applied Romain's comments

Author: lslusarczyk

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -2824,21 +2824,20 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 dnn_gemm(src1_f16, src0_f16, dst_ddf, ne12*ne13, ne02 * ne03);
             }
             else {
-
                 for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
-                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie03*nb03)/2); // div2 cuz nb is in bytes and pointer is in f16 (2 bytes)
+                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie03*nb03)/sizeof(sycl::half)); // nb is in bytes
                     const sycl::half* src1_f16_shifted = src1_f16 + ie03*s13;
-                    float* dst_shifted = dst_ddf + ((ie03*nb3)/4); // div4 cuz nb is in bytes and pointer is float (4 bytes)
+                    float* dst_shifted = dst_ddf + ((ie03*nb3)/sizeof(float));
                     dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, ne12, ne02);
                 }
             }
         } else {
             // iterate over batches from smaller set of matrices (matrix 0)
             for (int64_t ie02 = 0; ie02 < ne02; ++ie02) {
                 for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
-                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie02*nb02 + ie03*nb03)/2); // div2 cuz nb is in bytes and pointer is in f16 (2 bytes)
+                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie02*nb02 + ie03*nb03)/sizeof(sycl::half));
                     const sycl::half* src1_f16_shifted = src1_f16 + ie02*s12*r2 + ie03*s13*r3;
-                    float* dst_shifted = dst_ddf + ((ie02*nb2*r2 + ie03*nb3*r3)/4); // div4 cuz nb is in bytes and pointer is float (4 bytes)
+                    float* dst_shifted = dst_ddf + ((ie02*nb2*r2 + ie03*nb3*r3)/sizeof(float));
                     dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, r2*r3, 1);
                 }
             }

tests/test-backend-ops.cpp

@@ -3919,7 +3919,7 @@ static const ggml_type other_types[] = {
 // Test cases for evaluation: should try to cover edge cases while using small input sizes to keep the runtime low
 static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     std::vector<std::unique_ptr<test_case>> test_cases;
-    [[maybe_unused]] std::default_random_engine rng(0);
+    std::default_random_engine rng(0);
 
     // unary ops
     for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
@@ -4242,8 +4242,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
 
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  16, 256, {2, 1}, {1, 1}, {0, 2, 1, 3}));
-
             // test cases with large ne00/ne10 to cover stream-k fixup
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 1024, {3, 2}, {1, 1}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 1024, {3, 2}, {1, 1}));