XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 5 | // |
| 6 | // Auto-generated file. Do not edit! |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 7 | // Specification: test/f32-vhswish.yaml |
Marat Dukhan | 949b6e7 | 2021-05-13 11:21:06 -0700 | [diff] [blame] | 8 | // Generator: tools/generate-vunary-test.py |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 9 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 10 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 11 | #include <gtest/gtest.h> |
| 12 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 13 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <xnnpack/isa-checks.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 15 | |
Marat Dukhan | a91559a | 2021-05-05 23:58:21 -0700 | [diff] [blame] | 16 | #include <xnnpack/vunary.h> |
Marat Dukhan | 949b6e7 | 2021-05-13 11:21:06 -0700 | [diff] [blame] | 17 | #include "vunary-microkernel-tester.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 18 | |
| 19 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 20 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 21 | TEST(F32_VHSWISH__NEON_X4, batch_eq_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 23 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 24 | .batch_size(4) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 25 | .Test(xnn_f32_vhswish_ukernel__neon_x4, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | } |
| 27 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 28 | TEST(F32_VHSWISH__NEON_X4, batch_div_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 30 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 31 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 32 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 33 | .Test(xnn_f32_vhswish_ukernel__neon_x4, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 34 | } |
| 35 | } |
| 36 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 37 | TEST(F32_VHSWISH__NEON_X4, batch_lt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 38 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 39 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 40 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 41 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 42 | .Test(xnn_f32_vhswish_ukernel__neon_x4, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 43 | } |
| 44 | } |
| 45 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 46 | TEST(F32_VHSWISH__NEON_X4, batch_gt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 48 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 49 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 50 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 51 | .Test(xnn_f32_vhswish_ukernel__neon_x4, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 52 | } |
| 53 | } |
| 54 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 55 | TEST(F32_VHSWISH__NEON_X4, inplace) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 56 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 57 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 58 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 59 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 60 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 61 | .Test(xnn_f32_vhswish_ukernel__neon_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 62 | } |
| 63 | } |
| 64 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 65 | |
| 66 | |
| 67 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 68 | TEST(F32_VHSWISH__NEON_X8, batch_eq_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 69 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 70 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 71 | .batch_size(8) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 72 | .Test(xnn_f32_vhswish_ukernel__neon_x8, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 73 | } |
| 74 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 75 | TEST(F32_VHSWISH__NEON_X8, batch_div_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 76 | TEST_REQUIRES_ARM_NEON; |
| 77 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 78 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 79 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 80 | .Test(xnn_f32_vhswish_ukernel__neon_x8, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | } |
| 82 | } |
| 83 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 84 | TEST(F32_VHSWISH__NEON_X8, batch_lt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 85 | TEST_REQUIRES_ARM_NEON; |
| 86 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 87 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 88 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 89 | .Test(xnn_f32_vhswish_ukernel__neon_x8, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 90 | } |
| 91 | } |
| 92 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 93 | TEST(F32_VHSWISH__NEON_X8, batch_gt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 94 | TEST_REQUIRES_ARM_NEON; |
| 95 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 96 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 97 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 98 | .Test(xnn_f32_vhswish_ukernel__neon_x8, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 99 | } |
| 100 | } |
| 101 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 102 | TEST(F32_VHSWISH__NEON_X8, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 103 | TEST_REQUIRES_ARM_NEON; |
| 104 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 105 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 106 | .batch_size(batch_size) |
| 107 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 108 | .Test(xnn_f32_vhswish_ukernel__neon_x8, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 109 | } |
| 110 | } |
| 111 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 112 | |
| 113 | |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 114 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 115 | TEST(F32_VHSWISH__NEON_X16, batch_eq_16) { |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 116 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 117 | VUnaryMicrokernelTester() |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 118 | .batch_size(16) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 119 | .Test(xnn_f32_vhswish_ukernel__neon_x16, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 120 | } |
| 121 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 122 | TEST(F32_VHSWISH__NEON_X16, batch_div_16) { |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 123 | TEST_REQUIRES_ARM_NEON; |
| 124 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 125 | VUnaryMicrokernelTester() |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 126 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 127 | .Test(xnn_f32_vhswish_ukernel__neon_x16, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 128 | } |
| 129 | } |
