| 1 | /* |
| 2 | * Copyright(c) 2023 Qualcomm Innovation Center, Inc. All Rights Reserved. |
| 3 | * |
| 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. |
| 8 | * |
| 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. |
| 13 | * |
| 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/>. |
| 16 | */ |
| 17 | |
| 18 | #include <stdio.h> |
| 19 | #include <stdint.h> |
| 20 | #include <stdbool.h> |
| 21 | #include <string.h> |
| 22 | #include <limits.h> |
| 23 | |
| 24 | int err; |
| 25 | |
| 26 | #include "hvx_misc.h" |
| 27 | |
| 28 | #define fVROUND(VAL, SHAMT) \ |
| 29 | ((VAL) + (((SHAMT) > 0) ? (1LL << ((SHAMT) - 1)) : 0)) |
| 30 | |
| 31 | #define fVSATUB(VAL) \ |
| 32 | ((((VAL) & 0xffLL) == (VAL)) ? \ |
| 33 | (VAL) : \ |
| 34 | ((((int32_t)(VAL)) < 0) ? 0 : 0xff)) |
| 35 | |
| 36 | #define fVSATUH(VAL) \ |
| 37 | ((((VAL) & 0xffffLL) == (VAL)) ? \ |
| 38 | (VAL) : \ |
| 39 | ((((int32_t)(VAL)) < 0) ? 0 : 0xffff)) |
| 40 | |
| 41 | static void test_vasrvuhubrndsat(void) |
| 42 | { |
| 43 | void *p0 = buffer0; |
| 44 | void *p1 = buffer1; |
| 45 | void *pout = output; |
| 46 | |
| 47 | memset(expect, 0xaa, sizeof(expect)); |
| 48 | memset(output, 0xbb, sizeof(output)); |
| 49 | |
| 50 | for (int i = 0; i < BUFSIZE / 2; i++) { |
| 51 | asm("v4 = vmem(%0 + #0)\n\t" |
| 52 | "v5 = vmem(%0 + #1)\n\t" |
| 53 | "v6 = vmem(%1 + #0)\n\t" |
| 54 | "v5.ub = vasr(v5:4.uh, v6.ub):rnd:sat\n\t" |
| 55 | "vmem(%2) = v5\n\t" |
| 56 | : : "r"(p0), "r"(p1), "r"(pout) |
| 57 | : "v4", "v5", "v6", "memory"); |
| 58 | p0 += sizeof(MMVector) * 2; |
| 59 | p1 += sizeof(MMVector); |
| 60 | pout += sizeof(MMVector); |
| 61 | |
| 62 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 2; j++) { |
| 63 | int shamt; |
| 64 | uint8_t byte0; |
| 65 | uint8_t byte1; |
| 66 | |
| 67 | shamt = buffer1[i].ub[2 * j + 0] & 0x7; |
| 68 | byte0 = fVSATUB(fVROUND(buffer0[2 * i + 0].uh[j], shamt) >> shamt); |
| 69 | shamt = buffer1[i].ub[2 * j + 1] & 0x7; |
| 70 | byte1 = fVSATUB(fVROUND(buffer0[2 * i + 1].uh[j], shamt) >> shamt); |
| 71 | expect[i].uh[j] = (byte1 << 8) | (byte0 & 0xff); |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | check_output_h(__LINE__, BUFSIZE / 2); |
| 76 | } |
| 77 | |
| 78 | static void test_vasrvuhubsat(void) |
| 79 | { |
| 80 | void *p0 = buffer0; |
| 81 | void *p1 = buffer1; |
| 82 | void *pout = output; |
| 83 | |
| 84 | memset(expect, 0xaa, sizeof(expect)); |
| 85 | memset(output, 0xbb, sizeof(output)); |
| 86 | |
| 87 | for (int i = 0; i < BUFSIZE / 2; i++) { |
| 88 | asm("v4 = vmem(%0 + #0)\n\t" |
| 89 | "v5 = vmem(%0 + #1)\n\t" |
| 90 | "v6 = vmem(%1 + #0)\n\t" |
| 91 | "v5.ub = vasr(v5:4.uh, v6.ub):sat\n\t" |
| 92 | "vmem(%2) = v5\n\t" |
| 93 | : : "r"(p0), "r"(p1), "r"(pout) |
| 94 | : "v4", "v5", "v6", "memory"); |
| 95 | p0 += sizeof(MMVector) * 2; |
| 96 | p1 += sizeof(MMVector); |
| 97 | pout += sizeof(MMVector); |
| 98 | |
| 99 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 2; j++) { |
| 100 | int shamt; |
| 101 | uint8_t byte0; |
| 102 | uint8_t byte1; |
| 103 | |
| 104 | shamt = buffer1[i].ub[2 * j + 0] & 0x7; |
| 105 | byte0 = fVSATUB(buffer0[2 * i + 0].uh[j] >> shamt); |
| 106 | shamt = buffer1[i].ub[2 * j + 1] & 0x7; |
| 107 | byte1 = fVSATUB(buffer0[2 * i + 1].uh[j] >> shamt); |
| 108 | expect[i].uh[j] = (byte1 << 8) | (byte0 & 0xff); |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | check_output_h(__LINE__, BUFSIZE / 2); |
