Fix too relaxed check on CUDA "fast copy" (can_be_transposed) condition (#17332)

* Fix too relaxed check on CUDA "fast copy" (can_be_transposed) condition

* Argh.

* Making CISC happy ;)

* Integrate CONT tests

* Use loopy loop

* Skip new tests for (B)F16 for now.

ggml/src/ggml-cuda/cpy.cu

@@ -384,7 +384,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src1_ddc = (char *) src1->data;
 
     const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
-    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) && src0->ne[3] == 1;
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) &&
+        src0->ne[3] == 1 && nb02 == ne00 * ne01 * (int64_t)ggml_element_size(src0);
 
     if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));

tests/test-backend-ops.cpp

@@ -2776,24 +2776,34 @@ struct test_cpy : public test_case {
 struct test_cont : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
+    bool use_view_slice;
 
     std::string vars() override {
-        return VARS_TO_STR2(type, ne);
+        return VARS_TO_STR3(type, ne, use_view_slice);
     }
 
     test_cont(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 10, 10, 1})
-        : type(type), ne(ne) {}
+            std::array<int64_t, 4> ne = {10, 10, 10, 1},
+            bool use_view_slice = false)
+        : type(type), ne(ne), use_view_slice(use_view_slice) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_param(src);
         ggml_set_name(src, "src");
 
-        src = ggml_transpose(ctx, src);
-        ggml_set_name(src, "src_transposed");
 
-        ggml_tensor * out = ggml_cont(ctx, src);
+        ggml_tensor * dst;
+        if (use_view_slice) {
+            dst = ggml_view_4d(ctx, src, src->ne[0], 1, src->ne[2], src->ne[3],
+                src->nb[1], src->nb[2], src->nb[3], src->nb[0] * (src->ne[1] - 1));
+            ggml_set_name(dst, "src_view_slice");
+        } else {
+            dst = ggml_transpose(ctx, src);
+            ggml_set_name(dst, "src_transposed");
+        }
+
+        ggml_tensor * out = ggml_cont(ctx, dst);
         ggml_set_name(out, "out");
 
         return out;
@@ -6945,16 +6955,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 1, 4, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
 
-    test_cases.emplace_back(new test_cont());
-    test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 1, 1 ,1}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 1, 3 ,5}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 3, 5 ,7}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_F16, {2, 1, 1 ,1}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_F16, {2, 1, 3 ,5}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_F16, {2, 3, 5 ,7}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 1, 1 ,1}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 1, 3 ,5}));
-    test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 3, 5 ,7}));
+    for (ggml_type type_dst : { GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16 }) {
+        for (bool use_view_slice : { true, false }) {
+            for (std::array<int64_t, 4> ne : std::initializer_list<std::array<int64_t, 4>>{ {2, 1, 1, 1}, {2, 1, 3, 5},
+                {2, 3, 5, 7}, {1, 4, 4, 1}, {1, 8, 17, 1}, {10, 10, 10, 1} }) {
+                if (use_view_slice && (type_dst == GGML_TYPE_F16 || type_dst == GGML_TYPE_BF16)) {
+                    continue; // TODO: add after WebGPU is fixed
+                }
+                test_cases.emplace_back(new test_cont(type_dst, ne, use_view_slice));
+            }
+        }
+    }
 
     auto add_test_bin_bcast = [&](ggml_type type, std::array<int64_t, 4> ne, std::array<int, 4> nr) {
         for (auto op : {ggml_add, ggml_sub, ggml_mul, ggml_div}) {