From 4afa38e7446f997d7034e239e718be26e524638f Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Wed, 31 May 2023 10:20:23 +0800
Subject: [PATCH] Revert "opencl : no need to allocate cl_mem on heap (#1612)"

This reverts commit bb051d9723d628414b9e929e5264e23262a2f1b2.
---
 ggml-opencl.cpp | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp
index 86a85d79b85aa..43650fd5f2b07 100644
--- a/ggml-opencl.cpp
+++ b/ggml-opencl.cpp
@@ -676,7 +676,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_CL) {
-        d_X = (cl_mem) src0->data;
+        d_X = *(cl_mem*) src0->data;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
     }
@@ -753,7 +753,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_CL) {
-        d_X = (cl_mem) src0->data;
+        d_X = *(cl_mem*) src0->data;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
     }
@@ -881,7 +881,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 events.emplace_back();
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
             } else if (src0->backend == GGML_BACKEND_CL) {
-                d_Q = (cl_mem) src0->data;
+                d_Q = *(cl_mem*) src0->data;
             } else {
                 GGML_ASSERT(false);
             }
@@ -1029,13 +1029,14 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
 
     size_t q_size;
-    cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
+    cl_mem* dst = (cl_mem*) malloc(sizeof(cl_mem));
+    *dst = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
 
     // copy tensor to device
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             int i = i3*ne2 + i2;
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, i*ne0*ne1, tensor, i3, i2, NULL));
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, *dst, i*ne0*ne1, tensor, i3, i2, NULL));
         }
     }