| 113 | } |
| 114 | |
| 115 | static void test_vasrvwuhrndsat(void) |
| 116 | { |
| 117 | void *p0 = buffer0; |
| 118 | void *p1 = buffer1; |
| 119 | void *pout = output; |
| 120 | |
| 121 | memset(expect, 0xaa, sizeof(expect)); |
| 122 | memset(output, 0xbb, sizeof(output)); |
| 123 | |
| 124 | for (int i = 0; i < BUFSIZE / 2; i++) { |
| 125 | asm("v4 = vmem(%0 + #0)\n\t" |
| 126 | "v5 = vmem(%0 + #1)\n\t" |
| 127 | "v6 = vmem(%1 + #0)\n\t" |
| 128 | "v5.uh = vasr(v5:4.w, v6.uh):rnd:sat\n\t" |
| 129 | "vmem(%2) = v5\n\t" |
| 130 | : : "r"(p0), "r"(p1), "r"(pout) |
| 131 | : "v4", "v5", "v6", "memory"); |
| 132 | p0 += sizeof(MMVector) * 2; |
| 133 | p1 += sizeof(MMVector); |
| 134 | pout += sizeof(MMVector); |
| 135 | |
| 136 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 4; j++) { |
| 137 | int shamt; |
| 138 | uint16_t half0; |
| 139 | uint16_t half1; |
| 140 | |
| 141 | shamt = buffer1[i].uh[2 * j + 0] & 0xf; |
| 142 | half0 = fVSATUH(fVROUND(buffer0[2 * i + 0].w[j], shamt) >> shamt); |
| 143 | shamt = buffer1[i].uh[2 * j + 1] & 0xf; |
| 144 | half1 = fVSATUH(fVROUND(buffer0[2 * i + 1].w[j], shamt) >> shamt); |
| 145 | expect[i].w[j] = (half1 << 16) | (half0 & 0xffff); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | check_output_w(__LINE__, BUFSIZE / 2); |
| 150 | } |
| 151 | |
| 152 | static void test_vasrvwuhsat(void) |
| 153 | { |
| 154 | void *p0 = buffer0; |
| 155 | void *p1 = buffer1; |
| 156 | void *pout = output; |
| 157 | |
| 158 | memset(expect, 0xaa, sizeof(expect)); |
| 159 | memset(output, 0xbb, sizeof(output)); |
| 160 | |
| 161 | for (int i = 0; i < BUFSIZE / 2; i++) { |
| 162 | asm("v4 = vmem(%0 + #0)\n\t" |
| 163 | "v5 = vmem(%0 + #1)\n\t" |
| 164 | "v6 = vmem(%1 + #0)\n\t" |
| 165 | "v5.uh = vasr(v5:4.w, v6.uh):sat\n\t" |
| 166 | "vmem(%2) = v5\n\t" |
| 167 | : : "r"(p0), "r"(p1), "r"(pout) |
| 168 | : "v4", "v5", "v6", "memory"); |
| 169 | p0 += sizeof(MMVector) * 2; |
| 170 | p1 += sizeof(MMVector); |
| 171 | pout += sizeof(MMVector); |
| 172 | |
| 173 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 4; j++) { |
| 174 | int shamt; |
| 175 | uint16_t half0; |
| 176 | uint16_t half1; |
| 177 | |
| 178 | shamt = buffer1[i].uh[2 * j + 0] & 0xf; |
| 179 | half0 = fVSATUH(buffer0[2 * i + 0].w[j] >> shamt); |
| 180 | shamt = buffer1[i].uh[2 * j + 1] & 0xf; |
| 181 | half1 = fVSATUH(buffer0[2 * i + 1].w[j] >> shamt); |
| 182 | expect[i].w[j] = (half1 << 16) | (half0 & 0xffff); |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | check_output_w(__LINE__, BUFSIZE / 2); |
| 187 | } |
| 188 | |
| 189 | static void test_vassign_tmp(void) |
| 190 | { |
| 191 | void *p0 = buffer0; |
| 192 | void *pout = output; |
| 193 | |
| 194 | memset(expect, 0xaa, sizeof(expect)); |
| 195 | memset(output, 0xbb, sizeof(output)); |
| 196 | |
| 197 | for (int i = 0; i < BUFSIZE; i++) { |
| 198 | /* |
| 199 | * Assign into v12 as .tmp, then use it in the next packet |
| 200 | * Should get the new value within the same packet and |
| 201 | * the old value in the next packet |
| 202 | */ |
| 203 | asm("v3 = vmem(%0 + #0)\n\t" |
| 204 | "r1 = #1\n\t" |
| 205 | "v12 = vsplat(r1)\n\t" |
| 206 | "r1 = #2\n\t" |
| 207 | "v13 = vsplat(r1)\n\t" |
| 208 | "{\n\t" |
| 209 | " v12.tmp = v13\n\t" |
| 210 | " v4.w = vadd(v12.w, v3.w)\n\t" |
