[Enhancement] Optimize BF16 casting performance (#689)

* use more efficient bf16 type related conversion

* update macro

src/target/codegen_cuda.cc

@@ -192,6 +192,9 @@ std::string CodeGenTileLangCUDA::Finish() {
   decl_stream << "#include <tl_templates/cuda/ldsm.h>\n";
   decl_stream << "#include <tl_templates/cuda/threadblock_swizzle.h>\n";
   decl_stream << "#include <tl_templates/cuda/debug.h>\n";
+  decl_stream << "#ifdef ENABLE_BF16\n";
+  decl_stream << "#include <tl_templates/cuda/cuda_bf16_fallbacks.cuh>\n";
+  decl_stream << "#endif\n";
 
   if (need_global_barrier_) {
     decl_stream << "__device__ unsigned " << vid_global_barrier_state_
@@ -734,8 +737,54 @@ void CodeGenTileLangCUDA::VisitExpr_(const CastNode *op, std::ostream &os) {
   this->PrintIndent();
   this->PrintType(target_ty, stream);
   stream << ' ' << sret << ";\n";
-  {
   std::string src = SSAGetID(PrintExpr(op->value), from_ty);
+
+  // Handle bfloat16 special cases with supported ops
+  bool used_bf16_op = false;
+  if (from_ty.is_bfloat16() || target_ty.is_bfloat16()) {
+    std::ostringstream func_name;
+    if (from_ty.is_bfloat16())
+      func_name << "bf16";
+    else if (from_ty.is_float())
+      func_name << "float";
+    if (from_ty.lanes() > 1)
+      func_name << from_ty.lanes();
+    func_name << "2";
+    if (target_ty.is_bfloat16())
+      func_name << "bf16";
+    else if (target_ty.is_float())
+      func_name << "float";
+    else if (target_ty == DataType::Int(16))
+      func_name << "int16";
+    if (target_ty.lanes() > 1)
+      func_name << target_ty.lanes();
+
+    auto fname = func_name.str();
+    if (bf16_supported_ops_.count(fname)) {
+      used_bf16_op = true;
+      stream << "#ifdef ENABLE_BF16\n";
+      PrintIndent();
+      stream << "reinterpret_cast<";
+      if (target_ty.is_bfloat16())
+        stream << "__nv_bfloat16";
+      else
+        PrintType(target_ty.element_of(), stream);
+      if (target_ty.lanes() > 1)
+        stream << target_ty.lanes();
+      stream << " &>(" << sret << ") = fastertransformer::" << fname
+             << "(reinterpret_cast<";
+      if (from_ty.is_bfloat16())
+        stream << "__nv_bfloat16";
+      else
+        PrintType(from_ty.element_of(), stream);
+      if (from_ty.lanes() > 1)
+        stream << from_ty.lanes();
+      stream << " const &>(" << src << "));\n";
+      stream << "#else\n";
+    }
+  }
+
+  // Fallback: elementwise cast
   for (int i = 0, lanes = from_ty.lanes(); i < lanes; ++i) {
     std::ostringstream val;
     val << "(";
@@ -745,6 +794,9 @@ void CodeGenTileLangCUDA::VisitExpr_(const CastNode *op, std::ostream &os) {
     val << ")";
     PrintVecElemStore(sret, target_ty, i, val.str());
   }
+
+  if (used_bf16_op) {
+    stream << "#endif\n";
   }
   os << sret;
 }

src/target/codegen_cuda.h

@@ -125,6 +125,9 @@ private:
                       const VarNode *variable, std::ostream &os);
   int32_t GetWmmaFragmentSize(const std::string &scope, const VarNode *variable,
                               int32_t size);
+
+  std::unordered_set<std::string> bf16_supported_ops_ = {
+      "bf1622float2", "bf1622int16", "float22bf162", "bf162bf162"};
 };
 
