Merge pull request #81 from ROCmSoftwarePlatform/conv_batchnorm_opt_fix

Conv batchnorm opt fix

src/fwd_conv_batchnorm_rewrite.cpp

@@ -52,7 +52,8 @@ void fwd_conv_batchnorm_rewrite::apply(program& p) const
                             gamma2(k) / std::sqrt(variance2(k) + epsilon) * weights2(k, c, h, w);
                     });
                 dfor(new_bias.get_shape().elements())([&](std::size_t c) {
-                    new_bias2(c) = bias2(c) - (mean2(c) / std::sqrt(variance2(c) + epsilon));
+                    new_bias2(c) =
+                        bias2(c) - (gamma2(c) * mean2(c) / std::sqrt(variance2(c) + epsilon));
                 });
             });
         // Replace convolution instruction with updated weights

test/fwd_conv_batchnorm_rewrite_test.cpp

+#include <migraph/fwd_conv_batchnorm_rewrite.hpp>
+#include <migraph/program.hpp>
+#include <migraph/cpu/cpu_target.hpp>
+#include <migraph/operators.hpp>
+#include <migraph/instruction.hpp>
+#include <test.hpp>
+#include <migraph/verify.hpp>
+
+void fwd_conv_batchnorm_rewrite_test()
+{
+    std::vector<float> xdata = {
+        0.26485917, 0.61703885, 0.32762103, 0.2503367,  0.6552712,  0.07947932, 0.95442678,
+        0.70892651, 0.890563,   0.80808088, 0.89540492, 0.52657048, 0.94614791, 0.64371508,
+        0.0971229,  0.2475562,  0.47405955, 0.85538928, 0.05428386, 0.993078,   0.72771973,
+        0.18312255, 0.3091522,  0.51396558, 0.35158192, 0.2419852,  0.83691474, 0.36355352,
+        0.04769134, 0.08312604, 0.61804092, 0.0508887,  0.30987137, 0.81307629, 0.16398955,
+        0.69886166, 0.02415926, 0.60608918, 0.81907569, 0.13208211, 0.48303735, 0.87533734,
+        0.92998813, 0.65553674, 0.73223327, 0.99401001, 0.09850688, 0.76972609, 0.11118327,
+        0.04392097, 0.39252306, 0.91129653, 0.89078693, 0.60571206, 0.98410397, 0.15290698,
+        0.86992609, 0.7575111,  0.80583525, 0.23649562, 0.7478029,  0.62888878, 0.39886601,
+        0.37066793, 0.72627947, 0.8745595,  0.13568234, 0.7413787,  0.5039495,  0.18945697,
+        0.87046838, 0.63970494, 0.01124038, 0.27459063, 0.65745586, 0.69182619, 0.80470603,
+        0.58039348, 0.36950583, 0.43634225, 0.01694425, 0.14099377, 0.77015849, 0.35809292,
+        0.40547674, 0.46538817, 0.65835358, 0.2266954,  0.39057646, 0.64642207, 0.84491134,
+        0.20998067, 0.41074121, 0.73055221, 0.26424874, 0.10612507, 0.24478521, 0.24091282,
+        0.52536754, 0.57292341, 0.82190903, 0.51858515, 0.17162996, 0.52048114, 0.96624787,
+        0.17527163, 0.56384485, 0.91991603};
+    std::vector<float> wdata = {
+        -1.12125056, 0.50228441,  1.12719446,  -2.61705068, -0.2027315,  -0.82199441, 0.05337102,
+        -0.62146691, -2.40572931, -1.47175612, 1.49654601,  -1.07070376, -0.65908074, -0.28457694,
+        1.60046717,  0.20677642,  -1.51844486, 0.41203847,  -0.01285751, 0.07948031,  -0.91507006,
+        -1.59481079, -0.12856238, 0.39970482,  -1.89015158, 0.66969754,  0.10312618};
+    migraph::shape xs{migraph::shape::float_type, {1, 3, 6, 6}};
+    migraph::shape ws{migraph::shape::float_type, {1, 3, 3, 3}};
+    migraph::shape vars{migraph::shape::float_type, {1}};
+
+    auto create_program = [&]() {
+        migraph::program p;
+        auto x        = p.add_literal(xs, xdata);
+        auto w        = p.add_literal(ws, wdata);
+        auto conv     = p.add_instruction(migraph::op::convolution{{0, 0}, {1, 1}, {1, 1}}, x, w);
+        auto scale    = p.add_literal(migraph::literal{vars, {3.0f}});
+        auto bias     = p.add_literal(migraph::literal{vars, {8.1f}});
+        auto mean     = p.add_literal(migraph::literal{vars, {4.0f}});
+        auto variance = p.add_literal(migraph::literal{vars, {37.11f}});
+        p.add_instruction(migraph::op::batch_norm_inference{}, conv, scale, bias, mean, variance);
+        return p;
+    };
+
+    migraph::program p1 = create_program();
+    migraph::program p2 = create_program();
+    migraph::fwd_conv_batchnorm_rewrite opt;
+    opt.apply(p2);
+    p1.compile(migraph::cpu::cpu_target{});
+    p2.compile(migraph::cpu::cpu_target{});
+
+    auto result1 = p1.eval({});
+    auto result2 = p2.eval({});
+
+    std::vector<float> results_vector1;
+    std::vector<float> results_vector2;
+    result1.visit([&](auto output) { results_vector1.assign(output.begin(), output.end()); });
+    result2.visit([&](auto output) { results_vector2.assign(output.begin(), output.end()); });
+    EXPECT(migraph::verify_range(results_vector1, results_vector2));
+}
+
+int main()
+{
+    fwd_conv_batchnorm_rewrite_test();
+    return 0;
+}