CUDA: Factor out and re-use block_reduce function (#18785)

* CUDA: Refactor and expose two_stage_warp_reduce_* function

* Use `two_stage_warp_reduce` also in softmax kernel, move smem out of it

Moving smem out of `__device__` function to `__global__` function
allows for explicit smem reuse, as either compiler or cuda rt seem to not
free it afterwards (`cudaFuncSetAttribute` fails when not accounting for
it once for each call to two_stage_warp_reduce)

* Update ggml/src/ggml-cuda/common.cuh

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

* Use two_stage_warp_reduce in group_norm_f32

* Use two_stage_warp_reduce in rms_norm_f32

* Fix smem calculation which expects bytes

* Make `two_stage_warp_reduce` accept all values warp_reduce accepts

Also integrate it into norm_f32 function

* Use two_stage_warp_reduce in l2_norm_f32

* Use type traits for block reduction for better legibility

Also adresss other requests by @am17an such as variable renaming

* Make norm tests cover all cuda paths

* Mark columns % WARP_SIZE !=0 as supported for RMS_NORM_BACK

Unit-tests passed locally, let's see if they pass in the CI as well

* Use `enum class` for `block_reduce_method`

This is more type-safe than plain enum

* Rename variables as suggested in code review by @am17an

* Rename two_stage_warp_reduce -> block_reduce

* Fix trailing whitespace in common.cuh

* Make condition of static_assert type-dependent

This delays evaluation until the template is actually instantiated.
Otherwise, some compilers may evaluate the assert when parsing the
template, resulting in build errors as observed here:

https://github.com/ggml-org/llama.cpp/actions/runs/20960323123/job/60235530068?pr=18785

* Inline definitions

---------

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

ggml/src/ggml-cuda/common.cuh

@@ -530,6 +530,86 @@ static __device__ __forceinline__ half2 warp_prefix_inclusive_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+enum class block_reduce_method {
+    MAX,
+    SUM,
+};
+
+template<block_reduce_method method_t, typename T>
+struct block_reduce_policy;
+
+template <typename T, typename... Ts>
+inline constexpr bool is_any = (std::is_same_v<T, Ts> || ...);
+
+template<typename...>
+inline constexpr bool ggml_cuda_dependent_false_v = false;
+
+template <typename T> struct block_reduce_policy<block_reduce_method::SUM, T> {
+    static __device__ T reduce(T val) {
+        if constexpr(is_any<T, float, float2, half2, int>) {
+            return warp_reduce_sum(val);
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce sum");
+        }
+    }
+
+    static __device__ T sentinel() {
+        if constexpr (std::is_same_v<T, float>) {
+            return 0.0f;
+        } else if constexpr (std::is_same_v<T, float2>) {
+            return make_float2(0.0f, 0.0f);
+        } else if constexpr (std::is_same_v<T, half2>) {
+            return make_half2(0.0f, 0.0f);
+        } else if constexpr (std::is_same_v<T, int>) {
+            return 0;
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce sum");
+        }
+    }
+};
+
+template <typename T> struct block_reduce_policy<block_reduce_method::MAX, T> {
+    static __device__ T reduce(T val) {
+        if constexpr (is_any<T, float, half2>) {
+            return warp_reduce_max(val);
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce max");
+        }
+    }
+
+    static __device__ T sentinel() {
+        if constexpr (std::is_same_v<T, float>) {
+            return -INFINITY;
+        } else if constexpr (std::is_same_v<T, half2>) {
+            return make_half2(-INFINITY, -INFINITY);
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce max");
+        }
+    }
+};
+
+template <block_reduce_method reduce_method_t, const unsigned int block_size_template = 0, typename T>
+static __device__ T block_reduce(T val, T * shared_vals) {
+    val                           = block_reduce_policy<reduce_method_t, T>::reduce(val);
+    const unsigned int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
+    if (block_size > WARP_SIZE) {
+        assert((block_size <= 1024) && (block_size % WARP_SIZE) == 0);
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            shared_vals[warp_id] = val;
+        }
+        __syncthreads();
+        val = block_reduce_policy<reduce_method_t, T>::sentinel();
+        if (lane_id < (static_cast<int>(block_size) / WARP_SIZE)) {
+            val = shared_vals[lane_id];
+        }
+        return block_reduce_policy<reduce_method_t, T>::reduce(val);
+    }
+
+    return val;
+}
+
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE