sse approx

VowpalWabbit · Nov 10, 2023 · f60ce8f · f60ce8f
1 parent 3f88060
commit f60ce8f
Show file tree

Hide file tree

Showing 3 changed files with 36 additions and 57 deletions.
diff --git a/test/train-sets/ref/active_interactive.stdout b/test/train-sets/ref/active_interactive.stdout
@@ -4,25 +4,6 @@ Connecting to localhost:12353...
 Done
 Sending unlabeled examples...
 Request for example 0: tag="", prediction=0.0: 
-Provide? [0/1/skip]: Request for example 1: tag="", prediction=0.0: 
-Provide? [0/1/skip]: Request for example 2: tag="", prediction=0.076144: 
-Provide? [0/1/skip]: Request for example 3: tag="", prediction=0.147194: 
-Provide? [0/1/skip]: Request for example 4: tag="", prediction=0.060332: 
-Provide? [0/1/skip]: Request for example 5: tag="", prediction=0.060998: 
-Provide? [0/1/skip]: Request for example 6: tag="", prediction=0.090767: 
-Provide? [0/1/skip]: Request for example 7: tag="", prediction=0.049165: 
-Provide? [0/1/skip]: Request for example 8: tag="", prediction=0.147871: 
-Provide? [0/1/skip]: Request for example 9: tag="", prediction=0.099147: 
-Provide? [0/1/skip]: Request for example 10: tag="", prediction=0.006728: 
-Provide? [0/1/skip]: Request for example 11: tag="", prediction=0.16017: 
-Provide? [0/1/skip]: Request for example 12: tag="", prediction=0.117594: 
-Provide? [0/1/skip]: Request for example 13: tag="", prediction=0.052956: 
-Provide? [0/1/skip]: Request for example 14: tag="", prediction=0.17299: 
-Provide? [0/1/skip]: Request for example 15: tag="", prediction=0.521811: 
-Provide? [0/1/skip]: Request for example 16: tag="", prediction=0.761672: 
-Provide? [0/1/skip]: Request for example 17: tag="", prediction=0.992553: 
-Provide? [0/1/skip]: Request for example 18: tag="", prediction=0.360711: 
-Provide? [0/1/skip]: Request for example 19: tag="", prediction=0.195309: 
 Provide? [0/1/skip]: 
 active_interactor stderr: 
 
@@ -44,17 +25,14 @@ average  since         example        example        current        current  cur
 loss     last          counter         weight          label        predict features
 n.a.     n.a.                1            1.0        unknown         0.0000       41
 0.000000 0.000000            2            2.0         0.0000         0.0000       41
-0.500000 1.000000            4            4.0         1.0000         0.0000       75
-0.256866 0.013732            8            8.0         0.0000         0.1472       44
-0.130685 0.004504           16           16.0         0.0000         0.0492       42
-0.182888 0.235091           32           32.0         1.0000         0.5218       34
+0.000000 n.a.                4            4.0        unknown         0.0000       62
+0.000000 n.a.                8            8.0        unknown         0.0000      108
+0.000000 n.a.               16           16.0        unknown         0.0000       31
 
 finished run
-number of examples = 40
-weighted example sum = 40.000000
-weighted label sum = 6.000000
-average loss = 0.157566
-best constant = 0.300000
-best constant's loss = 0.210000
-total feature number = 2522
+number of examples = 21
+weighted example sum = 21.000000
+weighted label sum = 0.000000
+average loss = 0.000000
+total feature number = 1302
 
diff --git a/test/train-sets/ref/big_feature_poison.stdout b/test/train-sets/ref/big_feature_poison.stdout
@@ -4,3 +4,4 @@
 [warning] NAN prediction in example 2, forcing 0
 [error] The features have too much magnitude
 [warning] update is NAN, replacing with 0
+[warning] update is NAN, replacing with 0
diff --git a/vowpalwabbit/core/src/reductions/gd.cc b/vowpalwabbit/core/src/reductions/gd.cc
@@ -153,33 +153,33 @@ VW_WARNING_STATE_POP
 
 static inline float inv_sqrt(float x)
 {
-  return 1.0f / std::sqrt(x);
-  /*
-  #if !defined(VW_NO_INLINE_SIMD)
-  #  if defined(__ARM_NEON__)
-    // Propagate into vector
-    float32x2_t v1 = vdup_n_f32(x);
-    // Estimate
-    float32x2_t e1 = vrsqrte_f32(v1);
-    // N-R iteration 1
-    float32x2_t e2 = vmul_f32(e1, vrsqrts_f32(v1, vmul_f32(e1, e1)));
-    // N-R iteration 2
-    float32x2_t e3 = vmul_f32(e2, vrsqrts_f32(v1, vmul_f32(e2, e2)));
-    // Extract result
-    return vget_lane_f32(e3, 0);
-  #  elif defined(__SSE2__)
-    __m128 eta = _mm_load_ss(&x);
-    eta = _mm_rsqrt_ss(eta);
-    _mm_store_ss(&x, eta);
-  #  else
-    x = quake_inv_sqrt(x);
-  #  endif
-  #else
-    x = quake_inv_sqrt(x);
-  #endif
-
-    return x;
-  */
+#if !defined(VW_NO_INLINE_SIMD)
+#  if defined(__ARM_NEON__)
+  // Propagate into vector
+  float32x2_t v1 = vdup_n_f32(x);
+  // Estimate
+  float32x2_t e1 = vrsqrte_f32(v1);
+  // N-R iteration 1
+  float32x2_t e2 = vmul_f32(e1, vrsqrts_f32(v1, vmul_f32(e1, e1)));
+  // N-R iteration 2
+  float32x2_t e3 = vmul_f32(e2, vrsqrts_f32(v1, vmul_f32(e2, e2)));
+  // Extract result
+  return vget_lane_f32(e3, 0);
+#  elif defined(__SSE2__)
+  __m128 eta = _mm_load_ss(&x);
+  eta = _mm_rsqrt_ss(eta);  // Fast approximate inverse square root
+  // One iteration of Newton-Raphson refinement:
+  __m128 half_x = _mm_set_ss(0.5f * x);
+  eta = _mm_mul_ss(eta, _mm_sub_ss(_mm_set_ss(1.5f), _mm_mul_ss(half_x, _mm_mul_ss(eta, eta))));
+  _mm_store_ss(&x, eta);
+#  else
+  x = quake_inv_sqrt(x);
+#  endif
+#else
+  x = quake_inv_sqrt(x);
+#endif
+
+  return x;
 }
 
 VW_WARNING_STATE_PUSH