// Copyright 2024 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 DATATYPE in ["QC8", "QD8", "QS8"]
$assert DATATYPE == "QD8" or REQUANTIZATION == "FP32"
$assert DATATYPE != "QD8" or not REQUANTIZATION
$assert NR == 16
$UNROLL = locals().get("UNROLL", 1)
$assert ACCUMULATORS == UNROLL or ACCUMULATORS == 1
#include <assert.h>

#include <wasm_simd128.h>

#include "xnnpack/gemm.h"
#include "xnnpack/math.h"


$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUV"
$DATATYPE_SPEC = {"QC8": "qs8_qc8w", "QD8": "qd8_f32_qc8w", "QS8": "qs8"}[DATATYPE]
$REQUANTIZATION_SPEC = "_" + REQUANTIZATION.lower() if REQUANTIZATION else ""
$PARAMS_TYPE = {"QC8": "union xnn_qs8_qc8w_conv_minmax_params", "QD8": "union xnn_f32_minmax_params", "QS8": "union xnn_qs8_conv_minmax_params"}[DATATYPE]
$OUT_T = "float" if DATATYPE == "QD8" else "int8_t"
$def VACC(M,K=0):
$  return f"vacc{M}x{K}" if UNROLL > 1 else f"vacc{M}"
$ACC_POSTFIX=f"_acc{ACCUMULATORS}" if ACCUMULATORS > 1 else ""
void xnn_${DATATYPE_SPEC}_igemm_minmax${REQUANTIZATION_SPEC}_ukernel_${MR}x${NR}c4__wasm${"sdot" if SDOT else "usdot"}${"_u" + str(UNROLL) if UNROLL > 1 else ""}${ACC_POSTFIX}(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const int8_t** restrict a,
    const void* restrict w,
    ${OUT_T}* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const int8_t* zero,
    $if DATATYPE in ["QD8"]:
      const int8_t* zero_data,
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)],
      const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
    $else:
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  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, 4 * sizeof(int8_t));
  ${OUT_T}* c0 = c;
  $for M in range(1, MR):
    ${OUT_T}* c${M} = (${OUT_T}*) ((uintptr_t) c${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};
      }

  $if DATATYPE == "QD8":
    const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
    const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
    XNN_FORCE_REALIZATION(vmin);
    XNN_FORCE_REALIZATION(vmax);
  $else:
    $if DATATYPE != "QC8":
      const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
      XNN_FORCE_REALIZATION(vscale);
    const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
    const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
    const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
    const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
    const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
    //XNN_FORCE_REALIZATION(vmagic_bias);
    //XNN_FORCE_REALIZATION(vmagic_min);
    //XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
    //XNN_FORCE_REALIZATION(voutput_max);

  $if not SDOT:
    const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
    XNN_FORCE_REALIZATION(vsign_mask);
  do {
    $if DATATYPE == "QD8":
      v128_t vksum${ABC[0:4]} = wasm_v128_load((const int32_t*) w);
      $for N in range(4, NR, 4):
        v128_t vksum${ABC[N:N+4]} = wasm_v128_load((const int32_t*) w + ${N});
      $if SDOT:
        const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
      $else:
        const v128_t vinput_zero_point = wasm_i32x4_splat((int32_t) quantization_params->zero_point + 128);
      $for M in range(MR):
        $for N in range(0, NR, 4):
          v128_t ${VACC(M)}x${ABC[N:N+4]} = wasm_i32x4_mul(vksum${ABC[N:N+4]}, vinput_zero_point);
          $if ACCUMULATORS > 1:
            $for K in range(1, UNROLL):
              v128_t ${VACC(M, K)}x${ABC[N:N+4]} = wasm_u32x4_const(0, 0, 0, 0);
    $else:
      v128_t ${VACC(0)}x0123 = wasm_v128_load(w);
      $for N in range(4, NR, 4):
        v128_t ${VACC(0)}x${ABC[N:N+4]} = wasm_v128_load((const int32_t*) w + ${N});
      $if ACCUMULATORS > 1:
        $for K in range(1, UNROLL):
          $for N in range(0, NR, 4):
            v128_t ${VACC(0, K)}x${ABC[N:N+4]} = wasm_u32x4_const(0, 0, 0, 0);
      $for M in range(1, MR, 1):
        $for N in range(0, NR, 4):
          $if ACCUMULATORS > 1:
            $for K in range(UNROLL):
              v128_t ${VACC(M, K)}x${ABC[N:N+4]}= ${VACC(0, K)}x${ABC[N:N+4]};
          $else:
            v128_t ${VACC(M)}x${ABC[N:N+4]}= ${VACC(0)}x${ABC[N:N+4]};
    w = (const int32_t*) w + ${NR};

