Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 1 | // Copyright 2018 The Wuffs Authors. |
| 2 | // |
| 3 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | // you may not use this file except in compliance with the License. |
| 5 | // You may obtain a copy of the License at |
| 6 | // |
| 7 | // https://www.apache.org/licenses/LICENSE-2.0 |
| 8 | // |
| 9 | // Unless required by applicable law or agreed to in writing, software |
| 10 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | // See the License for the specific language governing permissions and |
| 13 | // limitations under the License. |
| 14 | |
| 15 | // This file contains a hand-written C benchmark of different strategies for |
| 16 | // decoding PNG data. |
| 17 | // |
| 18 | // For a PNG image with width W and height H, the H rows can be decompressed |
| 19 | // one-at-a-time or all-at-once. Roughly speaking, this corresponds to H versus |
| 20 | // 1 call into the zlib decoder. The former (call it "fragmented dst") requires |
| 21 | // less scratch-space memory than the latter ("full dst"): 2 * bytes_per_row |
| 22 | // instead of H * bytes_per row, but the latter can be faster. |
| 23 | // |
| 24 | // The zlib-compressed data can be split into multiple IDAT chunks. Similarly, |
| 25 | // these chunks can be decompressed separately ("fragmented IDAT") or together |
| 26 | // ("full IDAT"), again providing a memory vs speed trade-off. |
| 27 | // |
| 28 | // This program reports the speed of combining the independent frag/full dst |
| 29 | // and frag/full IDAT techniques. |
| 30 | // |
| 31 | // For example, with gcc 7.3 (and -O3) as of April 2018: |
| 32 | // |
| 33 | // On ../test/data/hat.png (90 × 112 pixels): |
| 34 | // name time/op relative |
| 35 | // FragDstFragIDAT/gcc 203µs ± 1% 1.00x |
| 36 | // FragDstFullIDAT/gcc 203µs ± 0% 1.00x |
| 37 | // FullDstFragIDAT/gcc 170µs ± 0% 1.19x |
| 38 | // FullDstFullIDAT/gcc 147µs ± 0% 1.38x |
| 39 | // |
| 40 | // On ../test/data/hibiscus.png (312 × 442 pixels): |
| 41 | // name time/op relative |
| 42 | // FragDstFragIDAT/gcc 2.62ms ± 1% 1.00x |
| 43 | // FragDstFullIDAT/gcc 2.61ms ± 1% 1.00x |
| 44 | // FullDstFragIDAT/gcc 2.44ms ± 0% 1.07x |
| 45 | // FullDstFullIDAT/gcc 2.03ms ± 1% 1.29x |
| 46 | // |
| 47 | // On ../test/data/harvesters.png (1165 × 859 pixels): |
| 48 | // name time/op relative |
| 49 | // FragDstFragIDAT/gcc 18.3ms ± 0% 1.00x |
| 50 | // FragDstFullIDAT/gcc 18.3ms ± 1% 1.00x |
| 51 | // FullDstFragIDAT/gcc 17.1ms ± 0% 1.07x |
| 52 | // FullDstFullIDAT/gcc 13.9ms ± 0% 1.32x |
| 53 | |
| 54 | #include <errno.h> |
| 55 | #include <inttypes.h> |
| 56 | #include <stdio.h> |
| 57 | #include <string.h> |
| 58 | #include <sys/time.h> |
| 59 | #include <unistd.h> |
| 60 | |
Nigel Tao | a0bafff | 2018-07-14 08:59:36 +1000 | [diff] [blame] | 61 | // Wuffs ships as a "single file C library" or "header file library" as per |
