// 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 MR % 4 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/spmm.h>

$ARCH_SUFFIX = "_x86" if X86 else "_arm"

void xnn_f32_spmm_minmax_ukernel_${MR}x${NR}__wasmsimd${ARCH_SUFFIX}_pipelined${"_x" + str(UNROLL) if UNROLL > 1 else ""}(
    size_t mc,
    size_t nc,
    const float*restrict input,
    const float*restrict weights,
    const int32_t*restrict widx_dmap,
    const uint32_t*restrict nidx_nnzmap,
    float*restrict output,
    size_t output_stride,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mc != 0);
  assert(mc % sizeof(float) == 0);
  assert(nc != 0);

  const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
  const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
  size_t output_decrement = output_stride * nc - ${MR} * sizeof(float);
  while XNN_LIKELY(mc >= ${MR} * sizeof(float)) {
    const float*restrict w = weights;
    const int32_t* dmap = widx_dmap;
    const uint32_t* nnzmap = nidx_nnzmap;
    v128_t vw = wasm_v32x4_load_splat(w); w += 1;
    intptr_t diff = *dmap++;
    $for M in range(0, MR, 4):
      v128_t vi${ABC[M:M+4]} = wasm_v128_load(input + ${M});
    size_t n = nc;
    do {
      uint32_t nnz = *nnzmap++;
       $for M in range(0, MR, 4):
        v128_t vacc${ABC[M:M+4]} = vw;
      vw = wasm_v32x4_load_splat(w); w += 1;

      $if UNROLL > 1:
        for (; nnz >= ${UNROLL}; nnz -= ${UNROLL}) {
          $for K in range(0, UNROLL):
            $for M in range(0, MR, 4):
              vacc${ABC[M:M+4]} = wasm_f32x4_add(vacc${ABC[M:M+4]}, wasm_f32x4_mul(vi${ABC[M:M+4]},   vw));
            input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
            diff = *dmap++;
            vw = wasm_v32x4_load_splat(w); w += 1;
            $for M in range(0, MR, 4):
              vi${ABC[M:M+4]} = wasm_v128_load(input + ${M});
        }

      if XNN_LIKELY(nnz != 0) {
        do {
          $for M in range(0, MR, 4):
            vacc${ABC[M:M+4]} = wasm_f32x4_add(vacc${ABC[M:M+4]}, wasm_f32x4_mul(vi${ABC[M:M+4]}, vw));
          input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);

