2 * Copyright(c) 2021-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, see <http://www.gnu.org/licenses/>.
28 static void test_load_tmp(void)
34 for (int i
= 0; i
< BUFSIZE
; i
++) {
36 * Load into v12 as .tmp, then use it in the next packet
37 * Should get the new value within the same packet and
38 * the old value in the next packet
40 asm("v3 = vmem(%0 + #0)\n\t"
42 "v12 = vsplat(r1)\n\t"
44 " v12.tmp = vmem(%1 + #0)\n\t"
45 " v4.w = vadd(v12.w, v3.w)\n\t"
47 "v4.w = vadd(v4.w, v12.w)\n\t"
48 "vmem(%2 + #0) = v4\n\t"
49 : : "r"(p0
), "r"(p1
), "r"(pout
)
50 : "r1", "v12", "v3", "v4", "v6", "memory");
51 p0
+= sizeof(MMVector
);
52 p1
+= sizeof(MMVector
);
53 pout
+= sizeof(MMVector
);
55 for (int j
= 0; j
< MAX_VEC_SIZE_BYTES
/ 4; j
++) {
56 expect
[i
].w
[j
] = buffer0
[i
].w
[j
] + buffer1
[i
].w
[j
] + 1;
60 check_output_w(__LINE__
, BUFSIZE
);
63 static void test_load_cur(void)
68 for (int i
= 0; i
< BUFSIZE
; i
++) {
70 " v2.cur = vmem(%0 + #0)\n\t"
71 " vmem(%1 + #0) = v2\n\t"
73 : : "r"(p0
), "r"(pout
) : "v2", "memory");
74 p0
+= sizeof(MMVector
);
75 pout
+= sizeof(MMVector
);
77 for (int j
= 0; j
< MAX_VEC_SIZE_BYTES
/ 4; j
++) {
78 expect
[i
].uw
[j
] = buffer0
[i
].uw
[j
];
82 check_output_w(__LINE__
, BUFSIZE
);
85 static void test_load_aligned(void)
87 /* Aligned loads ignore the low bits of the address */
90 const size_t offset
= 13;
92 p0
+= offset
; /* Create an unaligned address */
93 asm("v2 = vmem(%0 + #0)\n\t"
94 "vmem(%1 + #0) = v2\n\t"
95 : : "r"(p0
), "r"(pout
) : "v2", "memory");
97 expect
[0] = buffer0
[0];
99 check_output_w(__LINE__
, 1);
102 static void test_load_unaligned(void)
106 const size_t offset
= 12;
108 p0
+= offset
; /* Create an unaligned address */
109 asm("v2 = vmemu(%0 + #0)\n\t"
110 "vmem(%1 + #0) = v2\n\t"
111 : : "r"(p0
), "r"(pout
) : "v2", "memory");
113 memcpy(expect
, &buffer0
[0].ub
[offset
], sizeof(MMVector
));
115 check_output_w(__LINE__
, 1);
118 static void test_store_aligned(void)
120 /* Aligned stores ignore the low bits of the address */
123 const size_t offset
= 13;
125 pout
+= offset
; /* Create an unaligned address */
126 asm("v2 = vmem(%0 + #0)\n\t"
127 "vmem(%1 + #0) = v2\n\t"
128 : : "r"(p0
), "r"(pout
) : "v2", "memory");
130 expect
[0] = buffer0
[0];
132 check_output_w(__LINE__
, 1);
135 static void test_store_unaligned(void)
139 const size_t offset
= 12;
141 pout
+= offset
; /* Create an unaligned address */
142 asm("v2 = vmem(%0 + #0)\n\t"
143 "vmemu(%1 + #0) = v2\n\t"
144 : : "r"(p0
), "r"(pout
) : "v2", "memory");
146 memcpy(expect
, buffer0
, 2 * sizeof(MMVector
));
147 memcpy(&expect
[0].ub
[offset
], buffer0
, sizeof(MMVector
));
149 check_output_w(__LINE__
, 2);
152 static void test_masked_store(bool invert
)
158 memset(expect
, 0xff, sizeof(expect
));
159 memset(output
, 0xff, sizeof(expect
));
161 for (int i
= 0; i
< BUFSIZE
; i
++) {
