commit f44f0228165b2e8595d1535379fd77658eda46c9
author Marat Dukhan <maratek@google.com> Mon Dec 14 11:53:27 2020 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Mon Dec 14 11:53:59 2020 -0800
tree 81835d8eaaa82ec06a122e1a480ac3444f6350f0
parent cf67ec6d79a412063e413a93720e37a10fd46d3f

Accuracy verification test for EXPM1MINUS evaluation stubs

PiperOrigin-RevId: 347436156

eval/f32-expm1minus.cc

diff --git a/eval/f32-expm1minus.cc b/eval/f32-expm1minus.cc
new file mode 100644
index 0000000..b2746c5
--- /dev/null
+++ b/eval/f32-expm1minus.cc
@@ -0,0 +1,1805 @@
+// 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 <algorithm>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <iomanip>
+#include <ios>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include <fp16.h>
+
+#include <xnnpack/AlignedAllocator.h>
+#include <xnnpack/common.h>
+#include <xnnpack/isa-checks.h>
+#include <xnnpack/math-stubs.h>
+
+
+constexpr int kBlockSize = 1024;
+
+
+#if XNN_ARCH_ARM || XNN_ARCH_ARM64
+  TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, negative_zero) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, negative_saturation) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, positive_nan) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, negative_nan) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM || XNN_ARCH_ARM64
+  TEST(EXPM1MINUS__NEON_RR2_P6, negative_zero) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__NEON_RR2_P6, negative_saturation) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEON_RR2_P6, positive_nan) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEON_RR2_P6, negative_nan) {
+    TEST_REQUIRES_ARM_NEON;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM || XNN_ARCH_ARM64
+  TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, negative_zero) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, negative_saturation) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, positive_nan) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, negative_nan) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM || XNN_ARCH_ARM64
+  TEST(EXPM1MINUS__NEONFMA_RR1_P6, negative_zero) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__NEONFMA_RR1_P6, negative_saturation) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEONFMA_RR1_P6, positive_nan) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__NEONFMA_RR1_P6, negative_nan) {
+    TEST_REQUIRES_ARM_NEON_FMA;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, negative_zero) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, negative_saturation) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, positive_nan) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, negative_nan) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX512F_RR1_P6, negative_zero) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX512F_RR1_P6, negative_saturation) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX512F_RR1_P6, positive_nan) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX512F_RR1_P6, negative_nan) {
+    TEST_REQUIRES_X86_AVX512F;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, negative_zero) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, negative_saturation) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, positive_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, negative_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, negative_zero) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, negative_saturation) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, positive_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, negative_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, negative_zero) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, negative_saturation) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, positive_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, negative_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX2_RR1_P6, negative_zero) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_P6, negative_saturation) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_P6, positive_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX2_RR1_P6, negative_nan) {
+    TEST_REQUIRES_X86_AVX2;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, negative_zero) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, negative_saturation) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, positive_nan) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, negative_nan) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, negative_zero) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, negative_saturation) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, positive_nan) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, negative_nan) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__AVX_RR2_P6, negative_zero) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_P6, negative_saturation) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_P6, positive_nan) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__AVX_RR2_P6, negative_nan) {
+    TEST_REQUIRES_X86_AVX;
+
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, negative_zero) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, negative_saturation) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, positive_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, negative_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  TEST(EXPM1MINUS__SSE2_RR2_P6, negative_zero) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__SSE2_RR2_P6, negative_saturation) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__SSE2_RR2_P6, positive_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__SSE2_RR2_P6, negative_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+
+#if XNN_ARCH_WASMSIMD
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, negative_zero) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, negative_saturation) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, positive_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, negative_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_WASMSIMD
+
+
+#if XNN_ARCH_WASMSIMD
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, negative_zero) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, negative_saturation) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, positive_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, negative_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_WASMSIMD
+
+
+#if XNN_ARCH_WASMSIMD
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, negative_zero) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, negative_saturation) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, positive_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, negative_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_WASMSIMD
+
+
+#if XNN_ARCH_WASMSIMD
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, negative_zero) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    std::fill(inputs.begin(), inputs.end(), -0.0f);
+    xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    const float reference_output = 0.0f;
+    ASSERT_EQ(reference_output, outputs[0])
+      << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+      << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+      << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, negative_saturation) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        const float reference_output = -1.0f;
+        ASSERT_EQ(reference_output, outputs[i])
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, positive_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+
+  TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, negative_nan) {
+    std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+    std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+    for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+      }
+      xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+      for (uint32_t i = 0; i < kBlockSize; i++) {
+        ASSERT_TRUE(std::isnan(outputs[i]))
+          << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+          << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+      }
+    }
+  }
+#endif  // XNN_ARCH_WASMSIMD
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_P5, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_P5, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_P5, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_P5, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+
+TEST(EXPM1MINUS__SCALAR_RR2_P6, negative_zero) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  std::fill(inputs.begin(), inputs.end(), -0.0f);
+  xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+  const float reference_output = 0.0f;
+  ASSERT_EQ(reference_output, outputs[0])
+    << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0])
+    << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+    << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]);
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_P6, negative_saturation) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      const float reference_output = -1.0f;
+      ASSERT_EQ(reference_output, outputs[i])
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output)
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_P6, positive_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}
+
+TEST(EXPM1MINUS__SCALAR_RR2_P6, negative_nan) {
+  std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize);
+  std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize);
+  for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) {
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i)));
+    }
+    xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data());
+    for (uint32_t i = 0; i < kBlockSize; i++) {
+      ASSERT_TRUE(std::isnan(outputs[i]))
+        << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i])
+        << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]);
+    }
+  }
+}