blob: 5bcea291f3c85ff60c29c7175ac654e523d7b628 [file] [log] [blame]
Marat Dukhan2b9efd82020-06-08 01:09:31 -07001// Copyright 2019 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5//
6// Auto-generated file. Do not edit!
7// Specification: test/f32-vsqr.yaml
8// Generator: tools/generate-vunary-test.py
9
10
11#include <gtest/gtest.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/isa-checks.h>
15
16#include <xnnpack/vunary.h>
17#include "vunary-microkernel-tester.h"
18
19
20#if XNN_ARCH_ARM || XNN_ARCH_ARM64
21 TEST(F32_VSQR__NEON_X4, batch_eq_4) {
22 TEST_REQUIRES_ARM_NEON;
23 VUnOpMicrokernelTester()
24 .batch_size(4)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070025 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070026 }
27
28 TEST(F32_VSQR__NEON_X4, batch_div_4) {
29 TEST_REQUIRES_ARM_NEON;
30 for (size_t batch_size = 8; batch_size < 40; batch_size += 4) {
31 VUnOpMicrokernelTester()
32 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070033 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070034 }
35 }
36
37 TEST(F32_VSQR__NEON_X4, batch_lt_4) {
38 TEST_REQUIRES_ARM_NEON;
39 for (size_t batch_size = 1; batch_size < 4; batch_size++) {
40 VUnOpMicrokernelTester()
41 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070042 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070043 }
44 }
45
46 TEST(F32_VSQR__NEON_X4, batch_gt_4) {
47 TEST_REQUIRES_ARM_NEON;
48 for (size_t batch_size = 5; batch_size < 8; batch_size++) {
49 VUnOpMicrokernelTester()
50 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070051 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070052 }
53 }
54
55 TEST(F32_VSQR__NEON_X4, inplace) {
56 TEST_REQUIRES_ARM_NEON;
57 for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) {
58 VUnOpMicrokernelTester()
59 .batch_size(batch_size)
60 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070061 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070062 }
63 }
64#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
65
66
67#if XNN_ARCH_ARM || XNN_ARCH_ARM64
68 TEST(F32_VSQR__NEON_X8, batch_eq_8) {
69 TEST_REQUIRES_ARM_NEON;
70 VUnOpMicrokernelTester()
71 .batch_size(8)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070072 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070073 }
74
75 TEST(F32_VSQR__NEON_X8, batch_div_8) {
76 TEST_REQUIRES_ARM_NEON;
77 for (size_t batch_size = 16; batch_size < 80; batch_size += 8) {
78 VUnOpMicrokernelTester()
79 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070080 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070081 }
82 }
83
84 TEST(F32_VSQR__NEON_X8, batch_lt_8) {
85 TEST_REQUIRES_ARM_NEON;
86 for (size_t batch_size = 1; batch_size < 8; batch_size++) {
87 VUnOpMicrokernelTester()
88 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070089 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070090 }
91 }
92
93 TEST(F32_VSQR__NEON_X8, batch_gt_8) {
94 TEST_REQUIRES_ARM_NEON;
95 for (size_t batch_size = 9; batch_size < 16; batch_size++) {
96 VUnOpMicrokernelTester()
97 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -070098 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -070099 }
100 }
101
102 TEST(F32_VSQR__NEON_X8, inplace) {
103 TEST_REQUIRES_ARM_NEON;
104 for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) {
105 VUnOpMicrokernelTester()
106 .batch_size(batch_size)
107 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700108 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700109 }
110 }
111#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
112
113
114#if XNN_ARCH_X86 || XNN_ARCH_X86_64
115 TEST(F32_VSQR__SSE_X4, batch_eq_4) {
116 TEST_REQUIRES_X86_SSE;
117 VUnOpMicrokernelTester()
118 .batch_size(4)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700119 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700120 }
121
122 TEST(F32_VSQR__SSE_X4, batch_div_4) {
123 TEST_REQUIRES_X86_SSE;
124 for (size_t batch_size = 8; batch_size < 40; batch_size += 4) {
125 VUnOpMicrokernelTester()
126 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700127 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700128 }
129 }
130
131 TEST(F32_VSQR__SSE_X4, batch_lt_4) {
132 TEST_REQUIRES_X86_SSE;
133 for (size_t batch_size = 1; batch_size < 4; batch_size++) {
134 VUnOpMicrokernelTester()
135 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700136 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700137 }
138 }
139
140 TEST(F32_VSQR__SSE_X4, batch_gt_4) {
141 TEST_REQUIRES_X86_SSE;
142 for (size_t batch_size = 5; batch_size < 8; batch_size++) {
143 VUnOpMicrokernelTester()
144 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700145 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700146 }
147 }
148
149 TEST(F32_VSQR__SSE_X4, inplace) {
150 TEST_REQUIRES_X86_SSE;
151 for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) {
152 VUnOpMicrokernelTester()
153 .batch_size(batch_size)
154 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700155 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x4), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700156 }
157 }
158#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
159
160
