From ff374afd0dd304d860edf9bf451d3f7e4253e1fc Mon Sep 17 00:00:00 2001
From: steven varga <steven.varga@gmail.com>
Date: Wed, 17 Jun 2026 20:09:03 +0000
Subject: [PATCH 1/3] [#16]:svarga:refactor, replace bench::types marker with
 explicit std::tuple type axis

---
 include/bench/bench.hpp | 39 ++++++++++++++++++++++++++-------------
 1 file changed, 26 insertions(+), 13 deletions(-)

diff --git a/include/bench/bench.hpp b/include/bench/bench.hpp
index e342c0e..a67f48a 100644
--- a/include/bench/bench.hpp
+++ b/include/bench/bench.hpp
@@ -119,16 +119,22 @@ namespace bench::impl {
 		}) / static_cast<T>(data.size());
 		return std::pair<T,T>{mean, std::sqrt(var)};
 	}
+
+	// Detects whether `T` is a specialization of the class template `Tmpl`.
+	// Used to constrain the type-axis throughput overload to `std::tuple<...>`.
+	template <class T, template <class...> class Tmpl>
+	struct is_specialization_of : std::false_type {};
+	template <template <class...> class Tmpl, class... Us>
+	struct is_specialization_of<Tmpl<Us...>, Tmpl> : std::true_type {};
+
+	template <class T>
+	concept is_tuple = is_specialization_of<std::remove_cvref_t<T>, std::tuple>::value;
 }
 
 namespace bench {
 	constexpr std::size_t max_dims = 32;
 	namespace arg = bench::meta::arg;
 
-	// ---- compile-time type axis -------------------------------------------
-	// Marker carrying the list of types to fold the benchmark body over.
-	template <class... Ts> struct types {};
-
 	// ---- named, order-independent arguments -------------------------------
 	using name   = impl::value_t<std::string,   impl::tag::name_t>;
 	using warmup = impl::value_t<std::uint16_t, impl::tag::warmup_t>;
@@ -250,8 +256,8 @@ namespace bench {
 
 	// ---- the throughput driver --------------------------------------------
 	// body signature: double(std::size_t idx, std::size_t n) -> bytes moved.
-	// (The compile-time type-axis overload `throughput(types<Ts...>{}, ...)`
-	//  is added by a separate lane.)
+	// The compile-time type-axis form is the explicit-tuple overload below
+	// (`throughput<std::tuple<Ts...>>(...)`).
 	template <class... args_t>
 	void throughput(args_t... args) {
 		using callback_t = std::function<double(std::size_t, std::size_t)>;
@@ -286,14 +292,22 @@ namespace bench {
 	}
 
 	// ---- the type-axis throughput driver ----------------------------------
-	// Folds the benchmark body over the compile-time typelist `types<Ts...>`.
+	// Folds the benchmark body over the compile-time typelist supplied as an
+	// explicit template argument, a `std::tuple<Ts...>`. Invoke as
+	//   bench::throughput<std::tuple<A,B,C>>(bench::name{...}, ...);
+	// The leading template parameter `Tuple` is constrained (impl::is_tuple) to
+	// be a `std::tuple` specialization so this never collides with the scalar
+	// `throughput(args...)` overload — ordinary calls deduce nothing for `Tuple`
+	// and select the scalar form, while an explicit tuple selects this one.
+	//
 	// `body` is a C++20 generic lambda invoked once per type as
 	//   body.template operator()<T>(std::size_t idx, std::size_t n) -> double
 	// returning bytes moved (same contract as the scalar driver). Honours the
 	// same named args/hooks; one result row per (type, x-point), the row name
 	// suffixed with the type label so types are distinguishable.
-	template <class... Ts, class... args_t>
-	void throughput(types<Ts...>, args_t... args) {
+	template <class Tuple, class... args_t>
+		requires impl::is_tuple<Tuple>
+	void throughput(args_t... args) {
 		using name_t     = typename arg::tpos<impl::tag::name_t,   args_t...>;
 		using x_t        = typename arg::tpos<impl::tag::x_t,      args_t...>;
 		using warmup_t   = typename arg::tpos<impl::tag::warmup_t, args_t...>;
@@ -303,7 +317,7 @@ namespace bench {
 
