ggml: add CUDA backend to sched-issue

danbev · Dec 17, 2024 · a9c7deb · a9c7deb
1 parent 8b967b5
commit a9c7deb
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 1 deletion.
diff --git a/fundamentals/ggml/Makefile b/fundamentals/ggml/Makefile
@@ -40,7 +40,7 @@ CXXFLAGS += -fopenmp
 WHOLE_ARCHIVE  = -Wl,--whole-archive ggml/build/src/libggml.a -Wl,--no-whole-archive 
 WHOLE_ARCHIVE += -Wl,--whole-archive ggml/build/src/libggml-base.a -Wl,--no-whole-archive 
 WHOLE_ARCHIVE += -Wl,--whole-archive ggml/build/src/libggml-cpu.a -Wl,--no-whole-archive 
-#WHOLE_ARCHIVE += -Wl,--whole-archive ggml/build/src/ggml-cuda/libggml-cuda.a -Wl,--no-whole-archive 
+WHOLE_ARCHIVE += -Wl,--whole-archive ggml/build/src/ggml-cuda/libggml-cuda.a -Wl,--no-whole-archive 
 else ifeq ($(OS),Darwin)
 CFLAGS += -framework Metal -framework Foundation -framework MetalKit -framework Accelerate
 CXXFLAGS += -framework Metal -framework Foundation -framework MetalKit -framework Accelerate
@@ -90,6 +90,11 @@ conv2d-sched-cuda: CUDA_FLAGS := -L${CUDA_DIR} -lcuda -lcublas -lculibos -lcudar
 conv2d-sched-cuda: CC = $(CXX)
 conv2d-sched-cuda: conv2d-sched
 
+#.PHONY: sched-issue-cuda
+sched-issue-cuda: CUDA_FLAGS := -L${CUDA_DIR} -lcuda -lcublas -lculibos -lcudart -lcublasLt
+sched-issue-cuda: CC = $(CXX)
+sched-issue-cuda: sched-issue
+
 .PHONY: run-backend-vulkan
 run-backend-vulkan: backend-vulkan
 	env GGML_VULKAN_DEVICE=1 ./bin/backend

diff --git a/fundamentals/ggml/src/sched-issue.c b/fundamentals/ggml/src/sched-issue.c
@@ -30,6 +30,9 @@ int main(int argc, char **argv) {
     };
     struct ggml_context* ctx = ggml_init(params);
 
+    size_t count = ggml_backend_reg_count();
+    printf("Number of backends registered: %ld\n", count);
+
     // Tensors for the first graph (simulating the language model):
     struct ggml_tensor* l_a = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
     ggml_set_input(l_a);
@@ -49,6 +52,11 @@ int main(int argc, char **argv) {
 
     // Create a scheduler.
     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. Did you forget to compile ggml with GGML_CUDA?\n");
+        return 1;
+    }
 
     auto threadpool_params = ggml_threadpool_params_default(1);
     auto threadpool = ggml_threadpool_new(&threadpool_params);