Add 16-bit support for AVX2
Equivalent to 16-bit SSE support added in:
https://github.com/talumbau/gemmlowp/commit/76de7f790a2d65b88a06a2fbe9cfb5e7a70aebb6#diff-a414deff691f120bc86c429de8e68319
diff --git a/fixedpoint/fixedpoint_avx.h b/fixedpoint/fixedpoint_avx.h
index b6e8917..f3fe732 100644
--- a/fixedpoint/fixedpoint_avx.h
+++ b/fixedpoint/fixedpoint_avx.h
@@ -23,63 +23,133 @@
namespace gemmlowp {
+struct int16x16_m256i {
+ __m256i v;
+};
+
+// Keep int16x16_m256i trivially constructible/destructible and provide
+// easily optimized helper function.
+inline int16x16_m256i to_int16x16_m256i(__m256i w) {
+ int16x16_m256i r;
+ r.v = w;
+ return r;
+}
+
template <>
struct FixedPointRawTypeTraits<__m256i> {
typedef std::int32_t ScalarRawType;
+ // TODO: This can actually support up to 8 lanes, so we should either
+ // change to 8 or create int32x8_m256i struct to handle that case.
static const int kLanes = 4;
};
template <>
+struct FixedPointRawTypeTraits<int16x16_m256i> {
+ typedef std::int16_t ScalarRawType;
+ static const int kLanes = 16;
+};
+
+template <>
inline __m256i BitAnd(__m256i a, __m256i b) {
return _mm256_and_si256(a, b);
}
template <>
+inline int16x16_m256i BitAnd(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_and_si256(a.v, b.v));
+}
+
+template <>
inline __m256i BitOr(__m256i a, __m256i b) {
return _mm256_or_si256(a, b);
}
template <>
+inline int16x16_m256i BitOr(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_or_si256(a.v, b.v));
+}
+
+template <>
inline __m256i BitXor(__m256i a, __m256i b) {
return _mm256_xor_si256(a, b);
}
template <>
+inline int16x16_m256i BitXor(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_xor_si256(a.v, b.v));
+}
+
+template <>
inline __m256i BitNot(__m256i a) {
return _mm256_andnot_si256(a, _mm256_set1_epi32(-1));
}
template <>
+inline int16x16_m256i BitNot(int16x16_m256i a) {
+ return to_int16x16_m256i(_mm256_andnot_si256(a.v, _mm256_set1_epi16(-1)));
+}
+
+template <>
inline __m256i Add(__m256i a, __m256i b) {
return _mm256_add_epi32(a, b);
}
template <>
+inline int16x16_m256i Add(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_add_epi16(a.v, b.v));
+}
+
+template <>
inline __m256i Mul(__m256i a, __m256i b) {
return _mm256_mullo_epi32(a, b);
}
template <>
+inline int16x16_m256i Mul(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_mullo_epi16(a.v, b.v));
+}
+
+template <>
inline __m256i Sub(__m256i a, __m256i b) {
return _mm256_sub_epi32(a, b);
}
template <>
+inline int16x16_m256i Sub(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_sub_epi16(a.v, b.v));
+}
+
+template <>
inline __m256i Neg(__m256i a) {
return _mm256_sign_epi32(a, _mm256_set1_epi32(-1));
}
template <>
+inline int16x16_m256i Neg(int16x16_m256i a) {
+ return to_int16x16_m256i(_mm256_sign_epi16(a.v, _mm256_set1_epi16(-1)));
+}
+
+template <>
inline __m256i ShiftLeft(__m256i a, int offset) {
return _mm256_slli_epi32(a, offset);
}
template <>
+inline int16x16_m256i ShiftLeft(int16x16_m256i a, int offset) {
+ return to_int16x16_m256i(_mm256_slli_epi16(a.v, offset));
+}
+
+template <>
inline __m256i ShiftRight(__m256i a, int offset) {
return _mm256_srai_epi32(a, offset);
}
template <>
+inline int16x16_m256i ShiftRight(int16x16_m256i a, int offset) {
+ return to_int16x16_m256i(_mm256_srai_epi16(a.v, offset));
+}
+
+template <>
inline __m256i SelectUsingMask(__m256i if_mask, __m256i then_val,
__m256i else_val) {
return _mm256_castps_si256(_mm256_blendv_ps(_mm256_castsi256_ps(else_val),
@@ -88,45 +158,97 @@
}
template <>
+inline int16x16_m256i SelectUsingMask(int16x16_m256i if_mask,
+ int16x16_m256i then_val,
+ int16x16_m256i else_val) {
+ // Borrowed from Intel's arm_neon_sse.h header.