| 130 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 131 | TEST(F32_VHSWISH__NEON_X16, batch_lt_16) { |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 132 | TEST_REQUIRES_ARM_NEON; |
| 133 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 134 | VUnaryMicrokernelTester() |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 135 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 136 | .Test(xnn_f32_vhswish_ukernel__neon_x16, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 137 | } |
| 138 | } |
| 139 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 140 | TEST(F32_VHSWISH__NEON_X16, batch_gt_16) { |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 141 | TEST_REQUIRES_ARM_NEON; |
| 142 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 143 | VUnaryMicrokernelTester() |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 144 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 145 | .Test(xnn_f32_vhswish_ukernel__neon_x16, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 146 | } |
| 147 | } |
| 148 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 149 | TEST(F32_VHSWISH__NEON_X16, inplace) { |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 150 | TEST_REQUIRES_ARM_NEON; |
| 151 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 152 | VUnaryMicrokernelTester() |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 153 | .batch_size(batch_size) |
| 154 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 155 | .Test(xnn_f32_vhswish_ukernel__neon_x16, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 156 | } |
| 157 | } |
| 158 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 159 | |
| 160 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 161 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 162 | TEST(F32_VHSWISH__SSE_X4, batch_eq_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 163 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 164 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 165 | .batch_size(4) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 166 | .Test(xnn_f32_vhswish_ukernel__sse_x4, xnn_init_f32_hswish_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 167 | } |
| 168 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 169 | TEST(F32_VHSWISH__SSE_X4, batch_div_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 170 | TEST_REQUIRES_X86_SSE; |
| 171 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 172 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 173 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 174 | .Test(xnn_f32_vhswish_ukernel__sse_x4, xnn_init_f32_hswish_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 175 | } |
| 176 | } |
| 177 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 178 | TEST(F32_VHSWISH__SSE_X4, batch_lt_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 179 | TEST_REQUIRES_X86_SSE; |
| 180 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 181 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 182 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 183 | .Test(xnn_f32_vhswish_ukernel__sse_x4, xnn_init_f32_hswish_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 184 | } |
| 185 | } |
| 186 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 187 | TEST(F32_VHSWISH__SSE_X4, batch_gt_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 188 | TEST_REQUIRES_X86_SSE; |
| 189 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 190 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 191 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 192 | .Test(xnn_f32_vhswish_ukernel__sse_x4, xnn_init_f32_hswish_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 193 | } |
| 194 | } |
| 195 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 196 | TEST(F32_VHSWISH__SSE_X4, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 197 | TEST_REQUIRES_X86_SSE; |
| 198 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 199 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 200 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 201 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 202 | .Test(xnn_f32_vhswish_ukernel__sse_x4, xnn_init_f32_hswish_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 203 | } |
| 204 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 205 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 206 | |
| 207 | |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 208 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 209 | TEST(F32_VHSWISH__SSE_X8, batch_eq_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 210 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 211 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 212 | .batch_size(8) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 213 | .Test(xnn_f32_vhswish_ukernel__sse_x8, xnn_init_f32_hswish_sse_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 214 | } |
| 215 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 216 | TEST(F32_VHSWISH__SSE_X8, batch_div_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 217 | TEST_REQUIRES_X86_SSE; |
| 218 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 219 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 220 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 221 | .Test(xnn_f32_vhswish_ukernel__sse_x8, xnn_init_f32_hswish_sse_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 222 | } |
| 223 | } |
| 224 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 225 | TEST(F32_VHSWISH__SSE_X8, batch_lt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 226 | TEST_REQUIRES_X86_SSE; |
| 227 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 228 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 229 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 230 | .Test(xnn_f32_vhswish_ukernel__sse_x8, xnn_init_f32_hswish_sse_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 231 | } |
| 232 | } |
| 233 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 234 | TEST(F32_VHSWISH__SSE_X8, batch_gt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 235 | TEST_REQUIRES_X86_SSE; |
| 236 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 237 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 238 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 239 | .Test(xnn_f32_vhswish_ukernel__sse_x8, xnn_init_f32_hswish_sse_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 240 | } |
| 241 | } |
| 242 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 243 | TEST(F32_VHSWISH__SSE_X8, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 244 | TEST_REQUIRES_X86_SSE; |
| 245 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 246 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 247 | .batch_size(batch_size) |
| 248 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 249 | .Test(xnn_f32_vhswish_ukernel__sse_x8, xnn_init_f32_hswish_sse_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 250 | } |
| 251 | } |
| 252 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 253 | |
| 254 | |
