Fix priority of args -use vs. per-FFT config

Author: preda

src/Gpu.cpp

@@ -151,11 +151,19 @@ constexpr bool isInList(const string& s, initializer_list<string> list) {
 }
 
 string clDefines(const Args& args, cl_device_id id, FFTConfig fft, const vector<KeyVal>& extraConf, u32 E, bool doLog) {
-  map<string, string> config{args.flags};
+  map<string, string> config;
+
+  // Highest priority is the requested "extra" conf
+  config.insert(extraConf.begin(), extraConf.end());
+
+  // Next, args config
+  config.insert(args.flags.begin(), args.flags.end());
+
+  // Lowest priority: the per-FFT config if any
   if (auto it = args.perFftConfig.find(fft.shape.spec()); it != args.perFftConfig.end()) {
+    // log("Found %s\n", fft.shape.spec().c_str());
     config.insert(it->second.begin(), it->second.end());
   }
-  config.insert(extraConf.begin(), extraConf.end());
 
   for (const auto& [k, v] : config) {
     bool isValid = isInList(k, {

src/tune.cpp

@@ -225,7 +225,8 @@ void Tune::ctune() {
   }
 
   for (FFTShape shape : shapes) {
-    u32 exponent = primes.prevPrime(FFTConfig{shape, 0, CARRY_AUTO}.maxExp());
+    FFTConfig fft{shape, 0, CARRY_32};
+    u32 exponent = primes.prevPrime(fft.maxExp());
     // log("tuning %10s with exponent %u\n", fft.shape.spec().c_str(), exponent);
 
     vector<int> bestPos(configsVect.size());
@@ -242,7 +243,7 @@ void Tune::ctune() {
         for (u32 k = i + 1; k < configsVect.size(); ++k) {
           add(c, configsVect[k][bestPos[k]]);
         }
-        auto cost = Gpu::make(q, exponent, shared, FFTConfig{shape, 0, CARRY_AUTO}, c, false)->timePRP();
+        auto cost = Gpu::make(q, exponent, shared, fft, c, false)->timePRP();
 
         bool isBest = (cost < best.cost);
         if (isBest) {