| 211 | "}\n\t" |
| 212 | "v4.w = vadd(v4.w, v12.w)\n\t" |
| 213 | "vmem(%1 + #0) = v4\n\t" |
| 214 | : : "r"(p0), "r"(pout) |
| 215 | : "r1", "v3", "v4", "v12", "v13", "memory"); |
| 216 | p0 += sizeof(MMVector); |
| 217 | pout += sizeof(MMVector); |
| 218 | |
| 219 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 4; j++) { |
| 220 | expect[i].w[j] = buffer0[i].w[j] + 3; |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | check_output_w(__LINE__, BUFSIZE); |
| 225 | } |
| 226 | |
| 227 | static void test_vcombine_tmp(void) |
| 228 | { |
| 229 | void *p0 = buffer0; |
| 230 | void *p1 = buffer1; |
| 231 | void *pout = output; |
| 232 | |
| 233 | memset(expect, 0xaa, sizeof(expect)); |
| 234 | memset(output, 0xbb, sizeof(output)); |
| 235 | |
| 236 | for (int i = 0; i < BUFSIZE; i++) { |
| 237 | /* |
| 238 | * Combine into v13:12 as .tmp, then use it in the next packet |
| 239 | * Should get the new value within the same packet and |
| 240 | * the old value in the next packet |
| 241 | */ |
| 242 | asm("v3 = vmem(%0 + #0)\n\t" |
| 243 | "r1 = #1\n\t" |
| 244 | "v12 = vsplat(r1)\n\t" |
| 245 | "r1 = #2\n\t" |
| 246 | "v13 = vsplat(r1)\n\t" |
| 247 | "r1 = #3\n\t" |
| 248 | "v14 = vsplat(r1)\n\t" |
| 249 | "r1 = #4\n\t" |
| 250 | "v15 = vsplat(r1)\n\t" |
| 251 | "{\n\t" |
| 252 | " v13:12.tmp = vcombine(v15, v14)\n\t" |
| 253 | " v4.w = vadd(v12.w, v3.w)\n\t" |
| 254 | " v16 = v13\n\t" |
| 255 | "}\n\t" |
| 256 | "v4.w = vadd(v4.w, v12.w)\n\t" |
| 257 | "v4.w = vadd(v4.w, v13.w)\n\t" |
| 258 | "v4.w = vadd(v4.w, v16.w)\n\t" |
| 259 | "vmem(%2 + #0) = v4\n\t" |
| 260 | : : "r"(p0), "r"(p1), "r"(pout) |
| 261 | : "r1", "v3", "v4", "v12", "v13", "v14", "v15", "v16", "memory"); |
| 262 | p0 += sizeof(MMVector); |
| 263 | p1 += sizeof(MMVector); |
| 264 | pout += sizeof(MMVector); |
| 265 | |
| 266 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 4; j++) { |
| 267 | expect[i].w[j] = buffer0[i].w[j] + 10; |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | check_output_w(__LINE__, BUFSIZE); |
| 272 | } |
| 273 | |
| 274 | static void test_vmpyuhvs(void) |
| 275 | { |
| 276 | void *p0 = buffer0; |
| 277 | void *p1 = buffer1; |
| 278 | void *pout = output; |
| 279 | |
| 280 | memset(expect, 0xaa, sizeof(expect)); |
| 281 | memset(output, 0xbb, sizeof(output)); |
| 282 | |
| 283 | for (int i = 0; i < BUFSIZE; i++) { |
| 284 | asm("v4 = vmem(%0 + #0)\n\t" |
| 285 | "v5 = vmem(%1 + #0)\n\t" |
| 286 | "v4.uh = vmpy(V4.uh, v5.uh):>>16\n\t" |
| 287 | "vmem(%2) = v4\n\t" |
| 288 | : : "r"(p0), "r"(p1), "r"(pout) |
| 289 | : "v4", "v5", "memory"); |
| 290 | p0 += sizeof(MMVector); |
| 291 | p1 += sizeof(MMVector); |
| 292 | pout += sizeof(MMVector); |
| 293 | |
| 294 | for (int j = 0; j < MAX_VEC_SIZE_BYTES / 2; j++) { |
| 295 | expect[i].uh[j] = (buffer0[i].uh[j] * buffer1[i].uh[j]) >> 16; |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | check_output_h(__LINE__, BUFSIZE); |
| 300 | } |
| 301 | |
| 302 | int main() |
| 303 | { |
| 304 | init_buffers(); |
| 305 | |
| 306 | test_vasrvuhubrndsat(); |
| 307 | test_vasrvuhubsat(); |
| 308 | test_vasrvwuhrndsat(); |
| 309 | test_vasrvwuhsat(); |
| 310 | |
| 311 | test_vassign_tmp(); |
| 312 | test_vcombine_tmp(); |
| 313 | |
| 314 | test_vmpyuhvs(); |
| 315 | |
| 316 | puts(err ? "FAIL" : "PASS"); |
| 317 | return err ? 1 : 0; |
| 318 | } |