 		static_assert( name_t::present, "benchmark name must be specified (bench::name{...})" );
 		static_assert( x_t::present,    "x axis must be specified (bench::arg_x{...})" );
-		static_assert( sizeof...(Ts) > 0, "type axis must list at least one type" );
+		static_assert( std::tuple_size_v<Tuple> > 0, "type axis must list at least one type" );
 
 		auto tuple = std::forward_as_tuple(args...);
 		store_t& store = store_t::get();
@@ -320,9 +334,8 @@ namespace bench {
 		// the generic body is the only positional (un-tagged) argument
 		auto&& body = arg::getn<sizeof...(args_t) - 1>(args...);
 
-		using list_t = std::tuple<Ts...>;
-		meta::impl::static_for<list_t>([&](auto I){
-			using T = std::tuple_element_t<decltype(I)::value, list_t>;
+		meta::impl::static_for<Tuple>([&](auto I){
+			using T = std::tuple_element_t<decltype(I)::value, Tuple>;
 			std::string label = impl::type_name<T>();
 			if (label.empty()) label = "T" + std::to_string(decltype(I)::value);
 			const std::string row_name = nm.value + "/" + label;

From 476eff6094367aeb611fbd22e11e05c80db4ade7 Mon Sep 17 00:00:00 2001
From: steven varga <steven.varga@gmail.com>
Date: Wed, 17 Jun 2026 20:09:03 +0000
Subject: [PATCH 2/3] [#16]:svarga:refactor, migrate type-axis call sites to
 throughput<std::tuple<...>>

---
 README.md                  | 15 ++++++++-------
 examples/heatmap.cpp       |  4 ++--
 examples/type_dispatch.cpp |  6 +++---
 test/type_dispatch.cpp     |  7 +++----
 4 files changed, 16 insertions(+), 16 deletions(-)

diff --git a/README.md b/README.md
index a8a89d6..92343e9 100644
--- a/README.md
+++ b/README.md
@@ -10,12 +10,12 @@ fully specialized, no type erasure — crossed with runtime parameter sweeps.
 
 ```cpp
 #include <bench/all>
+#include <tuple>
 
 int main() {
   bench::arg_x record_size{10'000, 100'000, 1'000'000};
 
-  bench::throughput(
-    bench::types<int, double, float>{},      // compile-time type axis
+  bench::throughput<std::tuple<int, double, float>>(  // compile-time type axis
     bench::name{"write"}, record_size,        // named args, any order
     bench::warmup{3}, bench::sample{10},
     [&]<class T>(std::size_t idx, std::size_t n) -> double {
@@ -35,10 +35,11 @@ runtime and throughput.
   the bytes it moved; the engine reports MiB/s. Order-independent named args:
   `bench::name`, `bench::arg_x` (size sweep), `bench::warmup`, `bench::sample`,
   and the untimed/timed hooks `bench::before_sample` / `bench::after_sample`.
-- **Type-axis dispatch** — the headline feature. `bench::throughput(`
-  `bench::types<Ts...>{}, ...)` folds the *same* generic-lambda body over a
-  compile-time type list, each fully specialized, yielding one row per
-  `(type, size)` (rows named `"<name>/<typelabel>"`).
+- **Type-axis dispatch** — the headline feature. `bench::throughput<`
+  `std::tuple<Ts...>>(...)` folds the *same* generic-lambda body over a
+  compile-time type list passed as an explicit template argument, each fully
+  specialized, yielding one row per `(type, size)` (rows named
+  `"<name>/<typelabel>"`).
 - **Sinks** — results flow through a pluggable `bench::sink`. The default
   `bench::stdout_sink` prints a table; opt-in `bench::csv_sink` and
   `bench::json_sink` (in `<bench/sink/csv.hpp>` / `<bench/sink/json.hpp>`, not
@@ -57,7 +58,7 @@ default (toggle with `-DBENCH_BUILD_EXAMPLES=OFF`):
 | Example | Shows |
 | --- | --- |
 | `memcpy_bandwidth.cpp` | scalar `throughput` over an `arg_x` size sweep; default stdout table |
-| `type_dispatch.cpp` | `bench::types<...>` running one body across several element types, fed by `get_test_data<T>` |
+| `type_dispatch.cpp` | `bench::throughput<std::tuple<...>>` running one body across several element types, fed by `get_test_data<T>` |
 | `sinks.cpp` | emitting results to CSV and JSON via `csv_sink` / `json_sink` and `set_sink(...)` |
 
