[RUNTIME] Add fp16 functions

src/runtime/builtin_fp16.cc

@@ -0,0 +1,21 @@
+/*!
+ * Copyright (c) 2018 by Contributors
+ * \file builtin_fp16.cc
+ * \brief Functions for conversion between fp32 and fp16
+*/
+
+#include <builtin_fp16.h>
+
+namespace tvm {
+namespace runtime {
+
+extern "C"  uint16_t __gnu_f2h_ieee(float a) {
+  return __truncXfYf2__<float, uint32_t, 23, uint16_t, uint16_t, 10>(a);
+}
+
+extern "C" float __gnu_h2f_ieee(uint16_t a) {
+  return __extendXfYf2__<uint16_t, uint16_t, 10, float, uint32_t, 23>(a);
+}
+
+}  // namespace runtime
+}  // namespace tvm

tests/python/unittest/test_runtime_ndarray.py

@@ -35,5 +35,26 @@ def test_nd_create():
         ctx.sync()
 
 
+def test_fp16_conversion():
+    n = 100
+
+    for (src, dst) in [('float32', 'float16'), ('float16', 'float32')]:
+        A = tvm.placeholder((n,), dtype=src)
+        B = tvm.compute((n,), lambda i: A[i].astype(dst))
+
+        s = tvm.create_schedule([B.op])
+        func = tvm.build(s, [A, B], 'llvm')
+
+        x_tvm = tvm.nd.array(100 * np.random.randn(n).astype(src) - 50)
+        y_tvm = tvm.nd.array(100 * np.random.randn(n).astype(dst) - 50)
+
+        func(x_tvm, y_tvm)
+
+        expected = x_tvm.asnumpy().astype(dst)
+        real = y_tvm.asnumpy()
+
+        np.testing.assert_allclose(expected, real)
+
 if __name__ == "__main__":
     test_nd_create()
+    test_fp16_conversion()