| 62 | // https://github.com/nothings/stb/blob/master/docs/stb_howto.txt |
| 63 | // |
| 64 | // To use that single file as a "foo.c"-like implementation, instead of a |
| 65 | // "foo.h"-like header, #define WUFFS_IMPLEMENTATION before #include'ing or |
| 66 | // compiling it. |
| 67 | #define WUFFS_IMPLEMENTATION |
| 68 | |
Jimmy Casey | 3c50ada | 2018-07-26 17:12:19 +0000 | [diff] [blame] | 69 | // If building this program in an environment that doesn't easily accommodate |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 70 | // relative includes, you can use the script/inline-c-relative-includes.go |
| 71 | // program to generate a stand-alone C file. |
Nigel Tao | bc82e51 | 2018-07-21 09:20:10 +1000 | [diff] [blame] | 72 | #include "../release/c/wuffs-unsupported-snapshot.h" |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 73 | |
| 74 | // The order matters here. Clang also defines "__GNUC__". |
| 75 | #if defined(__clang__) |
| 76 | const char* cc = "clang"; |
| 77 | const char* cc_version = __clang_version__; |
| 78 | #elif defined(__GNUC__) |
| 79 | const char* cc = "gcc"; |
| 80 | const char* cc_version = __VERSION__; |
| 81 | #elif defined(_MSC_VER) |
| 82 | const char* cc = "cl"; |
| 83 | const char* cc_version = "???"; |
| 84 | #else |
| 85 | const char* cc = "cc"; |
| 86 | const char* cc_version = "???"; |
| 87 | #endif |
| 88 | |
| 89 | static inline uint32_t load_u32be(uint8_t* p) { |
| 90 | return ((uint32_t)(p[0]) << 24) | ((uint32_t)(p[1]) << 16) | |
| 91 | ((uint32_t)(p[2]) << 8) | ((uint32_t)(p[3]) << 0); |
| 92 | } |
| 93 | |
| 94 | // Limit the input PNG image (and therefore its IDAT data) to (64 MiB - 1 byte) |
| 95 | // compressed, in up to 1024 IDAT chunks, and 256 MiB and 16384 × 16384 pixels |
| 96 | // uncompressed. This is a limitation of this program (which uses the Wuffs |
| 97 | // standard library), not a limitation of Wuffs per se. |
| 98 | #define DST_BUFFER_SIZE (256 * 1024 * 1024) |
| 99 | #define SRC_BUFFER_SIZE (64 * 1024 * 1024) |
| 100 | #define MAX_DIMENSION (16384) |
| 101 | #define MAX_IDAT_CHUNKS (1024) |
| 102 | |
| 103 | uint8_t dst_buffer[DST_BUFFER_SIZE] = {0}; |
| 104 | size_t dst_len = 0; |
| 105 | uint8_t src_buffer[SRC_BUFFER_SIZE] = {0}; |
| 106 | size_t src_len = 0; |
| 107 | uint8_t idat_buffer[SRC_BUFFER_SIZE] = {0}; |
| 108 | // The n'th IDAT chunk data (where n is a zero-based count) is in |
| 109 | // idat_buffer[i:j], where i = idat_splits[n+0] and j = idat_splits[n+1]. |
| 110 | size_t idat_splits[MAX_IDAT_CHUNKS + 1] = {0}; |
| 111 | uint32_t num_idat_chunks = 0; |
| 112 | |
| 113 | uint32_t width = 0; |
| 114 | uint32_t height = 0; |
| 115 | uint64_t bytes_per_pixel = 0; |
| 116 | uint64_t bytes_per_row = 0; |
| 117 | uint64_t bytes_per_frame = 0; |
| 118 | |
| 119 | const char* read_stdin() { |
| 120 | while (src_len < SRC_BUFFER_SIZE) { |
| 121 | const int stdin_fd = 0; |