 ```sh
diff --git a/examples/heatmap.cpp b/examples/heatmap.cpp
index 116cfa8..a96446b 100644
--- a/examples/heatmap.cpp
+++ b/examples/heatmap.cpp
@@ -13,6 +13,7 @@
 #include <vector>
 #include <cstddef>
 #include <memory>
+#include <tuple>
 
 int main(){
 	// colour each (type, size) cell by mean throughput; write report.svg.
@@ -22,8 +23,7 @@ int main(){
 
 	std::vector<unsigned char> buf(1'000'000 * sizeof(double), 1);
 
-	bench::throughput(
-		bench::types<int, float, double>{},
+	bench::throughput<std::tuple<int, float, double>>(
 		bench::name{"sum"},
 		sizes,
 		bench::warmup{2}, bench::sample{10},
diff --git a/examples/type_dispatch.cpp b/examples/type_dispatch.cpp
index 22bc103..640e30a 100644
--- a/examples/type_dispatch.cpp
+++ b/examples/type_dispatch.cpp
@@ -17,6 +17,7 @@
 
 #include <cstddef>
 #include <cstdint>
+#include <tuple>
 #include <vector>
 
 int main() {
@@ -24,9 +25,8 @@ int main() {
 	// so each type reports the bytes it actually moves.
 	bench::arg_x counts{ 1'000, 100'000, 1'000'000 };
 
-	bench::throughput(
-		// the compile-time type axis: one row-group per type
-		bench::types<std::uint8_t, std::uint32_t, double>{},
+	// the compile-time type axis: one row-group per type
+	bench::throughput<std::tuple<std::uint8_t, std::uint32_t, double>>(
 		bench::name{"sum"},
 		counts,
 		bench::warmup{3}, bench::sample{20},
diff --git a/test/type_dispatch.cpp b/test/type_dispatch.cpp
index 68d882f..9446d91 100644
--- a/test/type_dispatch.cpp
+++ b/test/type_dispatch.cpp
@@ -1,7 +1,7 @@
 /* Copyright (c) 2026 Steven Varga, Toronto, ON, Canada
  * MIT License — see LICENSE
  *
- * Type-axis test: the compile-time `throughput(types<Ts...>{}, ...)` overload
+ * Type-axis test: the compile-time `throughput<std::tuple<Ts...>>(...)` overload
  * folds a generic benchmark body over a typelist and emits one result row per
  * (type, x-point), with the type label distinguishing the rows.
  */
@@ -21,9 +21,8 @@ int main() {
 
 	std::vector<unsigned char> buf(100'000 * sizeof(double), 1);
 
-	// args intentionally out of order; types<...> is the leading positional arg.
-	bench::throughput(
-		bench::types<int, double, float>{},
+	// args intentionally out of order; the type axis is an explicit tuple arg.
+	bench::throughput<std::tuple<int, double, float>>(
 		bench::sample{su}, sizes, bench::warmup{wu},
 		bench::before_sample{ [&]{ ++before_calls; } },
 		bench::after_sample{  [&]{ ++after_calls;  } },

