ggml : fix q4xx mat mul, increase ggml_aligned_malloc alignment (llama/10167)

ggml/src/ggml-cpu.c

@@ -304,6 +304,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
     },
     [GGML_TYPE_Q8_0] = {
+        .from_float_to_mat        = quantize_mat_q8_0,
         .vec_dot                  = ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -13692,9 +13693,7 @@ void ggml_cpu_init(void) {
                     uint16_t u16;
                     ggml_fp16_t fp16;
                 } u = {i};
-                // FIXME: this table is used in conversion functions outside of compute
-                // current code depends on ggml_init initializing this table
-                float f = ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
+                float f = GGML_FP16_TO_FP32(u.fp16);
                 ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
                 ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }

ggml/src/ggml.c

@@ -220,8 +220,10 @@ void ggml_log_callback_default(enum ggml_log_level level, const char * text, voi
 
 
 void * ggml_aligned_malloc(size_t size) {
+    const int alignment = 64;
+
 #if defined(_MSC_VER) || defined(__MINGW32__)
-    return _aligned_malloc(size, TENSOR_ALIGNMENT);
+    return _aligned_malloc(size, alignment);
 #else
     if (size == 0) {
         GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
@@ -229,8 +231,9 @@ void * ggml_aligned_malloc(size_t size) {
     }
     void * aligned_memory = NULL;
   #ifdef GGML_USE_CPU_HBM
-    int result = hbw_posix_memalign(&aligned_memory, TENSOR_ALIGNMENT, size);
+    int result = hbw_posix_memalign(&aligned_memory, alignment, size);
   #elif TARGET_OS_OSX
+    GGML_UNUSED(alignment);
     kern_return_t alloc_status = vm_allocate((vm_map_t) mach_task_self(), (vm_address_t *) &aligned_memory, size, VM_FLAGS_ANYWHERE);
     int result = EFAULT;
     switch (alloc_status) {
@@ -248,7 +251,7 @@ void * ggml_aligned_malloc(size_t size) {
             break;
     }
   #else
-    int result = posix_memalign(&aligned_memory, TENSOR_ALIGNMENT, size);
+    int result = posix_memalign(&aligned_memory, alignment, size);
   #endif
     if (result != 0) {
         // Handle allocation failure