third_party/compiler-rt/builtin_fp16.h

@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2009-2015 by llvm/compiler-rt contributors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+
+ * Copyright (c) 2018 by Contributors
+ * \file builtin_fp16.cc
+ * \brief Functions for conversion between fp32 and fp16, adopted from compiler-rt.
+ */
+
+#include <cstdint>
+
+static inline uint32_t __clz(uint32_t x) {
+  // count leading zeros
+  int n = 32;
+  uint32_t y;
+
+  y = x >>16; if (y) { n = n -16; x = y; }
+  y = x >> 8; if (y) { n = n - 8; x = y; }
+  y = x >> 4; if (y) { n = n - 4; x = y; }
+  y = x >> 2; if (y) { n = n - 2; x = y; }
+  y = x >> 1; if (y) return n - 2;
+  return n - x;
+}
+
+template <typename SRC_T, typename SRC_REP_T, int SRC_SIG_BITS,
+          typename DST_T, typename DST_REP_T, int DST_SIG_BITS>
+static inline DST_T __truncXfYf2__(SRC_T a) {
+  // Various constants whose values follow from the type parameters.
+  // Any reasonable optimizer will fold and propagate all of these.
+  const int srcBits = sizeof(SRC_T) * 8;
+  const int srcExpBits = srcBits - SRC_SIG_BITS - 1;
+  const int srcInfExp = (1 << srcExpBits) - 1;
+  const int srcExpBias = srcInfExp >> 1;
+
+  const SRC_REP_T srcMinNormal = SRC_REP_T(1) << SRC_SIG_BITS;
+  const SRC_REP_T srcSignificandMask = srcMinNormal - 1;
+  const SRC_REP_T srcInfinity = (SRC_REP_T)srcInfExp << SRC_SIG_BITS;
+  const SRC_REP_T srcSignMask = SRC_REP_T(1) << (SRC_SIG_BITS + srcExpBits);
+  const SRC_REP_T srcAbsMask = srcSignMask - 1;
+  const SRC_REP_T roundMask = (SRC_REP_T(1) << (SRC_SIG_BITS - DST_SIG_BITS)) - 1;
+  const SRC_REP_T halfway = SRC_REP_T(1) << (SRC_SIG_BITS - DST_SIG_BITS - 1);
+  const SRC_REP_T srcQNaN = SRC_REP_T(1) << (SRC_SIG_BITS - 1);
+  const SRC_REP_T srcNaNCode = srcQNaN - 1;
+
+  const int dstBits = sizeof(DST_T) * 8;
+  const int dstExpBits = dstBits - DST_SIG_BITS - 1;
+  const int dstInfExp = (1 << dstExpBits) - 1;
+  const int dstExpBias = dstInfExp >> 1;
+
+  const int underflowExponent = srcExpBias + 1 - dstExpBias;
+  const int overflowExponent = srcExpBias + dstInfExp - dstExpBias;
+  const SRC_REP_T underflow = (SRC_REP_T)underflowExponent << SRC_SIG_BITS;
+  const SRC_REP_T overflow = (SRC_REP_T)overflowExponent << SRC_SIG_BITS;
+
+  const DST_REP_T dstQNaN = DST_REP_T(1) << (DST_SIG_BITS - 1);
+  const DST_REP_T dstNaNCode = dstQNaN - 1;
+
+  // Break a into a sign and representation of the absolute value
+  const union { SRC_T f; SRC_REP_T i; } src_rep = {.f = a};
+  const SRC_REP_T aRep = src_rep.i;
+  const SRC_REP_T aAbs = aRep & srcAbsMask;
+  const SRC_REP_T sign = aRep & srcSignMask;
+  DST_REP_T absResult;
+
+  if (aAbs - underflow < aAbs - overflow) {
+    // The exponent of a is within the range of normal numbers in the
+    // destination format.  We can convert by simply right-shifting with
+    // rounding and adjusting the exponent.
+    absResult = aAbs >> (SRC_SIG_BITS - DST_SIG_BITS);
+    absResult -= (DST_REP_T)(srcExpBias - dstExpBias) << DST_SIG_BITS;
+
+    const SRC_REP_T roundBits = aAbs & roundMask;
+    // Round to nearest
+    if (roundBits > halfway)
+      absResult++;
+      // Ties to even
+    else if (roundBits == halfway)
+      absResult += absResult & 1;
+  }
+  else if (aAbs > srcInfinity) {
+    // a is NaN.
+    // Conjure the result by beginning with infinity, setting the qNaN
+    // bit and inserting the (truncated) trailing NaN field.
+    absResult = (DST_REP_T)dstInfExp << DST_SIG_BITS;
+    absResult |= dstQNaN;
+    absResult |= ((aAbs & srcNaNCode) >> (SRC_SIG_BITS - DST_SIG_BITS)) & dstNaNCode;
+  }
+  else if (aAbs >= overflow) {
+    // a overflows to infinity.
+    absResult = (DST_REP_T)dstInfExp << DST_SIG_BITS;
+  }
+  else {
+    // a underflows on conversion to the destination type or is an exact
+    // zero.  The result may be a denormal or zero.  Extract the exponent
+    // to get the shift amount for the denormalization.
+    const int aExp = aAbs >> SRC_SIG_BITS;
+    const int shift = srcExpBias - dstExpBias - aExp + 1;