164 "v4 = vsplat(r4)\n\t"
165 "v5 = vmem(%0 + #0)\n\t"
166 "q0 = vcmp.eq(v4.w, v5.w)\n\t"
168 "if (!q0) vmem(%2) = v5\n\t" /* Inverted test */
169 : : "r"(pmask
), "r"(p0
), "r"(pout
)
170 : "r4", "v4", "v5", "q0", "memory");
173 "v4 = vsplat(r4)\n\t"
174 "v5 = vmem(%0 + #0)\n\t"
175 "q0 = vcmp.eq(v4.w, v5.w)\n\t"
177 "if (q0) vmem(%2) = v5\n\t" /* Non-inverted test */
178 : : "r"(pmask
), "r"(p0
), "r"(pout
)
179 : "r4", "v4", "v5", "q0", "memory");
181 p0
+= sizeof(MMVector
);
182 pmask
+= sizeof(MMVector
);
183 pout
+= sizeof(MMVector
);
185 for (int j
= 0; j
< MAX_VEC_SIZE_BYTES
/ 4; j
++) {
187 if (i
+ j
% MASKMOD
!= 0) {
188 expect
[i
].w
[j
] = buffer0
[i
].w
[j
];
191 if (i
+ j
% MASKMOD
== 0) {
192 expect
[i
].w
[j
] = buffer0
[i
].w
[j
];
198 check_output_w(__LINE__
, BUFSIZE
);
201 static void test_new_value_store(void)
207 " v2 = vmem(%0 + #0)\n\t"
208 " vmem(%1 + #0) = v2.new\n\t"
210 : : "r"(p0
), "r"(pout
) : "v2", "memory");
212 expect
[0] = buffer0
[0];
214 check_output_w(__LINE__
, 1);
217 static void test_max_temps()
222 asm("v0 = vmem(%0 + #0)\n\t"
223 "v1 = vmem(%0 + #1)\n\t"
224 "v2 = vmem(%0 + #2)\n\t"
225 "v3 = vmem(%0 + #3)\n\t"
226 "v4 = vmem(%0 + #4)\n\t"
228 " v1:0.w = vadd(v3:2.w, v1:0.w)\n\t"
229 " v2.b = vshuffe(v3.b, v2.b)\n\t"
230 " v3.w = vadd(v1.w, v4.w)\n\t"
231 " v4.tmp = vmem(%0 + #5)\n\t"
233 "vmem(%1 + #0) = v0\n\t"
234 "vmem(%1 + #1) = v1\n\t"
235 "vmem(%1 + #2) = v2\n\t"
236 "vmem(%1 + #3) = v3\n\t"
237 "vmem(%1 + #4) = v4\n\t"
238 : : "r"(p0
), "r"(pout
) : "memory");
240 /* The first two vectors come from the vadd-pair instruction */
241 for (int i
= 0; i
< MAX_VEC_SIZE_BYTES
/ 4; i
++) {
242 expect
[0].w
[i
] = buffer0
[0].w
[i
] + buffer0
[2].w
[i
];
243 expect
[1].w
[i
] = buffer0
[1].w
[i
] + buffer0
[3].w
[i
];
245 /* The third vector comes from the vshuffe instruction */
246 for (int i
= 0; i
< MAX_VEC_SIZE_BYTES
/ 2; i
++) {
247 expect
[2].uh
[i
] = (buffer0
[2].uh
[i
] & 0xff) |
248 (buffer0
[3].uh
[i
] & 0xff) << 8;
250 /* The fourth vector comes from the vadd-single instruction */
251 for (int i
= 0; i
< MAX_VEC_SIZE_BYTES
/ 4; i
++) {
252 expect
[3].w
[i
] = buffer0
[1].w
[i
] + buffer0
[5].w
[i
];
255 * The fifth vector comes from the load to v4
256 * make sure the .tmp is dropped
258 expect
[4] = buffer0
[4];
260 check_output_b(__LINE__
, 5);
263 TEST_VEC_OP2(vadd_w
, vadd
, .w
, w
, 4, +)
264 TEST_VEC_OP2(vadd_h
, vadd
, .h
, h
, 2, +)
265 TEST_VEC_OP2(vadd_b
, vadd
, .b
, b
, 1, +)
266 TEST_VEC_OP2(vsub_w
, vsub
, .w
, w
, 4, -)
267 TEST_VEC_OP2(vsub_h
, vsub
, .h
, h
, 2, -)
268 TEST_VEC_OP2(vsub_b
, vsub
, .b
, b
, 1, -)
269 TEST_VEC_OP2(vxor
, vxor
, , d
, 8, ^)
270 TEST_VEC_OP2(vand
, vand
, , d
, 8, &)
271 TEST_VEC_OP2(vor
, vor
, , d
, 8, |)
272 TEST_VEC_OP1(vnot
, vnot
, , d
, 8, ~)
274 TEST_PRED_OP2(pred_or
, or, |, "")
275 TEST_PRED_OP2(pred_or_n
, or, |, "!")