 } // namespace codegen

src/tl_templates/cuda/cuda_bf16_fallbacks.cuh

+// Downloaded from from FasterTransformer v5.2.1
+// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.2.1_tag/src/fastertransformer/utils/cuda_bf16_fallbacks.cuh
+/*
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include "cuda_bf16_wrapper.h"
+#include <cuda_fp16.h>
+
+namespace fastertransformer {
+
+#ifdef ENABLE_BF16
+inline __device__ float2 bf1622float2(const __nv_bfloat162 val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float2 f_val;
+    f_val.x = __low2float(val);
+    f_val.y = __high2float(val);
+    return f_val;
+#else
+    return __bfloat1622float2(val);
+#endif
+}
+
+inline __device__ int16_t bf1622int16(__nv_bfloat162 val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float2 f_val;
+    f_val.x = max(min(__low2float(val), 127.f), -128.f);
+    f_val.y = max(min(__high2float(val), 127.f), -128.f);
+    union { int8_t int8[2]; int16_t int16; };
+    int8[0] = static_cast<int8_t>(static_cast<short>(f_val.x));
+    int8[1] = static_cast<int8_t>(static_cast<short>(f_val.y));
+    return int16;
+#else
+    val = __hmin2(val, make_bfloat162(127., 127.));
+    val = __hmax2(val, make_bfloat162(-128., -128.));
+    union { int8_t int8[2]; int16_t int16; };
+    int8[0] = static_cast<int8_t>(static_cast<short>(val.x));
+    int8[1] = static_cast<int8_t>(static_cast<short>(val.y));
+    return int16;
+#endif
+}
+
+inline __device__ __nv_bfloat162 float22bf162(const float2 val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __floats2bfloat162_rn(val.x, val.y);
+#else
+    return __float22bfloat162_rn(val);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf162bf162(const __nv_bfloat16 val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    __nv_bfloat162 val2;
+    val2.x = val;
+    val2.y = val;
+    return val2;
+#else
+    return __bfloat162bfloat162(val);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hadd2(const __nv_bfloat162 x, const __nv_bfloat162 y) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fxl, fxh, fyl, fyh;
+    fxl = __low2float(x);
+    fxh = __high2float(x);
+    fyl = __low2float(y);
+    fyh = __high2float(y);
+    return __floats2bfloat162_rn(fxl + fyl, fxh + fyh);
+#else
+    return __hadd2(x, y);
+#endif
+}
+
+inline __device__ __nv_bfloat16 bf16hadd(const __nv_bfloat16 x, const __nv_bfloat16 y) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16( __bfloat162float(x) + __bfloat162float(y) );
+#else
+    return __hadd(x, y);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hsub2(const __nv_bfloat162 x, const __nv_bfloat162 y) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fxl, fxh, fyl, fyh;
+    fxl = __low2float(x);
+    fxh = __high2float(x);
+    fyl = __low2float(y);
+    fyh = __high2float(y);
+    return __floats2bfloat162_rn(fxl - fyl, fxh - fyh);
+#else
+    return __hsub2(x, y);
+#endif
+}
+
+inline __device__ __nv_bfloat16 bf16hsub(const __nv_bfloat16 x, const __nv_bfloat16 y) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16( __bfloat162float(x) - __bfloat162float(y) );
+#else
+    return __hsub(x, y);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hmul2(const __nv_bfloat162 x, const __nv_bfloat162 y) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fxl, fxh, fyl, fyh;
+    fxl = __low2float(x);
+    fxh = __high2float(x);
+    fyl = __low2float(y);
+    fyh = __high2float(y);
+    return __floats2bfloat162_rn(fxl * fyl, fxh * fyh);
+#else
+    return __hmul2(x, y);
+#endif
+}
+
+inline __device__ __nv_bfloat16 bf16hmul(const __nv_bfloat16 x, const __nv_bfloat16 y) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16( __bfloat162float(x) * __bfloat162float(y) );
+#else
+    return __hmul(x, y);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hfma2(const __nv_bfloat162 x, const __nv_bfloat162 y, const __nv_bfloat162 z) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fxl, fxh, fyl, fyh, fzl, fzh;
+    fxl = __low2float(x);
+    fxh = __high2float(x);
