C  INT8_CONV2D — symmetric per-tensor int8 2D direct convolution
C  with int32 accumulation. Stride 1, no padding, no dilation.
C  Single-batch, multi-channel.
C
C  Reference: B. Jacob et al., "Quantization and Training of Neural
C  Networks for Efficient Integer-Arithmetic-Only Inference",
C  CVPR 2018 (arXiv:1712.05877). The classic NHWC/NCHW direct loop
C  form, written in the textbook six-nested-loop arrangement
C  (cf. Sze et al., "Efficient Processing of Deep Neural Networks",
C  MIT Press 2020, Ch. 4).
C
C  Hand-written reference for the Dark Factory's Phase 3
C  inference-kernel ladder, 2026-05-23. Public domain.
C
C  Shapes (column-major, NCHW layout):
C    X is W_IN x H_IN x C_IN          (INTEGER*1)  input feature map
C    W is K_W  x K_H  x C_IN x C_OUT  (INTEGER*1)  kernel weights
C    Y is W_OUT x H_OUT x C_OUT       (INTEGER)    output, int32
C
C  Output spatial dims (valid convolution, stride 1):
C    W_OUT = W_IN - K_W + 1
C    H_OUT = H_IN - K_H + 1
C
C  Overflow note: each output element accumulates K_W * K_H * C_IN
C  signed int8 x int8 products. Worst case product magnitude is
C  127 * 127 = 16129, so a 32-bit accumulator is safe while
C  K_W * K_H * C_IN <= 133164.

      SUBROUTINE INT8_CONV2D (W_IN, H_IN, C_IN, K_W, K_H, C_OUT,
     +                       X, W, Y)
C  Inputs:
C    W_IN, H_IN, C_IN  — input width, height, channels (positive)
C    K_W, K_H          — kernel width, height (positive, <= input)
C    C_OUT             — number of output channels (positive)
C    X                 — input tensor, W_IN x H_IN x C_IN int8
C    W                 — weight tensor, K_W x K_H x C_IN x C_OUT int8
C  Output:
C    Y                 — output tensor, (W_IN-K_W+1) x (H_IN-K_H+1)
C                        x C_OUT int32
      INTEGER W_IN, H_IN, C_IN, K_W, K_H, C_OUT
      INTEGER*1 X(W_IN, H_IN, C_IN)
      INTEGER*1 W(K_W, K_H, C_IN, C_OUT)
      INTEGER Y(W_IN - K_W + 1, H_IN - K_H + 1, C_OUT)
      INTEGER OC, OY, OX, IC, KY, KX, ACC
      DO 60 OC = 1, C_OUT
         DO 50 OY = 1, H_IN - K_H + 1
            DO 40 OX = 1, W_IN - K_W + 1
               ACC = 0
               DO 30 IC = 1, C_IN
                  DO 20 KY = 1, K_H
                     DO 10 KX = 1, K_W
                        ACC = ACC + X(OX + KX - 1, OY + KY - 1, IC)
     +                            * W(KX, KY, IC, OC)
   10                CONTINUE
   20             CONTINUE
   30          CONTINUE
               Y(OX, OY, OC) = ACC
   40       CONTINUE
   50    CONTINUE
   60 CONTINUE
      RETURN
      END