src/f32-raddstoreexpminusmax/neon-p5.c.in - platform/external/XNNPACK - Git at Google

 // Copyright 2020 Google LLC
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.

 $assert ELEMENTS_TILE % 4 == 0
 $assert ELEMENTS_TILE >= 4
 $SIMD_TILE = ELEMENTS_TILE // 4
 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
 $VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
 #include <assert.h>

 #include <arm_neon.h>

 #include <xnnpack/common.h>
 #include <xnnpack/raddstoreexpminusmax.h>


 void xnn_f32_raddstoreexpminusmax_ukernel__${"neonfma" if FMA else "neon"}_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
     size_t elements,
     const float* input,
     float* output,
     float* sum,
     float max)
 {
   assert(elements % sizeof(float) == 0);

   const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
   // The smallest x for which expf(x) is normalized.
   const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep6f);
   const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
   $if FMA:
     const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E43p-1f);
     const float32x4_t vminus_ln2_lo = vmovq_n_f32(0x1.05C61p-29f);
   $else:
     // Last 7 bits are zeroes
     const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E400p-1f);
     const float32x4_t vminus_ln2_lo = vmovq_n_f32(-0x1.7F7D1Cp-20f);

   const float32x4_t vc1 = vmovq_n_f32(0x1.FFFFF6p-1f);
   const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f);
   const float32x4_t vc3 = vmovq_n_f32(0x1.555A80p-3f);
   const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f);
   const float32x4_t vc5 = vmovq_n_f32(0x1.0F9F9Cp-7f);

   const float32x4_t vi_max = vdupq_n_f32(max);

   $if ELEMENTS_TILE > 4:
     $for K in range(ACCUMULATORS):
       float32x4_t vacc${K} = vmovq_n_f32(0.0f);
     for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
       // Load ${ELEMENTS_TILE} (${SIMD_TILE}x4) inputs at a time.
       $for N in range(0, ELEMENTS_TILE, 4):
         const float32x4_t vi${ABC[N:N+4]} = vld1q_f32(input); input += 4;

       // Subtract maximum input x := i - i_max. This implies x <= 0.
       $for N in range(0, ELEMENTS_TILE, 4):
         const float32x4_t vx${ABC[N:N+4]} = vsubq_f32(vi${ABC[N:N+4]}, vi_max);

       // Compute reduced argument n := round(x / log(2)).
       // We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
       // large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
       // The trick with adding large number is valid only within certain bounds (|x| <= 2**22), but thats ok, because
       // inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very end
       // of the algorithm.
       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vn${ABC[N:N+4]} = ${VMULADDQ_F32}(vmagic_bias, vx${ABC[N:N+4]}, vlog2e);

       // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
       // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
       $for N in range(0, ELEMENTS_TILE, 4):
         const float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), 23));

       // Subtract the large number back to get final n := round(x / log(2)).
       $for N in range(0, ELEMENTS_TILE, 4):
         vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

       // Compute reduced argument t := z - n * log(2).
       // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vx${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_hi);

       $for N in range(0, ELEMENTS_TILE, 4):
         vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_lo);

       // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc4, vc5, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc3, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc2, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc1, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       // Reconstruct the final f value:
       //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
       //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
       //     = s + (t * s) * p
       $for N in range(0, ELEMENTS_TILE, 4):
         vt${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vf${ABC[N:N+4]} = ${VMULADDQ_F32}(vs${ABC[N:N+4]}, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       // For inputs below denormal cutoff, replace output with +0.0f.
       // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
       $for N in range(0, ELEMENTS_TILE, 4):
         vf${ABC[N:N+4]} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf${ABC[N:N+4]}), vcltq_f32(vx${ABC[N:N+4]}, vdenorm_cutoff)));

       // Store ${ELEMENTS_TILE} (${SIMD_TILE}x4) outputs at a time.
       $for N in range(0, ELEMENTS_TILE, 4):
         vst1q_f32(output, vf${ABC[N:N+4]}); output += 4;

