From: Yann Collet <cyan@fb.com>
Date: Wed, 18 Dec 2024 01:25:57 +0000 (-0800)
Subject: added benchmark for ZSTD_convertBlockSequences_wBlockDelim()
X-Git-Tag: v1.5.7^2~48^2~15
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=95ad9e47ffa930d2facb5d2dd2de511bf8171e5d;p=thirdparty%2Fzstd.git

added benchmark for ZSTD_convertBlockSequences_wBlockDelim()
---

diff --git a/lib/compress/zstd_compress.c b/lib/compress/zstd_compress.c
index 3c5cb209c..c4f699eb0 100644
--- a/lib/compress/zstd_compress.c
+++ b/lib/compress/zstd_compress.c
@@ -3240,12 +3240,6 @@ static size_t ZSTD_fastSequenceLengthSum(ZSTD_Sequence const* seqBuf, size_t seq
     return litLenSum + matchLenSum;
 }
 
-typedef struct {
-    U32 idx;            /* Index in array of ZSTD_Sequence */
-    U32 posInSequence;  /* Position within sequence at idx */
-    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
-} ZSTD_SequencePosition;
-
 static size_t
 ZSTD_transferSequences_wBlockDelim(ZSTD_CCtx* cctx,
                                    ZSTD_SequencePosition* seqPos,
@@ -7103,7 +7097,7 @@ size_t ZSTD_compressSequences(ZSTD_CCtx* cctx,
  * @blockSize must be exactly correct.
  */
 FORCE_INLINE_TEMPLATE size_t
-ZSTD_convertSequences_wBlockDelim_internal(ZSTD_CCtx* cctx,
+ZSTD_convertBlockSequences_wBlockDelim_internal(ZSTD_CCtx* cctx,
                         ZSTD_SequencePosition* seqPos,
                         const ZSTD_Sequence* const inSeqs, size_t nbSequences,
                         size_t blockSize,
@@ -7200,32 +7194,39 @@ ZSTD_convertSequences_wBlockDelim_internal(ZSTD_CCtx* cctx,
     return litConsumed;
 }
 
-typedef size_t (*ZSTD_transferSequencesOnly_f) (ZSTD_CCtx* cctx,
+/* for tests only */
+void CCTX_resetSeqStore(ZSTD_CCtx* cctx)
+{
+    cctx->seqStore.sequences = cctx->seqStore.sequencesStart;
+    cctx->seqStore.lit = cctx->seqStore.litStart;
+}
+
+typedef size_t (*ZSTD_convertBlockSequences_f) (ZSTD_CCtx* cctx,
                         ZSTD_SequencePosition* seqPos,
                         const ZSTD_Sequence* const inSeqs, size_t nbSequences,
                         size_t blockSize,
                         int const repcodeResolution);
 
-static size_t
-ZSTD_convertSequences_wBlockDelim(ZSTD_CCtx* cctx,
+size_t
+ZSTD_convertBlockSequences_wBlockDelim(ZSTD_CCtx* cctx,
                         ZSTD_SequencePosition* seqPos,
                         const ZSTD_Sequence* const inSeqs, size_t nbSequences,
                         size_t blockSize,
                         int const repcodeResolution)
 {
-    return ZSTD_convertSequences_wBlockDelim_internal(cctx,
+    return ZSTD_convertBlockSequences_wBlockDelim_internal(cctx,
                 seqPos, inSeqs, nbSequences, blockSize,
                 repcodeResolution, 0);
 }
 