    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);
        $if DATATYPE == "QD8":
          } else {
            a${M} = zero_data;
        }
      a += ${MR};

      size_t k = kc;
      $if UNROLL > 1:
        while (k >= ${UNROLL * 4} * sizeof(int8_t)) {
          $for M in range(MR):
            v128_t va${M}x0x0123 = wasm_v128_load32_splat(a${M});
            $for K in range(1, UNROLL):
              v128_t va${M}x${K}x0123 = wasm_v128_load32_splat(a${M} + ${4 * K});
            a${M} += ${4 * UNROLL};

          $if not SDOT:
            $for M in range(MR):
              $for K in range(UNROLL):
                va${M}x${K}x0123 = wasm_v128_xor(va${M}x${K}x0123, vsign_mask);

          $for K in range(UNROLL):
            $for N in range(0, NR, 4):
              $if N == 0 and K == 0:
                const v128_t vb${K}x${ABC[N:N+4]} = wasm_v128_load((const int8_t*) w);
              $else:
                const v128_t vb${K}x${ABC[N:N+4]} = wasm_v128_load((const int8_t*) w + ${4 * N + 4 * NR * K});

          $if ACCUMULATORS > 1:
            $for K in range(UNROLL):
              $for M in range(MR):
                $for N in range(0, NR, 4):
                  ${VACC(M, K)}x${ABC[N:N+4]} = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb${K}x${ABC[N:N+4]}, va${M}x${K}x0123, ${VACC(M, K)}x${ABC[N:N+4]});
          $else:
            $for K in range(UNROLL):
              $for M in range(MR):
                $for N in range(0, NR, 4):
                  ${VACC(M)}x${ABC[N:N+4]} = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb${K}x${ABC[N:N+4]}, va${M}x${K}x0123, ${VACC(M)}x${ABC[N:N+4]});

          w = (const int8_t*) w + ${NR * 4 * UNROLL};
          k -= ${UNROLL * 4} * sizeof(int8_t);
        }
        $if ACCUMULATORS > 1:
          $PAIRS = [(i,) for i in range(UNROLL)]
          $while len(PAIRS) > 1:
            $TPLS=[PAIRS[i:i+2] for i in range(0, len(PAIRS), 2)]
            $PAIRS = [(P1[0],P2[0]) for P1, P2 in TPLS]
            $for K1, K2 in PAIRS:
              $for M in range(MR):
                $for N in range(0, NR, 4):
                  ${VACC(M)}x${ABC[N:N+4]} = wasm_i32x4_add(${VACC(M, K1)}x${ABC[N:N+4]}, ${VACC(M, K2)}x${ABC[N:N+4]});

      while (k != 0) {
        $for M in range(MR):
          v128_t va${M}x0123 = wasm_v128_load32_splat(a${M});
          a${M} += 4;

        $if not SDOT:
          $for M in range(MR):
            va${M}x0123 = wasm_v128_xor(va${M}x0123, vsign_mask);

        $for N in range(0, NR, 4):
          $if N == 0:
            const v128_t vb${ABC[N:N+4]} = wasm_v128_load((const int8_t*) w);
          $else:
            const v128_t vb${ABC[N:N+4]} = wasm_v128_load((const int8_t*) w + ${4 * N});

        $for M in range(MR):
          $for N in range(0, NR, 4):
            ${VACC(M)}x${ABC[N:N+4]} = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb${ABC[N:N+4]}, va${M}x0123, ${VACC(M)}x${ABC[N:N+4]});

        w = (const int8_t*) w + ${NR * 4};
        k -= 4 * sizeof(int8_t);
      }

      p -= ${MR} * sizeof(void*);
    } while (p != 0);

    $for M in range(MR):
      $for N in range(0, NR, 4):
        ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_convert_i32x4(${VACC(M)}x${ABC[N:N+4]});

    $if DATATYPE == "QD8":
      const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_mul(${VACC(M)}x${ABC[N:N+4]}, vinput_scale);

      $for N in range(0, NR, 4):
        const v128_t vfilter_output_scale${ABC[N:N+4]} = wasm_v128_load(w);
        w = (const float*) w + 4;
      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_mul(${VACC(M)}x${ABC[N:N+4]}, vfilter_output_scale${ABC[N:N+4]});

      $for N in range(0, NR, 4):
        const v128_t vbias${ABC[N:N+4]} = wasm_v128_load(w);
        w = (const float*) w + 4;
      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_add(${VACC(M)}x${ABC[N:N+4]}, vbias${ABC[N:N+4]});

      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_pmax(${VACC(M)}x${ABC[N:N+4]}, vmin);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_pmin(${VACC(M)}x${ABC[N:N+4]}, vmax);

      if XNN_LIKELY(nc >= ${NR}) {
        $for M in reversed(range(MR)):
          wasm_v128_store(c${M}, ${VACC(M)}x${ABC[0:4]});
          $for N in range(4, NR, 4):
            wasm_v128_store(c${M} + ${N}, ${VACC(M)}x${ABC[N:N+4]});

