src/f32-vsigmoid/avx-p5.c.in - platform/external/XNNPACK - Gitiles

 // 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 BATCH_TILE % 8 == 0
 $assert BATCH_TILE >= 8
 $assert RR_STEPS in [1, 2]
 $assert DIV_ALGO in ["div", "nr2"]
 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
 $SIMD_TILE = BATCH_TILE // 8
 #include <assert.h>

 #include <immintrin.h>

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


 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

 void xnn_f32_vsigmoid_ukernel__avx_rr${RR_STEPS}_p5_${DIV_ALGO}_x${BATCH_TILE}(
     size_t n,
     const float* x,
     float* y,
     const void* params)
 {
   assert(n % sizeof(float) == 0);

   const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
   $if RR_STEPS == 1:
     const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
   $else:
     const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
     const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
   const __m256 vone = _mm256_set1_ps(1.0f);
   $if DIV_ALGO == "nr2":
     const __m256 vtwo = _mm256_set1_ps(2.0f);
   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);

   $if BATCH_TILE > 8:
     for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
       const __m256 vx${ABC[0]} = _mm256_loadu_ps(x);
       $for N in range(1, SIMD_TILE):
         const __m256 vx${ABC[N]} = _mm256_loadu_ps(x + ${N * 8});
       x += ${BATCH_TILE};

       $for N in range(SIMD_TILE):
         const __m256 vz${ABC[N]} = _mm256_or_ps(vx${ABC[N]}, vsign_mask);

       $for N in range(SIMD_TILE):
         __m256 vn${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vz${ABC[N]}, vlog2e), vmagic_bias);

       $for N in range(SIMD_TILE):
         const __m128 vs_lo${ABC[N]} = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn${ABC[N]})), 23));
         const __m128 vs_hi${ABC[N]} = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn${ABC[N]}, 1)), 23));
         const __m256 vs${ABC[N]} = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo${ABC[N]}), vs_hi${ABC[N]}, 1);

       $for N in range(SIMD_TILE):
         vn${ABC[N]} = _mm256_sub_ps(vn${ABC[N]}, vmagic_bias);

       $if RR_STEPS == 1:
         $for N in range(SIMD_TILE):
           __m256 vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2), vz${ABC[N]});
       $else:
         $for N in range(SIMD_TILE):
           __m256 vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2_hi), vz${ABC[N]});

         $for N in range(SIMD_TILE):
           vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2_lo), vt${ABC[N]});

       $for N in range(SIMD_TILE):
         __m256 vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vc5, vt${ABC[N]}), vc4);

       $for N in range(SIMD_TILE):
         vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc3);

       $for N in range(SIMD_TILE):
         vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc2);

       $for N in range(SIMD_TILE):
         vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc1);

       $for N in range(SIMD_TILE):
         vt${ABC[N]} = _mm256_mul_ps(vt${ABC[N]}, vs${ABC[N]});

       $for N in range(SIMD_TILE):
         const __m256 ve${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vt${ABC[N]}, vp${ABC[N]}), vs${ABC[N]});

       $for N in range(SIMD_TILE):
         const __m256 vd${ABC[N]} = _mm256_add_ps(ve${ABC[N]}, vone);

       $if DIV_ALGO == "div":
         $for N in range(SIMD_TILE):
           __m256 vf${ABC[N]} = _mm256_div_ps(ve${ABC[N]}, vd${ABC[N]});
       $else:
         $for N in range(SIMD_TILE):
           __m256 vr${ABC[N]} = _mm256_rcp_ps(vd${ABC[N]});

         $for N in range(SIMD_TILE):
           vr${ABC[N]} = _mm256_mul_ps(vr${ABC[N]}, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr${ABC[N]}, vd${ABC[N]})));
           vr${ABC[N]} = _mm256_mul_ps(vr${ABC[N]}, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr${ABC[N]}, vd${ABC[N]})));

         $for N in range(SIMD_TILE):
           __m256 vf${ABC[N]} = _mm256_mul_ps(ve${ABC[N]}, vr${ABC[N]});

       $for N in range(SIMD_TILE):
         vf${ABC[N]} = _mm256_andnot_ps(_mm256_cmp_ps(vz${ABC[N]}, vdenorm_cutoff, _CMP_LT_OS), vf${ABC[N]});

       $for N in range(SIMD_TILE):
         vf${ABC[N]} = _mm256_blendv_ps(_mm256_sub_ps(vone, vf${ABC[N]}), vf${ABC[N]}, vx${ABC[N]});

       _mm256_storeu_ps(y, vf${ABC[0]});
       $for N in range(1, SIMD_TILE):
         _mm256_storeu_ps(y + ${N * 8}, vf${ABC[N]});
       y += ${BATCH_TILE};
     }
   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
     const __m256 vx = _mm256_loadu_ps(x);
     x += 8;

     const __m256 vz = _mm256_or_ps(vx, vsign_mask);

