// 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>


extern XNN_INTERNAL const float xnn_table_exp2_k_over_64[64];

void xnn_f32_raddstoreexpminusmax_ukernel__${"neonfma" if FMA else "neon"}_lut64_p2_x${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(0x1.800000p23f);
  // The smallest x for which expf(x) is normalized.
  const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep6f);
  const float32x4_t vlog2e_x64  = vmovq_n_f32(0x1.715476p6f);
  $if FMA:
    const float32x4_t vminus_ln2_o64_hi = vmovq_n_f32(-0x1.62e43p-7f);
    const float32x4_t vminus_ln2_o64_lo = vmovq_n_f32(0x1.05c61p-35f);
  $else:
    // Last 13 bits are zeroes
    const float32x4_t vminus_ln2_o64_hi = vmovq_n_f32(-0x1.630000p-7f);
    const float32x4_t vminus_ln2_o64_lo = vmovq_n_f32(0x1.BD0106p-19f);

  const float32x4_t vc2 = vmovq_n_f32(0x1.FFFF0Ap-2f);

  const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));

  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 * 64 / log(2)).
      // We do it by adding a large number (magic bias), which cause rounding of the 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 * 64 / log(2)| <= 2**22, i.e.
      // |x| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs outside of [-87.336540, 0.0]
      // result in denormalized or underflown expf(x). 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_x64);

      // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that expf(x) is normalized,
      // i.e. -87.33642 <= x <= 0.0. As n has 6 fractional bits, we split s == 2**(n / 64) = 2**e * 2**(n / 64 - e), where
      // e := int(n / 64). We create s in two steps:
      // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
      //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
      // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
      //    number, because for -87.33642 <= x <= 0.0 (inputs for which expf(x) is normalized) we have -126 <= e <= 0,
      //    and thus the adjusted exponent is not lower than -126.
      //
      // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
      $for N in range(0, ELEMENTS_TILE, 4):
        const int32x4_t ve${ABC[N:N+4]} = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), vmovq_n_s32(INT32_C(0x3F))), 17);

      // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
      $for N in range(0, ELEMENTS_TILE, 4):
        const uint64x2_t vidx${ABC[N:N+4]} = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), vindex_mask));
        const uint64_t vidx${ABC[N:N+2]} = vgetq_lane_u64(vidx${ABC[N:N+4]}, 0);
        const uint64_t vidx${ABC[N+2:N+4]} = vgetq_lane_u64(vidx${ABC[N:N+4]}, 1);

      $for N in range(0, ELEMENTS_TILE, 4):
        float32x2_t vl${ABC[N:N+2]} = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx${ABC[N:N+2]}]);
        float32x2_t vl${ABC[N+2:N+4]} = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx${ABC[N+2:N+4]}]);

      $for N in range(0, ELEMENTS_TILE, 4):
        vl${ABC[N:N+2]} = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx${ABC[N:N+2]} >> 32)], vl${ABC[N:N+2]}, 1);
        vl${ABC[N+2:N+4]} = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx${ABC[N+2:N+4]} >> 32)], vl${ABC[N+2:N+4]}, 1);
        const float32x4_t vl${ABC[N:N+4]} = vcombine_f32(vl${ABC[N:N+2]}, vl${ABC[N+2:N+4]});

      // Adjust exponent of the value l fetched from the table to get the final s value.
      $for N in range(0, ELEMENTS_TILE, 4):
        const float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl${ABC[N:N+4]}), ve${ABC[N:N+4]}));

      // Subtract the large number back to get final n := round(x * 64 / log(2)) as a floating-point number.
      $for N in range(0, ELEMENTS_TILE, 4):
        vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

      // Compute reduced argument t := x - n * log(2) / 64.
      // Use Cody-Waite range reduction method (note the two constants representing log(2) / 64) 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_o64_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_o64_lo);

      // Compute degree-2 polynomial approxiatmion for exp(t) on [-log(2)/128, log(2)/128].
      $for N in range(0, ELEMENTS_TILE, 4):
        float32x4_t vp${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vc2);

