Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 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 | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [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-vclamp.yaml |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 8 | // Generator: tools/generate-vunary-test.py |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [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 | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 16 | #include <xnnpack/vunary.h> |
| 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_VCLAMP__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 | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 24 | .batch_size(4) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 25 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
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_VCLAMP__NEON_X4, batch_div_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [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 | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 32 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 33 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
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_VCLAMP__NEON_X4, batch_lt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 38 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [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 | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 41 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 42 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
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_VCLAMP__NEON_X4, batch_gt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [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 | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 50 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 51 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
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_VCLAMP__NEON_X4, inplace) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 56 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [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 | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 59 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 60 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 61 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 62 | } |
| 63 | } |
| 64 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 65 | TEST(F32_VCLAMP__NEON_X4, qmin) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 66 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 67 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 68 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 69 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 70 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 71 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 72 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 73 | } |
| 74 | } |
| 75 | } |
| 76 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 77 | TEST(F32_VCLAMP__NEON_X4, qmax) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 78 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 79 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 80 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 81 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 82 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 83 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 84 | .Test(xnn_f32_vclamp_ukernel__neon_x4, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 85 | } |
| 86 | } |
| 87 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 88 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 89 | |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 90 | |
| 91 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 92 | TEST(F32_VCLAMP__NEON_X8, batch_eq_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 93 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 94 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 95 | .batch_size(8) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 96 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 97 | } |
| 98 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 99 | TEST(F32_VCLAMP__NEON_X8, batch_div_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 100 | TEST_REQUIRES_ARM_NEON; |
| 101 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 102 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 103 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 104 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 105 | } |
| 106 | } |
| 107 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 108 | TEST(F32_VCLAMP__NEON_X8, batch_lt_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 109 | TEST_REQUIRES_ARM_NEON; |
| 110 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 111 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 112 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 113 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 114 | } |
| 115 | } |
| 116 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 117 | TEST(F32_VCLAMP__NEON_X8, batch_gt_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 118 | TEST_REQUIRES_ARM_NEON; |
| 119 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 120 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 121 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 122 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 123 | } |
| 124 | } |
| 125 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 126 | TEST(F32_VCLAMP__NEON_X8, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 127 | TEST_REQUIRES_ARM_NEON; |
| 128 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 129 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 130 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 131 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 132 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 133 | } |
| 134 | } |
| 135 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 136 | TEST(F32_VCLAMP__NEON_X8, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 137 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 138 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 139 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 140 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 141 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 142 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 143 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 144 | } |
| 145 | } |
| 146 | } |
| 147 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 148 | TEST(F32_VCLAMP__NEON_X8, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 149 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 150 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 151 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 152 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 153 | .batch_size(batch_size) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 154 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 155 | .Test(xnn_f32_vclamp_ukernel__neon_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 156 | } |
| 157 | } |
| 158 | } |
| 159 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 160 | |
| 161 | |
| 162 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 163 | TEST(F32_VCLAMP__SSE_X4, batch_eq_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 164 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 165 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 166 | .batch_size(4) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 167 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 168 | } |
| 169 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 170 | TEST(F32_VCLAMP__SSE_X4, batch_div_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 171 | TEST_REQUIRES_X86_SSE; |
| 172 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 173 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 174 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 175 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 176 | } |
| 177 | } |
| 178 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 179 | TEST(F32_VCLAMP__SSE_X4, batch_lt_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 180 | TEST_REQUIRES_X86_SSE; |
| 181 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 182 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 183 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 184 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 185 | } |
