src/qs8-igemm/MRx4c2-sse.c.in

// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$SSE_HEADER = {2: "emmintrin.h", 3: "tmmintrin.h", 4: "smmintrin.h", 5: "ammintrin.h"}[SSE]
$assert MR <= 4
#include <assert.h>

$if SSE == 5:
  #ifdef __GNUC__
    #include <x86intrin.h>
  #else
    #include <immintrin.h>
    #include <${SSE_HEADER}>
  #endif
$else:
  #include <${SSE_HEADER}>

#include <xnnpack/igemm.h>
#include <xnnpack/math.h>


$ISA = {2: "sse2", 3: "ssse3", 4: "sse41", 5: "xop"}[SSE]
void xnn_qs8_igemm_minmax_ukernel_${MR}x4c2__${ISA}_${"ld128" if LD128 else "ld64"}(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const int8_t** restrict a,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const int8_t* zero,
    const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(ks != 0);
  assert(ks % (${MR} * sizeof(void*)) == 0);
  assert(a_offset % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 2);
  int8_t* c0 = c;
  $for M in range(1, MR):
    int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        c${M} = c${M-1};
      }

  do {
    __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
    $for M in range(1, MR):
      __m128i vacc${M}x0123 = vacc0x0123;
    w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));

    size_t p = ks;
    do {
      $for M in range(MR):
        const int8_t* restrict a${M} = a[${M}];
        if XNN_UNPREDICTABLE(a${M} != zero) {
          a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
        }
      a += ${MR};

      size_t k = kc;
      while (k >= 8 * sizeof(int8_t)) {
        $for M in range(MR):
          const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M});
          $if SSE >= 4:
            const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
          $else:
            const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, _mm_cmpgt_epi8(_mm_setzero_si128(), va${M}));
          a${M} += 8;

        $if LD128:
          $for K in range(0, 4, 2):
            $if K == 0:
              const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) w);
            $else:
              const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${K * 8}));
            const __m128i vsb${K}${K+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${K}${K+1});
            const __m128i vxb${K} = _mm_unpacklo_epi8(vb${K}${K+1}, vsb${K}${K+1});
            const __m128i vxb${K+1} = _mm_unpackhi_epi8(vb${K}${K+1}, vsb${K}${K+1});

            $for M in range(MR):
              $if SSE == 5:
                vacc${M}x0123 = _mm_maddd_epi16(
                  _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}, vacc${M}x0123);
              $else:
                vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
                  _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}));

            $for M in range(MR):
              $if SSE == 5:
                vacc${M}x0123 = _mm_maddd_epi16(
                  _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K+1}, ${K+1}, ${K+1}, ${K+1})), vxb${K+1}, vacc${M}x0123);
              $else:
                vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
                  _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K+1}, ${K+1}, ${K+1}, ${K+1})), vxb${K+1}));
        $else:
          $for K in range(4):
            $if K == 0:
              const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) w);
            $else:
              const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${K * 8}));
            $if SSE >= 4:
              const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K});
            $else:
              const __m128i vxb${K} = _mm_unpacklo_epi8(vb${K}, _mm_cmpgt_epi8(_mm_setzero_si128(), vb${K}));

            $for M in range(MR):
              $if SSE == 5:
                vacc${M}x0123 = _mm_maddd_epi16(
                  _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}, vacc${M}x0123);
              $else:
                vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
                  _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}));

        w = (const void*) ((uintptr_t) w + 32);
        k -= 8 * sizeof(int8_t);
      }
      if (k != 0) {
        $for M in range(MR):
          const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M});
          $if SSE >= 4:
            const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
          $else:
            const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, _mm_cmpgt_epi8(_mm_setzero_si128(), va${M}));
          a${M} = (const int8_t*) ((uintptr_t) a${M} + k);

        const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
        w = (const void*) ((uintptr_t) w + 8);
        const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));

        $for M in range(MR):
          $if SSE == 5:
            vacc${M}x0123 = _mm_maddd_epi16(
              _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc${M}x0123);
          $else:
            vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
              _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));

        if (k > 2 * sizeof(int8_t)) {
          const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
          w = (const void*) ((uintptr_t) w + 8);
          const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));

          $for M in range(MR):
            $if SSE == 5:
              vacc${M}x0123 = _mm_maddd_epi16(
                _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc${M}x0123);
            $else:
              vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
                _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));

          if (k > 4 * sizeof(int8_t)) {
            const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
            w = (const void*) ((uintptr_t) w + 8);
            const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));

            $for M in range(MR):
              $if SSE == 5:
                vacc${M}x0123 = _mm_maddd_epi16(
                  _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc${M}x0123);
              $else:
                vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
                  _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
          }
        }
      }
      p -= ${MR} * sizeof(void*);
    } while (p != 0);

    const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
    const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);

    $if SSE >= 4:
      $for M in range(MR):
        const __m128i vacc${M}x1133 = _mm_shuffle_epi32(vacc${M}x0123, _MM_SHUFFLE(3, 3, 1, 1));

      $for M in range(MR):
        const __m128i vprod${M}x02 = _mm_add_epi64(_mm_mul_epi32(vacc${M}x0123, vmultiplier), vrounding);