From ab9c7596f4c423f2a0ad19c58d697b1ef9be7ce0 Mon Sep 17 00:00:00 2001
From: steven varga <steven.varga@gmail.com>
Date: Wed, 17 Jun 2026 20:09:03 +0000
Subject: [PATCH 3/3] [#16]:svarga:feature, rewrite memcpy_bandwidth example to
 typed tuple form

---
 examples/memcpy_bandwidth.cpp | 54 ++++++++++++++---------------------
 1 file changed, 22 insertions(+), 32 deletions(-)

diff --git a/examples/memcpy_bandwidth.cpp b/examples/memcpy_bandwidth.cpp
index 143ce90..78d079b 100644
--- a/examples/memcpy_bandwidth.cpp
+++ b/examples/memcpy_bandwidth.cpp
@@ -1,45 +1,35 @@
 /* Copyright (c) 2026 Steven Varga, Toronto, ON, Canada
  * MIT License — see LICENSE
  *
- * memcpy_bandwidth — the simplest possible bench: a memory-bandwidth sweep.
+ * memcpy_bandwidth — a type-aware memory-bandwidth sweep.
  *
- * Uses the *scalar* throughput driver. We sweep a set of buffer sizes via
- * `bench::arg_x{...}` and, for each size, copy a source buffer into a
- * destination buffer with std::memcpy. The body returns the number of bytes
- * touched (read + written), which the engine divides by the measured time to
- * report MiB/s. Output goes to the default stdout table sink.
+ * Uses the *type-axis* throughput driver: the same body is folded over a
+ * compile-time tuple of element widths, with a typed source/destination buffer
+ * per type. We sweep a set of buffer sizes via `bench::arg_x{...}` and, for each
+ * (type, size) point, copy a source buffer into a destination buffer with
+ * std::memcpy. The body returns the number of bytes touched (read + written),
+ * which the engine divides by the measured time to report MiB/s. Output goes to
+ * the default stdout table sink.
  *
  * Dependency-free: only the bench core + standard library.
  */
 #include <bench/all>
-
 #include <cstddef>
+#include <cstdint>
 #include <cstring>
+#include <tuple>
 #include <vector>
-
 int main() {
-	// x-axis: buffer sizes in bytes, small -> large (crosses cache levels).
-	bench::arg_x sizes{ 4 * 1024, 64 * 1024, 1 * 1024 * 1024, 16 * 1024 * 1024 };
-
-	// One source/destination pair, sized to the largest sweep point so the
-	// body never reallocates inside the timed region.
-	std::vector<char> src(16 * 1024 * 1024, 'x');
-	std::vector<char> dst(src.size());
-
-	bench::throughput(
-		bench::name{"memcpy"},
-		sizes,
-		bench::warmup{5}, bench::sample{50},
-		// idx = sweep index, n = the size in bytes for this point.
-		[&](std::size_t /*idx*/, std::size_t n) -> double {
-			std::memcpy(dst.data(), src.data(), n);
-			// Defeat dead-store elimination by reading one byte back.
-			volatile char sink = dst[n - 1];
-			(void)sink;
-			// Bytes moved across the bus: n read + n written.
-			return static_cast<double>(2 * n);
-		});
-
-	// Results print as a table when the global store flushes at program exit.
-	return 0;
+    using types = std::tuple<uint8_t, uint16_t, uint32_t, uint64_t>;
+    constexpr std::size_t max_bytes = 16 * 1024 * 1024;
+    bench::arg_x sizes{ 4*1024, 64*1024, 1*1024*1024, max_bytes };
+    bench::throughput<types>(
+        bench::name{"memcpy"}, sizes, bench::warmup{5}, bench::sample{50},
+        [&]<class T>(std::size_t /*idx*/, std::size_t n) -> double {
+            static std::vector<T> src(max_bytes/sizeof(T), T{1}), dst(src.size());
+            std::memcpy(dst.data(), src.data(), n);
+            volatile T sink = dst[n/sizeof(T) - 1]; (void)sink;
+            return static_cast<double>(2 * n);
+        });
+    return 0;
 }