From 9fdd415367244ab63a2150dbf4a12c2bc4ac3b97 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 12 Sep 2023 20:10:53 +0300
Subject: [PATCH] ggml : fix ggml_nbytes (probably temp solution)

---
 ggml.c | 39 ++++++++++++++++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 3 deletions(-)

diff --git a/ggml.c b/ggml.c
index 2b4d3455..f1e847b1 100644
--- a/ggml.c
+++ b/ggml.c
@@ -4303,10 +4303,43 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
 }
 
 size_t ggml_nbytes(const struct ggml_tensor * tensor) {
-    size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type);
-    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
-        nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+    // original:
+    //size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type);
+    //for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+    //    nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+    //}
+    //return nbytes;
+
+    // TODO: the imlpementation below is stupid - need something better
+
+    // sort ne and nb
+    int64_t sne[GGML_MAX_DIMS];
+    size_t  snb[GGML_MAX_DIMS];
+
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        sne[i] = tensor->ne[i];
+        snb[i] = tensor->nb[i];
     }
+
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        for (int j = i + 1; j < GGML_MAX_DIMS; ++j) {
+            if ((snb[i] > snb[j]) || (snb[i] == snb[j] && sne[i] < sne[j])) {
+                size_t tmp = snb[i];
+                snb[i] = snb[j];
+                snb[j] = tmp;
+
+                int64_t tmp2 = sne[i];
+                sne[i] = sne[j];
+                sne[j] = tmp2;
+            }
+        }
+    }
+
+    size_t nbytes = (sne[0]/ggml_blck_size(tensor->type))*snb[0];
+    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+        nbytes += (sne[i] - 1)*snb[i];
+    }
+
     return nbytes;
 }