      $for M in range(MR):
        const __m128i vprod${M}x13 = _mm_add_epi64(_mm_mul_epi32(vacc${M}x1133, vmultiplier), vrounding);

      $for M in range(MR):
        const __m128i vq31prod${M}x02 = _mm_srli_epi64(vprod${M}x02, 31);
        const __m128i vq31prod${M}x13 = _mm_add_epi64(vprod${M}x13, vprod${M}x13);

      $for M in range(MR):
        const __m128i vq31prod${M}x0123 = _mm_blend_epi16(vq31prod${M}x02, vq31prod${M}x13, 0xCC);
    $else:
      $for M in range(MR):
        const __m128i vnmask${M}x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${M}x0123);

      $for M in range(MR):
        $if SSE >= 3:
          const __m128i vabsacc${M}x0123 = _mm_abs_epi32(vacc${M}x0123);
        $else:
          const __m128i vabsacc${M}x0123 = _mm_sub_epi32(_mm_xor_si128(vacc${M}x0123, vnmask${M}x0123), vnmask${M}x0123);

      $for M in range(MR):
        const __m128i vabsacc${M}x1133 = _mm_shuffle_epi32(vabsacc${M}x0123, _MM_SHUFFLE(3, 3, 1, 1));

      $for M in range(MR):
        const __m128i vabsprod${M}x02 = _mm_mul_epu32(vabsacc${M}x0123, vmultiplier);

      $for M in range(MR):
        const __m128i vnmask${M}x02 = _mm_shuffle_epi32(vnmask${M}x0123, _MM_SHUFFLE(2, 2, 0, 0));

      $for M in range(MR):
        const __m128i vprod${M}x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod${M}x02, vnmask${M}x02), vnmask${M}x02);

      $for M in range(MR):
        const __m128i vq31prod${M}x02 = _mm_srli_epi64(_mm_add_epi64(vprod${M}x02, vrounding), 31);

      $for M in range(MR):
        const __m128i vabsprod${M}x13 = _mm_mul_epu32(vabsacc${M}x1133, vmultiplier);

      $for M in range(MR):
        const __m128i vnmask${M}x13 = _mm_shuffle_epi32(vnmask${M}x0123, _MM_SHUFFLE(3, 3, 1, 1));

      $for M in range(MR):
        const __m128i vprod${M}x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod${M}x13, vnmask${M}x13), vnmask${M}x13);

      $for M in range(MR):
        const __m128i vq31prod${M}x13 = _mm_srli_epi64(_mm_add_epi64(vprod${M}x13, vrounding), 31);

      $for M in range(MR):
        const __m128i vq31prod${M}x0213 = _mm_castps_si128(_mm_shuffle_ps(
            _mm_castsi128_ps(vq31prod${M}x02), _mm_castsi128_ps(vq31prod${M}x13), _MM_SHUFFLE(2, 0, 2, 0)));

      $for M in range(MR):
        const __m128i vq31prod${M}x0123 = _mm_shuffle_epi32(vq31prod${M}x0213, _MM_SHUFFLE(3, 1, 2, 0));

    const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
    $for M in range(MR):
      const __m128i vrem${M}x0123 =
        _mm_add_epi32(_mm_and_si128(vq31prod${M}x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod${M}x0123));

    const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
    const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
    $for M in range(MR):
      vacc${M}x0123 =
        _mm_sub_epi32(_mm_sra_epi32(vq31prod${M}x0123, vshift), _mm_cmpgt_epi32(vrem${M}x0123, vremainder_threshold));

    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
    $for M in range(0, MR, 2):
      __m128i vacc${M}${min(M+1, MR-1)}x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123), voutput_zero_point);

    const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
    const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
    $for M in range(0, MR, 2):
      vacc${M}${min(M+1, MR-1)}x0123 = _mm_min_epi16(_mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min), voutput_max);

    __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);

    if (nc >= 4) {
      $for M in reversed(range(1, MR)):
        $if SSE >= 4:
          *((uint32_t*) c${M}) = (uint32_t) _mm_extract_epi32(vout, ${M});
        $else:
          *((uint32_t*) c${M}) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(${M}, ${M}, ${M}, ${M})));
        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
      *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
      c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

      a = (const int8_t**restrict) ((uintptr_t) a - ks);

      nc -= 4;
    } else {
      if (nc & 2) {
        $for M in reversed(range(MR)):
          *((uint16_t*) c${M}) = (uint16_t) _mm_extract_epi16(vout, ${M * 2});
          c${M} += 2;
        vout = _mm_srli_epi32(vout, 16);
      }
      if (nc & 1) {
        $if SSE >= 4:
          $for M in reversed(range(MR)):
            *((int8_t*) c${M}) = (int8_t) _mm_extract_epi8(vout, ${M * 4});
        $else:
          $for M in reversed(range(1, MR)):
            *((int8_t*) c${M}) = (int8_t) _mm_extract_epi16(vout, ${M * 2});
          *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
      }

      nc = 0;
    }
  } while (nc != 0);
}