add compress128 function and implementation

diff --git a/src/util/arch/arm/bitutils.h b/src/util/arch/arm/bitutils.h
index 514ddc5c61d09465974643a753a5d72895542978..0b579dc94b3c7026a1a0c202e121f1a234a622c6 100644 (file)
--- a/src/util/arch/arm/bitutils.h
+++ b/src/util/arch/arm/bitutils.h
@@ -104,6 +104,108 @@ u64a compress64_impl(u64a x, u64a m) {
     return compress64_impl_c(x, m);
 }
 
+static really_inline
+m128 compress128_impl(m128 x, m128 m) {
+
+/*    x = and128(x, m); // clear irrelevant bits
+
+    // Return zero quickly on trivial cases
+    if (diff128(x, zeroes128()) == 0) {
+        return zeroes128();
+    }*/
+
+
+    u64a ALIGN_ATTR(16) xv[2];
+    u64a ALIGN_ATTR(16) mv[2];
+    u64a ALIGN_ATTR(16) res[2];
+    u64a ALIGN_ATTR(16) t[2];
+    u64a ALIGN_ATTR(16) bbv[2];
+    store128(xv, x);
+    store128(mv, m);
+    res[0] = 0;
+    res[1] = 0;
+    printf("x[%d] = %0llx\n", 0, xv[0]);
+    printf("x[%d] = %0llx\n", 1, xv[1]);
+
+    m128 one = set1_2x64(1);
+    m128 bitset = one;
+    m128 vres = zeroes128();
+    for (u64a bb = 1; mv[0] | mv[1]; bb <<= 1) {
+        printf("bb = %lld\n", bb);
+       store128(bbv, bitset);
+        printf("bb[%d] = %0lld\n", 0, bbv[0]);
+        printf("bb[%d] = %0lld\n", 1, bbv[1]);
+        printf("m[%d] = %0llx\n", 0, mv[0]);
+        printf("m[%d] = %0llx\n", 1, mv[1]);
+        printf("scalar: -m[%d] = %0llx\n", 0, -mv[0]);
+        printf("scalar: -m[%d] = %0llx\n", 1, -mv[1]);
+       m128 mm = sub_2x64(zeroes128(), m);
+       store128(t, mm);
+        printf("vector: -m[0] = %0llx\n", t[0]);
+        printf("vector: -m[1] = %0llx\n", t[1]);
+       m128 tv = and128(x, m);
+       store128(t, tv);
+        printf("vector: x[0] & m[0] = %0llx\n", t[0]);
+        printf("vector: x[1] & m[1] = %0llx\n", t[1]);
+       tv = and128(tv, mm);
+       store128(t, tv);
+        printf("vector: x[0] & m[0] & -m[0] = %0llx\n", t[0]);
+        printf("vector: x[1] & m[1] & -m[1] = %0llx\n", t[1]);
+        t[0] = xv[0] & mv[0];
+        t[1] = xv[1] & mv[1];
+        printf("scalar: x[0] & m[0] = %0llx\n", t[0]);
+        printf("scalar: x[1] & m[1] = %0llx\n", t[1]);
+        t[0] = xv[0] & mv[0] & -mv[0];
+        t[1] = xv[1] & mv[1] & -mv[1];
+        printf("scalar: x[0] & m[0] & -m[0] = %0llx\n", t[0]);
+        printf("scalar: x[1] & m[1] & -m[1] = %0llx\n", t[1]);
+        
+        if ( t[0] ) {
+            printf("x & m & -m != 0\n");
+            res[0] |= bb;
+            printf("x[%d] = %0llx\n", 0, xv[0]);
+        }
+        if ( t[1] ) {
+            printf("x & m & -m != 0\n");
+            res[1] |= bb;
+            printf("x[%d] = %0llx\n", 1, xv[1]);
+        }
+
+       m128 mask = not128(eq64_m128(tv, zeroes128()));
+       store128(t, mask);
+        printf("mask: x[0] & m[0] & -m[0] != 0 : %0llx\n", t[0]);
+        printf("mask: x[1] & m[1] & -m[1] != 0 : %0llx\n", t[1]);
+
+       mask = vandq_s64(bitset, mask);
+       store128(t, mask);
+        printf("mask: mask[0] & bitset[1] != 0 : %0llx\n", t[0]);
+        printf("mask: mask[1] & bitset[1] != 0 : %0llx\n", t[1]);
+
+        vres = or128(vres, mask);
+       store128(t, vres);
+        printf("res: res[0] != 0 : %0llx\n", t[0]);
+        printf("res: res[1] != 0 : %0llx\n", t[1]);
+       if (t[0] != res[0]) {
+            printf("mismatch: t[0] != res[0]: %0llx != %0llx\n", t[0], res[0]);
+        }
+       if (t[1] != res[1]) {
+            printf("mismatch: t[1] != res[1]: %0llx != %0llx\n", t[1], res[1]);
+        }
+
+        mv[0] &= mv[0] - 1;
+        mv[1] &= mv[1] - 1;
+       m = and128(m, sub_2x64(m, set1_2x64(1)));
+        printf("x[%d] = %0llx\n", 0, xv[0]);
+        printf("x[%d] = %0llx\n", 1, xv[1]);
+        bitset = lshift64_m128(bitset, 1);
+    }
+    store128(res, vres);
+    printf("final x[%d] = %0llx\n", 0, res[0]);
+    printf("final x[%d] = %0llx\n", 1, res[1]);
+//    x = load128(res);
+    return vres;
+}
+
 static really_inline
 u32 expand32_impl(u32 x, u32 m) {
     return expand32_impl_c(x, m);

