CLBlast: Fix uploading tensor data to device

ggml_cl_transform_tensor:
* Fix offset of 3D and 4D tensors.

ggml_cl_h2d_tensor_2d:
* Fix parameters for non-contiguous data.
* Correct handling of OpenCL events when multiple commands are queued.

Author: shibe2

ggml-opencl.cpp

@@ -1349,30 +1349,42 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
     const enum ggml_type type = src->type;
     const size_t ts = ggml_type_size(type);
     const size_t bs = ggml_blck_size(type);
+    const uint64_t row_size = ts*ne0/bs;
 
-    const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
-    if (nb0 == ts && nb1 == ts*ne0/bs) {
-        err = clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*nb1, x, 0, NULL, ev);
-        return err;
+    const char * x = (const char *) src->data + i2*nb2 + i3*nb3;
+    if (nb0 == ts && nb1 == row_size) {
+        return clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*row_size, x, 0, NULL, ev);
     }
     if (nb0 == ts) {
         const size_t buffer_origin[3] = { offset, 0, 0 };
         const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts*ne0/bs, ne1, 1 };
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts*ne0/bs, 0, nb1, 0, x, 0, NULL, ev);
-        return err;
+        const size_t region[3] = { row_size, ne1, 1 };
+        return clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, row_size, 0, nb1, 0, x, 0, NULL, ev);
     }
+    std::vector<cl_event> events;
+    if (ev && ne1>1) events.reserve(ne1-1);
     for (uint64_t i1 = 0; i1 < ne1; i1++) {
         // pretend the row is a matrix with cols=1
-        const size_t buffer_origin[3] = { offset, i1, 0 };
+        const size_t buffer_origin[3] = { offset + i1*row_size, 0, 0 };
         const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts/bs, ne0, 1 };
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, 0, 0, nb0, 0, ((const char *)x) + i1*nb0, 0, NULL, ev);
+        const size_t region[3] = { ts, ne0/bs, 1 };
+        // if an event is requested, make the last write wait for all previous writes to complete
+        if (ev && i1) {
+            events.push_back(*ev);
+        }
+        cl_uint nevents = i1 == ne1-1 ? events.size() : 0U;
+        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts, 0, nb0, 0, x + i1*nb1, nevents, nevents ? events.data() : nullptr, ev);
         if (err != CL_SUCCESS) {
-            break;
+            for (auto event : events) {
+                clReleaseEvent(event);
+            }
+            return err;
         }
     }
-    return err;
+    for (auto event : events) {
+        CL_CHECK(clReleaseEvent(event));
+    }
+    return CL_SUCCESS;
 }
 
 static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -1888,17 +1900,19 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     const int64_t ne3 = tensor->ne[3];
 
     const ggml_type type = tensor->type;
-    const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
+    const size_t s_sz = ggml_type_size(type) * (size_t) (ne0 * ne1 / ggml_blck_size(type));
+    const size_t q_sz = s_sz * (size_t) (ne2 * ne3);
 
     size_t q_size;
     cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
 
     tensor->data = data;
     // copy tensor to device
+    size_t offset = 0;
     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, offset, tensor, i3, i2, NULL));
+            offset += s_sz;
         }
     }