Changeset 4552

Timestamp:

May 14, 2015, 11:51:32 AM (4 years ago)

Author:

nmedfort

Message:

Possible fix for 256-bit mode

Location:

icGREP/icgrep-devel/icgrep

Files:

• include/simd-lib/idisa_cpp/idisa_avx2.cpp (modified) (7 diffs)
• pablo/pablo_compiler.cpp (modified) (1 diff)

icGREP/icgrep-devel/icgrep/include/simd-lib/idisa_cpp/idisa_avx2.cpp

@@ -610 +610 @@
 #define avx_general_combine256(x, y) \
     (_mm256_insertf128_si256(_mm256_castsi128_si256(y), x, 1))
+
+// Prevents erroneous clang template expansion compile errors. Should not be used unless necessary!
+#define TEMPLATE_SUBTRACT(x, y) \
+    (((x) > (y)) ? ((x) - (y)) : (y))
+
 //The total number of operations is 2.0
 IDISA_ALWAYS_INLINE bitblock256_t simd_nor(bitblock256_t arg1, bitblock256_t arg2)
@@ -1109 +1114 @@
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srli(bitblock256_t arg1)
 {
-        return ((sh < 128) ? simd_or(simd256<128>::srli<sh>(arg1), simd256<128>::slli<(128-sh)>(_mm256_castsi128_si256(avx_select_hi128(arg1)))) : simd256<128>::srli<(sh-128)>(avx_move_hi128_to_lo128(arg1)));
+    return ((sh < 128) ? simd_or(simd256<128>::srli<sh>(arg1), simd256<128>::slli<(128-sh)>(_mm256_castsi128_si256(avx_select_hi128(arg1)))) : simd256<128>::srli<TEMPLATE_SUBTRACT(sh,128)>(avx_move_hi128_to_lo128(arg1)));
 }
 
@@ -1581 +1586 @@
 
 //The total number of operations is 9.5
-template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::slli(bitblock256_t arg1)
-{
-        return ((sh < 128) ? simd_or(simd256<128>::slli<sh>(arg1), avx_move_lo128_to_hi128(simd256<128>::srli<(128-sh)>(arg1))) : simd256<128>::slli<(sh-128)>(avx_move_lo128_to_hi128(arg1)));
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::slli(bitblock256_t arg1) {
+    return (sh < 128) ? simd_or(simd256<128>::slli<sh>(arg1), avx_move_lo128_to_hi128(simd256<128>::srli<(sh < 128) ? (128 - sh) : 0>(arg1))) : simd256<128>::slli<TEMPLATE_SUBTRACT(sh, 128)>(avx_move_lo128_to_hi128(arg1));
 }
 
@@ -2091 +2095 @@
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srai(bitblock256_t arg1)
 {
-        return simd_or(simd_and(simd256<64>::himask(), simd256<(32)>::srai<((sh < (32)) ? sh : (32))>(arg1)), ((sh <= (32)) ? simd256<64>::srli<sh>(arg1) : simd256<(32)>::srai<(sh-(32))>(simd256<64>::srli<(32)>(arg1))));
+    return simd_or(simd_and(simd256<64>::himask(), simd256<(32)>::srai<((sh < (32)) ? sh : (32))>(arg1)), ((sh <= (32)) ? simd256<64>::srli<sh>(arg1) : simd256<(32)>::srai<TEMPLATE_SUBTRACT(sh,32)>(simd256<64>::srli<(32)>(arg1))));
 }
 
@@ -2097 +2101 @@
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srai(bitblock256_t arg1)
 {
-        return simd_or(simd_and(simd256<128>::himask(), simd256<(64)>::srai<((sh < (64)) ? sh : (64))>(arg1)), ((sh <= (64)) ? simd256<128>::srli<sh>(arg1) : simd256<(64)>::srai<(sh-(64))>(simd256<128>::srli<(64)>(arg1))));
+    return simd_or(simd_and(simd256<128>::himask(), simd256<(64)>::srai<((sh < (64)) ? sh : (64))>(arg1)), ((sh <= (64)) ? simd256<128>::srli<sh>(arg1) : simd256<(64)>::srai<TEMPLATE_SUBTRACT(sh,64)>(simd256<128>::srli<(64)>(arg1))));
 }
 
@@ -2103 +2107 @@
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srai(bitblock256_t arg1)
 {
-        return simd_or(simd_and(simd256<256>::himask(), simd256<(128)>::srai<((sh < (128)) ? sh : (128))>(arg1)), ((sh <= (128)) ? simd256<256>::srli<sh>(arg1) : simd256<(128)>::srai<(sh-(128))>(simd256<256>::srli<(128)>(arg1))));
+    return simd_or(simd_and(simd256<256>::himask(), simd256<(128)>::srai<((sh < (128)) ? sh : (128))>(arg1)), ((sh <= (128)) ? simd256<256>::srli<sh>(arg1) : simd256<(128)>::srai<TEMPLATE_SUBTRACT(sh,128)>(simd256<256>::srli<(128)>(arg1))));
 }
 
@@ -3673 +3677 @@
 }
 
+#undef TEMPLATE_SUBTRACT
+
 #endif

icGREP/icgrep-devel/icgrep/pablo/pablo_compiler.cpp

@@ -1025 +1025 @@
     Value* advanceq_longint = b.CreateBitCast(genCarryDataLoad(advanceIndex), b.getIntNTy(BLOCK_SIZE));
     Value* strm_longint = b.CreateBitCast(strm_value, b.getIntNTy(BLOCK_SIZE));
-    Value* adv_longint = b.CreateOr(b.CreateShl(strm_longint, block_shift), b.CreateLShr(advanceq_longint, BLOCK_SIZE - shift_amount), "advance");
+    Value* adv_longint = b.CreateOr(b.CreateShl(strm_longint, shift_amount), b.CreateLShr(advanceq_longint, BLOCK_SIZE - shift_amount), "advance");
     result_value = b.CreateBitCast(adv_longint, mBitBlockType);