161#if XNN_ARCH_X86 || XNN_ARCH_X86_64
162 TEST(F32_VSQR__SSE_X8, batch_eq_8) {
163 TEST_REQUIRES_X86_SSE;
164 VUnOpMicrokernelTester()
165 .batch_size(8)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700166 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700167 }
168
169 TEST(F32_VSQR__SSE_X8, batch_div_8) {
170 TEST_REQUIRES_X86_SSE;
171 for (size_t batch_size = 16; batch_size < 80; batch_size += 8) {
172 VUnOpMicrokernelTester()
173 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700174 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700175 }
176 }
177
178 TEST(F32_VSQR__SSE_X8, batch_lt_8) {
179 TEST_REQUIRES_X86_SSE;
180 for (size_t batch_size = 1; batch_size < 8; batch_size++) {
181 VUnOpMicrokernelTester()
182 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700183 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700184 }
185 }
186
187 TEST(F32_VSQR__SSE_X8, batch_gt_8) {
188 TEST_REQUIRES_X86_SSE;
189 for (size_t batch_size = 9; batch_size < 16; batch_size++) {
190 VUnOpMicrokernelTester()
191 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700192 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700193 }
194 }
195
196 TEST(F32_VSQR__SSE_X8, inplace) {
197 TEST_REQUIRES_X86_SSE;
198 for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) {
199 VUnOpMicrokernelTester()
200 .batch_size(batch_size)
201 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700202 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__sse_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700203 }
204 }
205#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
206
207
208#if XNN_ARCH_X86 || XNN_ARCH_X86_64
209 TEST(F32_VSQR__AVX_X8, batch_eq_8) {
210 TEST_REQUIRES_X86_AVX;
211 VUnOpMicrokernelTester()
212 .batch_size(8)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700213 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700214 }
215
216 TEST(F32_VSQR__AVX_X8, batch_div_8) {
217 TEST_REQUIRES_X86_AVX;
218 for (size_t batch_size = 16; batch_size < 80; batch_size += 8) {
219 VUnOpMicrokernelTester()
220 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700221 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700222 }
223 }
224
225 TEST(F32_VSQR__AVX_X8, batch_lt_8) {
226 TEST_REQUIRES_X86_AVX;
227 for (size_t batch_size = 1; batch_size < 8; batch_size++) {
228 VUnOpMicrokernelTester()
229 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700230 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700231 }
232 }
233
234 TEST(F32_VSQR__AVX_X8, batch_gt_8) {
235 TEST_REQUIRES_X86_AVX;
236 for (size_t batch_size = 9; batch_size < 16; batch_size++) {
237 VUnOpMicrokernelTester()
238 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700239 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700240 }
241 }
242
243 TEST(F32_VSQR__AVX_X8, inplace) {
244 TEST_REQUIRES_X86_AVX;
245 for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) {
246 VUnOpMicrokernelTester()
247 .batch_size(batch_size)
248 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700249 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700250 }
251 }
252#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
253
254
255#if XNN_ARCH_X86 || XNN_ARCH_X86_64
256 TEST(F32_VSQR__AVX_X16, batch_eq_16) {
257 TEST_REQUIRES_X86_AVX;
258 VUnOpMicrokernelTester()
259 .batch_size(16)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700260 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700261 }
262
263 TEST(F32_VSQR__AVX_X16, batch_div_16) {
264 TEST_REQUIRES_X86_AVX;
265 for (size_t batch_size = 32; batch_size < 160; batch_size += 16) {
266 VUnOpMicrokernelTester()
267 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700268 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700269 }
270 }
271
272 TEST(F32_VSQR__AVX_X16, batch_lt_16) {
273 TEST_REQUIRES_X86_AVX;
274 for (size_t batch_size = 1; batch_size < 16; batch_size++) {
275 VUnOpMicrokernelTester()
276 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700277 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700278 }
279 }
280
281 TEST(F32_VSQR__AVX_X16, batch_gt_16) {
282 TEST_REQUIRES_X86_AVX;
283 for (size_t batch_size = 17; batch_size < 32; batch_size++) {
284 VUnOpMicrokernelTester()
285 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700286 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700287 }
288 }
289
290 TEST(F32_VSQR__AVX_X16, inplace) {
291 TEST_REQUIRES_X86_AVX;
292 for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) {
293 VUnOpMicrokernelTester()
294 .batch_size(batch_size)
295 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700296 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700297 }
298 }
299#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
300
301
302#if XNN_ARCH_X86 || XNN_ARCH_X86_64
303 TEST(F32_VSQR__AVX512F_X16, batch_eq_16) {
304 TEST_REQUIRES_X86_AVX512F;
305 VUnOpMicrokernelTester()
306 .batch_size(16)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700307 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700308 }
309
310 TEST(F32_VSQR__AVX512F_X16, batch_div_16) {
311 TEST_REQUIRES_X86_AVX512F;
312 for (size_t batch_size = 32; batch_size < 160; batch_size += 16) {
