// 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.

#include <assert.h>

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


$assert ACTIVATION in ["LINEAR", "RELU", "MINMAX"]
$MIN_F32 = "__builtin_wasm_min_f32" if WASM else "math_min_f32"
$MAX_F32 = "__builtin_wasm_max_f32" if WASM else "math_max_f32"
$KERNEL = "gemminc" if INC else "gemm"
$SUFFIX = {"LINEAR": "", "RELU": "_relu", "MINMAX": "_minmax"}[ACTIVATION]
$PARAMS = {"LINEAR": "xnn_f32_default_params", "RELU": "xnn_f32_relu_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION]
void xnn_f32_${KERNEL}${SUFFIX}_ukernel_${MR}x${NR}__${"wasm" if WASM else "scalar"}(
    size_t mr,
    size_t nc,
    size_t kc,
    const float* restrict a,
    size_t a_stride,
    const float* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const float*restrict acc,
    const union ${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 ACTIVATION == "MINMAX":
    const float vmin = params->scalar.min;
    const float vmax = params->scalar.max;
  do {
    $if INC:
      $for M in range(MR):
        $for N in range(NR):
          float vacc${M}${N} = acc[${M*NR+N}];
      acc += ${MR*NR};
    $else:
      $for N in range(NR):
        float vacc0${N} = w[${N}];
      w += ${NR};
      $for M in range(1, MR):
        $for N in range(NR):
          float vacc${M}${N} = vacc0${N};

    size_t k = kc;
    do {
      $for M in range(MR):
        const float va${M} = *a${M}++;

      $for N in range(NR):
        const float vb${N} = w[${N}];
      w += ${NR};

      $for M in range(MR):
        $for N in range(NR):
          vacc${M}${N} += va${M} * vb${N};

      k -= sizeof(float);
    } while (k != 0);

    $if ACTIVATION == "MINMAX":
      $for M in range(MR):
        $for N in range(NR):
          vacc${M}${N} = ${MAX_F32}(vacc${M}${N}, vmin);

      $for M in range(MR):
        $for N in range(NR):
          vacc${M}${N} = ${MIN_F32}(vacc${M}${N}, vmax);
    $elif ACTIVATION == "RELU":
      $for M in range(MR):
        $for N in range(NR):
          vacc${M}${N} = ${MAX_F32}(vacc${M}${N}, 0.0f);

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

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

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

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