     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);

     const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
     const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
     const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

     vn = _mm256_sub_ps(vn, vmagic_bias);

     $if RR_STEPS == 1:
       __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2), vz);
     $else:
       __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
       vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);

     vt = _mm256_mul_ps(vt, vs);
     const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);

     const __m256 vd = _mm256_add_ps(ve, vone);
     $if DIV_ALGO == "div":
       __m256 vf = _mm256_div_ps(ve, vd);
     $else:
       __m256 vr = _mm256_rcp_ps(vd);
       vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
       vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
       __m256 vf = _mm256_mul_ps(ve, vr);

     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

     _mm256_storeu_ps(y, vf);
     y += 8;
   }
   if XNN_UNLIKELY(n != 0) {
     assert(n >= 1 * sizeof(float));
     assert(n <= 7 * sizeof(float));
     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));

     const __m256 vx = _mm256_maskload_ps(x, vmask);

     const __m256 vz = _mm256_or_ps(vx, vsign_mask);

     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
     const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
     const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
     const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

     vn = _mm256_sub_ps(vn, vmagic_bias);

     $if RR_STEPS == 1:
       __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2), vz);
     $else:
       __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
       vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);

     vt = _mm256_mul_ps(vt, vs);
     const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);

     const __m256 vd = _mm256_add_ps(ve, vone);
     $if DIV_ALGO == "div":
       __m256 vf = _mm256_div_ps(ve, vd);
     $else:
       __m256 vr = _mm256_rcp_ps(vd);
       vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
       vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
       __m256 vf = _mm256_mul_ps(ve, vr);

     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
     __m128 vf_lo = _mm256_castps256_ps128(vf);
     if (n & (4 * sizeof(float))) {
       _mm_storeu_ps(y, vf_lo);
       vf_lo = _mm256_extractf128_ps(vf, 1);
       y += 4;
     }
     if (n & (2 * sizeof(float))) {
       _mm_storel_pi((__m64*) y, vf_lo);
       vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
       y += 2;
     }
     if (n & (1 * sizeof(float))) {
       _mm_store_ss(y, vf_lo);
     }
   }
 }
	// 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 BATCH_TILE % 8 == 0
	$assert BATCH_TILE >= 8
	$assert RR_STEPS in [1, 2]
	$assert DIV_ALGO in ["div", "nr2"]
	$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
	$SIMD_TILE = BATCH_TILE // 8
	#include <assert.h>

	#include <immintrin.h>

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


	static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

	void xnn_f32_vsigmoid_ukernel__avx_rr${RR_STEPS}_p5_${DIV_ALGO}_x${BATCH_TILE}(
	size_t n,
	const float* x,
	float* y,
	const void* params)
	{
	assert(n % sizeof(float) == 0);

	const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
	const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
	const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
	$if RR_STEPS == 1:
	const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
	$else:
	const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
	const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
	const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
	const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
	const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
	const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
	const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
	const __m256 vone = _mm256_set1_ps(1.0f);
	$if DIV_ALGO == "nr2":
	const __m256 vtwo = _mm256_set1_ps(2.0f);
	const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);

	$if BATCH_TILE > 8:
	for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
	const __m256 vx${ABC[0]} = _mm256_loadu_ps(x);
	$for N in range(1, SIMD_TILE):
	const __m256 vx${ABC[N]} = _mm256_loadu_ps(x + ${N * 8});
	x += ${BATCH_TILE};

	$for N in range(SIMD_TILE):
	const __m256 vz${ABC[N]} = _mm256_or_ps(vx${ABC[N]}, vsign_mask);

	$for N in range(SIMD_TILE):
	__m256 vn${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vz${ABC[N]}, vlog2e), vmagic_bias);

	$for N in range(SIMD_TILE):
	const __m128 vs_lo${ABC[N]} = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn${ABC[N]})), 23));
	const __m128 vs_hi${ABC[N]} = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn${ABC[N]}, 1)), 23));
	const __m256 vs${ABC[N]} = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo${ABC[N]}), vs_hi${ABC[N]}, 1);

	$for N in range(SIMD_TILE):
	vn${ABC[N]} = _mm256_sub_ps(vn${ABC[N]}, vmagic_bias);

	$if RR_STEPS == 1:
	$for N in range(SIMD_TILE):
	__m256 vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2), vz${ABC[N]});
	$else:
	$for N in range(SIMD_TILE):
	__m256 vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2_hi), vz${ABC[N]});

	$for N in range(SIMD_TILE):
	vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2_lo), vt${ABC[N]});