 static size_t
-ZSTD_convertSequences_wBlockDelim_andCheckSequences(ZSTD_CCtx* cctx,
+ZSTD_convertBlockSequences_wBlockDelim_andCheckSequences(ZSTD_CCtx* cctx,
                         ZSTD_SequencePosition* seqPos,
                         const ZSTD_Sequence* const inSeqs, size_t nbSequences,
                         size_t blockSize,
                         int const repcodeResolution)
 {
-    return ZSTD_convertSequences_wBlockDelim_internal(cctx,
+    return ZSTD_convertBlockSequences_wBlockDelim_internal(cctx,
                 seqPos, inSeqs, nbSequences, blockSize,
                 repcodeResolution, 1);
 }
@@ -7241,9 +7242,9 @@ ZSTD_compressSequencesAndLiterals_internal(ZSTD_CCtx* cctx,
     ZSTD_SequencePosition seqPos = {0, 0, 0};
     BYTE* op = (BYTE*)dst;
     int const repcodeResolution = (cctx->appliedParams.searchForExternalRepcodes == ZSTD_ps_enable);
-    const ZSTD_transferSequencesOnly_f transfer = cctx->appliedParams.validateSequences ?
-            ZSTD_convertSequences_wBlockDelim_andCheckSequences
-            : ZSTD_convertSequences_wBlockDelim;
+    const ZSTD_convertBlockSequences_f convertBlockSequences = cctx->appliedParams.validateSequences ?
+            ZSTD_convertBlockSequences_wBlockDelim_andCheckSequences
+            : ZSTD_convertBlockSequences_wBlockDelim;
 
     DEBUGLOG(4, "ZSTD_compressSequencesAndLiterals_internal: nbSeqs=%zu, litSize=%zu", nbSequences, litSize);
     if (cctx->appliedParams.blockDelimiters == ZSTD_sf_noBlockDelimiters) {
@@ -7271,7 +7272,7 @@ ZSTD_compressSequencesAndLiterals_internal(ZSTD_CCtx* cctx,
         assert(blockSize <= remaining);
         ZSTD_resetSeqStore(&cctx->seqStore);
 
-        litConsumed = transfer(cctx,
+        litConsumed = convertBlockSequences(cctx,
                             &seqPos,
                             inSeqs, nbSequences,
                             blockSize,
diff --git a/lib/compress/zstd_compress_internal.h b/lib/compress/zstd_compress_internal.h
index 10bd8ca70..bbcf4a7d5 100644
--- a/lib/compress/zstd_compress_internal.h
+++ b/lib/compress/zstd_compress_internal.h
@@ -1522,6 +1522,22 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
 
 void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);
 
+typedef struct {
+    U32 idx;            /* Index in array of ZSTD_Sequence */
+    U32 posInSequence;  /* Position within sequence at idx */
+    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
+} ZSTD_SequencePosition;
+
+size_t
+ZSTD_convertBlockSequences_wBlockDelim(ZSTD_CCtx* cctx,
+                        ZSTD_SequencePosition* seqPos,
+                        const ZSTD_Sequence* const inSeqs, size_t nbSequences,
+                        size_t blockSize,
+                        int const repcodeResolution);
+
+/* for tests only */
+void CCTX_resetSeqStore(ZSTD_CCtx* cctx);
+
 /* ==============================================================
  * Private declarations
  * These prototypes shall only be called from within lib/compress
diff --git a/tests/fullbench.c b/tests/fullbench.c
index e89a2b326..c42c923c1 100644
--- a/tests/fullbench.c
+++ b/tests/fullbench.c
@@ -21,6 +21,7 @@
 #include <assert.h>
 
 #include "mem.h"         /* U32 */
+#include "compress/zstd_compress_internal.h"
 #ifndef ZSTD_DLL_IMPORT
     #include "zstd_internal.h"   /* ZSTD_decodeSeqHeaders, ZSTD_blockHeaderSize, ZSTD_getcBlockSize, blockType_e, KB, MB */
     #include "decompress/zstd_decompress_internal.h"   /* ZSTD_DCtx struct */
@@ -627,6 +628,64 @@ local_compressSequencesAndLiterals(const void* input, size_t inputSize,
     return ZSTD_compressSequencesAndLiterals(g_zcc, dst, dstCapacity, seqs, nbSeqs, literals, nbLiterals, srcSize);
 }
 
