Hacking.

fast-pack · Aug 25, 2016 · 88eddb0 · 88eddb0
1 parent ee32efc
commit 88eddb0
Show file tree

Hide file tree

Showing 3 changed files with 2,116 additions and 5,712 deletions.
diff --git a/scripts/avx512dict.py b/scripts/avx512dict.py
@@ -74,16 +74,16 @@ def plurial(number):
           if(firstshift + bit <> 32):
             wfirst  = maskstr.format(wfirst)
           print("  wout = {0}; // 512-bit word to be output".format(wfirst));
-          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(dictionary,_mm512_castsi512_si256(wout), 8)); // load from dictionary and store".format(2*j))
-          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(dictionary,_mm512_extracti64x4_epi64(wout,1), 8)); // load from dictionary and store".format(2*j+1))
+          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(_mm512_castsi512_si256(wout),dictionary, 8)); // load from dictionary and store".format(2*j))
+          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(_mm512_extracti64x4_epi64(wout,1),dictionary, 8)); // load from dictionary and store".format(2*j+1))
       else:
           secondshift = (32-firstshift)
           wsecond = "_mm512_slli_epi32( w{0} , {1} ) ".format((firstword+1)%2,secondshift)
           wfirstorsecond = " _mm512_or_si512 ({0},{1}) ".format(wfirst,wsecond)
           wfirstorsecond = maskstr.format(wfirstorsecond)
           print("  wout = {0}; // 512-bit word to be output".format(wfirstorsecond));
-          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(dictionary,_mm512_castsi512_si256(wout), 8)); // load from dictionary and store".format(2*j))
-          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(dictionary,_mm512_extracti64x4_epi64(wout,1), 8)); // load from dictionary and store".format(2*j+1))
+          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(_mm512_castsi512_si256(wout),dictionary, 8)); // load from dictionary and store".format(2*j))
+          print("  _mm512_storeu_si512(out + {0},_mm512_i32gather_epi64(_mm512_extracti64x4_epi64(wout,1),dictionary, 8)); // load from dictionary and store".format(2*j+1))
     print("}");
     print("")
 

diff --git a/src/avx512bpacking.h b/src/avx512bpacking.h
@@ -7770,18 +7770,21 @@ static avx512unpackblockfnc avx512funcUnpackArr[] = {
 
 
 /* reads 512 values from "in", writes  "bit" 512-bit vectors to "out" */
-void avx512pack(const uint32_t *  in,__m512i *  out, const uint32_t bit) {
-	avx512funcPackMaskArr[bit](in,out);
+void avx512pack(const uint32_t *  in,__m512i *  out, const uint32_t number, const uint32_t bit) {
+    for(uint32_t i = 0; i < number / 512; ++i)
+        avx512funcPackMaskArr[bit](in + i * 512,out + i * bit);
 }
 
 /* reads 512 values from "in", writes  "bit" 512-bit vectors to "out" */
-void avx512packwithoutmask(const uint32_t *  in,__m512i *  out, const uint32_t bit) {
-	avx512funcPackArr[bit](in,out);
+void avx512packwithoutmask(const uint32_t *  in,__m512i *  out, const uint32_t number, const uint32_t bit) {
+    for(uint32_t i = 0; i < number / 512; ++i)
+        avx512funcPackArr[bit](in + i * 512,out + i * bit);
 }
 
 /* reads  "bit" 512-bit vectors from "in", writes  512 values to "out" */
-void avx512unpack(const __m512i *  in,uint32_t *  out, const uint32_t bit) {
-	avx512funcUnpackArr[bit](in,out);
+void avx512unpack(const __m512i *  in,uint32_t *  out, const uint32_t number, const uint32_t bit) {
+    for(uint32_t i = 0; i < number / 512; ++i)
+        avx512funcUnpackArr[bit](in + i * bit,out + i * 512);
 }
 
 #endif