perf[gpu]: reduce register pressure in dyn dispatch (#7489)

We decrease the number of values per tile in the output stage each GPU
thread uses, as well as limit the register count to 32 in the launch
bounds. This brings the dynamic dispatch kernel into a reasonably close
range compared to the standalone kernel for now.

Type | Dynamic dispatch | Standalone | Ratio |
|---|---|---|---|
| u8 bw6 | 172 µs | 79 µs | 2.17× |
| u16 bw6 | 140 µs | 88 µs | 1.59× |
| u32 bw6 | 184 µs | 148 µs | 1.24× |
| u64 bw8 | 303 µs | 276 µs | 1.10×|

Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

Author: 0ax1

vortex-cuda/kernels/src/dynamic_dispatch.cu

@@ -279,7 +279,8 @@ __device__ void execute_output_stage(T *__restrict output,
                                      char *__restrict smem,
                                      uint64_t block_start,
                                      uint32_t block_len) {
-    constexpr uint32_t VALUES_PER_TILE = 32 / sizeof(T);
+    // Cap at 4 values per thread per tile to minimise register pressure.
+    constexpr uint32_t VALUES_PER_TILE = (32 / sizeof(T)) < 4 ? (32 / sizeof(T)) : 4;
     const uint32_t tile_size = blockDim.x * VALUES_PER_TILE;
     const auto &src = stage.source;
     const void *raw_input = reinterpret_cast<const void *>(stage.input_ptr);
@@ -472,9 +473,10 @@ dynamic_dispatch(T *__restrict output, uint64_t array_len, const uint8_t *__rest
 // matters is load_element(), which dispatches on the per-op PTypeTag to
 // sign-extend or zero-extend when widening a narrow source to T.
 #define GENERATE_KERNEL(suffix, Type)                                                                        \
-    extern "C" __global__ void dynamic_dispatch_##suffix(Type *__restrict output,                            \
-                                                         uint64_t array_len,                                 \
-                                                         const uint8_t *__restrict packed_plan) {            \
+    extern "C" __global__ void __launch_bounds__(BLOCK_SIZE, 32)                                             \
+        dynamic_dispatch_##suffix(Type *__restrict output,                                                   \
+                                  uint64_t array_len,                                                        \
+                                  const uint8_t *__restrict packed_plan) {                                   \
         dynamic_dispatch<Type>(output, array_len, packed_plan);                                              \
     }