276 TEST_PRED_OP2(pred_and
, and, &, "")
277 TEST_PRED_OP2(pred_and_n
, and, &, "!")
278 TEST_PRED_OP2(pred_xor
, xor, ^, "")
280 static void test_vadduwsat(void)
283 * Test for saturation by adding two numbers that add to more than UINT_MAX
284 * and make sure the result saturates to UINT_MAX
286 const uint32_t x
= 0xffff0000;
287 const uint32_t y
= 0x000fffff;
289 memset(expect
, 0x12, sizeof(MMVector
));
290 memset(output
, 0x34, sizeof(MMVector
));
292 asm volatile ("v10 = vsplat(%0)\n\t"
293 "v11 = vsplat(%1)\n\t"
294 "v21.uw = vadd(v11.uw, v10.uw):sat\n\t"
295 "vmem(%2+#0) = v21\n\t"
297 : "r"(x
), "r"(y
), "r"(output
)
298 : "v10", "v11", "v21", "memory");
300 for (int j
= 0; j
< MAX_VEC_SIZE_BYTES
/ 4; j
++) {
301 expect
[0].uw
[j
] = UINT_MAX
;
304 check_output_w(__LINE__
, 1);
307 static void test_vsubuwsat_dv(void)
310 * Test for saturation by subtracting two numbers where the result is
311 * negative and make sure the result saturates to zero
313 * vsubuwsat_dv operates on an HVX register pair, so we'll have a
314 * pair of subtractions
318 const uint32_t w
= 0x000000b7;
319 const uint32_t x
= 0xffffff4e;
320 const uint32_t y
= 0x31fe88e7;
321 const uint32_t z
= 0x7fffff79;
323 memset(expect
, 0x12, sizeof(MMVector
) * 2);
324 memset(output
, 0x34, sizeof(MMVector
) * 2);
326 asm volatile ("v16 = vsplat(%0)\n\t"
327 "v17 = vsplat(%1)\n\t"
328 "v26 = vsplat(%2)\n\t"
329 "v27 = vsplat(%3)\n\t"
330 "v25:24.uw = vsub(v17:16.uw, v27:26.uw):sat\n\t"
331 "vmem(%4+#0) = v24\n\t"
332 "vmem(%4+#1) = v25\n\t"
334 : "r"(w
), "r"(y
), "r"(x
), "r"(z
), "r"(output
)
335 : "v16", "v17", "v24", "v25", "v26", "v27", "memory");
337 for (int j
= 0; j
< MAX_VEC_SIZE_BYTES
/ 4; j
++) {
338 expect
[0].uw
[j
] = 0x00000000;
339 expect
[1].uw
[j
] = 0x00000000;
342 check_output_w(__LINE__
, 2);
345 static void test_vshuff(void)
347 /* Test that vshuff works when the two operands are the same register */
348 const uint32_t splat
= 0x089be55c;
349 const uint32_t shuff
= 0x454fa926;
352 memset(expect
, 0x12, sizeof(MMVector
));
353 memset(output
, 0x34, sizeof(MMVector
));
355 asm volatile("v25 = vsplat(%0)\n\t"
356 "vshuff(v25, v25, %1)\n\t"
357 "vmem(%2 + #0) = v25\n\t"
359 : "r"(splat
), "r"(shuff
), "r"(output
)
363 * The semantics of Hexagon are the operands are pass-by-value, so create
364 * two copies of the vsplat result.