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

      // Reconstruct the final f value:
      //   f = s * (1 + t * (1 + t * c2))
      //     = s * (1 + t + t * (t * c2))
      //     = s + s * (t + t * (t * c2))
      //     = s + s * p
      $for N in range(0, ELEMENTS_TILE, 4):
        float32x4_t vf${ABC[N:N+4]} = ${VMULADDQ_F32}(vs${ABC[N:N+4]}, vs${ABC[N:N+4]}, vp${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 * 64 / log(2)).
    // We do it by adding a large number (magic bias), which cause rounding of the 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 * 64 / log(2)| <= 2**22, i.e.
    // |x| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs outside of [-87.336540, 0.0]
    // result in denormalized or underflown expf(x). We fixup the result for such inputs at the very end of the
    // algorithm.
    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e_x64);

    // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that expf(x) is normalized,
    // i.e. -87.33642 <= x <= 0.0. As n has 6 fractional bits, we split s == 2**(n / 64) = 2**e * 2**(n / 64 - e), where
    // e := int(n / 64). We create s in two steps:
    // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
    //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
    // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
    //    number, because for -87.33642 <= x <= 0.0 (inputs for which expf(x) is normalized) we have -126 <= e <= 0,
    //    and thus the adjusted exponent is not lower than -126.
    //
    // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
    const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);

    // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
    const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
    const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
    const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
    float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_lo]);
    float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_hi]);
    vl_lo = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
    vl_hi = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
    const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
    // Adjust exponent of the value l fetched from the table to get the final s value.
    const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));

    // Subtract the large number back to get final n := round(x * 64 / log(2)) as a floating-point number.
    vn = vsubq_f32(vn, vmagic_bias);

    // Compute reduced argument t := x - n * log(2) / 64.
    // Use Cody-Waite range reduction method (note the two constants representing log(2) / 64) to improve accuracy.
    float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_o64_hi);
    vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_o64_lo);

    // Compute degree-2 polynomial approxiatmion for exp(t) on [-log(2)/128, log(2)/128].
    float32x4_t vp = vmulq_f32(vt, vc2);
    vp = ${VMULADDQ_F32}(vt, vt, vp);

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

    // 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 * 64 / log(2)).
    // We do it by adding a large number (magic bias), which cause rounding of the 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 * 64 / log(2)| <= 2**22, i.e.
    // |x| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs outside of [-87.336540, 0.0]
    // result in denormalized or underflown expf(x). We fixup the result for such inputs at the very end of the
    // algorithm.
    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e_x64);

    // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that expf(x) is normalized,
    // i.e. -87.33642 <= x <= 0.0. As n has 6 fractional bits, we split s == 2**(n / 64) = 2**e * 2**(n / 64 - e), where
    // e := int(n / 64). We create s in two steps:
    // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
    //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
    // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
    //    number, because for -87.33642 <= x <= 0.0 (inputs for which expf(x) is normalized) we have -126 <= e <= 0,
    //    and thus the adjusted exponent is not lower than -126.
    //
    // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
    const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);

    // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
    const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
    const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
    const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
    float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_lo]);
    float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_hi]);
    vl_lo = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
    vl_hi = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
    const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
    // Adjust exponent of the value l fetched from the table to get the final s value.
    const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));

    // Subtract the large number back to get final n := round(x * 64 / log(2)) as a floating-point number.
    vn = vsubq_f32(vn, vmagic_bias);

    // Compute reduced argument t := x - n * log(2) / 64.
    // Use Cody-Waite range reduction method (note the two constants representing log(2) / 64) to improve accuracy.
    float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_o64_hi);
    vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_o64_lo);

    // Compute degree-2 polynomial approxiatmion for exp(t) on [-log(2)/128, log(2)/128].
    float32x4_t vp = vmulq_f32(vt, vc2);
    vp = ${VMULADDQ_F32}(vt, vt, vp);

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

    // 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
}