| 122 | ssize_t n = read(stdin_fd, src_buffer + src_len, SRC_BUFFER_SIZE - src_len); |
| 123 | if (n > 0) { |
| 124 | src_len += n; |
| 125 | } else if (n == 0) { |
| 126 | return NULL; |
| 127 | } else if (errno == EINTR) { |
| 128 | // No-op. |
| 129 | } else { |
| 130 | return strerror(errno); |
| 131 | } |
| 132 | } |
| 133 | return "input is too large"; |
| 134 | } |
| 135 | |
| 136 | const char* process_png_chunks(uint8_t* p, size_t n) { |
| 137 | while (n > 0) { |
| 138 | // Process the 8 byte chunk header. |
| 139 | if (n < 8) { |
| 140 | return "invalid PNG chunk"; |
| 141 | } |
| 142 | uint32_t chunk_len = load_u32be(p + 0); |
| 143 | uint32_t chunk_type = load_u32be(p + 4); |
| 144 | p += 8; |
| 145 | n -= 8; |
| 146 | |
| 147 | // Process the chunk payload. |
| 148 | if (n < chunk_len) { |
| 149 | return "short PNG chunk data"; |
| 150 | } |
| 151 | switch (chunk_type) { |
| 152 | case 0x49484452: // "IHDR" |
| 153 | if (chunk_len != 13) { |
| 154 | return "invalid PNG IDAT chunk"; |
| 155 | } |
| 156 | width = load_u32be(p + 0); |
| 157 | height = load_u32be(p + 4); |
| 158 | if ((width == 0) || (height == 0)) { |
| 159 | return "image dimensions are too small"; |
| 160 | } |
| 161 | if ((width > MAX_DIMENSION) || (height > MAX_DIMENSION)) { |
| 162 | return "image dimensions are too large"; |
| 163 | } |
| 164 | if (p[8] != 8) { |
| 165 | return "unsupported PNG bit depth"; |
| 166 | } |
| 167 | if (bytes_per_pixel != 0) { |
| 168 | return "duplicate PNG IHDR chunk"; |
| 169 | } |
| 170 | // Process the color type, as per the PNG spec table 11.1. |
| 171 | switch (p[9]) { |
| 172 | case 0: |
| 173 | bytes_per_pixel = 1; |
| 174 | break; |
| 175 | case 2: |
| 176 | bytes_per_pixel = 3; |
| 177 | break; |
| 178 | case 3: |
| 179 | bytes_per_pixel = 1; |
| 180 | break; |
| 181 | case 4: |
| 182 | bytes_per_pixel = 2; |
| 183 | break; |
| 184 | case 6: |
| 185 | bytes_per_pixel = 4; |
| 186 | break; |
| 187 | default: |
| 188 | return "unsupported PNG color type"; |
| 189 | } |
| 190 | if (p[12] != 0) { |
| 191 | return "unsupported PNG interlacing"; |
| 192 | } |
| 193 | break; |
| 194 | |
| 195 | case 0x49444154: // "IDAT" |
| 196 | if (num_idat_chunks == MAX_IDAT_CHUNKS - 1) { |
| 197 | return "too many IDAT chunks"; |
| 198 | } |
| 199 | memcpy(idat_buffer + idat_splits[num_idat_chunks], p, chunk_len); |
| 200 | idat_splits[num_idat_chunks + 1] = |
| 201 | idat_splits[num_idat_chunks] + chunk_len; |
| 202 | num_idat_chunks++; |
| 203 | break; |
| 204 | } |
| 205 | p += chunk_len; |
| 206 | n -= chunk_len; |
| 207 | |
| 208 | // Process (and ignore) the 4 byte chunk footer (a checksum). |
| 209 | if (n < 4) { |
| 210 | return "invalid PNG chunk"; |
| 211 | } |
| 212 | p += 4; |
| 213 | n -= 4; |
| 214 | } |
| 215 | return NULL; |
| 216 | } |
| 217 | |
| 218 | const char* decode_once(bool frag_dst, bool frag_idat) { |
