| FazBrowse GitHub Viewer | Trending | | Home |
| Tools: [Download Repo ZIP] [Original HTTPS Page] |
1 parent e7db639 commit d0c23f3
21 files changed
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
@@ -24,6 +24,7 @@ check_include_file(stddef.h HAVE_STDDEF_H) | |||
| 24 | 24 | option(ENABLE_SIMD_OPTIMIZATIONS "Enable all SIMD optimizations" OFF) | |
| 25 | 25 | option(ENABLE_SIMD_AVX512 "Enable SIMD AXV512 optimizations" OFF) | |
| 26 | 26 | option(USE_ZLIB_RABIN_KARP_HASH "Enable bitstream compatibility with canonical zlib" OFF) | |
| 27 | + option(ENABLE_INTEL_QAT_COMPRESSION "Enable Intel Quick Assist Technology use for compression" OFF) | ||
| 27 | 28 | option(BUILD_UNITTESTS "Enable standalone unit tests build" OFF) | |
| 28 | 29 | option(BUILD_MINIZIP_BIN "Enable building minzip_bin tool" OFF) | |
| 29 | 30 | option(BUILD_ZPIPE "Enable building zpipe tool" OFF) | |
@@ -228,6 +229,22 @@ if (ENABLE_SIMD_OPTIMIZATIONS) | |||
| 228 | 229 | endif() | |
| 229 | 230 | endif() | |
| 230 | 231 | ||
| 232 | + if (ENABLE_INTEL_QAT_COMPRESSION) | ||
| 233 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/deflate_qat.cpp) | ||
| 234 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/io_buffers.cpp) | ||
| 235 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/memory.cpp) | ||
| 236 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/qat_buffer_list.cpp) | ||
| 237 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/qat.cpp) | ||
| 238 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/qat_instance.cpp) | ||
| 239 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/session.cpp) | ||
| 240 | + list(APPEND ZLIB_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp/qat_task.cpp) | ||
| 241 | + | ||
| 242 | + # TODO(gustavoa): Find a way to include the qatzpp headers without having the | ||
| 243 | + # presubmit check throw errors. | ||
| 244 | + include_directories(${CMAKE_CURRENT_SOURCE_DIR}/contrib/qat/qatzpp) | ||
| 245 | + add_compile_definitions(QAT_COMPRESSION_ENABLED) | ||
| 246 | + endif() | ||
| 247 | + | ||
| 231 | 248 | # parse the full version number from zlib.h and include in ZLIB_FULL_VERSION | |
| 232 | 249 | file(READ ${CMAKE_CURRENT_SOURCE_DIR}/zlib.h _zlib_h_contents) | |
| 233 | 250 | string(REGEX REPLACE ".*#define[ \t]+ZLIB_VERSION[ \t]+\"([-0-9A-Za-z.]+)\".*" | |
@@ -254,6 +271,15 @@ add_library(zlibstatic STATIC ${ZLIB_SRCS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HD | |||
| 254 | 271 | set_target_properties(zlib PROPERTIES DEFINE_SYMBOL ZLIB_DLL) | |
| 255 | 272 | set_target_properties(zlib PROPERTIES SOVERSION 1) | |
| 256 | 273 | ||
| 274 | + if (ENABLE_INTEL_QAT_COMPRESSION) | ||
| 275 | + target_include_directories(zlib PUBLIC ${QATZPP_INCLUDE_DIRS}) | ||
| 276 | + target_link_libraries(zlib ${QATZPP_LIBRARY}) | ||
| 277 | + target_link_libraries(zlib qat) | ||
| 278 | + target_include_directories(zlibstatic PUBLIC ${QATZPP_INCLUDE_DIRS}) | ||
