feat(ops): implement CausalSoftmax operator with Moore backend

    - add the Moore operator specialization
    - limit block size in the shared CUDA-style kernel for Moore

Author: gongchensu

src/cuda/causal_softmax/kernel.h

@@ -1,7 +1,9 @@
 #ifndef INFINI_OPS_CUDA_CAUSAL_SOFTMAX_KERNEL_H_
 #define INFINI_OPS_CUDA_CAUSAL_SOFTMAX_KERNEL_H_
 
+#include <algorithm>
 #include <cstdint>
+#include <type_traits>
 
 #include "base/causal_softmax.h"
 #include "cuda/causal_softmax/kernel.cuh"
@@ -11,6 +13,49 @@
 
 namespace infini::ops {
 
+namespace causal_softmax::detail {
+
+template <typename Backend, typename = void>
+struct MaxBlockSize : std::integral_constant<int, 2048> {};
+
+template <typename Backend>
+struct MaxBlockSize<Backend, std::void_t<decltype(Backend::max_block_size)>>
+    : std::integral_constant<int, Backend::max_block_size> {};
+
+template <int max_block_size>
+struct SupportedBlockSizes;
+
+template <>
+struct SupportedBlockSizes<2048> {
+  using type = AllCudaBlockSizes;
+};
+
+template <>
+struct SupportedBlockSizes<1024> {
+  using type = List<128, 256, 512, 1024>;
+};
+
+template <>
+struct SupportedBlockSizes<512> {
+  using type = List<128, 256, 512>;
+};
+
+template <>
+struct SupportedBlockSizes<256> {
+  using type = List<128, 256>;
+};
+
+template <>
+struct SupportedBlockSizes<128> {
+  using type = List<128>;
+};
+
+template <typename Backend>
+using SupportedBlockSizesFor =
+    typename SupportedBlockSizes<MaxBlockSize<Backend>::value>::type;
+
+}  // namespace causal_softmax::detail
+
 template <typename Backend>
 class CudaCausalSoftmax : public CausalSoftmax {
  public:
@@ -32,10 +77,13 @@ class CudaCausalSoftmax : public CausalSoftmax {
       std::abort();
     }
 
-    int block_size = GetOptimalBlockSize();
+    constexpr int kMaxBlockSize =
+        causal_softmax::detail::MaxBlockSize<Backend>::value;
+    int block_size = std::min(GetOptimalBlockSize(),
+                              kMaxBlockSize);
 
     DispatchFunc<ConcatType<List<DataType::kFloat32>, ReducedFloatTypes>,
-                 AllCudaBlockSizes>(
+                 causal_softmax::detail::SupportedBlockSizesFor<Backend>>(
         // TODO: Output dtype should use the one passed in during construction.
         {static_cast<int64_t>(out.dtype()), block_size},
         [&](auto list_tag) {

src/moore/causal_softmax/kernel.h

@@ -0,0 +1,35 @@
+#ifndef INFINI_OPS_MOORE_CAUSAL_SOFTMAX_KERNEL_H_
+#define INFINI_OPS_MOORE_CAUSAL_SOFTMAX_KERNEL_H_
+
+// clang-format off
+#include <musa_runtime.h>
+// clang-format on
+
+// clang-format off
+#include "moore/device_.h"
+// clang-format on
+
+#include "cuda/causal_softmax/kernel.h"
+
+namespace infini::ops {
+
+namespace causal_softmax {
+
+struct MooreBackend {
+  using stream_t = musaStream_t;
+
+  static constexpr int max_block_size = 256;
+};
+
+}  // namespace causal_softmax
+
+template <>
+class Operator<CausalSoftmax, Device::Type::kMoore>
+    : public CudaCausalSoftmax<causal_softmax::MooreBackend> {
+ public:
+  using CudaCausalSoftmax<causal_softmax::MooreBackend>::CudaCausalSoftmax;
+};
+
+}  // namespace infini::ops
+
+#endif