       // Accumulate computed exponents.
       $for N in range(0, ELEMENTS_TILE, 4):
         vacc${N % ACCUMULATORS} = vaddq_f32(vacc${N % ACCUMULATORS}, vf${ABC[N:N+4]});
     }
     $if ACCUMULATORS > 1:
       // Add up all accumulators to vacc0
       $ACC_SLICE = 1
       $while ACC_SLICE < ACCUMULATORS:
         $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
           $if A + ACC_SLICE < ACCUMULATORS:
             vacc${A} = vaddq_f32(vacc${A}, vacc${A + ACC_SLICE});
         $ACC_SLICE *= 2

     float32x4_t vacc = vacc0;
   $else:
     float32x4_t vacc = vmovq_n_f32(0.0f);
   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
     // Load 4 inputs at a time.
     const float32x4_t vi = vld1q_f32(input); input += 4;

     // Subtract maximum input x := i - i_max. This implies x <= 0.
     const float32x4_t vx = vsubq_f32(vi, vi_max);

     // Compute reduced argument n := round(x / log(2)).
     // We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
     // large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
     // The trick with adding large number is valid only within certain bounds (|x| <= 2**22), but thats ok, because
     // inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very end
     // of the algorithm.
     float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
     // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

     // Subtract the large number back to get final n := round(x / log(2)).
     vn = vsubq_f32(vn, vmagic_bias);

     // Compute reduced argument t := z - n * log(2).
     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
     float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
     vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
     float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
     vp = ${VMULADDQ_F32}(vc3, vp, vt);
     vp = ${VMULADDQ_F32}(vc2, vp, vt);
     vp = ${VMULADDQ_F32}(vc1, vp, vt);

     // Reconstruct the final f value:
     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
     //     = s + (t * s) * p
     vt = vmulq_f32(vt, vs);
     float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

     // For inputs below denormal cutoff, replace output with +0.0f.
     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

     // Store 4 outputs at a time.
     vst1q_f32(output, vf); output += 4;

     // Accumulate computed exponents.
     vacc = vaddq_f32(vacc, vf);
   }
 #if XNN_ARCH_ARM64
   float vacc_lo = vaddvq_f32(vacc);
 #else
   float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
 #endif
   if (elements != 0) {
     assert(elements >= 1 * sizeof(float));
     assert(elements <= 3 * sizeof(float));
     // Load 4 inputs at a time.
     const float32x4_t vi = vld1q_f32(input); input += 4;

     // Subtract maximum input x := i - i_max. This implies x <= 0.
     const float32x4_t vx = vsubq_f32(vi, vi_max);

     // Compute reduced argument n := round(x / log(2)).
     // We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
     // large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
     // The trick with adding large number is valid only within certain bounds (|x| <= 2**22), but thats ok, because
     // inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very end
     // of the algorithm.
     float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
     // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

     // Subtract the large number back to get final n := round(x / log(2)).
     vn = vsubq_f32(vn, vmagic_bias);

     // Compute reduced argument t := z - n * log(2).
     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
     float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
     vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
     float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
     vp = ${VMULADDQ_F32}(vc3, vp, vt);
     vp = ${VMULADDQ_F32}(vc2, vp, vt);
     vp = ${VMULADDQ_F32}(vc1, vp, vt);

     // Reconstruct the final f value:
     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
     //     = s + (t * s) * p
     vt = vmulq_f32(vt, vs);
     float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

     // For inputs below denormal cutoff, replace output with +0.0f.
     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

     float32x2_t vf_lo = vget_low_f32(vf);
     if (elements & (2 * sizeof(float))) {
       // Store 2 outputs at a time.
       vst1_f32(output, vf_lo); output += 2;

       // Accumulate 2 computed exponents.
       #if XNN_ARCH_ARM64
         vacc_lo += vaddv_f32(vf_lo);
       #else
         vacc_lo = vadd_f32(vacc_lo, vf_lo);
       #endif

       vf_lo = vget_high_f32(vf);
     }
     if (elements & (1 * sizeof(float))) {
       // Store 1 output at a time.
       vst1_lane_f32(output, vf_lo, 0);

       // Accumulate 1 computed exponent.
       #if XNN_ARCH_ARM64
         vacc_lo += vget_lane_f32(vf_lo, 0);
       #else
         vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
       #endif
     }
   }
   // Reduce 4 elements in the SIMD register
 #if XNN_ARCH_ARM64
   *sum = vacc_lo;
 #else
   vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
 #endif
 }
	// Copyright 2020 Google LLC
	//
	// This source code is licensed under the BSD-style license found in the
	// LICENSE file in the root directory of this source tree.