313 VUnOpMicrokernelTester()
314 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700315 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700316 }
317 }
318
319 TEST(F32_VSQR__AVX512F_X16, batch_lt_16) {
320 TEST_REQUIRES_X86_AVX512F;
321 for (size_t batch_size = 1; batch_size < 16; batch_size++) {
322 VUnOpMicrokernelTester()
323 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700324 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700325 }
326 }
327
328 TEST(F32_VSQR__AVX512F_X16, batch_gt_16) {
329 TEST_REQUIRES_X86_AVX512F;
330 for (size_t batch_size = 17; batch_size < 32; batch_size++) {
331 VUnOpMicrokernelTester()
332 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700333 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700334 }
335 }
336
337 TEST(F32_VSQR__AVX512F_X16, inplace) {
338 TEST_REQUIRES_X86_AVX512F;
339 for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) {
340 VUnOpMicrokernelTester()
341 .batch_size(batch_size)
342 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700343 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700344 }
345 }
346#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
347
348
349#if XNN_ARCH_X86 || XNN_ARCH_X86_64
350 TEST(F32_VSQR__AVX512F_X32, batch_eq_32) {
351 TEST_REQUIRES_X86_AVX512F;
352 VUnOpMicrokernelTester()
353 .batch_size(32)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700354 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700355 }
356
357 TEST(F32_VSQR__AVX512F_X32, batch_div_32) {
358 TEST_REQUIRES_X86_AVX512F;
359 for (size_t batch_size = 64; batch_size < 320; batch_size += 32) {
360 VUnOpMicrokernelTester()
361 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700362 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700363 }
364 }
365
366 TEST(F32_VSQR__AVX512F_X32, batch_lt_32) {
367 TEST_REQUIRES_X86_AVX512F;
368 for (size_t batch_size = 1; batch_size < 32; batch_size++) {
369 VUnOpMicrokernelTester()
370 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700371 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700372 }
373 }
374
375 TEST(F32_VSQR__AVX512F_X32, batch_gt_32) {
376 TEST_REQUIRES_X86_AVX512F;
377 for (size_t batch_size = 33; batch_size < 64; batch_size++) {
378 VUnOpMicrokernelTester()
379 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700380 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700381 }
382 }
383
384 TEST(F32_VSQR__AVX512F_X32, inplace) {
385 TEST_REQUIRES_X86_AVX512F;
386 for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) {
387 VUnOpMicrokernelTester()
388 .batch_size(batch_size)
389 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700390 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::Square);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700391 }
392 }
393#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
394
395
396#if !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC
397 TEST(F32_VSQR__PSIMD_X4, batch_eq_4) {
398 TEST_REQUIRES_PSIMD;
399 VUnOpMicrokernelTester()
400 .batch_size(4)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700401 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700402 }
403
404 TEST(F32_VSQR__PSIMD_X4, batch_div_4) {
405 TEST_REQUIRES_PSIMD;
406 for (size_t batch_size = 8; batch_size < 40; batch_size += 4) {
407 VUnOpMicrokernelTester()
408 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700409 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700410 }
411 }
412
413 TEST(F32_VSQR__PSIMD_X4, batch_lt_4) {
414 TEST_REQUIRES_PSIMD;
415 for (size_t batch_size = 1; batch_size < 4; batch_size++) {
416 VUnOpMicrokernelTester()
417 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700418 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700419 }
420 }
421
422 TEST(F32_VSQR__PSIMD_X4, batch_gt_4) {
423 TEST_REQUIRES_PSIMD;
424 for (size_t batch_size = 5; batch_size < 8; batch_size++) {
425 VUnOpMicrokernelTester()
426 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700427 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700428 }
429 }
430
431 TEST(F32_VSQR__PSIMD_X4, inplace) {
432 TEST_REQUIRES_PSIMD;
433 for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) {
434 VUnOpMicrokernelTester()
435 .batch_size(batch_size)
436 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700437 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700438 }
439 }
440#endif // !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC
441
442
443#if !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC
444 TEST(F32_VSQR__PSIMD_X8, batch_eq_8) {
445 TEST_REQUIRES_PSIMD;
446 VUnOpMicrokernelTester()
447 .batch_size(8)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700448 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700449 }
450
451 TEST(F32_VSQR__PSIMD_X8, batch_div_8) {
452 TEST_REQUIRES_PSIMD;
453 for (size_t batch_size = 16; batch_size < 80; batch_size += 8) {
454 VUnOpMicrokernelTester()
455 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700456 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700457 }
458 }
459
460 TEST(F32_VSQR__PSIMD_X8, batch_lt_8) {
461 TEST_REQUIRES_PSIMD;
462 for (size_t batch_size = 1; batch_size < 8; batch_size++) {