| 255 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 256 | TEST(F32_VHSWISH__AVX_X8, batch_eq_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 257 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 258 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 259 | .batch_size(8) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 260 | .Test(xnn_f32_vhswish_ukernel__avx_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 261 | } |
| 262 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 263 | TEST(F32_VHSWISH__AVX_X8, batch_div_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 264 | TEST_REQUIRES_X86_AVX; |
| 265 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 266 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 267 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 268 | .Test(xnn_f32_vhswish_ukernel__avx_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 269 | } |
| 270 | } |
| 271 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 272 | TEST(F32_VHSWISH__AVX_X8, batch_lt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 273 | TEST_REQUIRES_X86_AVX; |
| 274 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 275 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 276 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 277 | .Test(xnn_f32_vhswish_ukernel__avx_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 278 | } |
| 279 | } |
| 280 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 281 | TEST(F32_VHSWISH__AVX_X8, batch_gt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 282 | TEST_REQUIRES_X86_AVX; |
| 283 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 284 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 285 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 286 | .Test(xnn_f32_vhswish_ukernel__avx_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 287 | } |
| 288 | } |
| 289 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 290 | TEST(F32_VHSWISH__AVX_X8, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 291 | TEST_REQUIRES_X86_AVX; |
| 292 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 293 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 294 | .batch_size(batch_size) |
| 295 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 296 | .Test(xnn_f32_vhswish_ukernel__avx_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 297 | } |
| 298 | } |
| 299 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 300 | |
| 301 | |
| 302 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 303 | TEST(F32_VHSWISH__AVX_X16, batch_eq_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 304 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 305 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 306 | .batch_size(16) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 307 | .Test(xnn_f32_vhswish_ukernel__avx_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 308 | } |
| 309 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 310 | TEST(F32_VHSWISH__AVX_X16, batch_div_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 311 | TEST_REQUIRES_X86_AVX; |
| 312 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 313 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 314 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 315 | .Test(xnn_f32_vhswish_ukernel__avx_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 316 | } |
| 317 | } |
| 318 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 319 | TEST(F32_VHSWISH__AVX_X16, batch_lt_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 320 | TEST_REQUIRES_X86_AVX; |
| 321 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 322 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 323 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 324 | .Test(xnn_f32_vhswish_ukernel__avx_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 325 | } |
| 326 | } |
| 327 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 328 | TEST(F32_VHSWISH__AVX_X16, batch_gt_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 329 | TEST_REQUIRES_X86_AVX; |
| 330 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 331 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 332 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 333 | .Test(xnn_f32_vhswish_ukernel__avx_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 334 | } |
| 335 | } |
| 336 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 337 | TEST(F32_VHSWISH__AVX_X16, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 338 | TEST_REQUIRES_X86_AVX; |
| 339 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 340 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 341 | .batch_size(batch_size) |
| 342 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 343 | .Test(xnn_f32_vhswish_ukernel__avx_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 344 | } |
| 345 | } |
| 346 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 347 | |
| 348 | |
| 349 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 350 | TEST(F32_VHSWISH__FMA3_X8, batch_eq_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 351 | TEST_REQUIRES_X86_FMA3; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 352 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 353 | .batch_size(8) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 354 | .Test(xnn_f32_vhswish_ukernel__fma3_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 355 | } |
| 356 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 357 | TEST(F32_VHSWISH__FMA3_X8, batch_div_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 358 | TEST_REQUIRES_X86_FMA3; |
| 359 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 360 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 361 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 362 | .Test(xnn_f32_vhswish_ukernel__fma3_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 363 | } |
| 364 | } |
| 365 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 366 | TEST(F32_VHSWISH__FMA3_X8, batch_lt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 367 | TEST_REQUIRES_X86_FMA3; |
| 368 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 369 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 370 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 371 | .Test(xnn_f32_vhswish_ukernel__fma3_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 372 | } |
| 373 | } |
| 374 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 375 | TEST(F32_VHSWISH__FMA3_X8, batch_gt_8) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 376 | TEST_REQUIRES_X86_FMA3; |
| 377 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 378 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 379 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 380 | .Test(xnn_f32_vhswish_ukernel__fma3_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 381 | } |
| 382 | } |
| 383 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 384 | TEST(F32_VHSWISH__FMA3_X8, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 385 | TEST_REQUIRES_X86_FMA3; |
| 386 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 387 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 388 | .batch_size(batch_size) |
