Update stats tool for unbounded's method

Author: sw

examples/quantize-stats/quantize-stats.cpp

@@ -17,7 +17,7 @@
 static const char * type_strs[] = { "q4_0", "q4_1", "i8", "i16", "i32", "f16", "f32"  };
 static_assert(sizeof(type_strs) == GGML_TYPE_COUNT * sizeof(char *), "Incomplete type list");
 
-static const char * impl_strs[] = { "simd", "reference", "rmse" };
+static const char * impl_strs[] = { "simd", "reference", "rmse-sw", "rmse-unbounded" };
 static_assert(sizeof(impl_strs) == GGML_QUANTIZE_IMPL_COUNT * sizeof(char *), "Incomplete implementation list");
 
 struct quantize_stats_params {
@@ -52,7 +52,7 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) {
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "  -i, --implementation\n");
-    fprintf(stderr, "                        select implementation (simd, reference, rmse)\n");
+    fprintf(stderr, "                        select implementation (simd, reference, rmse-sw, rmse-unbounded)\n");
     fprintf(stderr, "  -v, --verbose\n");
     fprintf(stderr, "                        verbose output (default: false)\n");
     fprintf(stderr, "  -p, --per-layer-stats\n");
@@ -111,7 +111,7 @@ void print_error_stats(const std::string & name, ggml_quantize_impl_t impl, cons
     double rmse = sqrt(stats.total_error / (double) stats.num_samples);
     double median = find_quantile(stats, .5);
     double pct95 = find_quantile(stats, .95);
-    printf("%-4s %-10s: rmse %.8f, maxerr %.8f, 95pct<%.4f, median<%.4f\n",
+    printf("%-4s %-15s: rmse %.8f, maxerr %.8f, 95pct<%.4f, median<%.4f\n",
         name.c_str(), impl_strs[impl], rmse, stats.max_error, pct95, median);
     if (print_histogram) {
         printf("Error distribution:\n");

ggml.c

@@ -6799,18 +6799,20 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = {
         .dequantize_row_q         = dequantize_row_q4_0,
         .quantize_row_q           = {
-            [GGML_QUANTIZE_IMPL_SIMD]      = quantize_row_q4_0,
-            [GGML_QUANTIZE_IMPL_REFERENCE] = (quantize_row_q_t)quantize_row_q4_0_reference,
-            [GGML_QUANTIZE_IMPL_RMSE]      = (quantize_row_q_t)quantize_row_q4_0_rmse,
+            [GGML_QUANTIZE_IMPL_SIMD]           = quantize_row_q4_0,
+            [GGML_QUANTIZE_IMPL_REFERENCE]      = (quantize_row_q_t)quantize_row_q4_0_reference,
+            [GGML_QUANTIZE_IMPL_RMSE_SW]        = (quantize_row_q_t)quantize_row_q4_0_rmse,
+            [GGML_QUANTIZE_IMPL_RMSE_UNBOUNDED] = (quantize_row_q_t)quantize_row_q4_0_slow,
         },
         .vec_dot_q                = ggml_vec_dot_q4_0,
     },
     [GGML_TYPE_Q4_1] = {
         .dequantize_row_q         = dequantize_row_q4_1,
         .quantize_row_q           = {
-            [GGML_QUANTIZE_IMPL_SIMD]      = quantize_row_q4_1,
-            [GGML_QUANTIZE_IMPL_REFERENCE] = quantize_row_q4_1_reference,
-            [GGML_QUANTIZE_IMPL_RMSE]      = quantize_row_q_missing,
+            [GGML_QUANTIZE_IMPL_SIMD]           = quantize_row_q4_1,
+            [GGML_QUANTIZE_IMPL_REFERENCE]      = quantize_row_q4_1_reference,
+            [GGML_QUANTIZE_IMPL_RMSE_SW]        = quantize_row_q_missing,
+            [GGML_QUANTIZE_IMPL_RMSE_UNBOUNDED] = quantize_row_q_missing,
         },
         .vec_dot_q                = ggml_vec_dot_q4_1,
     },

ggml.h

@@ -795,7 +795,8 @@ typedef void (*vec_dot_q_t)(const int n, float * s, const void * x, const void *
 typedef enum {
     GGML_QUANTIZE_IMPL_SIMD,
     GGML_QUANTIZE_IMPL_REFERENCE,
-    GGML_QUANTIZE_IMPL_RMSE,
+    GGML_QUANTIZE_IMPL_RMSE_SW,
+    GGML_QUANTIZE_IMPL_RMSE_UNBOUNDED,
     GGML_QUANTIZE_IMPL_COUNT
 } ggml_quantize_impl_t;