          diff = *dmap++;
          vw = wasm_v32x4_load_splat(w); w += 1;
          $for M in range(0, MR, 4):
            vi${ABC[M:M+4]} = wasm_v128_load(input + ${M});
        } while (--nnz != 0);
      }
      $if X86:
        $for M in range(0, MR, 4):
          v128_t vout${ABC[M:M+4]} = wasm_v128_bitselect(vacc${ABC[M:M+4]}, vmax, wasm_f32x4_le(vacc${ABC[M:M+4]}, vmax));
        $for M in range(0, MR, 4):
          vout${ABC[M:M+4]} = wasm_v128_bitselect(vmin, vout${ABC[M:M+4]}, wasm_f32x4_lt(vout${ABC[M:M+4]}, vmin));
      $else:
        $for M in range(0, MR, 4):
          v128_t vout${ABC[M:M+4]} = wasm_f32x4_min(vacc${ABC[M:M+4]}, vmax);
        $for M in range(0, MR, 4):
          vout${ABC[M:M+4]} = wasm_f32x4_max(vout${ABC[M:M+4]}, vmin);
      wasm_v128_store(output, vout0123);
      $for M in range(4, MR, 4):
        wasm_v128_store(output + ${M}, vout${ABC[M:M+4]});
      output = (float*restrict) ((uintptr_t) output + output_stride);
    } while (--n != 0);
    output = (float*restrict) ((uintptr_t) output - output_decrement);
    input += ${MR};
    mc -= ${MR} * sizeof(float);
  }
  if XNN_UNLIKELY(mc != 0) {
    $for LOG2M in reversed(range((MR - 1).bit_length())):
      $SUBMR = 1 << LOG2M
      $if SUBMR * 2 >= MR:
        output_decrement += ${MR - SUBMR} * sizeof(float);
      $else:
        output_decrement += ${SUBMR} * sizeof(float);
      if (mc & (${SUBMR} * sizeof(float))) {
        const float*restrict w = weights;
        const int32_t* dmap = widx_dmap;
        const uint32_t* nnzmap = nidx_nnzmap;
        size_t n = nc;
        do {
          uint32_t nnz = *nnzmap++;
          $if SUBMR == 1:
            v128_t vacc0 = wasm_v32x4_load_splat(w); w += 1;
          $elif SUBMR == 2:
            v128_t vacc01 = wasm_v32x4_load_splat(w); w += 1;
          $else:
            v128_t vacc0123 = wasm_v32x4_load_splat(w); w += 1;
          $for M in range(4, SUBMR, 4):
            v128_t vacc${ABC[M:M+4]} = vacc0123;
          if XNN_LIKELY(nnz != 0) {
            do {
              const intptr_t diff = *dmap++;
              $if SUBMR >= 4:
                const v128_t vi0123 = wasm_v128_load(input);
              $elif SUBMR == 2:
                const v128_t vi01 = wasm_v64x2_load_splat(input);
              $elif SUBMR == 1:
                const v128_t vi0 = wasm_v32x4_load_splat(input);
              $for M in range(4, SUBMR, 4):
                const v128_t vi${ABC[M:M+4]} = wasm_v128_load(input + ${M});
              input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
              const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
              $if SUBMR == 1:
                vacc${ABC[0]} = wasm_f32x4_add(vacc${ABC[0]}, wasm_f32x4_mul(vi${ABC[0]}, vw));
              $else:
                $for M in range(0, SUBMR, 4):
                  vacc${ABC[M:min(M+4,SUBMR)]} = wasm_f32x4_add(vacc${ABC[M:min(M+4,SUBMR)]}, wasm_f32x4_mul(vi${ABC[M:min(M+4,SUBMR)]}, vw));
            } while (--nnz != 0);
          }
          $if SUBMR == 1:
            $if X86:
              v128_t vout${ABC[0]} = wasm_v128_bitselect(vacc${ABC[0]}, vmax, wasm_f32x4_le(vacc${ABC[0]}, vmax));
              vout${ABC[0]} = wasm_v128_bitselect(vmin, vout${ABC[0]}, wasm_f32x4_lt(vout${ABC[0]}, vmin));
            $else:
              v128_t vout${ABC[0]} = wasm_f32x4_min(vacc${ABC[0]}, vmax);
              vout${ABC[0]} = wasm_f32x4_max(vout${ABC[0]}, vmin);
          $else:
            $if X86:
              $for M in range(0, SUBMR, 4):
                v128_t vout${ABC[M:min(M+4,SUBMR)]} = wasm_v128_bitselect(vacc${ABC[M:min(M+4,SUBMR)]}, vmax, wasm_f32x4_le(vacc${ABC[M:min(M+4,SUBMR)]}, vmax));
              $for M in range(0, SUBMR, 4):
                vout${ABC[M:min(M+4,SUBMR)]} = wasm_v128_bitselect(vmin, vout${ABC[M:min(M+4,SUBMR)]}, wasm_f32x4_lt(vout${ABC[M:min(M+4,SUBMR)]}, vmin));
            $else:
              $for M in range(0, SUBMR, 4):
                v128_t vout${ABC[M:min(M+4,SUBMR)]} = wasm_f32x4_min(vacc${ABC[M:min(M+4,SUBMR)]}, vmax);
              $for M in range(0, SUBMR, 4):
                vout${ABC[M:min(M+4,SUBMR)]} = wasm_f32x4_max(vout${ABC[M:min(M+4,SUBMR)]}, vmin);
          $if SUBMR >= 4:
            wasm_v128_store(output, vout0123);
          $elif SUBMR == 2:
            *((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
          $elif SUBMR == 1:
            *output = wasm_f32x4_extract_lane(vout0, 0);

          $for M in range(4, SUBMR, 4):
            wasm_v128_store(output + ${M}, vout${ABC[M:M+4]});
          output = (float*restrict) ((uintptr_t) output + output_stride);
        } while (--n != 0);
        output = (float*restrict) ((uintptr_t) output - output_decrement);
        input += ${SUBMR};
      }
  }
}