| 186 | } |
| 187 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 188 | TEST(F32_VCLAMP__SSE_X4, batch_gt_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 189 | TEST_REQUIRES_X86_SSE; |
| 190 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 191 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 192 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 193 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 194 | } |
| 195 | } |
| 196 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 197 | TEST(F32_VCLAMP__SSE_X4, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 198 | TEST_REQUIRES_X86_SSE; |
| 199 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 200 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 201 | .batch_size(batch_size) |
| 202 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 203 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 204 | } |
| 205 | } |
| 206 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 207 | TEST(F32_VCLAMP__SSE_X4, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 208 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 209 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 210 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 211 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 212 | .batch_size(batch_size) |
| 213 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 214 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 219 | TEST(F32_VCLAMP__SSE_X4, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 220 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 221 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 222 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 223 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 224 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 225 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 226 | .Test(xnn_f32_vclamp_ukernel__sse_x4, VUnaryMicrokernelTester::OpType::Clamp); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 227 | } |
| 228 | } |
| 229 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 230 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 231 | |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 232 | |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 233 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 234 | TEST(F32_VCLAMP__SSE_X8, batch_eq_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 235 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 236 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 237 | .batch_size(8) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 238 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 239 | } |
| 240 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 241 | TEST(F32_VCLAMP__SSE_X8, batch_div_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 242 | TEST_REQUIRES_X86_SSE; |
| 243 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 244 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 245 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 246 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 247 | } |
| 248 | } |
| 249 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 250 | TEST(F32_VCLAMP__SSE_X8, batch_lt_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 251 | TEST_REQUIRES_X86_SSE; |
| 252 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 253 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 254 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 255 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 256 | } |
| 257 | } |
| 258 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 259 | TEST(F32_VCLAMP__SSE_X8, batch_gt_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 260 | TEST_REQUIRES_X86_SSE; |
| 261 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 262 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 263 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 264 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 265 | } |
| 266 | } |
| 267 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 268 | TEST(F32_VCLAMP__SSE_X8, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 269 | TEST_REQUIRES_X86_SSE; |
| 270 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 271 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 272 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 273 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 274 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 275 | } |
| 276 | } |
| 277 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 278 | TEST(F32_VCLAMP__SSE_X8, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 279 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 280 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 281 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 282 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 283 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 284 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 285 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 286 | } |
| 287 | } |
| 288 | } |
| 289 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 290 | TEST(F32_VCLAMP__SSE_X8, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 291 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 292 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 293 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 294 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 295 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 296 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 297 | .Test(xnn_f32_vclamp_ukernel__sse_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 298 | } |
| 299 | } |
| 300 | } |
| 301 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 302 | |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 303 | |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 304 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 305 | TEST(F32_VCLAMP__AVX_X8, batch_eq_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 306 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 307 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 308 | .batch_size(8) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 309 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 310 | } |
| 311 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 312 | TEST(F32_VCLAMP__AVX_X8, batch_div_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 313 | TEST_REQUIRES_X86_AVX; |
| 314 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 315 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 316 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 317 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 318 | } |
| 319 | } |
| 320 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 321 | TEST(F32_VCLAMP__AVX_X8, batch_lt_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 322 | TEST_REQUIRES_X86_AVX; |
| 323 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 324 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 325 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 326 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 327 | } |
| 328 | } |
| 329 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 330 | TEST(F32_VCLAMP__AVX_X8, batch_gt_8) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 331 | TEST_REQUIRES_X86_AVX; |
| 332 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 333 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 334 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 335 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 336 | } |
| 337 | } |
| 338 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 339 | TEST(F32_VCLAMP__AVX_X8, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 340 | TEST_REQUIRES_X86_AVX; |
| 341 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 342 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 343 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 344 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 345 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 346 | } |
| 347 | } |
| 348 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 349 | TEST(F32_VCLAMP__AVX_X8, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 350 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 351 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 352 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 353 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 354 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 355 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 356 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 357 | } |