| 279 | + target_link_libraries(zlibstatic ${QATZPP_LIBRARY}) | ||
| 280 | + target_link_libraries(zlibstatic qat) | ||
| 281 | + endif() | ||
| 282 | + | ||
| 257 | 283 | if(NOT CYGWIN) | |
| 258 | 284 | # This property causes shared libraries on Linux to have the full version | |
| 259 | 285 | # encoded into their final filename. We disable this on Cygwin because | |
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
@@ -0,0 +1,312 @@ | |||
| 1 | + /* | ||
| 2 | + * Copyright (C) 2024 Intel Corporation. All rights reserved. | ||
| 3 | + * Authors: | ||
| 4 | + * Gustavo A Espinoza <gustavo.adolfo.espinoza.quintero@intel.com> | ||
| 5 | + * <gustavoaespinozaq@hotmail.com> | ||
| 6 | + * | ||
| 7 | + * For conditions of distribution and use, see copyright notice in zlib.h | ||
| 8 | + */ | ||
| 9 | + #include "deflate_qat.h" | ||
| 10 | + #include "deflate.h" | ||
| 11 | + | ||
| 12 | + #include "session.hpp" | ||
| 13 | + #include "qat_instance.hpp" | ||
| 14 | + #include "qat_buffer_list.hpp" | ||
| 15 | + #include "qat.hpp" | ||
| 16 | + | ||
| 17 | + #include <memory> | ||
| 18 | + | ||
| 19 | + /* | ||
| 20 | + * TODO(gustavoa): Make the input size adjustable from the memlevel | ||
| 21 | + * attribute on deflateInit. | ||
| 22 | + */ | ||
| 23 | + static constexpr size_t kInputSize = 1024 * 1024; | ||
| 24 | + | ||
| 25 | + /* QAT Instances obtained available from the library. */ | ||
| 26 | + static std::vector<std::shared_ptr<qat::Instance>> qat_instances; | ||
| 27 | + | ||
| 28 | + /* | ||
| 29 | + * TODO(gustavoa): Verify if the ordering of the struct fields won't create | ||
| 30 | + * unnecessary holes in the structure that requires extraneous padding. | ||
| 31 | + */ | ||
| 32 | + struct qat_deflate { | ||
| 33 | + std::unique_ptr<qat::DeflateSession> qat_session; | ||
| 34 | + | ||
| 35 | + /* QAT requires contiguous physical pages. Cannot be allocated using | ||
| 36 | + * malloc/new. | ||
| 37 | + */ | ||
| 38 | + uint8_t *input_buffer; | ||
| 39 | + uint8_t *output_buffer; | ||
| 40 | + | ||
| 41 | + /* Pointer to the next byte in the output buffer. */ | ||
| 42 | + uint8_t *pending_out; | ||
| 43 | + | ||
| 44 | + unsigned input_buffer_size; | ||
| 45 | + unsigned output_buffer_size; | ||
| 46 | + | ||
| 47 | + unsigned pending_in_count; | ||
| 48 | + unsigned pending_out_count; | ||
| 49 | + }; | ||
| 50 | + | ||
| 51 | + static std::unique_ptr<qat::DeflateSession> qat_create_session(int level, int wrap) | ||
| 52 | + { | ||
| 53 | + CpaDcChecksum checksum = CPA_DC_NONE; | ||
| 54 | + | ||
| 55 | + switch(wrap) { | ||
| 56 | + case 1: | ||
| 57 | + checksum = CPA_DC_ADLER32; | ||
| 58 | + break; | ||
| 59 | + case 2: | ||
| 60 | + checksum = CPA_DC_CRC32; | ||
| 61 | + break; | ||
| 62 | + } | ||
| 63 | + | ||
| 64 | + return std::make_unique<qat::DeflateSession>( | ||
| 65 | + qat_instances[0], | ||
| 66 | + (CpaDcCompLvl)level, | ||
| 67 | + checksum, | ||
| 68 | + 0 | ||
| 69 | + ); | ||
| 70 | + } | ||
| 71 | + | ||