+ return to_int16x16_m256i(
+ _mm256_or_si256(_mm256_and_si256(if_mask.v, then_val.v),
+ _mm256_andnot_si256(if_mask.v, else_val.v)));
+}
+
+template <>
inline __m256i MaskIfEqual(__m256i a, __m256i b) {
return _mm256_cmpeq_epi32(a, b);
}
template <>
+inline int16x16_m256i MaskIfEqual(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_cmpeq_epi16(a.v, b.v));
+}
+
+template <>
inline __m256i MaskIfNotEqual(__m256i a, __m256i b) {
return BitNot(MaskIfEqual(a, b));
}
template <>
+inline int16x16_m256i MaskIfNotEqual(int16x16_m256i a, int16x16_m256i b) {
+ return BitNot(MaskIfEqual(a, b));
+}
+
+template <>
inline __m256i MaskIfZero(__m256i a) {
return MaskIfEqual(a, _mm256_set1_epi32(0));
}
template <>
+inline int16x16_m256i MaskIfZero(int16x16_m256i a) {
+ return MaskIfEqual(a, to_int16x16_m256i(_mm256_set1_epi16(0)));
+}
+
+template <>
inline __m256i MaskIfNonZero(__m256i a) {
return MaskIfNotEqual(a, _mm256_set1_epi32(0));
}
template <>
+inline int16x16_m256i MaskIfNonZero(int16x16_m256i a) {
+ return MaskIfNotEqual(a, to_int16x16_m256i(_mm256_set1_epi16(0)));
+}
+
+template <>
inline __m256i MaskIfGreaterThan(__m256i a, __m256i b) {
return _mm256_cmpgt_epi32(a, b);
}
template <>
+inline int16x16_m256i MaskIfGreaterThan(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_cmpgt_epi16(a.v, b.v));
+}
+
+template <>
inline __m256i MaskIfLessThan(__m256i a, __m256i b) {
return _mm256_cmpgt_epi32(b, a);
}
template <>
+inline int16x16_m256i MaskIfLessThan(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_cmpgt_epi16(b.v, a.v));
+}
+
+template <>
inline __m256i MaskIfGreaterThanOrEqual(__m256i a, __m256i b) {
return BitNot(MaskIfLessThan(a, b));
}
template <>
+inline int16x16_m256i MaskIfGreaterThanOrEqual(int16x16_m256i a,
+ int16x16_m256i b) {
+ return BitNot(MaskIfLessThan(a, b));
+}
+
+template <>
inline __m256i MaskIfLessThanOrEqual(__m256i a, __m256i b) {
return BitNot(MaskIfGreaterThan(a, b));
}
+template <>
+inline int16x16_m256i MaskIfLessThanOrEqual(int16x16_m256i a,
+ int16x16_m256i b) {
+ return BitNot(MaskIfGreaterThan(a, b));
+}
+
/* Assumptions:
- All and Any are used on masks.
- masks are all_ones for true lanes, all_zeroes otherwise.