| 389 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 390 | .Test(xnn_f32_vhswish_ukernel__fma3_x8, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 391 | } |
| 392 | } |
| 393 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 394 | |
| 395 | |
| 396 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 397 | TEST(F32_VHSWISH__FMA3_X16, batch_eq_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 398 | TEST_REQUIRES_X86_FMA3; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 399 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 400 | .batch_size(16) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 401 | .Test(xnn_f32_vhswish_ukernel__fma3_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 402 | } |
| 403 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 404 | TEST(F32_VHSWISH__FMA3_X16, batch_div_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 405 | TEST_REQUIRES_X86_FMA3; |
| 406 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 407 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 408 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 409 | .Test(xnn_f32_vhswish_ukernel__fma3_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 410 | } |
| 411 | } |
| 412 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 413 | TEST(F32_VHSWISH__FMA3_X16, batch_lt_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 414 | TEST_REQUIRES_X86_FMA3; |
| 415 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 416 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 417 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 418 | .Test(xnn_f32_vhswish_ukernel__fma3_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 419 | } |
| 420 | } |
| 421 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 422 | TEST(F32_VHSWISH__FMA3_X16, batch_gt_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 423 | TEST_REQUIRES_X86_FMA3; |
| 424 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 425 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 426 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 427 | .Test(xnn_f32_vhswish_ukernel__fma3_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 428 | } |
| 429 | } |
| 430 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 431 | TEST(F32_VHSWISH__FMA3_X16, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 432 | TEST_REQUIRES_X86_FMA3; |
| 433 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 434 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 435 | .batch_size(batch_size) |
| 436 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 437 | .Test(xnn_f32_vhswish_ukernel__fma3_x16, xnn_init_f32_hswish_avx_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 438 | } |
| 439 | } |
| 440 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 441 | |
| 442 | |
| 443 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 444 | TEST(F32_VHSWISH__AVX512F_X16, batch_eq_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 445 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 446 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 447 | .batch_size(16) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 448 | .Test(xnn_f32_vhswish_ukernel__avx512f_x16, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 449 | } |
| 450 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 451 | TEST(F32_VHSWISH__AVX512F_X16, batch_div_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 452 | TEST_REQUIRES_X86_AVX512F; |
| 453 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 454 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 455 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 456 | .Test(xnn_f32_vhswish_ukernel__avx512f_x16, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 457 | } |
| 458 | } |
| 459 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 460 | TEST(F32_VHSWISH__AVX512F_X16, batch_lt_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 461 | TEST_REQUIRES_X86_AVX512F; |
| 462 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 463 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 464 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 465 | .Test(xnn_f32_vhswish_ukernel__avx512f_x16, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 466 | } |
| 467 | } |
| 468 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 469 | TEST(F32_VHSWISH__AVX512F_X16, batch_gt_16) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 470 | TEST_REQUIRES_X86_AVX512F; |
| 471 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 472 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 473 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 474 | .Test(xnn_f32_vhswish_ukernel__avx512f_x16, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 475 | } |
| 476 | } |
| 477 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 478 | TEST(F32_VHSWISH__AVX512F_X16, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 479 | TEST_REQUIRES_X86_AVX512F; |
| 480 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 481 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 482 | .batch_size(batch_size) |
| 483 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 484 | .Test(xnn_f32_vhswish_ukernel__avx512f_x16, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 485 | } |
| 486 | } |
| 487 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 488 | |
| 489 | |
| 490 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 491 | TEST(F32_VHSWISH__AVX512F_X32, batch_eq_32) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 492 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 493 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 494 | .batch_size(32) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 495 | .Test(xnn_f32_vhswish_ukernel__avx512f_x32, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 496 | } |
| 497 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 498 | TEST(F32_VHSWISH__AVX512F_X32, batch_div_32) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 499 | TEST_REQUIRES_X86_AVX512F; |
| 500 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 501 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 502 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 503 | .Test(xnn_f32_vhswish_ukernel__avx512f_x32, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 504 | } |
| 505 | } |
| 506 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 507 | TEST(F32_VHSWISH__AVX512F_X32, batch_lt_32) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 508 | TEST_REQUIRES_X86_AVX512F; |
| 509 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 510 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 511 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 512 | .Test(xnn_f32_vhswish_ukernel__avx512f_x32, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 513 | } |
| 514 | } |
| 515 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 516 | TEST(F32_VHSWISH__AVX512F_X32, batch_gt_32) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 517 | TEST_REQUIRES_X86_AVX512F; |