Nigel Tao | a0bafff | 2018-07-14 08:59:36 +1000 | [diff] [blame] | 219 | wuffs_zlib__decoder dec = ((wuffs_zlib__decoder){}); |
| 220 | wuffs_zlib__decoder__check_wuffs_version(&dec, sizeof dec, WUFFS_VERSION); |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 221 | |
| 222 | wuffs_base__io_buffer dst = {.ptr = dst_buffer, .len = bytes_per_frame}; |
| 223 | wuffs_base__io_buffer idat = {.ptr = idat_buffer, |
| 224 | .len = SRC_BUFFER_SIZE, |
| 225 | .wi = idat_splits[num_idat_chunks], |
| 226 | .closed = true}; |
Nigel Tao | 8827e22 | 2018-04-27 20:54:44 +1000 | [diff] [blame] | 227 | wuffs_base__io_writer dst_writer = wuffs_base__io_buffer__writer(&dst); |
| 228 | wuffs_base__io_reader idat_reader = wuffs_base__io_buffer__reader(&idat); |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 229 | |
| 230 | uint32_t i = 0; // Number of dst fragments processed, if frag_dst. |
| 231 | if (frag_dst) { |
| 232 | dst.len = bytes_per_row; |
| 233 | } |
| 234 | |
| 235 | uint32_t j = 0; // Number of IDAT fragments processed, if frag_idat. |
| 236 | if (frag_idat) { |
| 237 | idat.wi = idat_splits[1]; |
| 238 | idat.closed = (num_idat_chunks == 1); |
| 239 | } |
| 240 | |
| 241 | while (true) { |
Nigel Tao | a0bafff | 2018-07-14 08:59:36 +1000 | [diff] [blame] | 242 | wuffs_base__status s = |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 243 | wuffs_zlib__decoder__decode(&dec, dst_writer, idat_reader); |
| 244 | |
Nigel Tao | a0bafff | 2018-07-14 08:59:36 +1000 | [diff] [blame] | 245 | if (s == WUFFS_BASE__STATUS_OK) { |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 246 | break; |
| 247 | } |
Nigel Tao | a0bafff | 2018-07-14 08:59:36 +1000 | [diff] [blame] | 248 | if ((s == WUFFS_BASE__SUSPENSION_SHORT_WRITE) && frag_dst && |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 249 | (i < height - 1)) { |
| 250 | i++; |
| 251 | dst.len = bytes_per_row * (i + 1); |
| 252 | continue; |
| 253 | } |
Nigel Tao | a0bafff | 2018-07-14 08:59:36 +1000 | [diff] [blame] | 254 | if ((s == WUFFS_BASE__SUSPENSION_SHORT_READ) && frag_idat && |
Nigel Tao | d2075ce | 2018-04-25 15:26:29 +1000 | [diff] [blame] | 255 | (j < num_idat_chunks - 1)) { |
| 256 | j++; |
| 257 | idat.wi = idat_splits[j + 1]; |
| 258 | idat.closed = (num_idat_chunks == j + 1); |
| 259 | continue; |
| 260 | } |
| 261 | return wuffs_zlib__status__string(s); |
| 262 | } |
| 263 | |
| 264 | if (dst.wi != bytes_per_frame) { |
| 265 | return "unexpected number of bytes decoded"; |
| 266 | } |
| 267 | return NULL; |
| 268 | } |
| 269 | |
| 270 | const char* decode(bool frag_dst, bool frag_idat) { |
| 271 | int reps; |
| 272 | if (bytes_per_frame < 100000) { |
| 273 | reps = 1000; |
| 274 | } else if (bytes_per_frame < 1000000) { |
| 275 | reps = 100; |
| 276 | } else if (bytes_per_frame < 10000000) { |
| 277 | reps = 10; |
| 278 | } else { |
| 279 | reps = 1; |
| 280 | } |
| 281 | |
| 282 | struct timeval bench_start_tv; |