	$assert ELEMENTS_TILE % 4 == 0
	$assert ELEMENTS_TILE >= 4
	$SIMD_TILE = ELEMENTS_TILE // 4
	$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
	$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
	#include <assert.h>

	#include <arm_neon.h>

	#include <xnnpack/common.h>
	#include <xnnpack/raddstoreexpminusmax.h>


	void xnn_f32_raddstoreexpminusmax_ukernel__${"neonfma" if FMA else "neon"}_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
	size_t elements,
	const float* input,
	float* output,
	float* sum,
	float max)
	{
	assert(elements % sizeof(float) == 0);

	const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
	// The smallest x for which expf(x) is normalized.
	const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep6f);
	const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
	$if FMA:
	const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E43p-1f);
	const float32x4_t vminus_ln2_lo = vmovq_n_f32(0x1.05C61p-29f);
	$else:
	// Last 7 bits are zeroes
	const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E400p-1f);
	const float32x4_t vminus_ln2_lo = vmovq_n_f32(-0x1.7F7D1Cp-20f);

	const float32x4_t vc1 = vmovq_n_f32(0x1.FFFFF6p-1f);
	const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f);
	const float32x4_t vc3 = vmovq_n_f32(0x1.555A80p-3f);
	const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f);
	const float32x4_t vc5 = vmovq_n_f32(0x1.0F9F9Cp-7f);

	const float32x4_t vi_max = vdupq_n_f32(max);

	$if ELEMENTS_TILE > 4:
	$for K in range(ACCUMULATORS):
	float32x4_t vacc${K} = vmovq_n_f32(0.0f);
	for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
	// Load ${ELEMENTS_TILE} (${SIMD_TILE}x4) inputs at a time.
	$for N in range(0, ELEMENTS_TILE, 4):
	const float32x4_t vi${ABC[N:N+4]} = vld1q_f32(input); input += 4;

	// Subtract maximum input x := i - i_max. This implies x <= 0.
	$for N in range(0, ELEMENTS_TILE, 4):
	const float32x4_t vx${ABC[N:N+4]} = vsubq_f32(vi${ABC[N:N+4]}, vi_max);

	// Compute reduced argument n := round(x / log(2)).
	// We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
	// large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
	// The trick with adding large number is valid only within certain bounds (\|x\| <= 2**22), but thats ok, because
	// inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very end
	// of the algorithm.
	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vn${ABC[N:N+4]} = ${VMULADDQ_F32}(vmagic_bias, vx${ABC[N:N+4]}, vlog2e);

	// Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
	// -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
	$for N in range(0, ELEMENTS_TILE, 4):
	const float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), 23));

	// Subtract the large number back to get final n := round(x / log(2)).
	$for N in range(0, ELEMENTS_TILE, 4):
	vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

	// Compute reduced argument t := z - n * log(2).
	// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vx${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_hi);

	$for N in range(0, ELEMENTS_TILE, 4):
	vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_lo);

	// Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc4, vc5, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc3, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc2, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc1, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	// Reconstruct the final f value:
	// f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
	// = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
	// = s + (t * s) * p
	$for N in range(0, ELEMENTS_TILE, 4):
	vt${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vf${ABC[N:N+4]} = ${VMULADDQ_F32}(vs${ABC[N:N+4]}, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	// For inputs below denormal cutoff, replace output with +0.0f.
	// Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
	$for N in range(0, ELEMENTS_TILE, 4):
	vf${ABC[N:N+4]} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf${ABC[N:N+4]}), vcltq_f32(vx${ABC[N:N+4]}, vdenorm_cutoff)));

	// Store ${ELEMENTS_TILE} (${SIMD_TILE}x4) outputs at a time.
	$for N in range(0, ELEMENTS_TILE, 4):
	vst1q_f32(output, vf${ABC[N:N+4]}); output += 4;

