return _mm512_reduce_add_epi32(sum) + bias;
};
- [[maybe_unused]] auto m512_haddx4 = [](__m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3, __m128i bias) -> __m128i {
+ // This function takes
+ // sum0 = [xmm0a, xmm0b, xmm0c, xmm0d]
+ // sum1 = [xmm1a, xmm1b, xmm1c, xmm1d]
+ // sum2 = [xmm2a, xmm2b, xmm2c, xmm2d]
+ // sum3 = [xmm3a, xmm3b, xmm3c, xmm3d]
+ // and returns
+ // ret = [
+ // reduce_add_epi32(xmm0a), reduce_add_epi32(xmm1a), reduce_add_epi32(xmm2a), reduce_add_epi32(xmm3a),
+ // reduce_add_epi32(xmm0b), reduce_add_epi32(xmm1b), reduce_add_epi32(xmm2b), reduce_add_epi32(xmm3b),
+ // reduce_add_epi32(xmm0c), reduce_add_epi32(xmm1c), reduce_add_epi32(xmm2c), reduce_add_epi32(xmm3c),
+ // reduce_add_epi32(xmm0d), reduce_add_epi32(xmm1d), reduce_add_epi32(xmm2d), reduce_add_epi32(xmm3d)
+ // ]
+ [[maybe_unused]] auto m512_hadd128x16_interleave = [](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3) -> __m512i {
+
__m512i sum01a = _mm512_unpacklo_epi32(sum0, sum1);
__m512i sum01b = _mm512_unpackhi_epi32(sum0, sum1);
__m512i sum0123a = _mm512_unpacklo_epi64(sum01, sum23);
__m512i sum0123b = _mm512_unpackhi_epi64(sum01, sum23);
- __m512i sum = _mm512_add_epi32(sum0123a, sum0123b);
+ return _mm512_add_epi32(sum0123a, sum0123b);
+ };
+
+ [[maybe_unused]] auto m512_haddx4 = [m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3, __m128i bias) -> __m128i {
+
+ __m512i sum = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
__m256i sum256lo = _mm512_castsi512_si256(sum);
__m256i sum256hi = _mm512_extracti64x4_epi64(sum, 1);
return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
};
+ [[maybe_unused]] auto m512_haddx8 = [m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3,
+ __m512i sum4, __m512i sum5, __m512i sum6, __m512i sum7, __m256i bias) -> __m256i {
+
+ __m512i suma = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+ __m512i sumb = m512_hadd128x16_interleave(sum4, sum5, sum6, sum7);
+
+ __m512i indices0 = _mm512_setr_epi64(0, 1, 8, 9, 4, 5, 12, 13);
+ __m512i indices1 = _mm512_setr_epi64(2, 3, 10, 11, 6, 7, 14, 15);
+ __m512i x = _mm512_add_epi32(
+ _mm512_permutex2var_epi64(suma, indices0, sumb),
+ _mm512_permutex2var_epi64(suma, indices1, sumb));
+
+ __m256i sum256lo = _mm512_castsi512_si256(x);
+ __m256i sum256hi = _mm512_extracti64x4_epi64(x, 1);
+
+ return _mm256_add_epi32(_mm256_add_epi32(sum256lo, sum256hi), bias);
+ };
+
+ [[maybe_unused]] auto m512_hadd256x8 =[m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3, __m256i bias) -> __m256i {
+
+ __m512i sum = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+
+ __m512i indices = _mm512_setr_epi32(
+ 0, 4, 8, 12, 2, 6, 10, 14,
+ 1, 5, 9, 13, 3, 7, 11, 15);
+ sum = _mm512_permutexvar_epi32(indices, sum);
+
+ __m256i sum256lo = _mm512_castsi512_si256(sum);
+ __m256i sum256hi = _mm512_extracti64x4_epi64(sum, 1);
+
+ return _mm256_add_epi32(_mm256_hadd_epi32(sum256lo, sum256hi), bias);
+ };
+
+ [[maybe_unused]] auto m512_hadd256x16 = [m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3,
+ __m512i sum4, __m512i sum5, __m512i sum6, __m512i sum7, __m512i bias) -> __m512i {
+
+ __m512i suma = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+ __m512i sumb = m512_hadd128x16_interleave(sum4, sum5, sum6, sum7);
+
+ __m512i indices0 = _mm512_setr_epi64(0, 1, 8, 9, 4, 5, 12, 13);
+ __m512i indices1 = _mm512_setr_epi64(2, 3, 10, 11, 6, 7, 14, 15);
+ __m512i x = _mm512_add_epi32(
+ _mm512_permutex2var_epi64(suma, indices0, sumb),
+ _mm512_permutex2var_epi64(suma, indices1, sumb));
+
+ __m512i indices = _mm512_setr_epi32(0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15);
+ return _mm512_add_epi32(_mm512_permutexvar_epi32(indices, x), bias);
+ };
+
[[maybe_unused]] auto m512_add_dpbusd_epi32 = [=](__m512i& acc, __m512i a, __m512i b) {
#if defined (USE_VNNI)
acc = _mm512_dpbusd_epi32(acc, a, b);
// kOutputDimensions is either 1 or a multiple of kSimdWidth
// because then it is also an input dimension.