| 72 | + | ||
| 73 | + int qat_deflate_init() | ||
| 74 | + { | ||
| 75 | + return (qat::Initialize()) ? Z_ERRNO : Z_OK; | ||
| 76 | + } | ||
| 77 | + | ||
| 78 | + struct qat_deflate* qat_deflate_state_init(int level, int wrap) | ||
| 79 | + { | ||
| 80 | + if (qat_instances.empty()) { | ||
| 81 | + qat_instances = qat::Instance::Create(); | ||
| 82 | + } | ||
| 83 | + if (qat_instances.empty()) { | ||
| 84 | + return nullptr; | ||
| 85 | + } | ||
| 86 | + | ||
| 87 | + struct qat_deflate *qat_deflate = new struct qat_deflate; | ||
| 88 | + if (!qat_deflate) { | ||
| 89 | + return nullptr; | ||
| 90 | + } | ||
| 91 | + | ||
| 92 | + /* TODO(gustavoa): Find a way to utilize all the available instances for the same | ||
| 93 | + * process. | ||
| 94 | + */ | ||
| 95 | + qat_instances[0]->Start(); | ||
| 96 | + | ||
| 97 | + qat_deflate->qat_session = qat_create_session(level, wrap); | ||
| 98 | + | ||
| 99 | + qat_deflate->input_buffer_size = kInputSize; | ||
| 100 | + qat_deflate->input_buffer = qat::AllocBlockArray<uint8_t>(kInputSize, 0); | ||
| 101 | + qat_deflate->output_buffer_size = | ||
| 102 | + qat_deflate->qat_session->GetDeflateBound(qat_deflate->input_buffer_size); | ||
| 103 | + qat_deflate->pending_out = qat_deflate->output_buffer = | ||
| 104 | + qat::AllocBlockArray<uint8_t>(qat_deflate->output_buffer_size, 0); | ||
| 105 | + | ||
| 106 | + qat_deflate->pending_in_count = qat_deflate->pending_out_count = 0; | ||
| 107 | + | ||
| 108 | + if (!qat_deflate->input_buffer || !qat_deflate->output_buffer) { | ||
| 109 | + return nullptr; | ||
| 110 | + } | ||
| 111 | + | ||
| 112 | + return qat_deflate; | ||
| 113 | + } | ||
| 114 | + | ||
| 115 | + static unsigned qat_read_buf(z_streamp strm, struct qat_deflate* qat, unsigned size) | ||
| 116 | + { | ||
| 117 | + unsigned len = strm->avail_in; | ||
| 118 | + | ||
| 119 | + if (len > size) { | ||
| 120 | + len = size; | ||
| 121 | + } | ||
| 122 | + if (len == 0) return 0; | ||
| 123 | + | ||
| 124 | + strm->avail_in -= len; | ||
| 125 | + strm->total_in += len; | ||
| 126 | + | ||
| 127 | + zmemcpy( | ||
| 128 | + qat->input_buffer + qat->pending_in_count, | ||
| 129 | + strm->next_in, | ||
| 130 | + len | ||
| 131 | + ); | ||
| 132 | + | ||
| 133 | + strm->next_in += len; | ||
| 134 | + qat->pending_in_count += len; | ||
| 135 | + | ||
| 136 | + return len; | ||
| 137 | + } | ||
| 138 | + | ||
| 139 | + void qat_flush_pending(deflate_state* s) | ||
| 140 | + { | ||
| 141 | + unsigned len; | ||
| 142 | + z_streamp strm = s->strm; | ||
| 143 | + struct qat_deflate* qat = s->qat_s; | ||
| 144 | + | ||
| 145 | + len = qat->pending_out_count; | ||
| 146 | + if (len > strm->avail_out) len = strm->avail_out; | ||
| 147 | + if (len == 0) return; | ||
| 148 | + | ||
| 149 | + zmemcpy(strm->next_out, qat->pending_out, len); | ||
| 150 | + | ||
| 151 | + qat->pending_out += len; | ||
| 152 | + qat->pending_out_count -= len; | ||
| 153 | + strm->next_out += len; | ||
