ggml: add CUDA backend to conv2d-sched

I'm trying to reproduce and issue I'm having with a program where the
input tensor that is convoluted over seems to be modified by the conv2d
operation which is not happening in my example program here. But this
will be useful for debugging the issue and stepping though a standlone
program as the other is pretty "heavy" with regard to startup and
running.

fundamentals/ggml/src/conv2d-sched.c

@@ -3,6 +3,7 @@
 
 #include "ggml.h"
 #include "ggml-cpu.h"
+#include "ggml-cuda.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
@@ -43,23 +44,38 @@ int main(int argc, char **argv) {
     ggml_build_forward_expand(c_graph, conv);
 
 
-    ggml_backend_t backend = ggml_backend_init_by_name("CPU", NULL);
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
-    ggml_backend_buffer_type_t buffer_type =ggml_backend_buffer_get_type(buffer);
+    ggml_backend_t cpu_backend = ggml_backend_init_by_name("CPU", NULL);
+    ggml_backend_t cuda_backend = ggml_backend_init_by_name("CUDA0", NULL);
+    if (!cuda_backend) {
+        fprintf(stderr, "Failed to initialize CUDA backend\n");
+        return 1;
+    }
+    ggml_backend_t backends[] = {cuda_backend, cpu_backend};
+
+    // Get buffer types
+    ggml_backend_buffer_type_t buffer_types[] = {
+        ggml_backend_get_default_buffer_type(cuda_backend),
+        ggml_backend_get_default_buffer_type(cpu_backend)
+    };
+
+    printf("CUDA buffer type: %s\n", ggml_backend_buft_name(buffer_types[0]));
+    printf("CPU buffer type: %s\n", ggml_backend_buft_name(buffer_types[1]));
 
     size_t graph_work_size = ggml_graph_size(c_graph);
     size_t graph_nodes = ggml_graph_n_nodes(c_graph);
-    ggml_backend_sched_t sched = ggml_backend_sched_new(&backend, &buffer_type, 1, graph_nodes, graph_work_size);
+    ggml_backend_sched_t sched = ggml_backend_sched_new(backends, buffer_types, 2, graph_nodes, graph_work_size);
     if (!sched) {
         fprintf(stderr, "Failed to create scheduler\n");
-        ggml_backend_free(backend);
+        ggml_backend_free(cpu_backend);
+        ggml_backend_free(cuda_backend);
         return 1;
     }
 
     if (!ggml_backend_sched_alloc_graph(sched, c_graph)) {
         fprintf(stderr, "Failed to allocate graph\n");
         ggml_backend_sched_free(sched);
-        ggml_backend_free(backend);
+        ggml_backend_free(cpu_backend);
+        ggml_backend_free(cuda_backend);
         return 1;
     }
 
@@ -95,12 +111,16 @@ int main(int argc, char **argv) {
     printf("\n");
 
     {
-        printf("\na:\n");
-        float* a_data = (float*)a->data;
-        for (int i = 0; i < ggml_nelements(a); i++) {
-            printf("%.2f ", a_data[i]);
+        struct ggml_tensor * tensor = ggml_graph_get_tensor(c_graph, "a");
+        float buf[4];
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        printf("Tensor type: %s\n", ggml_type_name(tensor->type));
+        printf("Backend type: %s\n", ggml_backend_name(backend));
+        ggml_backend_tensor_get_async(backend, tensor, buf, 0, sizeof(buf));
+        ggml_backend_sched_synchronize(sched);
+        for (int i = 0; i < 4; i++) {
+            printf("a[%d] = %f\n", i, buf[i]);
         }
-        printf("\n");
     }
 
     ggml_free(ctx);