| 518 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 519 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 520 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 521 | .Test(xnn_f32_vhswish_ukernel__avx512f_x32, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 522 | } |
| 523 | } |
| 524 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 525 | TEST(F32_VHSWISH__AVX512F_X32, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 526 | TEST_REQUIRES_X86_AVX512F; |
| 527 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 528 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 529 | .batch_size(batch_size) |
| 530 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 531 | .Test(xnn_f32_vhswish_ukernel__avx512f_x32, xnn_init_f32_hswish_avx512_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 532 | } |
| 533 | } |
| 534 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 535 | |
| 536 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 537 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 538 | TEST(F32_VHSWISH__WASMSIMD_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 539 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 540 | .batch_size(4) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 541 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x4, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 542 | } |
| 543 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 544 | TEST(F32_VHSWISH__WASMSIMD_X4, batch_div_4) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 545 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 546 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 547 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 548 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x4, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 549 | } |
| 550 | } |
| 551 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 552 | TEST(F32_VHSWISH__WASMSIMD_X4, batch_lt_4) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 553 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 554 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 555 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 556 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x4, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 557 | } |
| 558 | } |
| 559 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 560 | TEST(F32_VHSWISH__WASMSIMD_X4, batch_gt_4) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 561 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 562 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 563 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 564 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x4, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 565 | } |
| 566 | } |
| 567 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 568 | TEST(F32_VHSWISH__WASMSIMD_X4, inplace) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 569 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 570 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 571 | .batch_size(batch_size) |
| 572 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 573 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x4, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 574 | } |
| 575 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 576 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 577 | |
| 578 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 579 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 580 | TEST(F32_VHSWISH__WASMSIMD_X8, batch_eq_8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 581 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 582 | .batch_size(8) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 583 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x8, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 584 | } |
| 585 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 586 | TEST(F32_VHSWISH__WASMSIMD_X8, batch_div_8) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 587 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 588 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 589 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 590 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x8, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 591 | } |
| 592 | } |
| 593 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 594 | TEST(F32_VHSWISH__WASMSIMD_X8, batch_lt_8) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 595 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 596 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 597 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 598 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x8, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 599 | } |
| 600 | } |
| 601 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 602 | TEST(F32_VHSWISH__WASMSIMD_X8, batch_gt_8) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 603 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 604 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 605 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 606 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x8, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 607 | } |
| 608 | } |
| 609 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 610 | TEST(F32_VHSWISH__WASMSIMD_X8, inplace) { |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 611 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 612 | VUnaryMicrokernelTester() |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 613 | .batch_size(batch_size) |
| 614 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 615 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x8, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 616 | } |
| 617 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 618 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 619 | |
| 620 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 621 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 622 | TEST(F32_VHSWISH__WASMSIMD_X16, batch_eq_16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 623 | VUnaryMicrokernelTester() |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 624 | .batch_size(16) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 625 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x16, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 626 | } |
| 627 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 628 | TEST(F32_VHSWISH__WASMSIMD_X16, batch_div_16) { |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 629 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 630 | VUnaryMicrokernelTester() |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 631 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 632 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x16, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 633 | } |
| 634 | } |
