Merge branch 'add_dispatching'

Author: lemire

cppcon2025/cppcon_2025_slides.md

@@ -75,6 +75,34 @@ JSON can be *slow*. E.g., 20 MB/s.
 <img src="images/simdjson.png" width="10%" />
 
 
+---
+
+
+## SIMD
+
+- Stands for Single instruction, multiple data
+- Allows us to process 16 (or more) bytes or more with one instruction
+- Supported on all modern CPUs (phone, laptop)
+
+---
+
+# Superscalar vs. SIMD execution
+
+| processor       | year    | arithmetic logic units    | SIMD units     | simdjson |
+|-----------------|---------|---------------------------|----------------|----------|               
+| Apple M*       |  2019   |    6+                      | $4 \times 128$ | 🥉        |
+| Intel Lion Cove       |  2024   |    6                | $4 \times 256$ | 🥈🥈        |
+| AMD Zen 5       |  2024   |    6                      | $4 \times 512$ | 🥇🥇🥇        |
+
+
+---
+
+
+<img src="images/simdjson_benchmark.png" width="60%"/>
+
+https://openbenchmarking.org/test/pts/simdjson
+
+
 ---
 
 # Usage
@@ -445,7 +473,8 @@ template <typename T>
   requires(std::is_class_v<T>)  // For user-defined types
 error_code deserialize(auto& json_value, T& out) {
     simdjson::ondemand::object obj;
-    SIMDJSON_TRY(json_value.get_object().get(obj));
+    auto er = json_value.get_object().get(obj);
+    if(er) { return er; }
 
     // This [:expand:] happens at COMPILE TIME
     // It literally generates code for each member
@@ -822,14 +851,6 @@ simple_needs_escaping(std::string_view v) {
 }
 ```
 
----
-
-
-## SIMD
-
-- Stands for Single instruction, multiple data
-- Allows us to process 16 (or more) bytes or more with one instruction
-- Supported on all modern CPUs (phone, laptop)
 
 ---
 

cppcon2025/images/barsimdjson.py

@@ -0,0 +1,43 @@
+import matplotlib.pyplot as plt
+
+# Data from the user
+processors = [
+    "AMD Ryzen 9 9950X3D",
+    "Intel Core Ultra 9 285K",
+    "Snapdragon X Elite X1E78100",
+    "ARMv8 Neoverse-V2 72-Core"
+]
+gb_per_second = [13.64, 10.07, 4.27, 4.06]
+
+
+# Define colors for each processor type
+colors = ['#FF6F61', "#0655F1", '#2ECC71', '#2ECC71']  # AMD: coral, Intel: gray, ARM: green
+
+# Create bar chart
+plt.figure(figsize=(10, 6))
+bars = plt.bar(processors, gb_per_second, color=colors, edgecolor='black')
+
+# Remove top and right spines
+ax = plt.gca()
+ax.spines['top'].set_visible(False)
+ax.spines['right'].set_visible(False)
+
+# Add data labels on top of each bar
+for bar in bars:
+    yval = bar.get_height()
+    plt.text(bar.get_x() + bar.get_width()/2, yval + 0.2, f'{yval:.2g} GB/s',
+             ha='center', va='bottom', fontsize=14)
+
+plt.ylabel('Throughput (GB/s)', fontsize=14)
+
+# Add legend
+plt.legend(handles=[
+    plt.Rectangle((0,0),1,1,fc=colors[0],edgecolor='black'),
+    plt.Rectangle((0,0),1,1,fc=colors[1],edgecolor='black'),
+    plt.Rectangle((0,0),1,1,fc=colors[2],edgecolor='black')
+], labels=['AMD', 'Intel', 'ARM'], loc='upper right', fontsize=16, frameon=False)
+plt.tight_layout()
+
+# Save the plot as PNG
+plt.savefig('simdjson_benchmark.png', dpi=300, bbox_inches='tight')
+plt.close()

cppcon2025/images/simdjson_benchmark.png

@@ -26,9 +26,8 @@ float sum_avx2(const float *data, size_t n) {
 }
 
 SumFunc &get_sum_fnc();
-// Fonction d'initialisation pour le dispatching
+
 float sum_init(const float *data, size_t n) {
-  std::cout << "Initialisation de la fonction sum...\n";
   SumFunc &sum_impl = get_sum_fnc();
   if (has_avx2()) {
     sum_impl = sum_avx2;
@@ -40,21 +39,17 @@ float sum_init(const float *data, size_t n) {
   return sum_impl(data, n);
 }
 
-// Gestion du pointeur de fonction statique
 SumFunc &get_sum_fnc() {
   static SumFunc sum_impl = sum_init;
   return sum_impl;
 }
 
-// Fonction principale avec dispatching
 float sum(const float *data, size_t n) { return get_sum_fnc()(data, n); }
 
 int main() {
   float data[] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f};
   size_t n = sizeof(data) / sizeof(data[0]);
   float result = sum(data, n);
   std::cout << "sum : " << result << std::endl;
-  float result2 = sum(data, n);
-  std::cout << "sum : " << result2 << std::endl;
   return 0;
 }