[cmn/ggml/llama] reduced enum size using scoped enums (before C23 keyword is ignored)

This optimization technique reduces size structures where there is an scoped enum field, the smaller the size of the structure, the higher chance hitting to CPU cache L2 or L3.

Author: GermanAizek

common/chat.h

@@ -35,13 +35,13 @@ struct common_chat_tool {
     std::string parameters;
 };
 
-enum common_chat_tool_choice {
+enum common_chat_tool_choice : uint8_t  {
     COMMON_CHAT_TOOL_CHOICE_AUTO,
     COMMON_CHAT_TOOL_CHOICE_REQUIRED,
     COMMON_CHAT_TOOL_CHOICE_NONE,
 };
 
-enum common_chat_format {
+enum common_chat_format : uint8_t  {
     COMMON_CHAT_FORMAT_CONTENT_ONLY,
     COMMON_CHAT_FORMAT_GENERIC,
     COMMON_CHAT_FORMAT_MISTRAL_NEMO,

common/common.h

@@ -62,7 +62,7 @@ int32_t cpu_get_num_math();
 // Common params
 //
 
-enum llama_example {
+enum llama_example : uint8_t {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_MAIN,
@@ -84,7 +84,7 @@ enum llama_example {
     LLAMA_EXAMPLE_COUNT,
 };
 
-enum common_sampler_type {
+enum common_sampler_type : uint8_t {
     COMMON_SAMPLER_TYPE_NONE        = 0,
     COMMON_SAMPLER_TYPE_DRY         = 1,
     COMMON_SAMPLER_TYPE_TOP_K       = 2,
@@ -99,18 +99,18 @@ enum common_sampler_type {
 };
 
 // dimensionality reduction methods, used by cvector-generator
-enum dimre_method {
+enum dimre_method : uint8_t {
     DIMRE_METHOD_PCA,
     DIMRE_METHOD_MEAN,
 };
 
-enum common_conversation_mode {
+enum common_conversation_mode : uint8_t {
     COMMON_CONVERSATION_MODE_DISABLED = 0,
     COMMON_CONVERSATION_MODE_ENABLED  = 1,
     COMMON_CONVERSATION_MODE_AUTO     = 2,
 };
 
-enum common_grammar_trigger_type {
+enum common_grammar_trigger_type : uint8_t {
     COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
     COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
     COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
@@ -211,7 +211,7 @@ struct common_params_vocoder {
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
 
-enum common_reasoning_format {
+enum common_reasoning_format : uint8_t {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
 };

common/console.h

@@ -3,9 +3,10 @@
 #pragma once
 
 #include <string>
+#include <cstdint>
 
 namespace console {
-    enum display_t {
+    enum display_t : uint8_t {
         reset = 0,
         prompt,
         user_input,

common/log.cpp

@@ -20,7 +20,7 @@ static int64_t t_us() {
 }
 
 // colors
-enum common_log_col : int {
+enum common_log_col : uint8_t {
     COMMON_LOG_COL_DEFAULT = 0,
     COMMON_LOG_COL_BOLD,
     COMMON_LOG_COL_RED,

ggml/include/ggml-backend.h

@@ -46,7 +46,7 @@ extern "C" {
     // Backend buffer
     //
 
-    enum ggml_backend_buffer_usage {
+    enum ggml_backend_buffer_usage : uint8_t {
         GGML_BACKEND_BUFFER_USAGE_ANY = 0,
         GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
         GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2,
@@ -127,7 +127,7 @@ extern "C" {
     // Backend device
     //
 
-    enum ggml_backend_dev_type {
+    enum ggml_backend_dev_type : uint8_t {
         // CPU device using system memory
         GGML_BACKEND_DEVICE_TYPE_CPU,
         // GPU device using dedicated memory

ggml/include/ggml-cpu.h

@@ -22,7 +22,7 @@ extern "C" {
     };
 
     // numa strategies
-    enum ggml_numa_strategy {
+    enum ggml_numa_strategy : uint8_t {
         GGML_NUMA_STRATEGY_DISABLED   = 0,
         GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
         GGML_NUMA_STRATEGY_ISOLATE    = 2,

ggml/include/ggml-opt.h

@@ -27,7 +27,7 @@ extern "C" {
 
     // built-in loss types, i.e. the built-in quantities minimized by the optimizer
     // custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
-    enum ggml_opt_loss_type {
+    enum ggml_opt_loss_type : uint8_t {
         GGML_OPT_LOSS_TYPE_MEAN,
         GGML_OPT_LOSS_TYPE_SUM,
         GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
@@ -59,7 +59,7 @@ extern "C" {
 