| 358 | } |
| 359 | } |
| 360 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 361 | TEST(F32_VCLAMP__AVX_X8, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 362 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 363 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 364 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 365 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 366 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 367 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 368 | .Test(xnn_f32_vclamp_ukernel__avx_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 369 | } |
| 370 | } |
| 371 | } |
| 372 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 373 | |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 374 | |
| 375 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 376 | TEST(F32_VCLAMP__AVX_X16, batch_eq_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 377 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 378 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 379 | .batch_size(16) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 380 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 381 | } |
| 382 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 383 | TEST(F32_VCLAMP__AVX_X16, batch_div_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 384 | TEST_REQUIRES_X86_AVX; |
| 385 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 386 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 387 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 388 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 389 | } |
| 390 | } |
| 391 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 392 | TEST(F32_VCLAMP__AVX_X16, batch_lt_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 393 | TEST_REQUIRES_X86_AVX; |
| 394 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 395 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 396 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 397 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 398 | } |
| 399 | } |
| 400 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 401 | TEST(F32_VCLAMP__AVX_X16, batch_gt_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 402 | TEST_REQUIRES_X86_AVX; |
| 403 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 404 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 405 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 406 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 407 | } |
| 408 | } |
| 409 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 410 | TEST(F32_VCLAMP__AVX_X16, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 411 | TEST_REQUIRES_X86_AVX; |
| 412 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 413 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 414 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 415 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 416 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 417 | } |
| 418 | } |
| 419 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 420 | TEST(F32_VCLAMP__AVX_X16, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 421 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 422 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 423 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 424 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 425 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 426 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 427 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 428 | } |
| 429 | } |
| 430 | } |
| 431 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 432 | TEST(F32_VCLAMP__AVX_X16, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 433 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 434 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 435 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 436 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 437 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 438 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 439 | .Test(xnn_f32_vclamp_ukernel__avx_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 440 | } |
| 441 | } |
| 442 | } |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 443 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 444 | |
| 445 | |
| 446 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 447 | TEST(F32_VCLAMP__AVX512F_X16, batch_eq_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 448 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 449 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 450 | .batch_size(16) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 451 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 452 | } |
| 453 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 454 | TEST(F32_VCLAMP__AVX512F_X16, batch_div_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 455 | TEST_REQUIRES_X86_AVX512F; |
| 456 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 457 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 458 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 459 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 460 | } |
| 461 | } |
| 462 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 463 | TEST(F32_VCLAMP__AVX512F_X16, batch_lt_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 464 | TEST_REQUIRES_X86_AVX512F; |
| 465 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 466 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 467 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 468 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 469 | } |
| 470 | } |
| 471 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 472 | TEST(F32_VCLAMP__AVX512F_X16, batch_gt_16) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 473 | TEST_REQUIRES_X86_AVX512F; |
| 474 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 475 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 476 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 477 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 478 | } |
| 479 | } |
| 480 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 481 | TEST(F32_VCLAMP__AVX512F_X16, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 482 | TEST_REQUIRES_X86_AVX512F; |
| 483 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 484 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 485 | .batch_size(batch_size) |
| 486 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 487 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 488 | } |
| 489 | } |
| 490 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 491 | TEST(F32_VCLAMP__AVX512F_X16, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 492 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 493 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 494 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 495 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 496 | .batch_size(batch_size) |
| 497 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 498 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 499 | } |
| 500 | } |
| 501 | } |
| 502 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 503 | TEST(F32_VCLAMP__AVX512F_X16, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 504 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 505 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 506 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 507 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 508 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 509 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 510 | .Test(xnn_f32_vclamp_ukernel__avx512f_x16, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 511 | } |
| 512 | } |
| 513 | } |
| 514 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 515 | |
| 516 | |
| 517 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 518 | TEST(F32_VCLAMP__AVX512F_X32, batch_eq_32) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 519 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 520 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 521 | .batch_size(32) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 522 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 523 | } |
| 524 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 525 | TEST(F32_VCLAMP__AVX512F_X32, batch_div_32) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 526 | TEST_REQUIRES_X86_AVX512F; |