@@ -139,11 +261,21 @@
}
template <>
+inline bool All(int16x16_m256i a) {
+ return _mm256_testc_si256(a.v, a.v);
+}
+
+template <>
inline bool Any(__m256i a) {
return BitNot(_mm256_testz_si256(a, a));
}
template <>
+inline bool Any(int16x16_m256i a) {
+ return BitNot(_mm256_testz_si256(a.v, a.v));
+}
+
+template <>
inline __m256i RoundingHalfSum(__m256i a, __m256i b) {
/* __m256i round_bit_mask, a_over_2, b_over_2, round_bit, sum; */
/* We divide the inputs before the add to avoid the overflow and costly test
@@ -171,6 +303,17 @@
}
template <>
+inline int16x16_m256i RoundingHalfSum(int16x16_m256i a, int16x16_m256i b) {
+ // Borrowed from Intel's arm_neon_sse.h header.
+ __m256i constant_neg_32768 = _mm256_set1_epi16(-32768);
+ __m256i a_unsigned = _mm256_sub_epi16(a.v, constant_neg_32768);
+ __m256i b_unsigned = _mm256_sub_epi16(b.v, constant_neg_32768);
+ __m256i avg_unsigned = _mm256_avg_epu16(a_unsigned, b_unsigned);
+ __m256i avg = _mm256_add_epi16(avg_unsigned, constant_neg_32768);
+ return to_int16x16_m256i(avg);
+}
+
+template <>
inline __m256i SaturatingRoundingDoublingHighMul(__m256i a, __m256i b) {
__m256i min, saturation_mask, a0_a2, a1_a3, b0_b2, b1_b3;
__m256i a0b0_a2b2, a1b1_a3b3, a0b0_a2b2_rounded, a1b1_a3b3_rounded;
@@ -209,10 +352,33 @@
}
template <>
+inline int16x16_m256i SaturatingRoundingDoublingHighMul(int16x16_m256i a,
+ int16x16_m256i b) {
+ // Use _mm256_mulhrs_epi16 then saturate with a bit-operation,
+ // borrowed from Intel's arm_neon_sse.h header.
+ __m256i result_unsaturated = _mm256_mulhrs_epi16(a.v, b.v);
+ __m256i saturation_mask =
+ _mm256_cmpeq_epi16(result_unsaturated, _mm256_set1_epi16(0x8000));
+ __m256i result = _mm256_xor_si256(result_unsaturated, saturation_mask);
+ return to_int16x16_m256i(result);
+}
+
+template <>
inline __m256i Dup<__m256i>(std::int32_t x) {
return _mm256_set1_epi32(x);
}
+template <>
+inline int16x16_m256i Dup<int16x16_m256i>(std::int16_t x) {
+ return to_int16x16_m256i(_mm256_set1_epi16(x));
+}
+
+// So far this is only needed for int16.
+template <>
+inline int16x16_m256i SaturatingAdd(int16x16_m256i a, int16x16_m256i b) {
+ return to_int16x16_m256i(_mm256_adds_epi16(a.v, b.v));
+}
+
} // end namespace gemmlowp
#endif // GEMMLOWP_INTERNAL_FIXEDPOINT_AVX_H_
diff --git a/fixedpoint/fixedpoint_sse.h b/fixedpoint/fixedpoint_sse.h
index a1fae32..fbaa26a 100644
--- a/fixedpoint/fixedpoint_sse.h
+++ b/fixedpoint/fixedpoint_sse.h
@@ -32,13 +32,17 @@
// data type, int16x8_m128i, that wraps __m128i while being a separate
// type.
struct int16x8_m128i {
- int16x8_m128i() {}
- explicit int16x8_m128i(__m128i w) : v(w) {}
- ~int16x8_m128i() {}
-
__m128i v;
};
+// Keep int16x8_m128i trivially constructible/destructible and provide
+// easily optimized helper function.