| 154 | + strm->avail_out -= len; | ||
| 155 | + strm->total_out += len; | ||
| 156 | + if (qat->pending_out_count == 0) { | ||
| 157 | + qat->pending_out = qat->output_buffer; | ||
| 158 | + } | ||
| 159 | + } | ||
| 160 | + | ||
| 161 | + static int qat_compress_pending(deflate_state*s, int flush) | ||
| 162 | + { | ||
| 163 | + struct qat_deflate* qat = s->qat_s; | ||
| 164 | + uint32_t metadata_size; | ||
| 165 | + | ||
| 166 | + /* TODO(gustavoa): find a way to make qatzpp setup this number internally. */ | ||
| 167 | + cpaDcBufferListGetMetaSize(qat->qat_session->getInstance()->GetHandle(), 1, &metadata_size); | ||
| 168 | + | ||
| 169 | + auto job = qat->qat_session->Deflate( | ||
| 170 | + std::make_unique<qat::IOBuffers>( | ||
| 171 | + std::make_unique<qat::BufferListUser>( | ||
| 172 | + qat->input_buffer, | ||
| 173 | + qat->pending_in_count, | ||
| 174 | + metadata_size | ||
| 175 | + ), | ||
| 176 | + std::make_unique<qat::BufferListUser>( | ||
| 177 | + qat->output_buffer, | ||
| 178 | + qat->output_buffer_size, | ||
| 179 | + metadata_size | ||
| 180 | + ) | ||
| 181 | + ), (flush == Z_FINISH && s->strm->avail_in == 0) | ||
| 182 | + ); | ||
| 183 | + | ||
| 184 | + job->WaitCompletion(); | ||
| 185 | + | ||
| 186 | + /* | ||
| 187 | + * TODO(gustavoa): make QAT perform the checksum combine. | ||
| 188 | + */ | ||
| 189 | + if (s->wrap == 2) { | ||
| 190 | + s->strm->adler = crc32_combine( | ||
| 191 | + s->strm->adler, | ||
| 192 | + job->GetResults()->checksum, | ||
| 193 | + job->GetResults()->consumed | ||
| 194 | + ); | ||
| 195 | + } else if (s->wrap == 1) { | ||
| 196 | + s->strm->adler = adler32( | ||
| 197 | + s->strm->adler, | ||
| 198 | + qat->input_buffer, | ||
| 199 | + job->GetResults()->consumed | ||
| 200 | + ); | ||
| 201 | + } | ||
| 202 | + | ||
| 203 | + qat->pending_out_count = job->GetResults()->produced; | ||
| 204 | + qat->pending_in_count -= job->GetResults()->consumed; | ||
| 205 | + | ||
| 206 | + if(qat->pending_in_count != 0) { | ||
| 207 | + /* Copy any remaining bytes to the beginning of the buffer. */ | ||
| 208 | + zmemcpy( | ||
| 209 | + qat->input_buffer, | ||
| 210 | + qat->input_buffer + job->GetResults()->consumed, | ||
| 211 | + qat->pending_in_count | ||
| 212 | + ); | ||
| 213 | + } | ||
| 214 | + | ||
| 215 | + return 0; | ||
| 216 | + } | ||
| 217 | + | ||
| 218 | + qat_block_state qat_deflate_step(deflate_state* s, int flush) | ||
| 219 | + { | ||
| 220 | + z_streamp strm = s->strm; | ||
| 221 | + struct qat_deflate* qat_state = s->qat_s; | ||
| 222 | + | ||
| 223 | + for (;;) { | ||
| 224 | + if (qat_state->pending_in_count < qat_state->input_buffer_size) { | ||
| 225 | + qat_read_buf( | ||
| 226 | + strm, | ||
| 227 | + qat_state, | ||
| 228 | + qat_state->input_buffer_size - qat_state->pending_in_count | ||
| 229 | + ); | ||
| 230 | + if (qat_state->pending_in_count < qat_state->input_buffer_size && flush == Z_NO_FLUSH) { | ||
| 231 | + return qat_block_need_more; | ||