| 527 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 528 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 529 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 530 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 531 | } |
| 532 | } |
| 533 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 534 | TEST(F32_VCLAMP__AVX512F_X32, batch_lt_32) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 535 | TEST_REQUIRES_X86_AVX512F; |
| 536 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 537 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 538 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 539 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 540 | } |
| 541 | } |
| 542 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 543 | TEST(F32_VCLAMP__AVX512F_X32, batch_gt_32) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 544 | TEST_REQUIRES_X86_AVX512F; |
| 545 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 546 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 547 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 548 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 549 | } |
| 550 | } |
| 551 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 552 | TEST(F32_VCLAMP__AVX512F_X32, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 553 | TEST_REQUIRES_X86_AVX512F; |
| 554 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 555 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 556 | .batch_size(batch_size) |
| 557 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 558 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 559 | } |
| 560 | } |
| 561 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 562 | TEST(F32_VCLAMP__AVX512F_X32, qmin) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 563 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 564 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 565 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 566 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 567 | .batch_size(batch_size) |
| 568 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 569 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 570 | } |
| 571 | } |
| 572 | } |
| 573 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 574 | TEST(F32_VCLAMP__AVX512F_X32, qmax) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 575 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 576 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 577 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 578 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 579 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 580 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 581 | .Test(xnn_f32_vclamp_ukernel__avx512f_x32, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 582 | } |
| 583 | } |
| 584 | } |
| 585 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 586 | |
| 587 | |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 588 | #if XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 589 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 590 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 591 | .batch_size(4) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 592 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 593 | } |
| 594 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 595 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, batch_div_4) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 596 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 597 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 598 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 599 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 600 | } |
| 601 | } |
| 602 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 603 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, batch_lt_4) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 604 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 605 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 606 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 607 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 608 | } |
| 609 | } |
| 610 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 611 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, batch_gt_4) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 612 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 613 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 614 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 615 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 616 | } |
| 617 | } |
| 618 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 619 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, inplace) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 620 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 621 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 622 | .batch_size(batch_size) |
| 623 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 624 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 625 | } |
| 626 | } |
| 627 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 628 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 629 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 630 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 631 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 632 | .batch_size(batch_size) |
| 633 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 634 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 635 | } |
| 636 | } |
| 637 | } |
| 638 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 639 | TEST(F32_VCLAMP__WASMSIMD_ARM_X4, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 640 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 641 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 642 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 643 | .batch_size(batch_size) |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 644 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 645 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 646 | } |
| 647 | } |
| 648 | } |
| 649 | #endif // XNN_ARCH_WASMSIMD |
| 650 | |
| 651 | |
| 652 | #if XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 653 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, batch_eq_8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 654 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 655 | .batch_size(8) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 656 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 657 | } |
| 658 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 659 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, batch_div_8) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 660 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 661 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 662 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 663 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 664 | } |
| 665 | } |
| 666 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 667 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, batch_lt_8) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 668 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 669 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 670 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 671 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 672 | } |
| 673 | } |
| 674 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 675 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, batch_gt_8) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 676 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 677 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 678 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 679 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 680 | } |
| 681 | } |
| 682 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 683 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, inplace) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 684 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 685 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 686 | .batch_size(batch_size) |
| 687 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 688 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 689 | } |
| 690 | } |
| 691 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 692 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 693 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 694 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 695 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 696 | .batch_size(batch_size) |
