ggml : extend ggml_get_rows, ggml_repeat, ggml_concat (ggml/639)

* add more int ops

* ggml_compute_forward_dup_bytes

* add tests

* PR comments

* tests : minor indentations

---------

Co-authored-by: Georgi Gerganov <redacted>

diff --git a/ggml.c b/ggml.c
index a9e1ea9b40ec4eb7fd2d69fec1e4ef76255b85fa..4d8891c375aceb70ec93fde07874e3020931865d 100644 (file)
--- a/ggml.c
+++ b/ggml.c
@@ -4766,8 +4766,11 @@ struct ggml_tensor * ggml_get_rows(
     }
 
     // TODO: implement non F32 return
-    //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]);
-    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
+    enum ggml_type type = GGML_TYPE_F32;
+    if (a->type == GGML_TYPE_I32) {
+        type = a->type;
+    }
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
 
     result->op   = GGML_OP_GET_ROWS;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6938,14 +6941,165 @@ static void ggml_compute_forward_dup_f32(
     }
 }
 
-static void ggml_compute_forward_dup(
+// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy.
+static void ggml_compute_forward_dup_bytes(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
+    GGML_ASSERT(src0->type == dst->type);
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
         ggml_compute_forward_dup_same_cont(params, src0, dst);
         return;
     }
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
+    const size_t type_size = ggml_type_size(src0->type);
+    const int ith = params->ith; // thread index
+    const int nth = params->nth; // number of threads
+
+
+    // parallelize by rows
+    const int nr = ne01;
+    // number of rows per thread
+    const int dr = (nr + nth - 1) / nth;
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    if (src0->type == dst->type &&
+        ne00 == ne0 &&
+        nb00 == type_size && nb0 == type_size) {
+        // copy by rows
+        const size_t rs = ne00 * type_size;
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    memcpy(
+                        ((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3),
+                        ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03),
+                        rs);
+                }
+            }
+        }
+        return;
+    }
+
+    if (ggml_is_contiguous(dst)) {
+        size_t id = 0;
+        char * dst_ptr = (char *) dst->data;
+        const size_t rs = ne00 * type_size;
+
+        if (nb00 == type_size) {
+            // src0 is contigous on first dimension, copy by rows
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    id += rs * ir0;
+                    for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                        const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                        memcpy(dst_ptr + id, src0_ptr, rs);
+                        id += rs;
+                    }
+                    id += rs * (ne01 - ir1);
+                }
+            }
+        } else {
+            //printf("%s: this is not optimal - fix me\n", __func__);
+
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    id += rs * ir0;
+                    for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
+                            const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03;
+                            memcpy(dst_ptr + id, src0_ptr, type_size);
+
+                            id += type_size;
+                        }
+                    }
+                    id += rs * (ne01 - ir1);
+                }
+            }
+        }
+
+        return;
+    }
+
+    // dst counters
+
+    int64_t i10 = 0;
+    int64_t i11 = 0;
+    int64_t i12 = 0;
+    int64_t i13 = 0;
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            i10 += ne00 * ir0;
+            while (i10 >= ne0) {
+                i10 -= ne0;
+                if (++i11 == ne1) {
+                    i11 = 0;
+                    if (++i12 == ne2) {
+                        i12 = 0;
+                        if (++i13 == ne3) {
+                            i13 = 0;
+                        }
+                    }
+                }
+            }
+            for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                    memcpy(dst_ptr, src0_ptr, type_size);
+
+                    if (++i10 == ne0) {
+                        i10 = 0;
+                        if (++i11 == ne1) {
+                            i11 = 0;
+                            if (++i12 == ne2) {
+                                i12 = 0;
+                                if (++i13 == ne3) {
+                                    i13 = 0;
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            i10 += ne00 * (ne01 - ir1);
+            while (i10 >= ne0) {
+                i10 -= ne0;
+                if (++i11 == ne1) {
+                    i11 = 0;
+                    if (++i12 == ne2) {
+                        i12 = 0;
+                        if (++i13 == ne3) {
+                            i13 = 0;
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_dup(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    if (src0->type == dst->type) {
+        ggml_compute_forward_dup_bytes(params, src0, dst);
+        return;
+    }
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
@@ -8404,10 +8558,12 @@ static void ggml_compute_forward_repeat(
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F16:
+        case GGML_TYPE_I16:
             {
                 ggml_compute_forward_repeat_f16(params, src0, dst);
             } break;
         case GGML_TYPE_F32:
+        case GGML_TYPE_I32:
             {
                 ggml_compute_forward_repeat_f32(params, src0, dst);
             } break;
@@ -8550,6 +8706,7 @@ static void ggml_compute_forward_concat(
     struct ggml_tensor* dst) {
     switch (src0->type) {
         case GGML_TYPE_F32:
+        case GGML_TYPE_I32:
             {
                 ggml_compute_forward_concat_f32(params, src0, src1, dst);
             } break;
@@ -10674,6 +10831,7 @@ static void ggml_compute_forward_get_rows(
                 ggml_compute_forward_get_rows_f16(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F32:
+        case GGML_TYPE_I32:
             {
                 ggml_compute_forward_get_rows_f32(params, src0, src1, dst);
             } break;