Merge branch 'gpu-tail' into rc175

tests/MergeScaleTest.cpp

@@ -526,3 +526,32 @@ TEST_CASE("ErrorModel_BNotMeasuredOnWeakData") {
     CHECK(fit.b2 == 0.0);
     CHECK(fit.a == Catch::Approx(0.9).epsilon(0.03));
 }
+
+#ifdef JFJOCH_USE_CUDA
+#include "../common/CUDAWrapper.h"
+#include "../image_analysis/scale_merge/ErrorModelGPU.h"
+
+// The device fit against the host one. Same bins and same medians; only the order each bin's two sums
+// are added in differs, so a and b^2 agree to rounding and every flag agrees exactly. The same pool
+// twice on the device gives the same bits.
+TEST_CASE("ErrorModel_DeviceFitIsTheHostFit") {
+    if (get_gpu_count() == 0) {
+        WARN("No CUDA GPU present. Skipping ErrorModel_DeviceFitIsTheHostFit");
+        return;
+    }
+    for (const auto &pool : {SyntheticErrorModelSamples(1.3, 0.03, 300.0, 0.003),
+                             SyntheticErrorModelSamples(0.9, 0.05, 1.5, 0.0)}) {
+        std::vector<ErrorModelBinned> scratch;
+        const auto host = FitErrorModel(pool, scratch, 4);
+        const auto dev = FitErrorModelGPU(pool);
+        const auto again = FitErrorModelGPU(pool);
+        CHECK(dev.active == host.active);
+        CHECK(dev.b_measured == host.b_measured);
+        CHECK(dev.b_resolved == host.b_resolved);
+        CHECK(dev.a == Catch::Approx(host.a).epsilon(1e-12));
+        CHECK(dev.b2 == Catch::Approx(host.b2).epsilon(1e-12).margin(1e-300));
+        CHECK(again.a == dev.a);
+        CHECK(again.b2 == dev.b2);
+    }
+}
+#endif