+static PrepResult prepConvertSequences(const void* src, size_t srcSize, int cLevel)
+{
+    PrepResult r = PREPRESULT_INIT;
+    size_t const prepCapacity = srcSize * 4;
+    void* prepBuffer = malloc(prepCapacity);
+    void* sequencesStart = (char*)prepBuffer + 2*sizeof(unsigned);
+    ZSTD_Sequence* const seqs = sequencesStart;
+    size_t const seqsCapacity = prepCapacity / sizeof(ZSTD_Sequence);
+    size_t totalNbSeqs, nbSeqs, blockSize=0;
+    ZSTD_CCtx_reset(g_zcc, ZSTD_reset_session_and_parameters);
+    ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_compressionLevel, cLevel);
+    totalNbSeqs = ZSTD_generateSequences(g_zcc, seqs, seqsCapacity, src, srcSize);
+    CONTROL(!ZSTD_isError(totalNbSeqs));
+    /* find nb sequences in first block */
+    {   size_t n;
+        for (n=0; n<totalNbSeqs; n++) {
+            if (seqs[n].matchLength == 0) break;
+            blockSize += seqs[n].litLength + seqs[n].matchLength;
+        }
+        blockSize += seqs[n].litLength;
+        nbSeqs = n+1;
+#if 0
+        printf("found %zu sequences representing a first block of size %zu\n", nbSeqs, blockSize);
+#endif
+    }
+    /* generate benchmarked input */
+    CONTROL(blockSize < UINT_MAX);
+    MEM_write32(prepBuffer, (U32)blockSize);
+    MEM_write32((char*)prepBuffer+4, (U32)nbSeqs);
+    memcpy(seqs + nbSeqs, src, srcSize);
+    r.prepBuffer = prepBuffer;
+    r.prepSize = 8 + sizeof(ZSTD_Sequence)*nbSeqs + srcSize;
+    r.fixedOrigSize = blockSize;
+    return r;
+}
+
+static size_t
+local_convertSequences(const void* input, size_t inputSize,
+                       void* dst, size_t dstCapacity,
+                       void* payload)
+{
+    ZSTD_SequencePosition seqPos = { 0, 0 , 0 };
+    const char* ip = input;
+    size_t const blockSize = MEM_read32(ip);
+    size_t const nbSeqs = MEM_read32(ip+=4);
+    const ZSTD_Sequence* seqs = (const ZSTD_Sequence*)(const void*)(ip+=4);
+    ZSTD_CCtx_reset(g_zcc, ZSTD_reset_session_and_parameters);
+    CCTX_resetSeqStore(g_zcc);
+    ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_blockDelimiters, ZSTD_sf_explicitBlockDelimiters);
+# if 0 /* for tests */
+    ZSTD_CCtx_setParameter(g_zcc, ZSTD_c_repcodeResolution, ZSTD_ps_enable);
+#endif
+    assert(8 + nbSeqs * sizeof(ZSTD_Sequence) == inputSize); (void)inputSize;
+    (void)dst; (void)dstCapacity;
+    (void)payload;
+
+    return ZSTD_convertBlockSequences_wBlockDelim(g_zcc, &seqPos, seqs, nbSeqs, blockSize, 0);
+}
 
 
 static PrepResult prepCopy(const void* src, size_t srcSize, int cLevel)
@@ -684,6 +743,7 @@ static BenchScenario kScenarios[] = {
     { "compressStream2, -T2, end", NULL, local_ZSTD_compress_generic_T2_end },
     { "compressSequences", prepSequences, local_compressSequences },
     { "compressSequencesAndLiterals", prepSequencesAndLiterals, local_compressSequencesAndLiterals },
+    { "convertSequences (1st block)", prepConvertSequences, local_convertSequences },
 #ifndef ZSTD_DLL_IMPORT
     { "decodeLiteralsHeader (1st block)", prepLiterals, local_ZSTD_decodeLiteralsHeader },
     { "decodeLiteralsBlock (1st block)", prepLiterals, local_ZSTD_decodeLiteralsBlock },