+inline int16x8_m128i to_int16x8_m128i(__m128i w) {
+ int16x8_m128i r;
+ r.v = w;
+ return r;
+}
+
template <>
struct FixedPointRawTypeTraits<__m128i> {
typedef std::int32_t ScalarRawType;
@@ -58,7 +62,7 @@
template <>
inline int16x8_m128i BitAnd(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_and_si128(a.v, b.v));
+ return to_int16x8_m128i(_mm_and_si128(a.v, b.v));
}
template <>
@@ -68,7 +72,7 @@
template <>
inline int16x8_m128i BitOr(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_or_si128(a.v, b.v));
+ return to_int16x8_m128i(_mm_or_si128(a.v, b.v));
}
template <>
@@ -78,7 +82,7 @@
template <>
inline int16x8_m128i BitXor(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_xor_si128(a.v, b.v));
+ return to_int16x8_m128i(_mm_xor_si128(a.v, b.v));
}
template <>
@@ -88,7 +92,7 @@
template <>
inline int16x8_m128i BitNot(int16x8_m128i a) {
- return int16x8_m128i(_mm_andnot_si128(a.v, _mm_set1_epi16(-1)));
+ return to_int16x8_m128i(_mm_andnot_si128(a.v, _mm_set1_epi16(-1)));
}
template <>
@@ -98,7 +102,7 @@
template <>
inline int16x8_m128i Add(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_add_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_add_epi16(a.v, b.v));
}
template <>
@@ -108,7 +112,7 @@
template <>
inline int16x8_m128i Mul(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_mullo_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_mullo_epi16(a.v, b.v));
}
template <>
@@ -118,7 +122,7 @@
template <>
inline int16x8_m128i Sub(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_sub_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_sub_epi16(a.v, b.v));
}
template <>
@@ -128,7 +132,7 @@
template <>
inline int16x8_m128i Neg(int16x8_m128i a) {
- return int16x8_m128i(_mm_sign_epi16(a.v, _mm_set1_epi16(-1)));
+ return to_int16x8_m128i(_mm_sign_epi16(a.v, _mm_set1_epi16(-1)));
}
template <>
@@ -138,7 +142,7 @@
template <>
inline int16x8_m128i ShiftLeft(int16x8_m128i a, int offset) {
- return int16x8_m128i(_mm_slli_epi16(a.v, offset));
+ return to_int16x8_m128i(_mm_slli_epi16(a.v, offset));
}
template <>
@@ -148,7 +152,7 @@
template <>
inline int16x8_m128i ShiftRight(int16x8_m128i a, int offset) {
- return int16x8_m128i(_mm_srai_epi16(a.v, offset));
+ return to_int16x8_m128i(_mm_srai_epi16(a.v, offset));
}
template <>
@@ -164,7 +168,7 @@
int16x8_m128i then_val,
int16x8_m128i else_val) {
// borrowed from Intel's arm_neon_sse.h header.
- return int16x8_m128i(SelectUsingMask(if_mask.v, then_val.v, else_val.v));
+ return to_int16x8_m128i(SelectUsingMask(if_mask.v, then_val.v, else_val.v));
}
template <>
@@ -174,7 +178,7 @@
template <>
inline int16x8_m128i MaskIfEqual(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_cmpeq_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_cmpeq_epi16(a.v, b.v));
}
template <>
@@ -194,7 +198,7 @@
template <>
inline int16x8_m128i MaskIfZero(int16x8_m128i a) {
- return MaskIfEqual(a, int16x8_m128i(_mm_set1_epi16(0)));
+ return MaskIfEqual(a, to_int16x8_m128i(_mm_set1_epi16(0)));
}
template <>
@@ -204,7 +208,7 @@
template <>
inline int16x8_m128i MaskIfNonZero(int16x8_m128i a) {
- return MaskIfNotEqual(a, int16x8_m128i(_mm_set1_epi16(0)));
+ return MaskIfNotEqual(a, to_int16x8_m128i(_mm_set1_epi16(0)));
}
template <>
@@ -214,7 +218,7 @@
template <>
inline int16x8_m128i MaskIfGreaterThan(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_cmpgt_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_cmpgt_epi16(a.v, b.v));
}
template <>
@@ -224,7 +228,7 @@
template <>
inline int16x8_m128i MaskIfLessThan(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_cmplt_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_cmplt_epi16(a.v, b.v));
}
template <>
@@ -310,7 +314,7 @@
__m128i b_unsigned = _mm_sub_epi16(b.v, constant_neg_32768);
__m128i avg_unsigned = _mm_avg_epu16(a_unsigned, b_unsigned);
__m128i avg = _mm_add_epi16(avg_unsigned, constant_neg_32768);
- return int16x8_m128i(avg);
+ return to_int16x8_m128i(avg);
}
template <>
@@ -360,7 +364,7 @@
__m128i saturation_mask =
_mm_cmpeq_epi16(result_unsaturated, _mm_set1_epi16(0x8000));
__m128i result = _mm_xor_si128(result_unsaturated, saturation_mask);
- return int16x8_m128i(result);
+ return to_int16x8_m128i(result);
}
template <>
@@ -370,13 +374,13 @@
template <>
inline int16x8_m128i Dup<int16x8_m128i>(std::int16_t x) {
- return int16x8_m128i(_mm_set1_epi16(x));
+ return to_int16x8_m128i(_mm_set1_epi16(x));
}
// So far this is only needed for int16.