- if constexpr (kOutputDimensions % 4 == 0)
+ if constexpr (kOutputDimensions % 16 == 0 && kNumChunks256 == 1)
+ {
+ for (IndexType i = 0; i < kOutputDimensions; i += 16)
+ {
+ const IndexType offset01a = (i + 0) * kPaddedInputDimensions;
+ const IndexType offset23a = (i + 2) * kPaddedInputDimensions;
+ const IndexType offset45a = (i + 4) * kPaddedInputDimensions;
+ const IndexType offset67a = (i + 6) * kPaddedInputDimensions;
+ const IndexType offset01b = (i + 8) * kPaddedInputDimensions;
+ const IndexType offset23b = (i + 10) * kPaddedInputDimensions;
+ const IndexType offset45b = (i + 12) * kPaddedInputDimensions;
+ const IndexType offset67b = (i + 14) * kPaddedInputDimensions;
+
+ const __m512i bias = *reinterpret_cast<const __m512i*>(&biases_[i]);
+ __m512i* outptr = reinterpret_cast<__m512i*>(&output[i]);
+
+ __m512i sum01a = _mm512_setzero_si512();
+ __m512i sum23a = _mm512_setzero_si512();
+ __m512i sum45a = _mm512_setzero_si512();
+ __m512i sum67a = _mm512_setzero_si512();
+ __m512i sum01b = _mm512_setzero_si512();
+ __m512i sum23b = _mm512_setzero_si512();
+ __m512i sum45b = _mm512_setzero_si512();
+ __m512i sum67b = _mm512_setzero_si512();
+
+ const auto row01a = *reinterpret_cast<const __m512i*>(&weights_[offset01a]);
+ const auto row23a = *reinterpret_cast<const __m512i*>(&weights_[offset23a]);
+ const auto row45a = *reinterpret_cast<const __m512i*>(&weights_[offset45a]);
+ const auto row67a = *reinterpret_cast<const __m512i*>(&weights_[offset67a]);
+ const auto row01b = *reinterpret_cast<const __m512i*>(&weights_[offset01b]);
+ const auto row23b = *reinterpret_cast<const __m512i*>(&weights_[offset23b]);
+ const auto row45b = *reinterpret_cast<const __m512i*>(&weights_[offset45b]);
+ const auto row67b = *reinterpret_cast<const __m512i*>(&weights_[offset67b]);
+
+ const __m256i in256 = input_vector256[0];
+ const __m512i in = _mm512_inserti64x4(_mm512_castsi256_si512(in256), in256, 1);
+
+ m512_add_dpbusd_epi32(sum01a, in, row01a);
+ m512_add_dpbusd_epi32(sum23a, in, row23a);
+ m512_add_dpbusd_epi32(sum45a, in, row45a);
+ m512_add_dpbusd_epi32(sum67a, in, row67a);
+ m512_add_dpbusd_epi32(sum01b, in, row01b);
+ m512_add_dpbusd_epi32(sum23b, in, row23b);
+ m512_add_dpbusd_epi32(sum45b, in, row45b);
+ m512_add_dpbusd_epi32(sum67b, in, row67b);
+
+ *outptr = m512_hadd256x16(
+ sum01a, sum23a, sum45a, sum67a,
+ sum01b, sum23b, sum45b, sum67b, bias);
+ }
+ }
+ else if constexpr (kOutputDimensions % 4 == 0)
{
for (IndexType i = 0; i < kOutputDimensions; i += 4)
{