Put exp lookup table into run_time_lib/AIE2 (#604)

test/unit_tests/aievec_tests/bf16_exp_lut/exp_lut.h → aie_runtime_lib/AIE/exp_lut.cpp

@@ -1,6 +1,16 @@
-#ifndef __EXP_LUT_H__
-#define __EXP_LUT_H__
-#include "adf.h"
+//===---  exp_lut.cpp - exponential loopup tables ---===//
+//
+// This file is licensed under the Apache License v2.0 with LLVM Exceptions
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+// (c) Copyright 2023 Xilinx Inc.
+//
+//
+//===----------------------------------------------------------------------===//
+// These are exponential lookup tables for bfloat16 type
+//===----------------------------------------------------------------------===//
+
 #include "aie_api/aie.hpp"
 
 // The tables _ab and _cd are copies of each other
@@ -204,43 +214,3 @@ alignas(aie::vector_decl_align) int16 softmax_flut_cd[512] = {
     16422, 16423, 16423, 16424, 16419, 16420, 16421, 16421, 16422, 16423, 16423,
     16424, 16425, 16425, 16426, 16427, 16427, 16428, 16429, 16429, 16425, 16425,
     16426, 16427, 16427, 16428, 16429, 16429};
-
-__attribute__((always_inline)) v16accfloat getExpBf16(v16bfloat16 x) {
-  bfloat16 __aie_dm_resource_a *ilut_ab =
-      (bfloat16 __aie_dm_resource_a *)softmax_ilut_ab;
-  bfloat16 __aie_dm_resource_b *ilut_cd =
-      (bfloat16 __aie_dm_resource_b *)softmax_ilut_cd;
-  bfloat16 __aie_dm_resource_a *flut_ab =
-      (bfloat16 __aie_dm_resource_a *)softmax_flut_ab;
-  bfloat16 __aie_dm_resource_b *flut_cd =
-      (bfloat16 __aie_dm_resource_b *)softmax_flut_cd;
-
-  using lut_type = aie::lut<4, bfloat16, bfloat16>;
-  const int LUT_elems = 256;
-  const int step_i = 8;
-  const int step_f = 0;
-
-  lut_type lut_i(LUT_elems, ilut_ab, ilut_cd);
-  lut_type lut_f(LUT_elems, flut_ab, flut_cd);
-  aie::parallel_lookup<uint16, lut_type, aie::lut_oor_policy::truncate>
-      lookup_i(lut_i, step_i);
-  aie::parallel_lookup<uint16, lut_type, aie::lut_oor_policy::truncate>
-      lookup_f(lut_f, step_f);
-
-  aie::vector<bfloat16, 16> I_val_vec, F_val_vec;
-  aie::accum<accfloat, 16> exp_val;
-  aie::vector<bfloat16, 16> input_bf16 = x;
-
-  // position of output decimal point = 8, making input become 8 bits, and for
-  // LUT_elems = 256 lookup. aie::vector<int16, 16>
-  // input=aie::to_fixed<int16>(input_bf16,8);
-  aie::vector<int16, 32> input0 = v32int16(bfloat16_to_int(input_bf16, 8));
-  aie::vector<int16, 16> input = aie::filter_even(input0);
-
-  I_val_vec = lookup_i.fetch(input.cast_to<uint16>());
-  F_val_vec = lookup_f.fetch(input.cast_to<uint16>());
-  exp_val = aie::mul(I_val_vec, F_val_vec);
-  return v16accfloat(exp_val);
-}
-
-#endif //__EXP_LUT_H__

aie_runtime_lib/AIE/exp_lut.h

@@ -0,0 +1,61 @@
+//===---  exp_lut.h - get exponential values from loopup tables ---===//
+//
+// This file is licensed under the Apache License v2.0 with LLVM Exceptions
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+// (c) Copyright 2023 Xilinx Inc.
+//
+//
+//===----------------------------------------------------------------------===//
+// This is the implementation of getting exponential values for a bfloat16
+// vector from exponential lookup tables.
+//===----------------------------------------------------------------------===//
+#ifndef __EXP_LUT_H__
+#define __EXP_LUT_H__
+
+#include "aie_api/aie.hpp"
+
+alignas(aie::vector_decl_align) extern int16 softmax_ilut_ab[512];
+alignas(aie::vector_decl_align) extern int16 softmax_ilut_cd[512];
+alignas(aie::vector_decl_align) extern int16 softmax_flut_ab[512];
+alignas(aie::vector_decl_align) extern int16 softmax_flut_cd[512];
+
+__attribute__((always_inline)) v16accfloat getExpBf16(v16bfloat16 x) {
+  bfloat16 __aie_dm_resource_a *ilut_ab =
+      (bfloat16 __aie_dm_resource_a *)softmax_ilut_ab;
+  bfloat16 __aie_dm_resource_b *ilut_cd =
+      (bfloat16 __aie_dm_resource_b *)softmax_ilut_cd;
+  bfloat16 __aie_dm_resource_a *flut_ab =
+      (bfloat16 __aie_dm_resource_a *)softmax_flut_ab;
+  bfloat16 __aie_dm_resource_b *flut_cd =
+      (bfloat16 __aie_dm_resource_b *)softmax_flut_cd;
+
+  using lut_type = aie::lut<4, bfloat16, bfloat16>;
+  const int LUT_elems = 256;
+  const int step_i = 8;
+  const int step_f = 0;
+
+  lut_type lut_i(LUT_elems, ilut_ab, ilut_cd);
+  lut_type lut_f(LUT_elems, flut_ab, flut_cd);
+  aie::parallel_lookup<uint16, lut_type, aie::lut_oor_policy::truncate>
+      lookup_i(lut_i, step_i);
+  aie::parallel_lookup<uint16, lut_type, aie::lut_oor_policy::truncate>
+      lookup_f(lut_f, step_f);
+
+  aie::vector<bfloat16, 16> I_val_vec, F_val_vec;
+  aie::accum<accfloat, 16> exp_val;
+  aie::vector<bfloat16, 16> input_bf16 = x;
+
+  // position of output decimal point = 8, making input become 8 bits, and for
+  // LUT_elems = 256 lookup. aie::vector<int16, 16>
+  // input=aie::to_fixed<int16>(input_bf16,8);
+  aie::vector<int16, 32> input0 = v32int16(bfloat16_to_int(input_bf16, 8));
+  aie::vector<int16, 16> input = aie::filter_even(input0);
+
+  I_val_vec = lookup_i.fetch(input.cast_to<uint16>());
+  F_val_vec = lookup_f.fetch(input.cast_to<uint16>());
+  exp_val = aie::mul(I_val_vec, F_val_vec);
+  return v16accfloat(exp_val);
+}
+#endif //__EXP_LUT_H__