| 635 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 636 | TEST(F32_VHSWISH__WASMSIMD_X16, batch_lt_16) { |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 637 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 638 | VUnaryMicrokernelTester() |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 639 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 640 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x16, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 641 | } |
| 642 | } |
| 643 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 644 | TEST(F32_VHSWISH__WASMSIMD_X16, batch_gt_16) { |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 645 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 646 | VUnaryMicrokernelTester() |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 647 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 648 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x16, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 649 | } |
| 650 | } |
| 651 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 652 | TEST(F32_VHSWISH__WASMSIMD_X16, inplace) { |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 653 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 654 | VUnaryMicrokernelTester() |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 655 | .batch_size(batch_size) |
| 656 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 657 | .Test(xnn_f32_vhswish_ukernel__wasmsimd_x16, xnn_init_f32_hswish_wasmsimd_params); |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 658 | } |
| 659 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 660 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 661 | |
| 662 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 663 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 664 | TEST(F32_VHSWISH__WASM_X1, batch_eq_1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 665 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 666 | .batch_size(1) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 667 | .Test(xnn_f32_vhswish_ukernel__wasm_x1, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 668 | } |
| 669 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 670 | TEST(F32_VHSWISH__WASM_X1, batch_gt_1) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 671 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 672 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 673 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 674 | .Test(xnn_f32_vhswish_ukernel__wasm_x1, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 675 | } |
| 676 | } |
| 677 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 678 | TEST(F32_VHSWISH__WASM_X1, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 679 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 680 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 681 | .batch_size(batch_size) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 682 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 683 | .Test(xnn_f32_vhswish_ukernel__wasm_x1, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 684 | } |
| 685 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 686 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 687 | |
| 688 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 689 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 690 | TEST(F32_VHSWISH__WASM_X2, batch_eq_2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 691 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 692 | .batch_size(2) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 693 | .Test(xnn_f32_vhswish_ukernel__wasm_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 694 | } |
| 695 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 696 | TEST(F32_VHSWISH__WASM_X2, batch_div_2) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 697 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 698 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 699 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 700 | .Test(xnn_f32_vhswish_ukernel__wasm_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 701 | } |
| 702 | } |
| 703 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 704 | TEST(F32_VHSWISH__WASM_X2, batch_lt_2) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 705 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 706 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 707 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 708 | .Test(xnn_f32_vhswish_ukernel__wasm_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 709 | } |
| 710 | } |
| 711 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 712 | TEST(F32_VHSWISH__WASM_X2, batch_gt_2) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 713 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 714 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 715 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 716 | .Test(xnn_f32_vhswish_ukernel__wasm_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 717 | } |
| 718 | } |
| 719 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 720 | TEST(F32_VHSWISH__WASM_X2, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 721 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 722 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 723 | .batch_size(batch_size) |
| 724 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 725 | .Test(xnn_f32_vhswish_ukernel__wasm_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 726 | } |
| 727 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 728 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 729 | |
| 730 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 731 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 732 | TEST(F32_VHSWISH__WASM_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 733 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 734 | .batch_size(4) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 735 | .Test(xnn_f32_vhswish_ukernel__wasm_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 736 | } |
| 737 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 738 | TEST(F32_VHSWISH__WASM_X4, batch_div_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 739 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 740 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 741 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 742 | .Test(xnn_f32_vhswish_ukernel__wasm_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 743 | } |
| 744 | } |
| 745 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 746 | TEST(F32_VHSWISH__WASM_X4, batch_lt_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 747 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 748 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 749 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 750 | .Test(xnn_f32_vhswish_ukernel__wasm_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 751 | } |
| 752 | } |