diff --git a/src/util/arch/common/bitutils.h b/src/util/arch/common/bitutils.h
index e86b8d44cf2de282eedef1b112b28921c6b6f791..88e71bbaa82ec8a1ff8b721c93cf2541d10955aa 100644 (file)
--- a/src/util/arch/common/bitutils.h
+++ b/src/util/arch/common/bitutils.h
@@ -35,6 +35,7 @@
 
 #include "util/popcount.h"
 #include "util/unaligned.h"
+#include "util/simd_utils.h"
 
 static really_inline
 u32 clz32_impl_c(u32 x) {
@@ -177,7 +178,13 @@ u32 compress32_impl_c(u32 x, u32 m) {
 
 static really_inline
 u64a compress64_impl_c(u64a x, u64a m) {
-    // Return zero quickly on trivial cases
+  u64a res = 0;
+  for (u64a bb = 1; m != 0; bb += bb) {
+    if (x & m & -m) { res |= bb; }
+    m &= (m - 1);
+  }
+  return res;
+/*    // Return zero quickly on trivial cases
     if ((x & m) == 0) {
         return 0;
     }
@@ -202,7 +209,20 @@ u64a compress64_impl_c(u64a x, u64a m) {
         mk = mk & ~mp;
     }
 
-    return x;
+    return x;*/
+}
+
+static really_inline
+m128 compress128_impl_c(m128 xvec, m128 mvec) {
+    u64a ALIGN_ATTR(16) x[2];
+    u64a ALIGN_ATTR(16) m[2];
+    store128(x, xvec);
+    store128(m, mvec);
+
+    compress64_impl_c(x[0], m[0]);
+    compress64_impl_c(x[1], m[1]);
+
+    return xvec;
 }
 
 static really_inline
@@ -242,7 +262,13 @@ u32 expand32_impl_c(u32 x, u32 m) {
 static really_inline
 u64a expand64_impl_c(u64a x, u64a m) {
 
-    // Return zero quickly on trivial cases
+  u64a res = 0;
+  for (u64a bb = 1; m != 0; bb += bb) {
+    if (x & bb) { res |= m & (-m); }
+    m &= (m - 1);
+  }
+  return res;
+/*    // Return zero quickly on trivial cases
     if (!x || !m) {
         return 0;
     }
@@ -272,7 +298,7 @@ u64a expand64_impl_c(u64a x, u64a m) {
         x = (x & ~mv) | (t & mv);
     }
 
-    return x & m0; // clear out extraneous bits
+    return x & m0; // clear out extraneous bits*/
 }
 
 

diff --git a/src/util/arch/x86/bitutils.h b/src/util/arch/x86/bitutils.h
index ec4c95ad93a07f1f7533f361df2be5ef824fe163..a0769a5e5efe0082332ca3010221e95c7707085b 100644 (file)
--- a/src/util/arch/x86/bitutils.h
+++ b/src/util/arch/x86/bitutils.h
@@ -214,6 +214,11 @@ u64a compress64_impl(u64a x, u64a m) {
 #endif
 }
 
+static really_inline
+u64a compress128_impl(m128 x, m128 m) {
+    compress128_impl_c(x, m);
+}
+
 static really_inline
 u32 expand32_impl(u32 x, u32 m) {
 #if defined(HAVE_BMI2)

diff --git a/src/util/bitutils.h b/src/util/bitutils.h
index 556ba81856e3ac94a0394b93d15488c22a263611..21d35388527fdc9e90bcc3af37981d336418d986 100644 (file)
--- a/src/util/bitutils.h
+++ b/src/util/bitutils.h
@@ -120,6 +120,11 @@ u64a compress64(u64a x, u64a m) {
     return compress64_impl(x, m);
 }
 
+static really_inline
+m128 compress128(m128 x, m128 m) {
+    return compress128_impl(x, m);
+}
+
 static really_inline
 u32 expand32(u32 x, u32 m) {
     return expand32_impl(x, m);