// 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.

$assert VARIANT in ["LD64", "LD128", "EXTENDED"]
$assert MR <= 4
#include <assert.h>

#include <wasm_simd128.h>

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


$LOAD_SUFFIX = {"LD128": "_ld128", "LD64": "_ld64", "EXTENDED": ""}[VARIANT]
$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else ""
void xnn_qs8_gemm${GEMM_SUFFIX}_minmax_ukernel_${MR}x4c8__wasmsimd${LOAD_SUFFIX}(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_qs8_gemm${GEMM_SUFFIX}_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 8);
  const int8_t* a0 = a;
  int8_t* c0 = c;
  $for M in range(1, MR):
    const int8_t* a${M} = (const int8_t*) ((uintptr_t) a${M-1} + a_stride);
    int8_t* c${M} = (int8_t*) ((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};
      }

  const v128_t vzero = wasm_f64x2_splat(0.0);
  do {
    $for N in range(4):
      v128_t vacc0x${N} = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[${N}]);
    $for M in range(1, MR):
      $for N in range(4):
        v128_t vacc${M}x${N} = vacc0x${N};
    w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));

    size_t k = 0;
    while (k < kc) {
      $for M in range(MR):
        const v128_t vxa${M} = wasm_i16x8_load_8x8(a${M});
        a${M} += 8;

      $if VARIANT == "LD128":
        $for N in range(0, 4, 2):
          $if N == 0:
            const v128_t vb${N}${N+1} = wasm_v128_load(w);
          $else:
            const v128_t vb${N}${N+1} = wasm_v128_load((const void*) ((uintptr_t) w + ${N * 8} * sizeof(int8_t)));
          const v128_t vxb${N} = wasm_i16x8_widen_low_i8x16(vb${N}${N+1});
          const v128_t vxb${N+1} = wasm_i16x8_widen_high_i8x16(vb${N}${N+1});

          $for M in range(MR):
            const v128_t vprod${M}x${N} = wasm_i16x8_mul(vxb${N}, vxa${M});
            vacc${M}x${N} = wasm_i32x4_add(vacc${M}x${N}, wasm_i32x4_widen_low_i16x8(vprod${M}x${N}));

          $for M in range(MR):
            const v128_t vprod${M}x${N+1} = wasm_i16x8_mul(vxb${N+1}, vxa${M});
            vacc${M}x${N+1} = wasm_i32x4_add(vacc${M}x${N+1}, wasm_i32x4_widen_low_i16x8(vprod${M}x${N+1}));
            vacc${M}x${N} = wasm_i32x4_add(vacc${M}x${N}, wasm_i32x4_widen_high_i16x8(vprod${M}x${N}));

          $for M in range(MR):
            vacc${M}x${N+1} = wasm_i32x4_add(vacc${M}x${N+1}, wasm_i32x4_widen_high_i16x8(vprod${M}x${N+1}));
      $else:
        $for N in range(4):
          $if VARIANT == "LD64":
            $if N == 0:
              const v128_t vxb${N} = wasm_i16x8_load_8x8(w);
            $else:
              const v128_t vxb${N} = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + ${N * 8} * sizeof(int8_t)));
          $elif VARIANT == "EXTENDED":
            $if N == 0:
              const v128_t vxb${N} = wasm_v128_load(w);
            $else:
              const v128_t vxb${N} = wasm_v128_load((const void*) ((uintptr_t) w + ${N * 8} * sizeof(int16_t)));

          $for M in range(MR):
            const v128_t vprod${M}x${N} = wasm_i16x8_mul(vxa${M}, vxb${N});
            vacc${M}x${N} = wasm_i32x4_add(vacc${M}x${N}, wasm_i32x4_widen_low_i16x8(vprod${M}x${N}));
            vacc${M}x${N} = wasm_i32x4_add(vacc${M}x${N}, wasm_i32x4_widen_high_i16x8(vprod${M}x${N}));

      $if VARIANT == "EXTENDED":
        w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
      $else:
        w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
      k += 8 * sizeof(int8_t);
    }

    $for M in range(MR):
      const v128_t vacc${M}x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc${M}x0, vacc${M}x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc${M}x0, vacc${M}x2, 2, 6, 3, 7));
      const v128_t vacc${M}x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc${M}x1, vacc${M}x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc${M}x1, vacc${M}x3, 2, 6, 3, 7));

    $for M in range(MR):
      v128_t vacc${M}x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc${M}x02, vacc${M}x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc${M}x02, vacc${M}x13, 2, 6, 3, 7));

    $for M in range(MR):
      const v128_t vsign${M}x0123 = wasm_i32x4_lt(vacc${M}x0123, vzero);

    $for M in range(MR):
      const v128_t vacc${M}x01 = wasm_v32x4_shuffle(vacc${M}x0123, vsign${M}x0123, 0, 4, 1, 5);

    const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
    const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
    $for M in range(MR):
      const v128_t vprod${M}x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc${M}x01, vmultiplier), vrounding);
      const v128_t vacc${M}x23 = wasm_v32x4_shuffle(vacc${M}x0123, vsign${M}x0123, 2, 6, 3, 7);

    $for M in range(MR):
      const v128_t vprod${M}x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc${M}x23, vmultiplier), vrounding);

    $for M in range(MR):
      const v128_t vq31prod${M}x0123 = wasm_v32x4_shuffle(vprod${M}x01, vprod${M}x23, 1, 3, 5, 7);

    const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
    $for M in range(MR):
      const v128_t vrem${M}x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod${M}x0123, vremainder_mask), wasm_i32x4_lt(vq31prod${M}x0123, vzero));

    const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
    const int32_t vshift = params->wasmsimd.shift;
    $for M in range(MR):
      vacc${M}x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${M}x0123, vshift), wasm_i32x4_gt(vrem${M}x0123, vthreshold));

    const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
    $for M in range(0, MR, 2):
      v128_t vacc${M}${min(M+1, MR-1)}x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123), voutput_zero_point);

    $if MR > 2:
      v128_t vout = wasm_i8x16_narrow_i16x8(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
    $else:
      v128_t vout = wasm_i8x16_narrow_i16x8(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);

    const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
    vout = wasm_i8x16_max(vout, voutput_min);

    const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
    vout = wasm_i8x16_min(vout, voutput_max);

    if (nc >= 4) {
      $for M in range(MR):
        *((float*) c${M}) = (float) wasm_f32x4_extract_lane(vout, ${M});

      $for M in range(MR):
        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);

      nc -= 4;
    } else {
      if (nc & 2) {
        $for M in range(MR):
          *((uint16_t*) c${M}) = (uint16_t) wasm_i16x8_extract_lane(vout, ${M * 2});
          c${M} += 2;
        vout = wasm_u32x4_shr(vout, 16);
      }
      if (nc & 1) {
        $for M in range(MR):
          *c${M} = (int8_t) wasm_i8x16_extract_lane(vout, ${M * 4});
      }

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