| 232 | + } else { | ||
| 233 | + qat_compress_pending(s, flush); | ||
| 234 | + } | ||
| 235 | + if (strm->avail_in == 0) { | ||
| 236 | + break; | ||
| 237 | + } | ||
| 238 | + } else { | ||
| 239 | + qat_compress_pending(s, flush); | ||
| 240 | + } | ||
| 241 | + | ||
| 242 | + qat_flush_pending(s); | ||
| 243 | + if (strm->avail_out == 0) { | ||
| 244 | + return (flush == Z_FINISH) ? qat_block_finish_started : qat_block_need_more; | ||
| 245 | + } | ||
| 246 | + } | ||
| 247 | + | ||
| 248 | + if (flush == Z_FINISH) { | ||
| 249 | + qat_flush_pending(s); | ||
| 250 | + if (strm->avail_out == 0) { | ||
| 251 | + return qat_block_finish_started; | ||
| 252 | + } else { | ||
| 253 | + return qat_block_finish_done; | ||
| 254 | + } | ||
| 255 | + } | ||
| 256 | + | ||
| 257 | + qat_flush_pending(s); | ||
| 258 | + if (strm->avail_out == 0) { | ||
| 259 | + return qat_block_done; | ||
| 260 | + } | ||
| 261 | + | ||
| 262 | + return qat_block_need_more; | ||
| 263 | + } | ||
| 264 | + | ||
| 265 | + int qat_deflate_state_free(deflate_state* s) | ||
| 266 | + { | ||
| 267 | + struct qat_deflate* qat_state = s->qat_s; | ||
| 268 | + if (qat_state->input_buffer) { | ||
| 269 | + qat::Free(qat_state->input_buffer); | ||
| 270 | + } | ||
| 271 | + if (qat_state->output_buffer) { | ||
| 272 | + qat::Free(qat_state->output_buffer); | ||
| 273 | + } | ||
| 274 | + | ||
| 275 | + qat_state->qat_session.reset(); | ||
| 276 | + delete qat_state; | ||
| 277 | + s->qat_s = nullptr; | ||
| 278 | + | ||
| 279 | + return Z_OK; | ||
| 280 | + } | ||
| 281 | + | ||
| 282 | + struct qat_deflate *qat_deflate_copy(deflate_state *ss) | ||
| 283 | + { | ||
| 284 | + struct qat_deflate *sqat = ss->qat_s; | ||
| 285 | + struct qat_deflate *dqat = nullptr; | ||
| 286 | + | ||
| 287 | + if (!sqat) { | ||
| 288 | + return nullptr; | ||
| 289 | + } | ||
| 290 | + | ||
| 291 | + dqat = new struct qat_deflate; | ||
| 292 | + | ||
| 293 | + dqat->qat_session = qat_create_session(ss->level, ss->wrap); | ||
| 294 | + | ||
| 295 | + dqat->input_buffer_size = sqat->input_buffer_size; | ||
| 296 | + dqat->input_buffer = qat::AllocBlockArray<uint8_t>(dqat->input_buffer_size, 0); | ||
| 297 | + | ||
| 298 | + dqat->output_buffer_size = sqat->output_buffer_size; | ||
| 299 | + dqat->output_buffer = qat::AllocBlockArray<uint8_t>(dqat->output_buffer_size, 0); | ||
| 300 | + | ||
| 301 | + dqat->pending_in_count = sqat->pending_in_count; | ||
| 302 | + dqat->pending_out_count = sqat->pending_out_count; | ||
| 303 | + | ||
| 304 | + dqat->pending_out = | ||
| 305 | + dqat->output_buffer + (sqat->pending_out - sqat->output_buffer); | ||
| 306 | + | ||
| 307 | + zmemcpy(dqat->input_buffer, sqat->input_buffer, dqat->input_buffer_size); | ||
| 308 | + zmemcpy(dqat->output_buffer, sqat->output_buffer, dqat->output_buffer_size); | ||
| 309 | + | ||
| 310 | + return dqat; | ||
| 311 | + } | ||
| 312 | + | ||
| Back | FazBrowse Home | New Git URL |
0 commit comments