template <>
inline int16x8_m128i SaturatingAdd(int16x8_m128i a, int16x8_m128i b) {
- return int16x8_m128i(_mm_adds_epi16(a.v, b.v));
+ return to_int16x8_m128i(_mm_adds_epi16(a.v, b.v));
}
} // end namespace gemmlowp
diff --git a/test/test_fixedpoint.cc b/test/test_fixedpoint.cc
index a45ec17..44e6fae 100644
--- a/test/test_fixedpoint.cc
+++ b/test/test_fixedpoint.cc
@@ -17,14 +17,14 @@
#define GEMMLOWP_ENABLE_FIXEDPOINT_CONSTANTS_CHECKS
#include <algorithm>
+#include <cinttypes>
#include <cmath>
#include <cstdio>
-#include <cinttypes>
#include <random>
#include <vector>
-#include "test.h"
#include "../fixedpoint/fixedpoint.h"
+#include "test.h"
namespace gemmlowp {
@@ -67,7 +67,8 @@
}
template <>
int16x8_m128i Load<int16x8_m128i>(const std::int16_t* src) {
- return int16x8_m128i(_mm_loadu_si128(reinterpret_cast<const __m128i*>(src)));
+ return to_int16x8_m128i(
+ _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)));
}
template <>
void Store<int16x8_m128i>(std::int16_t* dst, int16x8_m128i v) {
@@ -93,6 +94,29 @@
}
#endif
+#ifdef GEMMLOWP_AVX2
+template <>
+__m256i Load<__m256i>(const std::int32_t* src) {
+ return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(src));
+}
+
+template <>
+int16x16_m256i Load<int16x16_m256i>(const std::int16_t* src) {
+ return to_int16x16_m256i(
+ _mm256_loadu_si256(reinterpret_cast<const __m256i*>(src)));
+}
+
+template <>
+void Store<__m256i>(std::int32_t* dst, __m256i v) {
+ _mm256_storeu_si256(reinterpret_cast<__m256i*>(dst), v);
+}
+
+template <>
+void Store<int16x16_m256i>(std::int16_t* dst, int16x16_m256i v) {
+ _mm256_storeu_si256(reinterpret_cast<__m256i*>(dst), v.v);
+}
+#endif
+
template <typename tSimdType>
class TestFixedPoint {
public:
@@ -571,4 +595,9 @@
gemmlowp::TestFixedPoint<v4i32>().RunTests("MSA v4i32");
gemmlowp::TestFixedPoint<v8i16>().RunTests("MSA v8i16");
#endif
+#ifdef GEMMLOWP_AVX2
+ gemmlowp::TestFixedPoint<__m256i>().RunTests("AVX __m256i");
+ gemmlowp::TestFixedPoint<gemmlowp::int16x16_m256i>().RunTests(
+ "AVX2 __m256i = int16x16");
+#endif
}