366 for (int i
= 0; i
< MAX_VEC_SIZE_BYTES
/ 4; i
++) {
370 /* Do the vshuff operation */
371 for (int offset
= 1; offset
< MAX_VEC_SIZE_BYTES
; offset
<<= 1) {
372 if (shuff
& offset
) {
373 for (int k
= 0; k
< MAX_VEC_SIZE_BYTES
; k
++) {
375 uint8_t tmp
= v0
.ub
[k
];
376 v0
.ub
[k
] = v1
.ub
[k
+ offset
];
377 v1
.ub
[k
+ offset
] = tmp
;
382 /* Put the result in the expect buffer for verification */
385 check_output_b(__LINE__
, 1);
388 static void test_load_tmp_predicated(void)
395 for (int i
= 0; i
< BUFSIZE
; i
++) {
397 * Load into v12 as .tmp with a predicate
398 * When the predicate is true, we get the vector from buffer1[i]
399 * When the predicate is false, we get a vector of all 1's
400 * Regardless of the predicate, the next packet should have
401 * a vector of all 1's
403 asm("v3 = vmem(%0 + #0)\n\t"
405 "v12 = vsplat(r1)\n\t"
406 "p1 = !cmp.eq(%3, #0)\n\t"
408 " if (p1) v12.tmp = vmem(%1 + #0)\n\t"
409 " v4.w = vadd(v12.w, v3.w)\n\t"
411 "v4.w = vadd(v4.w, v12.w)\n\t"
412 "vmem(%2 + #0) = v4\n\t"
413 : : "r"(p0
), "r"(p1
), "r"(pout
), "r"(pred
)
414 : "r1", "p1", "v12", "v3", "v4", "v6", "memory");
415 p0
+= sizeof(MMVector
);
416 p1
+= sizeof(MMVector
);
417 pout
+= sizeof(MMVector
);
419 for (int j
= 0; j
< MAX_VEC_SIZE_BYTES
/ 4; j
++) {
421 pred
? buffer0
[i
].w
[j
] + buffer1
[i
].w
[j
] + 1
422 : buffer0
[i
].w
[j
] + 2;
427 check_output_w(__LINE__
, BUFSIZE
);
430 static void test_load_cur_predicated(void)
433 for (int i
= 0; i
< BUFSIZE
; i
++) {
434 asm volatile("p0 = !cmp.eq(%3, #0)\n\t"
435 "v3 = vmem(%0+#0)\n\t"
437 * Preload v4 to make sure that the assignment from the
438 * packet below is not being ignored when pred is false.
440 "r0 = #0x01237654\n\t"
441 "v4 = vsplat(r0)\n\t"
443 " if (p0) v3.cur = vmem(%1+#0)\n\t"
446 "vmem(%2+#0) = v4\n\t"
448 : "r"(&buffer0
[i
]), "r"(&buffer1
[i
]),
449 "r"(&output
[i
]), "r"(pred
)
450 : "r0", "p0", "v3", "v4", "memory");
451 expect
[i
] = pred
? buffer1
[i
] : buffer0
[i
];
454 check_output_w(__LINE__
, BUFSIZE
);
464 test_load_unaligned();
465 test_store_aligned();
466 test_store_unaligned();
467 test_masked_store(false);
468 test_masked_store(true);
469 test_new_value_store();
484 test_pred_or_n(true);
485 test_pred_and(false);
486 test_pred_and_n(true);
487 test_pred_xor(false);
494 test_load_tmp_predicated();
495 test_load_cur_predicated();
497 puts(err
? "FAIL" : "PASS");