| 697 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 698 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 699 | } |
| 700 | } |
| 701 | } |
| 702 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 703 | TEST(F32_VCLAMP__WASMSIMD_ARM_X8, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 704 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 705 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 706 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 707 | .batch_size(batch_size) |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 708 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 709 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 710 | } |
| 711 | } |
| 712 | } |
| 713 | #endif // XNN_ARCH_WASMSIMD |
| 714 | |
| 715 | |
| 716 | #if XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 717 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 718 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 719 | .batch_size(4) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 720 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 721 | } |
| 722 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 723 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, batch_div_4) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 724 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 725 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 726 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 727 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 728 | } |
| 729 | } |
| 730 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 731 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, batch_lt_4) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 732 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 733 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 734 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 735 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 736 | } |
| 737 | } |
| 738 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 739 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, batch_gt_4) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 740 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 741 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 742 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 743 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 744 | } |
| 745 | } |
| 746 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 747 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, inplace) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 748 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 749 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 750 | .batch_size(batch_size) |
| 751 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 752 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 753 | } |
| 754 | } |
| 755 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 756 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 757 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 758 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 759 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 760 | .batch_size(batch_size) |
| 761 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 762 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 763 | } |
| 764 | } |
| 765 | } |
| 766 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 767 | TEST(F32_VCLAMP__WASMSIMD_X86_X4, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 768 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 769 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 770 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 771 | .batch_size(batch_size) |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 772 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 773 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 774 | } |
| 775 | } |
| 776 | } |
| 777 | #endif // XNN_ARCH_WASMSIMD |
| 778 | |
| 779 | |
| 780 | #if XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 781 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, batch_eq_8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 782 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 783 | .batch_size(8) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 784 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 785 | } |
| 786 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 787 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, batch_div_8) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 788 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 789 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 790 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 791 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 792 | } |
| 793 | } |
| 794 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 795 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, batch_lt_8) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 796 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 797 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 798 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 799 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 800 | } |
| 801 | } |
| 802 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 803 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, batch_gt_8) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 804 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 805 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 806 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 807 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 808 | } |
| 809 | } |
| 810 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 811 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, inplace) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 812 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 813 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 814 | .batch_size(batch_size) |
| 815 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 816 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 817 | } |
| 818 | } |
| 819 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 820 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 821 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 822 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 823 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 824 | .batch_size(batch_size) |
| 825 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 826 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 827 | } |
| 828 | } |
| 829 | } |
| 830 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 831 | TEST(F32_VCLAMP__WASMSIMD_X86_X8, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 832 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 833 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 834 | VUnaryMicrokernelTester() |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 835 | .batch_size(batch_size) |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 836 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 837 | .Test(xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 838 | } |
| 839 | } |
| 840 | } |
| 841 | #endif // XNN_ARCH_WASMSIMD |
| 842 | |
| 843 | |
Marat Dukhan | fb5b20a | 2020-06-26 13:14:50 -0700 | [diff] [blame] | 844 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 845 | TEST(F32_VCLAMP__WASM_X1, batch_eq_1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 846 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 847 | .batch_size(1) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 848 | .Test(xnn_f32_vclamp_ukernel__wasm_x1, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 849 | } |
| 850 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 851 | TEST(F32_VCLAMP__WASM_X1, batch_gt_1) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 852 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 853 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 854 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 855 | .Test(xnn_f32_vclamp_ukernel__wasm_x1, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 856 | } |
| 857 | } |
| 858 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 859 | TEST(F32_VCLAMP__WASM_X1, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 860 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 861 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 862 | .batch_size(batch_size) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 863 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 864 | .Test(xnn_f32_vclamp_ukernel__wasm_x1, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 865 | } |
