vulkan: improve topk perf for large k, fix overflow in unit tests (#17582)

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -10239,7 +10239,9 @@ static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, cons
 
         // Prefer going as small as num_topk_pipelines - 3 for perf reasons.
         // But if K is larger, then we need a larger workgroup
-        uint32_t max_pipeline = num_topk_pipelines - 3;
+        uint32_t max_pipeline = num_topk_pipelines - 1;
+        uint32_t preferred_pipeline = std::max(num_topk_pipelines - 3, (uint32_t)log2f(float(k)) + 2);
+        max_pipeline = std::min(preferred_pipeline, max_pipeline);
         uint32_t min_pipeline = (uint32_t)log2f(float(k)) + 1;
         // require full subgroup
         min_pipeline = std::max(min_pipeline, ctx->device->subgroup_size_log2);

tests/test-backend-ops.cpp

@@ -1446,14 +1446,14 @@ struct test_case {
             const uint64_t target_flops_cpu =   8ULL * GFLOP;
             const uint64_t target_flops_gpu = 100ULL * GFLOP;
             uint64_t target_flops = is_cpu ? target_flops_cpu : target_flops_gpu;
-            n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_flops / op_flops(out)) + 1;
+            n_runs = (int)std::min<int64_t>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_flops / op_flops(out)) + 1;
         } else {
             // based on memory size
             const size_t GB = 1ULL << 30;
             const size_t target_size_cpu =  8 * GB;
             const size_t target_size_gpu = 32 * GB;
             size_t target_size = is_cpu ? target_size_cpu : target_size_gpu;
-            n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_size / op_size(out)) + 1;
+            n_runs = (int)std::min<int64_t>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_size / op_size(out)) + 1;
         }
 
         // duplicate the op
@@ -8043,7 +8043,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     }
 
     test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {65000, 16, 1, 1}));
-    for (auto k : {1, 10, 40}) {
+
+    test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {2, 1, 1, 1}, 1));
+    for (auto k : {1, 10, 40, 400}) {
         for (auto nrows : {1, 16}) {
             for (auto cols : {k, 1000, 65000, 200000}) {
                 test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {cols, nrows, 1, 1}, k));