     // ====== Model / Context ======
 
-    enum ggml_opt_build_type {
+    enum ggml_opt_build_type : uint8_t {
         GGML_OPT_BUILD_TYPE_FORWARD,
         GGML_OPT_BUILD_TYPE_GRAD,
         GGML_OPT_BUILD_TYPE_OPT,

ggml/include/ggml.h

@@ -317,7 +317,7 @@ extern "C" {
     GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
     GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
 
-    enum ggml_status {
+    enum ggml_status : int8_t {
         GGML_STATUS_ALLOC_FAILED = -2,
         GGML_STATUS_FAILED = -1,
         GGML_STATUS_SUCCESS = 0,
@@ -348,7 +348,7 @@ extern "C" {
     struct ggml_cgraph;
 
     // NOTE: always add types at the end of the enum to keep backward compatibility
-    enum ggml_type {
+    enum ggml_type : uint8_t {
         GGML_TYPE_F32     = 0,
         GGML_TYPE_F16     = 1,
         GGML_TYPE_Q4_0    = 2,
@@ -392,13 +392,13 @@ extern "C" {
     };
 
     // precision
-    enum ggml_prec {
+    enum ggml_prec : uint8_t {
         GGML_PREC_DEFAULT,
         GGML_PREC_F32,
     };
 
     // model file types
-    enum ggml_ftype {
+    enum ggml_ftype : int8_t {
         GGML_FTYPE_UNKNOWN        = -1,
         GGML_FTYPE_ALL_F32        = 0,
         GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
@@ -426,7 +426,7 @@ extern "C" {
     };
 
     // available tensor operations:
-    enum ggml_op {
+    enum ggml_op : uint8_t {
         GGML_OP_NONE = 0,
 
         GGML_OP_DUP,
@@ -520,7 +520,7 @@ extern "C" {
         GGML_OP_COUNT,
     };
 
-    enum ggml_unary_op {
+    enum ggml_unary_op : uint8_t {
         GGML_UNARY_OP_ABS,
         GGML_UNARY_OP_SGN,
         GGML_UNARY_OP_NEG,
@@ -539,13 +539,13 @@ extern "C" {
         GGML_UNARY_OP_COUNT,
     };
 
-    enum ggml_object_type {
+    enum ggml_object_type : uint8_t {
         GGML_OBJECT_TYPE_TENSOR,
         GGML_OBJECT_TYPE_GRAPH,
         GGML_OBJECT_TYPE_WORK_BUFFER
     };
 
-    enum ggml_log_level {
+    enum ggml_log_level : uint8_t {
         GGML_LOG_LEVEL_NONE  = 0,
         GGML_LOG_LEVEL_DEBUG = 1,
         GGML_LOG_LEVEL_INFO  = 2,
@@ -555,7 +555,7 @@ extern "C" {
     };
 
     // this tensor...
-    enum ggml_tensor_flag {
+    enum ggml_tensor_flag : uint8_t {
         GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
         GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
         GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
@@ -1678,7 +1678,7 @@ extern "C" {
             struct ggml_tensor  * b,
             int                   stride);
 
-    enum ggml_op_pool {
+    enum ggml_op_pool : uint8_t {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
         GGML_OP_POOL_COUNT,
@@ -1717,7 +1717,7 @@ extern "C" {
             float                 p0,
             float                 p1);
 
-    enum ggml_scale_mode {
+    enum ggml_scale_mode : uint8_t {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
     };
@@ -1767,7 +1767,7 @@ extern "C" {
             int                   max_period);
 
     // sort rows
-    enum ggml_sort_order {
+    enum ggml_sort_order : uint8_t {
         GGML_SORT_ORDER_ASC,
         GGML_SORT_ORDER_DESC,
     };
@@ -2137,7 +2137,7 @@ extern "C" {
     // the goal should be to create an API that other backends can use move everything to the ggml base
 
     // scheduling priorities
-    enum ggml_sched_priority {
+    enum ggml_sched_priority : uint8_t  {
         GGML_SCHED_PRIO_NORMAL,
         GGML_SCHED_PRIO_MEDIUM,
         GGML_SCHED_PRIO_HIGH,

ggml/include/gguf.h

@@ -50,7 +50,7 @@ extern "C" {
 #endif
 
     // types that can be stored as GGUF KV data
-    enum gguf_type {
+    enum gguf_type : uint8_t {
         GGUF_TYPE_UINT8   = 0,
         GGUF_TYPE_INT8    = 1,
         GGUF_TYPE_UINT16  = 2,

ggml/src/ggml-cpu/ggml-cpu.c

@@ -114,7 +114,7 @@ typedef atomic_int atomic_flag;
 
 #define ATOMIC_FLAG_INIT 0
 
-typedef enum {
+typedef enum : uint8_t {
     memory_order_relaxed,
     memory_order_consume,
     memory_order_acquire,