vgg_fwd_bench.md

using Metalhead
using Metalhead: classify
using TimerOutputs

vgg = VGG19()
img = load("190px-Asian_elephant_-_melbourne_zoo.jpg")

begin
    reset_timer!()
    classify(vgg, img)
    print_timer()
end

Initial:

 ───────────────────────────────────────────────────────────────────────────────
                                        Time                   Allocations
                                ──────────────────────   ───────────────────────
        Tot / % measured:            766ms / 89.0%            675MiB / 96.4%

 Section                ncalls     time   %tot     avg     alloc   %tot      avg
 ───────────────────────────────────────────────────────────────────────────────
 conv                       16    466ms  68.3%  29.1ms    510MiB  78.4%  31.9MiB
   conv2d!                  16    366ms  53.7%  22.9ms    396MiB  60.9%  24.8MiB
     im2col2d!              16    195ms  28.6%  12.2ms   39.6MiB  6.09%  2.48MiB
       im2col_2d!           16    183ms  26.8%  11.4ms         -  0.00%        -
     gemm                   16    141ms  20.6%  8.79ms         -  0.00%        -
 maxpool2d!                  5    139ms  20.3%  27.7ms    140MiB  21.5%  28.0MiB
   max_pooling2d_fwd!        5    139ms  20.3%  27.7ms    140MiB  21.5%  28.0MiB
 dense                       3   77.3ms  11.3%  25.8ms   72.6KiB  0.01%  24.2KiB
 ───────────────────────────────────────────────────────────────────────────────

With diff

-    y[pw, ph, c, n] = maximum(x[wstart:wend, hstart:hend, c, n])
+    m = typemin(T)
+    for j in hstart:hend
+      for i in wstart:wend
+        m = max(x[i, j, c, n])
+      end
+    end
+
+    y[pw, ph, c, n] = m
  end
end

 ───────────────────────────────────────────────────────────────────────────────
                                       Time                   Allocations
                               ──────────────────────   ───────────────────────
       Tot / % measured:            643ms / 88.6%            532MiB / 95.9%

Section                ncalls     time   %tot     avg     alloc   %tot      avg
───────────────────────────────────────────────────────────────────────────────
conv                       16    487ms  85.4%  30.4ms    510MiB  100%   31.8MiB
  conv2d!                  16    406ms  71.1%  25.3ms    396MiB  77.7%  24.8MiB
    im2col2d!              16    195ms  34.1%  12.2ms   39.6MiB  7.78%  2.48MiB
      im2col_2d!           16    182ms  31.9%  11.4ms         -  0.00%        -
    gemm                   16    122ms  21.3%  7.60ms         -  0.00%        -
dense                       3   78.3ms  13.7%  26.1ms   72.6KiB  0.01%  24.2KiB
maxpool2d!                  5   4.64ms  0.81%   928μs   1.53KiB  0.00%        -
  max_pooling2d_fwd!        5   4.64ms  0.81%   927μs         -  0.00%        -
───────────────────────────────────────────────────────────────────────────────