src/f32-gemm/neon-ld64.c.in

// Copyright 2019 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.

$assert NR % 4 == 0
$assert DATATYPE in ["F32", "QC4", "QC8"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
$VMULADDQ_LANE_F32 = "vfmaq_lane_f32" if FMA else "vmlaq_lane_f32"

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/gemm.h>


$ISA = ("neonfma" if DUP else "aarch64_neonfma") if FMA else "neon"
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${ISA}_${"dup" if DUP else "lane"}_ld64(
    size_t mr,
    size_t nc,
    size_t kc,
    const float* restrict a,
    size_t a_stride,
    $if DATATYPE == "F32":
      const float* restrict w,
    $else:
      const void* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const float* restrict acc,
    $if DATATYPE == "QC4":
      const union xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
    $else:
      const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  $if INC:
    assert(acc != NULL);

  const float* a0 = a;
  float* c0 = c;
  $for M in range(1, MR):
    const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
  $if DATATYPE == "QC4":
    const int16x8_t vminus_kernel_zero_point = vld1q_dup_s16(&params->scalar.minus_kernel_zero_point[0]);
    const uint8x8_t vmask = vmov_n_u8(UINT8_C(0xF));

  do {
    $if INC:
      $for M in range(MR):
        $for N in range(0, NR, 4):
          float32x4_t vacc${M}x${ABC[N:N+4]} = vld1q_f32(acc); acc += 4;
    $else:
      $for N in range(0, NR, 4):
        $if DATATYPE == "F32":
          float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w += 4;
        $else:
          float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w = (const float*) w + 4;
      $for M in range(1, MR):
        $for N in range(0, NR, 4):
          float32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};

    size_t k = kc;
    for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
      $for M in range(MR):
        const float32x2_t va${M} = vld1_f32(a${M}); a${M} += 2;

      $if DATATYPE == "F32":
        $for L in range(2):
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;
      $elif DATATYPE == "QC4":
        $for N in range(0, NR, 8):
          const uint8x8_t vw${ABC[N:N+8]}c01 = vld1_u8(w); w = (const uint8_t*) w + 8;
          const uint8x8_t vw${ABC[N:N+8]}c0 = vand_u8(vw${ABC[N:N+8]}c01, vmask);
          const uint8x8_t vw${ABC[N:N+8]}c1 = vshr_n_u8(vw${ABC[N:N+8]}c01, 4);
          $for L in range(2):
            const int16x8_t vxw${ABC[N:N+8]}c${L} = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vw${ABC[N:N+8]}c${L}));
          $for L in range(2):
            const int32x4_t vxw${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vxw${ABC[N:N+8]}c${L}));
            const int32x4_t vxw${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vxw${ABC[N:N+8]}c${L}));
        $for N in range(0, NR, 4):
          $for L in range(2):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vxw${ABC[N:N+4]}c${L});
      $elif DATATYPE == "QC8":
        $for N in range(0, NR, 8):
          $for L in range(2):
            const int8x8_t vw${ABC[N:N+8]}c${L} = vld1_s8(w); w = (const int8_t*) w + 8;
          $for L in range(2):
            const int16x8_t vxw${ABC[N:N+8]}c${L} = vmovl_s8(vw${ABC[N:N+8]}c${L});
          $for L in range(2):
            const int32x4_t vxw${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vxw${ABC[N:N+8]}c${L}));
            const int32x4_t vxw${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vxw${ABC[N:N+8]}c${L}));
        $for N in range(0, NR, 4):
          $for L in range(2):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vxw${ABC[N:N+4]}c${L});

      $for L in range(2):
        $if DUP:
          $for M in range(MR):
            const float32x4_t va${M}c${L} = vdupq_lane_f32(va${M}, ${L});
          $for N in range(0, NR, 4):
            $for M in range(MR):
              vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}c${L}, vb${ABC[N:N+4]}c${L});
        $else:
           $for N in range(0, NR, 4):
             $for M in range(MR):
               vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_LANE_F32}(vacc${M}x${ABC[N:N+4]}, vb${ABC[N:N+4]}c${L}, va${M}, ${L});
    }
    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const float32x4_t va${M} = vld1q_dup_f32(a${M}); a${M} += 1;

