Log HAVE_FANCY_SIMD via LLAMA_LOG_INFO (#1492)

ggml/src/ggml.c

@@ -16865,17 +16865,6 @@ static int ggml_compute_forward_mul_mat(
     //   compute by src0 rows
 
 #if GGML_USE_IQK_MULMAT
-    if (ith == 0) {
-        static bool first_time = true;
-        if (first_time) {
-            first_time = false;
-#ifdef HAVE_FANCY_SIMD
-            printf("======================================= HAVE_FANCY_SIMD is defined\n");
-#else
-            printf("======================================= HAVE_FANCY_SIMD is NOT defined\n");
-#endif
-        }
-    }
     if (dst->type == GGML_TYPE_F32) {
         if (iqk_mul_mat_4d(ne01, ne11, ne00,
                     ne02, ne03, ne12, ne13, nb02, nb03, nb12, nb13, nb2/sizeof(float), nb3/sizeof(float),

ggml/src/iqk/iqk_cpu_ops.cpp

@@ -22,6 +22,14 @@
 #include <arm_neon.h>
 #endif
 
+bool iqk_has_fancy_simd(void) {
+#ifdef HAVE_FANCY_SIMD
+    return true;
+#else
+    return false;
+#endif
+}
+
 namespace {
 // Playing around with group scores: use sum of probabilities in the group
 inline float group_score(int n_per_group, const float * data) {

ggml/src/iqk/iqk_cpu_ops.h

@@ -14,6 +14,8 @@ extern "C" {
 
 struct ggml_tensor;
 
+bool iqk_has_fancy_simd(void);
+
 void iqk_sumrows_div(struct ggml_tensor * div, int ith, int nth);
 
 void iqk_grouped_top_k(struct ggml_tensor * dst, int ith, int nth);

src/llama.cpp

@@ -25,8 +25,9 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-// TODO: fix this include
+// TODO: fix these includes
 #include "iqk/iqk_quantize.h"
+#include "iqk/iqk_cpu_ops.h"
 
 #define IK_PRINT_TIMING 0
 
@@ -2082,6 +2083,12 @@ static bool llm_load_tensors(
         }
     }
 
+    if (iqk_has_fancy_simd()) {
+        LLAMA_LOG_INFO("======================================= HAVE_FANCY_SIMD is defined\n");
+    } else {
+        LLAMA_LOG_INFO("======================================= HAVE_FANCY_SIMD is NOT defined\n");
+    }
+
     model.split_mode   = split_mode;
     model.main_gpu     = main_gpu;
     model.max_gpu      = max_gpu;
@@ -4713,7 +4720,7 @@ struct llama_model * llama_model_load_from_file(
         if (has_rpc) {
             for (auto& it : model->rpc_servers) {
                 device_names.push_back(create_rpc_name(it.endpoint, it.device));
-            }     
+            }
         }
         device_names.insert(device_names.end(), gpu_names.begin(), gpu_names.end());
     }
@@ -4726,7 +4733,6 @@ struct llama_model * llama_model_load_from_file(
         }
     }
 
- 
     // no gpu used, so set layers offload to be 0
     if (!model->devices.size()) {
         params.n_gpu_layers = 0;
@@ -4737,7 +4743,7 @@ struct llama_model * llama_model_load_from_file(
             const char* name = ggml_backend_buft_name(buft);
             const char* description = name;
             size_t description_size = llama_get_device_memory(*model, i);
-            LLAMA_LOG_INFO("%s: using device %s - %zu MiB free\n", 
+            LLAMA_LOG_INFO("%s: using device %s - %zu MiB free\n",
                 name, description,
                 description_size / 1024 / 1024);
         }