+
+    const SRC_REP_T significand = (aRep & srcSignificandMask) | srcMinNormal;
+
+    // Right shift by the denormalization amount with sticky.
+    if (shift > SRC_SIG_BITS) {
+      absResult = 0;
+    } else {
+      const bool sticky = significand << (srcBits - shift);
+      SRC_REP_T denormalizedSignificand = significand >> shift | sticky;
+      absResult = denormalizedSignificand >> (SRC_SIG_BITS - DST_SIG_BITS);
+      const SRC_REP_T roundBits = denormalizedSignificand & roundMask;
+      // Round to nearest
+      if (roundBits > halfway)
+        absResult++;
+        // Ties to even
+      else if (roundBits == halfway)
+        absResult += absResult & 1;
+    }
+  }
+
+  // Apply the signbit to (DST_T)abs(a).
+  const DST_REP_T result = absResult | sign >> (srcBits - dstBits);
+  const union { DST_T f; DST_REP_T i; } dst_rep = {.i = result};
+  return dst_rep.f;
+}
+
+template<typename SRC_T, typename SRC_REP_T, int SRC_SIG_BITS,
+         typename DST_T, typename DST_REP_T, int DST_SIG_BITS>
+static inline DST_T __extendXfYf2__(SRC_T a) {
+  // Various constants whose values follow from the type parameters.
+  // Any reasonable optimizer will fold and propagate all of these.
+  const int srcBits = sizeof(SRC_T) * 8;
+  const int srcExpBits = srcBits - SRC_SIG_BITS - 1;
+  const int srcInfExp = (1 << srcExpBits) - 1;
+  const int srcExpBias = srcInfExp >> 1;
+
+  const SRC_REP_T srcMinNormal = SRC_REP_T(1) << SRC_SIG_BITS;
+  const SRC_REP_T srcInfinity = (SRC_REP_T)srcInfExp << SRC_SIG_BITS;
+  const SRC_REP_T srcSignMask = SRC_REP_T(1) << (SRC_SIG_BITS + srcExpBits);
+  const SRC_REP_T srcAbsMask = srcSignMask - 1;
+  const SRC_REP_T srcQNaN = SRC_REP_T(1) << (SRC_SIG_BITS - 1);
+  const SRC_REP_T srcNaNCode = srcQNaN - 1;
+
+  const int dstBits = sizeof(DST_T)*8;
+  const int dstExpBits = dstBits - DST_SIG_BITS - 1;
+  const int dstInfExp = (1 << dstExpBits) - 1;
+  const int dstExpBias = dstInfExp >> 1;
+
+  const DST_REP_T dstMinNormal = DST_REP_T(1) << DST_SIG_BITS;
+
+  // Break a into a sign and representation of the absolute value
+  const union { SRC_T f; SRC_REP_T i; } src_rep = {.f = a};
+  const SRC_REP_T aRep = src_rep.i;
+  const SRC_REP_T aAbs = aRep & srcAbsMask;
+  const SRC_REP_T sign = aRep & srcSignMask;
+  DST_REP_T absResult;
+
+  // If sizeof(SRC_REP_T) < sizeof(int), the subtraction result is promoted
+  // to (signed) int.  To avoid that, explicitly cast to SRC_REP_T.
+  if ((SRC_REP_T)(aAbs - srcMinNormal) < srcInfinity - srcMinNormal) {
+    // a is a normal number.
+    // Extend to the destination type by shifting the significand and
+    // exponent into the proper position and rebiasing the exponent.
+    absResult = (DST_REP_T)aAbs << (DST_SIG_BITS - SRC_SIG_BITS);
+    absResult += (DST_REP_T)(dstExpBias - srcExpBias) << DST_SIG_BITS;
+  }
+
+  else if (aAbs >= srcInfinity) {
+    // a is NaN or infinity.
+    // Conjure the result by beginning with infinity, then setting the qNaN
+    // bit (if needed) and right-aligning the rest of the trailing NaN
+    // payload field.
+    absResult = (DST_REP_T)dstInfExp << DST_SIG_BITS;
+    absResult |= (DST_REP_T)(aAbs & srcQNaN) << (DST_SIG_BITS - SRC_SIG_BITS);
+    absResult |= (DST_REP_T)(aAbs & srcNaNCode) << (DST_SIG_BITS - SRC_SIG_BITS);
+  }
+  else if (aAbs) {
+    // a is denormal.
+    // renormalize the significand and clear the leading bit, then insert
+    // the correct adjusted exponent in the destination type.
+    const int scale = __clz(aAbs) - __clz(srcMinNormal);
+    absResult = (DST_REP_T)aAbs << (DST_SIG_BITS - SRC_SIG_BITS + scale);
+    absResult ^= dstMinNormal;
+    const int resultExponent = dstExpBias - srcExpBias - scale + 1;
+    absResult |= (DST_REP_T)resultExponent << DST_SIG_BITS;
+  }
+  else {
+    // a is zero.
+    absResult = 0;
+  }
+
+  // Apply the signbit to (DST_T)abs(a).
+  const DST_REP_T result = absResult | (DST_REP_T)sign << (dstBits - srcBits);
+  const union { DST_T f; DST_REP_T i; } dst_rep = {.i = result};
+  return dst_rep.f;
+}