+    fyl = __low2float(y);
+    fyh = __high2float(y);
+    fzl = __low2float(z);
+    fzh = __high2float(z);
+    return __floats2bfloat162_rn(fxl * fyl + fzl, fxh * fyh + fzh);
+#else
+    return __hfma2(x, y, z);
+#endif
+}
+
+inline __device__ __nv_bfloat16 bf16hfma(const __nv_bfloat16 x, const __nv_bfloat16 y, const __nv_bfloat16 z) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16( __bfloat162float(x) * __bfloat162float(y) + __bfloat162float(z));
+#else
+    return __hfma(x, y, z);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16exp2(const __nv_bfloat162 x) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fxl, fxh;
+    fxl = __low2float(x);
+    fxh = __high2float(x);;
+    return __floats2bfloat162_rn(expf(fxl), expf(fxh));
+#else
+    return h2exp(x);
+#endif
+}
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 800)
+inline __device__ __nv_bfloat162 operator*(const __nv_bfloat162 x, const __nv_bfloat162 y) { return bf16hmul2(x, y); };
+inline __device__ __nv_bfloat162 operator+(const __nv_bfloat162 x, const __nv_bfloat162 y) { return bf16hadd2(x, y); };
+
+inline __device__ __nv_bfloat162 make_bfloat162(const __nv_bfloat16 x, const __nv_bfloat16 y)
+{
+    __nv_bfloat162 t; t.x = x; t.y = y; return t;
+}
+
+#endif
+
+inline __device__ __nv_bfloat16 bf16hadd(__nv_bfloat16 a, __nv_bfloat16 b, __nv_bfloat16 c) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16(__bfloat162float(a) + __bfloat162float(b) + __bfloat162float(c));
+#else
+    return a + b + c;
+#endif
+}
+
+inline __device__ __nv_bfloat16 bf16hadd(__nv_bfloat16 a, __nv_bfloat16 b, __nv_bfloat16 c, __nv_bfloat16 d) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16(__bfloat162float(a) + __bfloat162float(b) + __bfloat162float(c) + __bfloat162float(d));
+#else
+    return (__nv_bfloat16)((float)a + (float)b + (float)c + (float)d);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hadd2(__nv_bfloat162 a, __nv_bfloat162 b, __nv_bfloat162 c) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fal, fah, fbl, fbh, fcl, fch;
+    fal = __low2float(a);
+    fah = __high2float(a);
+    fbl = __low2float(b);
+    fbh = __high2float(b);
+    fcl = __low2float(c);
+    fch = __high2float(c);
+    return __floats2bfloat162_rn(fal + fbl + fcl, fah + fbh + fch);
+#else
+    return a + b + c;
+#endif
+}
+
+inline __device__ __nv_bfloat16 bf16hmul(__nv_bfloat16 a, __nv_bfloat16 b, __nv_bfloat16 c) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    return __float2bfloat16(__bfloat162float(a) * __bfloat162float(b) * __bfloat162float(c));
+#else
+    return a * b * c;
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hmul2(__nv_bfloat162 a, __nv_bfloat162 b, __nv_bfloat162 c) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fal, fah, fbl, fbh, fcl, fch;
+    fal = __low2float(a);
+    fah = __high2float(a);
+    fbl = __low2float(b);
+    fbh = __high2float(b);
+    fcl = __low2float(c);
+    fch = __high2float(c);
+    return __floats2bfloat162_rn(fal * fbl * fcl, fah * fbh * fch);
+#else
+    return a * b * c;
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf16hfma2(__nv_bfloat162 a, __nv_bfloat162 b, __nv_bfloat162 c, __nv_bfloat162 d) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+    float fal, fah, fbl, fbh, fcl, fch, fdl, fdh;
+    fal = __low2float(a);
+    fah = __high2float(a);
+    fbl = __low2float(b);
+    fbh = __high2float(b);
+    fcl = __low2float(c);
+    fch = __high2float(c);
+    fdl = __low2float(d);
+    fdh = __high2float(d);
+    return __floats2bfloat162_rn(fal * fbl * fcl + fdl, fah * fbh * fch + fdh);
+#else
+    return a * b * c + d;
+#endif
+}
+
+#endif // ENABLE_BF16
+
+}  // namespace fastertransformer

src/tl_templates/cuda/cuda_bf16_wrapper.h

+// Downloaded from from FasterTransformer v5.2.1
+// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.2.1_tag/src/fastertransformer/utils/cuda_bf16_wrapper.h
+/*
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#ifdef ENABLE_BF16
+#include <cuda_bf16.h>
+#endif