| 283 | gettimeofday(&bench_start_tv, NULL); |
| 284 | |
| 285 | int i; |
| 286 | for (i = 0; i < reps; i++) { |
| 287 | const char* msg = decode_once(frag_dst, frag_idat); |
| 288 | if (msg) { |
| 289 | return msg; |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | struct timeval bench_finish_tv; |
| 294 | gettimeofday(&bench_finish_tv, NULL); |
| 295 | int64_t micros = |
| 296 | (int64_t)(bench_finish_tv.tv_sec - bench_start_tv.tv_sec) * 1000000 + |
| 297 | (int64_t)(bench_finish_tv.tv_usec - bench_start_tv.tv_usec); |
| 298 | uint64_t nanos = 1; |
| 299 | if (micros > 0) { |
| 300 | nanos = (uint64_t)(micros)*1000; |
| 301 | } |
| 302 | |
| 303 | printf("Benchmark%sDst%sIDAT/%s\t%8d\t%8" PRIu64 " ns/op\n", |
| 304 | frag_dst ? "Frag" : "Full", // |
| 305 | frag_idat ? "Frag" : "Full", // |
| 306 | cc, reps, nanos / reps); |
| 307 | |
| 308 | return NULL; |
| 309 | } |
| 310 | |
| 311 | int fail(const char* msg) { |
| 312 | const int stderr_fd = 2; |
| 313 | write(stderr_fd, msg, strnlen(msg, 4095)); |
| 314 | write(stderr_fd, "\n", 1); |
| 315 | return 1; |
| 316 | } |
| 317 | |
| 318 | int main(int argc, char** argv) { |
| 319 | const char* msg = read_stdin(); |
| 320 | if (msg) { |
| 321 | return fail(msg); |
| 322 | } |
| 323 | if ((src_len < 8) || strncmp(src_buffer, "\x89PNG\x0D\x0A\x1A\x0A", 8)) { |
| 324 | return fail("invalid PNG"); |
| 325 | } |
| 326 | msg = process_png_chunks(src_buffer + 8, src_len - 8); |
| 327 | if (msg) { |
| 328 | return fail(msg); |
| 329 | } |
| 330 | if (bytes_per_pixel == 0) { |
| 331 | return fail("missing PNG IHDR chunk"); |
| 332 | } |
| 333 | if (num_idat_chunks == 0) { |
| 334 | return fail("missing PNG IDAT chunk"); |
| 335 | } |
| 336 | // The +1 here is for the per-row filter byte. |
| 337 | bytes_per_row = (uint64_t)width * bytes_per_pixel + 1; |
| 338 | bytes_per_frame = (uint64_t)height * bytes_per_row; |
| 339 | if (bytes_per_frame > DST_BUFFER_SIZE) { |
| 340 | return fail("decompressed data is too large"); |
| 341 | } |
| 342 | |
| 343 | printf("# %s version %s\n#\n", cc, cc_version); |
| 344 | printf( |
| 345 | "# The output format, including the \"Benchmark\" prefixes, is " |
| 346 | "compatible with the\n" |
| 347 | "# https://godoc.org/golang.org/x/perf/cmd/benchstat tool. To install " |
| 348 | "it, first\n" |
| 349 | "# install Go, then run \"go get golang.org/x/perf/cmd/benchstat\".\n"); |
| 350 | |
| 351 | int i; |
| 352 | for (i = 0; i < 5; i++) { |
| 353 | msg = decode(true, true); |
| 354 | if (msg) { |
| 355 | return fail(msg); |
| 356 | } |
| 357 | msg = decode(true, false); |
| 358 | if (msg) { |
| 359 | return fail(msg); |
| 360 | } |
| 361 | msg = decode(false, true); |
| 362 | if (msg) { |
| 363 | return fail(msg); |
| 364 | } |
| 365 | msg = decode(false, false); |
| 366 | if (msg) { |
| 367 | return fail(msg); |
| 368 | } |
| 369 | } |
| 370 | |
| 371 | return 0; |
| 372 | } |