tests: add mul_mat perf/functional tests for p021/nc vulkan shaders

Author: jeffbolznv

tests/test-backend-ops.cpp

@@ -1964,9 +1964,10 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> bs;  // dims 3 and 4
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
+    const bool v; // whether a is a non-contiguous view
 
     std::string vars() override {
-        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
+        return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
     }
 
     double max_nmse_err() override {
@@ -1986,8 +1987,9 @@ struct test_mul_mat : public test_case {
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             std::array<int64_t, 2> bs = {10, 10},
             std::array<int64_t, 2> nr = {2, 2},
-            std::array<int64_t, 4> per = {0, 1, 2, 3})
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
+            std::array<int64_t, 4> per = {0, 1, 2, 3},
+            bool v = false)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -1997,6 +1999,7 @@ struct test_mul_mat : public test_case {
         const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
         if (npermuted > 0) {
             GGML_ASSERT(npermuted == 2);
+            GGML_ASSERT(!v); // not handled
             GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
             GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
 
@@ -2020,7 +2023,13 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+
+            if (v) {
+                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
+                a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
+            } else {
+                a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+            }
             b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
@@ -4176,6 +4185,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
 
+    for (auto bs : {1,2,4,8}) {
+        for (auto nr : {1,4}) {
+            for (uint32_t m = 0; m < 2; ++m) {
+                for (uint32_t k = 0; k < 2; ++k) {
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+                }
+            }
+        }
+    }
+
     // sycl backend will limit task global_range < MAX_INT
     // test case for f16-type-convert-to-fp32 kernel with large k under fp32 compute dtype (occurs in stable-diffusion)
     // however this case needs to alloc more memory which may fail in some devices (Intel Arc770, etc.)
@@ -4444,6 +4464,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, true));
+
     for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {