_size   = wctx.model.buf->size();
-        params.mem_buffer = wctx.model.buf->data();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ wctx.model.buf->size(),
+            /*.mem_buffer =*/ wctx.model.buf->data(),
+            /*.no_alloc   =*/ false,
+        };
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -1254,10 +1260,12 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                 break;
             }
 
-            int32_t nelements = 1;
-            int32_t ne[3] = { 1, 1, 1 };
+            int64_t nelements = 1;
+            int64_t ne[3] = { 1, 1, 1 };
             for (int i = 0; i < n_dims; ++i) {
-                read_safe(loader, ne[i]);
+                int32_t ne_cur;
+                read_safe(loader, ne_cur);
+                ne[i] = ne_cur;
                 nelements *= ne[i];
             }
 
@@ -1278,7 +1286,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             }
 
             if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1] || tensor->ne[2] != ne[2]) {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d, %d], expected [%d, %d, %d]\n",
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%lld, %lld, %lld], expected [%lld, %lld, %lld]\n",
                         __func__, name.data(), tensor->ne[0], tensor->ne[1], tensor->ne[2], ne[0], ne[1], ne[2]);
                 return false;
             }
@@ -1286,7 +1294,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
 
             if (nelements*bpe != ggml_nbytes(tensor)) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %llu\n",
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
             }
@@ -1344,9 +1352,11 @@ static bool whisper_encode_internal(
     const int n_mels = hparams.n_mels;
     assert(mel_inp.n_mel == n_mels);
 
-    struct ggml_init_params params;
-    params.mem_size   = wstate.buf_compute.size();
-    params.mem_buffer = wstate.buf_compute.data();
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ wstate.buf_compute.size(),
+        /*.mem_buffer =*/ wstate.buf_compute.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
 
@@ -1501,8 +1511,7 @@ static bool whisper_encode_internal(
                                 Vcur,
                                 n_state/n_head, n_head, n_ctx),
                             1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, wctx.wtype, n_ctx, n_state/n_head, n_head)
-                        );
+                        ggml_new_tensor_3d(ctx0, wctx.wtype, n_ctx, n_state/n_head, n_head));
 
             struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, false);
 #else
@@ -1726,10 +1735,12 @@ static bool whisper_encode_internal(
 
             wstate.use_buf(ctx0, -1);
 
-            //struct ggml_tensor * k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*hparams.n_audio_ctx + iter*n_ctx));
-            //struct ggml_tensor * v = ggml_view_1d(ctx0, wstate.kv_cross.v, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.v)*n_state)*(il*hparams.n_audio_ctx + iter*n_ctx));
-            struct ggml_tensor* k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
-            struct ggml_tensor* v = ggml_view_1d(ctx0, wstate.kv_cross.v, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.v)*n_state)*(il*n_ctx));
+            Vcross = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));
+
+            struct ggml_tensor * k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
+            struct ggml_tensor * v = ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
+                    (   n_ctx)*ggml_element_size(wstate.kv_cross.v),
+                    (il*n_ctx)*ggml_element_size(wstate.kv_cross.v)*n_state);
 
             ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcross, k));
             ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcross, v));
@@ -1797,9 +1808,11 @@ static bool whisper_decode_internal(
 
     //WHISPER_PRINT_DEBUG("%s: n_past = %d, N = %d, M = %d, n_ctx = %d\n", __func__, n_past, N, M, n_ctx);
 
-    struct ggml_init_params params;
-    params.mem_size   = wstate.buf_compute.size();
-    params.mem_buffer = wstate.buf_compute.data();
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ wstate.buf_compute.size(),
+        /*.mem_buffer =*/ wstate.buf_compute.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
 
@@ -1862,20 +1875,24 @@ static bool whisper_decode_internal(
 
             Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
 
-            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
-                    layer.attn_v_w,
-                    cur);
-
-            Vcur = ggml_add(ctx0,
-                    ggml_repeat(ctx0,
-                        layer.attn_v_b,
-                        Vcur),
-                    Vcur);
-
             // store key and value to memory
             {
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
+                        layer.attn_v_w,
+                        cur);
+
+                Vcur = ggml_add(ctx0,
+                        ggml_repeat(ctx0,
+                            layer.attn_v_b,
+                            Vcur),
+                        Vcur);
+
+                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_state, N));
+
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_state, (ggml_element_size(kv_self.k)*n_state)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_state, (ggml_element_size(kv_self.v)*n_state)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_state,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_state + n_past*ggml_element_size(kv_self.v));
 
                 ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
                 ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
@@ -1914,16 +1931,14 @@ static bool whisper_decode_internal(
 
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
 
-            struct ggml_tensor * V_trans =
-                ggml_cpy(ctx0,
-                        ggml_permute(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_state, il*n_ctx*ggml_element_size(kv_self.v)*n_state),
-                                n_state/n_head, n_head, n_past + N),
-                            1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_state/n_head, n_head));
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_past + N, n_state/n_head, n_head,
+                        n_ctx*ggml_element_size(kv_self.v),
+                        n_ctx*ggml_element_size(kv_self.v)*n_state/n_head,
+                        il*n_ctx*ggml_element_size(kv_self.v)*n_state);
 
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
@@ -1986,15 +2001,22 @@ static bool whisper_decode_internal(
                         ggml_view_1d(ctx0, wstate.kv_cross.k, M*n_state, il*M*ggml_element_size(wstate.kv_cross.k)*n_state),
                         n_state/n_head, n_head, M);
 
-            struct ggml_tensor * Vcross =
-                ggml_reshape_3d(ctx0,
-                        ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
-                        n_state/n_head, n_head, M);
+            //struct ggml_tensor * Vcross =
+            //    ggml_reshape_3d(ctx0,
+            //            ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
+            //            n_state/n_head, n_head, M);
 
-            struct ggml_tensor * V_trans =
-                ggml_cpy(ctx0,
-                        ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, Vcross->type, M, n_state/n_head, n_head));
+            //struct ggml_tensor * V_trans =
+            //    ggml_cpy(ctx0,
+            //            ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
+            //            ggml_new_tensor_3d(ctx0, Vcross->type, M, n_state/n_head, n_head));
+
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, wstate.kv_cross.v,
+                        M, n_state/n_head, n_head,
+                        M*ggml_element_size(wstate.kv_cross.v),
+                        M*ggml_element_size(wstate.kv_cross.v)*n_state/n_head,
+                        il*M*ggml_element_size(wstate.kv_cross.v)*n_state);
 
             // ------
 
@@ -2021,7 +2043,7 @@ static bool whisper_decode_internal(
 
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);
 
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
@@ -4726,6 +4748,7 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
             struct ggml_init_params gparams = {
                 /*.mem_size   =*/ buf.size(),
                 /*.mem_buffer =*/ buf.data(),
+                /*.no_alloc   =*/ false,
             };
 
             struct ggml_context * ctx0 = ggml_init(gparams);