	$for N in range(SIMD_TILE):
	__m256 vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vc5, vt${ABC[N]}), vc4);

	$for N in range(SIMD_TILE):
	vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc3);

	$for N in range(SIMD_TILE):
	vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc2);

	$for N in range(SIMD_TILE):
	vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc1);

	$for N in range(SIMD_TILE):
	vt${ABC[N]} = _mm256_mul_ps(vt${ABC[N]}, vs${ABC[N]});

	$for N in range(SIMD_TILE):
	const __m256 ve${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vt${ABC[N]}, vp${ABC[N]}), vs${ABC[N]});

	$for N in range(SIMD_TILE):
	const __m256 vd${ABC[N]} = _mm256_add_ps(ve${ABC[N]}, vone);

	$if DIV_ALGO == "div":
	$for N in range(SIMD_TILE):
	__m256 vf${ABC[N]} = _mm256_div_ps(ve${ABC[N]}, vd${ABC[N]});
	$else:
	$for N in range(SIMD_TILE):
	__m256 vr${ABC[N]} = _mm256_rcp_ps(vd${ABC[N]});

	$for N in range(SIMD_TILE):
	vr${ABC[N]} = _mm256_mul_ps(vr${ABC[N]}, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr${ABC[N]}, vd${ABC[N]})));
	vr${ABC[N]} = _mm256_mul_ps(vr${ABC[N]}, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr${ABC[N]}, vd${ABC[N]})));

	$for N in range(SIMD_TILE):
	__m256 vf${ABC[N]} = _mm256_mul_ps(ve${ABC[N]}, vr${ABC[N]});

	$for N in range(SIMD_TILE):
	vf${ABC[N]} = _mm256_andnot_ps(_mm256_cmp_ps(vz${ABC[N]}, vdenorm_cutoff, _CMP_LT_OS), vf${ABC[N]});

	$for N in range(SIMD_TILE):
	vf${ABC[N]} = _mm256_blendv_ps(_mm256_sub_ps(vone, vf${ABC[N]}), vf${ABC[N]}, vx${ABC[N]});

	_mm256_storeu_ps(y, vf${ABC[0]});
	$for N in range(1, SIMD_TILE):
	_mm256_storeu_ps(y + ${N * 8}, vf${ABC[N]});
	y += ${BATCH_TILE};
	}
	for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
	const __m256 vx = _mm256_loadu_ps(x);
	x += 8;

	const __m256 vz = _mm256_or_ps(vx, vsign_mask);

	__m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);

	const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
	const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
	const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

	vn = _mm256_sub_ps(vn, vmagic_bias);

	$if RR_STEPS == 1:
	__m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2), vz);
	$else:
	__m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
	vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

	__m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
	vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
	vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
	vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);

	vt = _mm256_mul_ps(vt, vs);
	const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);

	const __m256 vd = _mm256_add_ps(ve, vone);
	$if DIV_ALGO == "div":
	__m256 vf = _mm256_div_ps(ve, vd);
	$else:
	__m256 vr = _mm256_rcp_ps(vd);
	vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
	vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
	__m256 vf = _mm256_mul_ps(ve, vr);

	vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
	vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

	_mm256_storeu_ps(y, vf);
	y += 8;
	}
	if XNN_UNLIKELY(n != 0) {
	assert(n >= 1 * sizeof(float));
	assert(n <= 7 * sizeof(float));
	__m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));

	const __m256 vx = _mm256_maskload_ps(x, vmask);

	const __m256 vz = _mm256_or_ps(vx, vsign_mask);

	__m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
	const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
	const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
	const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

	vn = _mm256_sub_ps(vn, vmagic_bias);

	$if RR_STEPS == 1:
	__m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2), vz);
	$else:
	__m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
	vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

	__m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
	vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
	vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
	vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);

	vt = _mm256_mul_ps(vt, vs);
	const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);

	const __m256 vd = _mm256_add_ps(ve, vone);
	$if DIV_ALGO == "div":
	__m256 vf = _mm256_div_ps(ve, vd);
	$else:
	__m256 vr = _mm256_rcp_ps(vd);
	vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
	vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
	__m256 vf = _mm256_mul_ps(ve, vr);

	vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
	vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

	// _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
	__m128 vf_lo = _mm256_castps256_ps128(vf);
	if (n & (4 * sizeof(float))) {
	_mm_storeu_ps(y, vf_lo);
	vf_lo = _mm256_extractf128_ps(vf, 1);
	y += 4;
	}
	if (n & (2 * sizeof(float))) {
	_mm_storel_pi((__m64*) y, vf_lo);
	vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
	y += 2;
	}
	if (n & (1 * sizeof(float))) {
	_mm_store_ss(y, vf_lo);
	}
	}
	}