      $if DATATYPE == "F32":
        $for N in range(0, NR, 4):
          const float32x4_t vb${ABC[N:N+4]} = vld1q_f32(w); w += 4;
      $elif DATATYPE == "QC4":
        $for N in range(0, NR, 8):
          const uint8x8_t vw${ABC[N:N+8]} = vld1_u8(w); w = (const uint8_t*) w + 8;
          const int16x8_t vxw${ABC[N:N+8]} = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vw${ABC[N:N+8]}));
          const int32x4_t vxw${ABC[N:N+4]} = vmovl_s16(vget_low_s16(vxw${ABC[N:N+8]}));
          const int32x4_t vxw${ABC[N+4:N+8]} = vmovl_s16(vget_high_s16(vxw${ABC[N:N+8]}));
        $for N in range(0, NR, 4):
          const float32x4_t vb${ABC[N:N+4]} = vcvtq_f32_s32(vxw${ABC[N:N+4]});
      $elif DATATYPE == "QC8":
        $for N in range(0, NR, 4):
          const int8x8_t vw${ABC[N:N+4]}${ABC[N:N+4]} = vreinterpret_s8_u32(vld1_dup_u32(w)); w = (const int8_t*) w + 4;
        $for N in range(0, NR, 4):
          const int16x8_t vxw${ABC[N:N+4]}${ABC[N:N+4]} = vmovl_s8(vw${ABC[N:N+4]}${ABC[N:N+4]});
        $for N in range(0, NR, 4):
          const int32x4_t vxw${ABC[N:N+4]} = vmovl_s16(vget_low_s16(vxw${ABC[N:N+4]}${ABC[N:N+4]}));
        $for N in range(0, NR, 4):
          const float32x4_t vb${ABC[N:N+4]} = vcvtq_f32_s32(vxw${ABC[N:N+4]});

      $for N in range(0, NR, 4):
        $for M in range(MR):
          vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]});
    }
    $if DATATYPE in ["QC8", "QC4"]:
      $for N in range(0, NR, 4):
        const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32(w); w = (const float*) w + 4;
        $for M in range(MR):
          vacc${M}x${ABC[N:N+4]} = vmulq_f32(vacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = vminq_f32(vacc${M}x${ABC[N:N+4]}, vmax);

    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = vmaxq_f32(vacc${M}x${ABC[N:N+4]}, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        vst1q_f32(c${M}, vacc${M}x${ABC[0:4]});
        $for N in range(4, NR, 4):
          vst1q_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

      $for M in reversed(range(MR)):
        a${M} = (const float*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};

    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 2:
              $for N in range(0, 1 << LOG2N, 4):
                $for M in reversed(range(MR)):
                  vst1q_f32(c${M}, vacc${M}x${ABC[N:N+4]}); c${M} += 4;

              $for M in reversed(range(MR)):
                $for N in range(0, NR - (1 << LOG2N), 4):
                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 <<   LOG2N)+4]};
            $elif LOG2N == 1:
              $for M in reversed(range(MR)):
                vst1_f32(c${M}, vacc${M}x${ABC[0:2]}); c${M} += 2;

              $for M in reversed(range(MR)):
                vacc${M}x${ABC[0:2]} = vget_high_f32(vacc${M}x${ABC[0:4]});
            $elif LOG2N == 0:
              $for M in reversed(range(MR)):
                vst1_lane_f32(c${M}, vacc${M}x${ABC[0:2]}, 0);
          }
          $if LOG2N == 2:
            $for M in reversed(range(MR)):
              float32x2_t vacc${M}x${ABC[0:2]} = vget_low_f32(vacc${M}x${ABC[0:4]});

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