// Computes part of matrix.vector v = Wu. Computes N=8 results.
// For details see PartialMatrixDotVector64 with N=8.
static void PartialMatrixDotVector8(const int8_t* wi, const double* scales,
const int8_t* u, int num_in, int num_out,
double* v) {
// Register containing 16-bit ones for horizontal add with 16->32 bit
// conversion.
__m256i ones =
_mm256_set_epi16(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1);
__m256i shift_id = _mm256_set_epi32(0, 7, 6, 5, 4, 3, 2, 1);
// Initialize all the results to 0.
__m256i result0 = _mm256_setzero_si256();
// Iterate over the input (u), one registerful at a time.
for (int j = 0; j < num_in;) {
__m256i inputs =
_mm256_loadu_si256(reinterpret_cast<const __m256i*>(u + j));
// Inputs are processed in groups of kNumInputsPerGroup, replicated
// kNumInputGroups times.
for (int ig = 0; ig < kNumInputGroups && j < num_in;
++ig, j += kNumInputsPerGroup) {
// Replicate the low 32 bits (4 inputs) 8 times.
__m256i rep_input =
_mm256_broadcastd_epi32(_mm256_castsi256_si128(inputs));
// Rotate the inputs in groups of 4, so the next 4 inputs are ready.
inputs = _mm256_permutevar8x32_epi32(inputs, shift_id);
__m256i weights, reps;
// Mul-add, with horizontal add of the 4 inputs to each of the results.
MultiplyGroup(rep_input, ones, wi, weights, reps, result0);
}
}
ExtractResults(result0, shift_id, wi, scales, num_out, v);
}
__m256i test_mm256_broadcastd_epi32(__m128i a) {
// CHECK: @llvm.x86.avx2.pbroadcastd.256
return _mm256_broadcastd_epi32(a);
}
__m256i test_mm256_broadcastd_epi32(__m128i a) {
// CHECK-LABEL: test_mm256_broadcastd_epi32
// CHECK-NOT: @llvm.x86.avx2.pbroadcastd.256
// CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> %{{.*}}, <8 x i32> zeroinitializer
return _mm256_broadcastd_epi32(a);
}