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

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

        nc -= ${NR};
      } else {
        $if NR == 16:
          if (nc & 8) {
            $for M in reversed(range(MR)):
              $for N in range(0, 8, 4):
                wasm_v128_store(c${M}, ${VACC(M)}x${ABC[N:N+4]});
                ${VACC(M)}x${ABC[N:N+4]} = ${VACC(M)}x${ABC[N+8:N+12]};
                c${M} += 4;
          }
        if (nc & 4) {
          $for M in reversed(range(MR)):
            wasm_v128_store(c${M}, ${VACC(M)}x0123);
            ${VACC(M)}x0123 = ${VACC(M)}x4567;
            c${M} += 4;
        }
        if (nc & 2) {
          $for M in reversed(range(MR)):
            wasm_v128_store64_lane(c${M}, ${VACC(M)}x0123, 0);
            ${VACC(M)}x0123 = wasm_v64x2_shuffle(${VACC(M)}x0123, ${VACC(M)}x0123, 1, 1);
            c${M} += 2;
        }
        if (nc & 1) {
          $for M in reversed(range(MR)):
            wasm_v128_store32_lane(c${M}, ${VACC(M)}x0123, 0);
        }
        nc = 0;
      }
    $else:
      $if DATATYPE == "QC8":
        $for N in range(0, NR, 4):
          const v128_t vscale${ABC[N:N+4]} = wasm_v128_load(w);
          w = (const float*) w + 4;
        $for M in range(MR):
          $for N in range(0, NR, 4):
            ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_mul(${VACC(M)}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
      $else:
        $for M in range(MR):
          ${VACC(M)}x0123 = wasm_f32x4_mul(${VACC(M)}x0123, vscale);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_f32x4_add(${VACC(M)}x${ABC[N:N+4]}, vmagic_bias);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_i32x4_max(${VACC(M)}x${ABC[N:N+4]}, vmagic_min);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          ${VACC(M)}x${ABC[N:N+4]} = wasm_i32x4_sub(${VACC(M)}x${ABC[N:N+4]}, vmagic_bias_less_output_zero_point);

      $for M in range(MR):
        $for N in range(0, NR, 8):
          v128_t vacc${M}x${ABC[N:N+8]} = wasm_i16x8_narrow_i32x4(${VACC(M)}x${ABC[N:N+4]}, ${VACC(M)}x${ABC[N+4:N+8]});

      $for M in range(MR):
        $for N in range(0, NR, 8):
          vacc${M}x${ABC[N:N+8]} = wasm_i8x16_narrow_i16x8(vacc${M}x${ABC[N:N+8]}, vacc${M}x${ABC[N:N+8]});

      $for M in range(MR):
        $for N in range(0, NR, 8):
          vacc${M}x${ABC[N:N+8]} = wasm_i8x16_min(vacc${M}x${ABC[N:N+8]}, voutput_max);

      if (nc >= ${NR}) {
        $for M in reversed(range(MR)):
          wasm_v128_store64_lane(c${M}, vacc${M}x${ABC[0:8]}, 0);
          $for N in range(8, NR, 8):
            wasm_v128_store64_lane(c${M} + ${N}, vacc${M}x${ABC[N:N+8]}, 0);

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

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

        nc -= ${NR};
      } else {
        $if NR == 16:
          if (nc & 8) {
            $for M in reversed(range(MR)):
              wasm_v128_store64_lane(c${M}, vacc${M}x${ABC[0:8]}, 0);
              c${M} += 8;

            $for M in range(MR):
              vacc${M}x${ABC[0:8]} = vacc${M}x${ABC[8:16]};
          }
        if (nc & 4) {
          $for M in reversed(range(MR)):
            wasm_v128_store32_lane(c${M}, vacc${M}x${ABC[0:8]}, 0);
            c${M} += 4;

          $for M in range(MR):
            vacc${M}x${ABC[0:8]} = wasm_u64x2_shr(vacc${M}x${ABC[0:8]}, 32);
        }
        if (nc & 2) {
          $for M in reversed(range(MR)):
            wasm_v128_store16_lane(c${M}, vacc${M}x${ABC[0:8]}, 0);
            c${M} += 2;

          $for M in range(MR):
            vacc${M}x${ABC[0:8]} = wasm_u32x4_shr(vacc${M}x${ABC[0:8]}, 16);
        }
        if (nc & 1) {
          $for M in reversed(range(MR)):
            wasm_v128_store8_lane(c${M}, vacc${M}x${ABC[0:8]}, 0);
        }

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