| 753 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 754 | TEST(F32_VHSWISH__WASM_X4, batch_gt_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 755 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 756 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 757 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 758 | .Test(xnn_f32_vhswish_ukernel__wasm_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 759 | } |
| 760 | } |
| 761 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 762 | TEST(F32_VHSWISH__WASM_X4, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 763 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 764 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 765 | .batch_size(batch_size) |
| 766 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 767 | .Test(xnn_f32_vhswish_ukernel__wasm_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 768 | } |
| 769 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 770 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 771 | |
| 772 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 773 | TEST(F32_VHSWISH__SCALAR_X1, batch_eq_1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 774 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 775 | .batch_size(1) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 776 | .Test(xnn_f32_vhswish_ukernel__scalar_x1, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 777 | } |
| 778 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 779 | TEST(F32_VHSWISH__SCALAR_X1, batch_gt_1) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 780 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 781 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 782 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 783 | .Test(xnn_f32_vhswish_ukernel__scalar_x1, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 784 | } |
| 785 | } |
| 786 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 787 | TEST(F32_VHSWISH__SCALAR_X1, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 788 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 789 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 790 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 791 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 792 | .Test(xnn_f32_vhswish_ukernel__scalar_x1, xnn_init_f32_hswish_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 793 | } |
| 794 | } |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 795 | |
Marat Dukhan | 949b6e7 | 2021-05-13 11:21:06 -0700 | [diff] [blame] | 796 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 797 | TEST(F32_VHSWISH__SCALAR_X2, batch_eq_2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 798 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 799 | .batch_size(2) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 800 | .Test(xnn_f32_vhswish_ukernel__scalar_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 801 | } |
| 802 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 803 | TEST(F32_VHSWISH__SCALAR_X2, batch_div_2) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 804 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 805 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 806 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 807 | .Test(xnn_f32_vhswish_ukernel__scalar_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 808 | } |
| 809 | } |
| 810 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 811 | TEST(F32_VHSWISH__SCALAR_X2, batch_lt_2) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 812 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 813 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 814 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 815 | .Test(xnn_f32_vhswish_ukernel__scalar_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 816 | } |
| 817 | } |
| 818 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 819 | TEST(F32_VHSWISH__SCALAR_X2, batch_gt_2) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 820 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 821 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 822 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 823 | .Test(xnn_f32_vhswish_ukernel__scalar_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 824 | } |
| 825 | } |
| 826 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 827 | TEST(F32_VHSWISH__SCALAR_X2, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 828 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 829 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 830 | .batch_size(batch_size) |
| 831 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 832 | .Test(xnn_f32_vhswish_ukernel__scalar_x2, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 833 | } |
| 834 | } |
| 835 | |
Marat Dukhan | 949b6e7 | 2021-05-13 11:21:06 -0700 | [diff] [blame] | 836 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 837 | TEST(F32_VHSWISH__SCALAR_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 838 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 839 | .batch_size(4) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 840 | .Test(xnn_f32_vhswish_ukernel__scalar_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 841 | } |
| 842 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 843 | TEST(F32_VHSWISH__SCALAR_X4, batch_div_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 844 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 845 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 846 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 847 | .Test(xnn_f32_vhswish_ukernel__scalar_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 848 | } |
| 849 | } |
| 850 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 851 | TEST(F32_VHSWISH__SCALAR_X4, batch_lt_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 852 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 853 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 854 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 855 | .Test(xnn_f32_vhswish_ukernel__scalar_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 856 | } |
| 857 | } |
| 858 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 859 | TEST(F32_VHSWISH__SCALAR_X4, batch_gt_4) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 860 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 861 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 862 | .batch_size(batch_size) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 863 | .Test(xnn_f32_vhswish_ukernel__scalar_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 864 | } |
| 865 | } |
| 866 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 867 | TEST(F32_VHSWISH__SCALAR_X4, inplace) { |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 868 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 869 | VUnaryMicrokernelTester() |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 870 | .batch_size(batch_size) |
| 871 | .inplace(true) |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 872 | .Test(xnn_f32_vhswish_ukernel__scalar_x4, xnn_init_f32_hswish_scalar_params); |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 873 | } |
Marat Dukhan | 949b6e7 | 2021-05-13 11:21:06 -0700 | [diff] [blame] | 874 | } |