| 866 | } |
| 867 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 868 | TEST(F32_VCLAMP__WASM_X1, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 869 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 870 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 871 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 872 | .batch_size(batch_size) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 873 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 874 | .Test(xnn_f32_vclamp_ukernel__wasm_x1, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 875 | } |
| 876 | } |
| 877 | } |
| 878 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 879 | TEST(F32_VCLAMP__WASM_X1, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 880 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 881 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 882 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 883 | .batch_size(batch_size) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 884 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 885 | .Test(xnn_f32_vclamp_ukernel__wasm_x1, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 886 | } |
| 887 | } |
| 888 | } |
Marat Dukhan | fb5b20a | 2020-06-26 13:14:50 -0700 | [diff] [blame] | 889 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 890 | |
| 891 | |
Marat Dukhan | fb5b20a | 2020-06-26 13:14:50 -0700 | [diff] [blame] | 892 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 893 | TEST(F32_VCLAMP__WASM_X2, batch_eq_2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 894 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 895 | .batch_size(2) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 896 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 897 | } |
| 898 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 899 | TEST(F32_VCLAMP__WASM_X2, batch_div_2) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 900 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 901 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 902 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 903 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 904 | } |
| 905 | } |
| 906 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 907 | TEST(F32_VCLAMP__WASM_X2, batch_lt_2) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 908 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 909 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 910 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 911 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 912 | } |
| 913 | } |
| 914 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 915 | TEST(F32_VCLAMP__WASM_X2, batch_gt_2) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 916 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 917 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 918 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 919 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 920 | } |
| 921 | } |
| 922 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 923 | TEST(F32_VCLAMP__WASM_X2, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 924 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 925 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 926 | .batch_size(batch_size) |
| 927 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 928 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 929 | } |
| 930 | } |
| 931 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 932 | TEST(F32_VCLAMP__WASM_X2, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 933 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 934 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 935 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 936 | .batch_size(batch_size) |
| 937 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 938 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 939 | } |
| 940 | } |
| 941 | } |
| 942 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 943 | TEST(F32_VCLAMP__WASM_X2, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 944 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 945 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 946 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 947 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 948 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 949 | .Test(xnn_f32_vclamp_ukernel__wasm_x2, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 950 | } |
| 951 | } |
| 952 | } |
Marat Dukhan | fb5b20a | 2020-06-26 13:14:50 -0700 | [diff] [blame] | 953 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 954 | |
| 955 | |
Marat Dukhan | fb5b20a | 2020-06-26 13:14:50 -0700 | [diff] [blame] | 956 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 957 | TEST(F32_VCLAMP__WASM_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 958 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 959 | .batch_size(4) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 960 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 961 | } |
| 962 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 963 | TEST(F32_VCLAMP__WASM_X4, batch_div_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 964 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 965 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 966 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 967 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 968 | } |
| 969 | } |
| 970 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 971 | TEST(F32_VCLAMP__WASM_X4, batch_lt_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 972 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 973 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 974 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 975 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 976 | } |
| 977 | } |
| 978 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 979 | TEST(F32_VCLAMP__WASM_X4, batch_gt_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 980 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 981 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 982 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 983 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 984 | } |
| 985 | } |
| 986 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 987 | TEST(F32_VCLAMP__WASM_X4, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 988 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 989 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 990 | .batch_size(batch_size) |
| 991 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 992 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 993 | } |
| 994 | } |
| 995 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 996 | TEST(F32_VCLAMP__WASM_X4, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 997 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 998 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 999 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1000 | .batch_size(batch_size) |
| 1001 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1002 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1003 | } |
| 1004 | } |
| 1005 | } |
| 1006 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1007 | TEST(F32_VCLAMP__WASM_X4, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1008 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 1009 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1010 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1011 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1012 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1013 | .Test(xnn_f32_vclamp_ukernel__wasm_x4, VUnaryMicrokernelTester::OpType::Clamp); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1014 | } |
| 1015 | } |
| 1016 | } |
Marat Dukhan | fb5b20a | 2020-06-26 13:14:50 -0700 | [diff] [blame] | 1017 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1018 | |
| 1019 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1020 | TEST(F32_VCLAMP__SCALAR_X1, batch_eq_1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1021 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1022 | .batch_size(1) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1023 | .Test(xnn_f32_vclamp_ukernel__scalar_x1, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1024 | } |
