AVX512F versions of Sigmoid microkernel
89% speedup on Skylake-X, 3.00 GElements/s -> 5.66 GElements/s
PiperOrigin-RevId: 333454419
diff --git a/src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c b/src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c
new file mode 100644
index 0000000..a399e95
--- /dev/null
+++ b/src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c
@@ -0,0 +1,183 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-sigmoid/avx512f-lut32-p2-perm2-scalef.c.in
+// Generator: tools/xngen
+//
+// 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.
+
+#include <assert.h>
+
+#include <immintrin.h>
+
+#include <xnnpack/common.h>
+#include <xnnpack/vunary.h>
+
+
+void xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64(
+ size_t n,
+ const float* x,
+ float* y,
+ const void* params)
+{
+ assert(n % sizeof(float) == 0);
+
+ const __m512i vsign_mask = _mm512_set1_epi32(0x80000000);
+ const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p18f);
+ const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
+ const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
+ const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
+
+ const __m512 vtable_hi = _mm512_set_ps(
+ 0x1.F50766p+0f, 0x1.EA4AFAp+0f, 0x1.DFC974p+0f, 0x1.D5818Ep+0f,
+ 0x1.CB720Ep+0f, 0x1.C199BEp+0f, 0x1.B7F770p+0f, 0x1.AE89FAp+0f,
+ 0x1.A5503Cp+0f, 0x1.9C4918p+0f, 0x1.93737Cp+0f, 0x1.8ACE54p+0f,
+ 0x1.82589Ap+0f, 0x1.7A1148p+0f, 0x1.71F75Ep+0f, 0x1.6A09E6p+0f);
+ const __m512 vtable_lo = _mm512_set_ps(
+ 0x1.6247ECp+0f, 0x1.5AB07Ep+0f, 0x1.5342B6p+0f, 0x1.4BFDAEp+0f,
+ 0x1.44E086p+0f, 0x1.3DEA64p+0f, 0x1.371A74p+0f, 0x1.306FE0p+0f,
+ 0x1.29E9E0p+0f, 0x1.2387A6p+0f, 0x1.1D4874p+0f, 0x1.172B84p+0f,
+ 0x1.11301Ep+0f, 0x1.0B5586p+0f, 0x1.059B0Ep+0f, 0x1.000000p+0f);
+
+ const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f);
+ const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f);
+ const __m512 vone = _mm512_set1_ps(1.0f);
+
+ for (; n >= 64 * sizeof(float); n -= 64 * sizeof(float)) {
+ const __m512 vx0 = _mm512_loadu_ps(x);
+ const __m512 vx1 = _mm512_loadu_ps(x + 16);
+ const __m512 vx2 = _mm512_loadu_ps(x + 32);
+ const __m512 vx3 = _mm512_loadu_ps(x + 48);
+ x += 64;
+
+ const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
+ const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
+ const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
+ const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
+
+ __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
+ __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
+ __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
+ __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
+
+ const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
+ const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
+ const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
+ const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
+
+ vn0 = _mm512_sub_ps(vn0, vmagic_bias);
+ vn1 = _mm512_sub_ps(vn1, vmagic_bias);
+ vn2 = _mm512_sub_ps(vn2, vmagic_bias);
+ vn3 = _mm512_sub_ps(vn3, vmagic_bias);
+
+ __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
+ __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
+ __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
+ __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
+
+ vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
+ vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
+ vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
+ vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
+
+ __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
+ __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
+ __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
+ __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
+
+ vt0 = _mm512_mul_ps(vt0, vl0);
+ vt1 = _mm512_mul_ps(vt1, vl1);
+ vt2 = _mm512_mul_ps(vt2, vl2);
+ vt3 = _mm512_mul_ps(vt3, vl3);
+
+ vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
+ vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
+ vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
+ vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
+
+ const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
+ const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
+ const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
+ const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
+
+ const __m512 vd0 = _mm512_add_ps(ve0, vone);
+ const __m512 vd1 = _mm512_add_ps(ve1, vone);
+ const __m512 vd2 = _mm512_add_ps(ve2, vone);
+ const __m512 vd3 = _mm512_add_ps(ve3, vone);
+
+ __m512 vf0 = _mm512_div_ps(ve0, vd0);
+ __m512 vf1 = _mm512_div_ps(ve1, vd1);
+ __m512 vf2 = _mm512_div_ps(ve2, vd2);
+ __m512 vf3 = _mm512_div_ps(ve3, vd3);
+
+ vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
+ vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
+ vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
+ vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
+
+ _mm512_storeu_ps(y, vf0);
+ _mm512_storeu_ps(y + 16, vf1);
+ _mm512_storeu_ps(y + 32, vf2);
+ _mm512_storeu_ps(y + 48, vf3);
+ y += 64;
+ }
+ for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
+ const __m512 vx = _mm512_loadu_ps(x);
+ x += 16;
+
+ const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
+
+ __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
+ const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
+ vn = _mm512_sub_ps(vn, vmagic_bias);
+
+ __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
+ vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
+
+ __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
+ vt = _mm512_mul_ps(vt, vl);
+ vp = _mm512_fmadd_ps(vt, vp, vl);
+
+ const __m512 ve = _mm512_scalef_ps(vp, vn);
+ const __m512 vd = _mm512_add_ps(ve, vone);
+
+ __m512 vf = _mm512_div_ps(ve, vd);
+
+ vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
+
+ _mm512_storeu_ps(y, vf);
+ y += 16;
+ }
+ if XNN_UNLIKELY(n != 0) {
+ assert(n >= 1 * sizeof(float));
+ assert(n <= 15 * sizeof(float));
+
+ // Prepare mask for valid 32-bit elements (depends on n).
+ n >>= 2 /* log2(sizeof(float)) */;
+ const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
+
+ const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
+ const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
+
+ __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
+ const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
+ vn = _mm512_sub_ps(vn, vmagic_bias);
+
+ __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
+ vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
+
+ __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
+ vt = _mm512_mul_ps(vt, vl);
+ vp = _mm512_fmadd_ps(vt, vp, vl);
+
+ const __m512 ve = _mm512_scalef_ps(vp, vn);
+ const __m512 vd = _mm512_add_ps(ve, vone);
+
+ __m512 vf = _mm512_div_ps(ve, vd);
+
+ vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
+
+ _mm512_mask_storeu_ps(y, vmask, vf);
+ }
+}