	// Accumulate computed exponents.
	$for N in range(0, ELEMENTS_TILE, 4):
	vacc${N % ACCUMULATORS} = vaddq_f32(vacc${N % ACCUMULATORS}, vf${ABC[N:N+4]});
	}
	$if ACCUMULATORS > 1:
	// Add up all accumulators to vacc0
	$ACC_SLICE = 1
	$while ACC_SLICE < ACCUMULATORS:
	$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
	$if A + ACC_SLICE < ACCUMULATORS:
	vacc${A} = vaddq_f32(vacc${A}, vacc${A + ACC_SLICE});
	$ACC_SLICE *= 2

	float32x4_t vacc = vacc0;
	$else:
	float32x4_t vacc = vmovq_n_f32(0.0f);
	for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
	// Load 4 inputs at a time.
	const float32x4_t vi = vld1q_f32(input); input += 4;

	// Subtract maximum input x := i - i_max. This implies x <= 0.
	const float32x4_t vx = vsubq_f32(vi, vi_max);

	// Compute reduced argument n := round(x / log(2)).
	// We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
	// large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
	// The trick with adding large number is valid only within certain bounds (\|x\| <= 2**22), but thats ok, because
	// inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very end
	// of the algorithm.
	float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

	// Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
	// -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
	const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

	// Subtract the large number back to get final n := round(x / log(2)).
	vn = vsubq_f32(vn, vmagic_bias);

	// Compute reduced argument t := z - n * log(2).
	// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
	float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
	vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

	// Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
	float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
	vp = ${VMULADDQ_F32}(vc3, vp, vt);
	vp = ${VMULADDQ_F32}(vc2, vp, vt);
	vp = ${VMULADDQ_F32}(vc1, vp, vt);

	// Reconstruct the final f value:
	// f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
	// = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
	// = s + (t * s) * p
	vt = vmulq_f32(vt, vs);
	float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

	// For inputs below denormal cutoff, replace output with +0.0f.
	// Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
	vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

	// Store 4 outputs at a time.
	vst1q_f32(output, vf); output += 4;

	// Accumulate computed exponents.
	vacc = vaddq_f32(vacc, vf);
	}
	#if XNN_ARCH_ARM64
	float vacc_lo = vaddvq_f32(vacc);
	#else
	float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
	#endif
	if (elements != 0) {
	assert(elements >= 1 * sizeof(float));
	assert(elements <= 3 * sizeof(float));
	// Load 4 inputs at a time.
	const float32x4_t vi = vld1q_f32(input); input += 4;

	// Subtract maximum input x := i - i_max. This implies x <= 0.
	const float32x4_t vx = vsubq_f32(vi, vi_max);

	// Compute reduced argument n := round(x / log(2)).
	// We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
	// large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
	// The trick with adding large number is valid only within certain bounds (\|x\| <= 2**22), but thats ok, because
	// inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very end
	// of the algorithm.
	float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

	// Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
	// -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
	const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

	// Subtract the large number back to get final n := round(x / log(2)).
	vn = vsubq_f32(vn, vmagic_bias);

	// Compute reduced argument t := z - n * log(2).
	// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
	float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
	vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

	// Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
	float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
	vp = ${VMULADDQ_F32}(vc3, vp, vt);
	vp = ${VMULADDQ_F32}(vc2, vp, vt);
	vp = ${VMULADDQ_F32}(vc1, vp, vt);

	// Reconstruct the final f value:
	// f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
	// = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
	// = s + (t * s) * p
	vt = vmulq_f32(vt, vs);
	float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

	// For inputs below denormal cutoff, replace output with +0.0f.
	// Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
	vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

	float32x2_t vf_lo = vget_low_f32(vf);
	if (elements & (2 * sizeof(float))) {
	// Store 2 outputs at a time.
	vst1_f32(output, vf_lo); output += 2;

	// Accumulate 2 computed exponents.
	#if XNN_ARCH_ARM64
	vacc_lo += vaddv_f32(vf_lo);
	#else
	vacc_lo = vadd_f32(vacc_lo, vf_lo);
	#endif

	vf_lo = vget_high_f32(vf);
	}
	if (elements & (1 * sizeof(float))) {
	// Store 1 output at a time.
	vst1_lane_f32(output, vf_lo, 0);

	// Accumulate 1 computed exponent.
	#if XNN_ARCH_ARM64
	vacc_lo += vget_lane_f32(vf_lo, 0);
	#else
	vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
	#endif
	}
	}
	// Reduce 4 elements in the SIMD register
	#if XNN_ARCH_ARM64
	*sum = vacc_lo;
	#else
	vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
	#endif
	}