| 1025 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1026 | TEST(F32_VCLAMP__SCALAR_X1, batch_gt_1) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1027 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1028 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1029 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1030 | .Test(xnn_f32_vclamp_ukernel__scalar_x1, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1031 | } |
| 1032 | } |
| 1033 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1034 | TEST(F32_VCLAMP__SCALAR_X1, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1035 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1036 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1037 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1038 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1039 | .Test(xnn_f32_vclamp_ukernel__scalar_x1, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1040 | } |
| 1041 | } |
| 1042 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1043 | TEST(F32_VCLAMP__SCALAR_X1, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1044 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 1045 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1046 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1047 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1048 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1049 | .Test(xnn_f32_vclamp_ukernel__scalar_x1, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1050 | } |
| 1051 | } |
| 1052 | } |
| 1053 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1054 | TEST(F32_VCLAMP__SCALAR_X1, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1055 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 1056 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1057 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1058 | .batch_size(batch_size) |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1059 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1060 | .Test(xnn_f32_vclamp_ukernel__scalar_x1, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1061 | } |
| 1062 | } |
| 1063 | } |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1064 | |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1065 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1066 | TEST(F32_VCLAMP__SCALAR_X2, batch_eq_2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1067 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1068 | .batch_size(2) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1069 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1070 | } |
| 1071 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1072 | TEST(F32_VCLAMP__SCALAR_X2, batch_div_2) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1073 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1074 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1075 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1076 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1077 | } |
| 1078 | } |
| 1079 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1080 | TEST(F32_VCLAMP__SCALAR_X2, batch_lt_2) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1081 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1082 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1083 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1084 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1085 | } |
| 1086 | } |
| 1087 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1088 | TEST(F32_VCLAMP__SCALAR_X2, batch_gt_2) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1089 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1090 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1091 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1092 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1093 | } |
| 1094 | } |
| 1095 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1096 | TEST(F32_VCLAMP__SCALAR_X2, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1097 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1098 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1099 | .batch_size(batch_size) |
| 1100 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1101 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1102 | } |
| 1103 | } |
| 1104 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1105 | TEST(F32_VCLAMP__SCALAR_X2, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1106 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 1107 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1108 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1109 | .batch_size(batch_size) |
| 1110 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1111 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1112 | } |
| 1113 | } |
| 1114 | } |
| 1115 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1116 | TEST(F32_VCLAMP__SCALAR_X2, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1117 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 1118 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1119 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1120 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1121 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1122 | .Test(xnn_f32_vclamp_ukernel__scalar_x2, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1123 | } |
| 1124 | } |
| 1125 | } |
| 1126 | |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1127 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1128 | TEST(F32_VCLAMP__SCALAR_X4, batch_eq_4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1129 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1130 | .batch_size(4) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1131 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1132 | } |
| 1133 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1134 | TEST(F32_VCLAMP__SCALAR_X4, batch_div_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1135 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1136 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1137 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1138 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1139 | } |
| 1140 | } |
| 1141 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1142 | TEST(F32_VCLAMP__SCALAR_X4, batch_lt_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1143 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1144 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1145 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1146 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1147 | } |
| 1148 | } |
| 1149 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1150 | TEST(F32_VCLAMP__SCALAR_X4, batch_gt_4) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1151 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1152 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1153 | .batch_size(batch_size) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1154 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1155 | } |
| 1156 | } |
| 1157 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1158 | TEST(F32_VCLAMP__SCALAR_X4, inplace) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1159 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1160 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1161 | .batch_size(batch_size) |
| 1162 | .inplace(true) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1163 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1164 | } |
| 1165 | } |
| 1166 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1167 | TEST(F32_VCLAMP__SCALAR_X4, qmin) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1168 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 1169 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1170 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1171 | .batch_size(batch_size) |
| 1172 | .qmin(qmin) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1173 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1174 | } |
| 1175 | } |
| 1176 | } |
| 1177 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1178 | TEST(F32_VCLAMP__SCALAR_X4, qmax) { |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1179 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 1180 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 1181 | VUnaryMicrokernelTester() |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1182 | .batch_size(batch_size) |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1183 | .qmax(qmax) |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 1184 | .Test(xnn_f32_vclamp_ukernel__scalar_x4, VUnaryMicrokernelTester::OpType::Clamp, VUnaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1185 | } |
| 1186 | } |
Marat Dukhan | 60d3f24 | 2021-05-13 11:59:02 -0700 | [diff] [blame] | 1187 | } |