463 VUnOpMicrokernelTester()
464 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700465 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700466 }
467 }
468
469 TEST(F32_VSQR__PSIMD_X8, batch_gt_8) {
470 TEST_REQUIRES_PSIMD;
471 for (size_t batch_size = 9; batch_size < 16; batch_size++) {
472 VUnOpMicrokernelTester()
473 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700474 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700475 }
476 }
477
478 TEST(F32_VSQR__PSIMD_X8, inplace) {
479 TEST_REQUIRES_PSIMD;
480 for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) {
481 VUnOpMicrokernelTester()
482 .batch_size(batch_size)
483 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700484 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700485 }
486 }
487#endif // !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC
488
489
490TEST(F32_VSQR__SCALAR_X1, batch_eq_1) {
491 VUnOpMicrokernelTester()
492 .batch_size(1)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700493 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x1), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700494}
495
496TEST(F32_VSQR__SCALAR_X1, batch_gt_1) {
497 for (size_t batch_size = 2; batch_size < 10; batch_size++) {
498 VUnOpMicrokernelTester()
499 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700500 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x1), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700501 }
502}
503
504TEST(F32_VSQR__SCALAR_X1, inplace) {
505 for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) {
506 VUnOpMicrokernelTester()
507 .batch_size(batch_size)
508 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700509 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x1), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700510 }
511}
512
Frank Barchard177fe2d2020-06-11 11:09:33 -0700513
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700514TEST(F32_VSQR__SCALAR_X2, batch_eq_2) {
515 VUnOpMicrokernelTester()
516 .batch_size(2)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700517 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x2), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700518}
519
520TEST(F32_VSQR__SCALAR_X2, batch_div_2) {
521 for (size_t batch_size = 4; batch_size < 20; batch_size += 2) {
522 VUnOpMicrokernelTester()
523 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700524 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x2), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700525 }
526}
527
528TEST(F32_VSQR__SCALAR_X2, batch_lt_2) {
529 for (size_t batch_size = 1; batch_size < 2; batch_size++) {
530 VUnOpMicrokernelTester()
531 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700532 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x2), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700533 }
534}
535
536TEST(F32_VSQR__SCALAR_X2, batch_gt_2) {
537 for (size_t batch_size = 3; batch_size < 4; batch_size++) {
538 VUnOpMicrokernelTester()
539 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700540 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x2), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700541 }
542}
543
544TEST(F32_VSQR__SCALAR_X2, inplace) {
545 for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) {
546 VUnOpMicrokernelTester()
547 .batch_size(batch_size)
548 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700549 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x2), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700550 }
551}
552
Frank Barchard177fe2d2020-06-11 11:09:33 -0700553
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700554TEST(F32_VSQR__SCALAR_X4, batch_eq_4) {
555 VUnOpMicrokernelTester()
556 .batch_size(4)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700557 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700558}
559
560TEST(F32_VSQR__SCALAR_X4, batch_div_4) {
561 for (size_t batch_size = 8; batch_size < 40; batch_size += 4) {
562 VUnOpMicrokernelTester()
563 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700564 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700565 }
566}
567
568TEST(F32_VSQR__SCALAR_X4, batch_lt_4) {
569 for (size_t batch_size = 1; batch_size < 4; batch_size++) {
570 VUnOpMicrokernelTester()
571 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700572 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700573 }
574}
575
576TEST(F32_VSQR__SCALAR_X4, batch_gt_4) {
577 for (size_t batch_size = 5; batch_size < 8; batch_size++) {
578 VUnOpMicrokernelTester()
579 .batch_size(batch_size)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700580 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700581 }
582}
583
584TEST(F32_VSQR__SCALAR_X4, inplace) {
585 for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) {
586 VUnOpMicrokernelTester()
587 .batch_size(batch_size)
588 .inplace(true)
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700589 .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqr_ukernel__scalar_x4), VUnOpMicrokernelTester::OpType::Square, VUnOpMicrokernelTester::Variant::Scalar);
Marat Dukhan2b9efd82020-06-08 01:09:31 -0700590 }
Frank Barchard177fe2d2020-06-11 11:09:33 -0700591}