requires all optimizer variables to be created at the first
        `apply_gradients()` or `minimize()` call. If your workflow calls
        optimizer to update different parts of model in multiple stages,
        please call `optimizer.build(model.trainable_variables)` before the
        training loop.
    *   **Performance regression on `ParameterServerStrategy`.** This could be

VSLDOI $8, T2, T1, T1 // VSLDB $8, T2, T1, T1 VADDCUQ T0, RED1, CAR1 // VACCQ T0, RED1, CAR1 VADDUQM T0, RED1, T0 // VAQ T0, RED1, T0 VADDECUQ T1, RED2, CAR1, CAR2 // VACCCQ T1, RED2, CAR1, CAR2 VADDEUQM T1, RED2, CAR1, T1 // VACQ T1, RED2, CAR1, T1 VADDUQM T2, CAR2, T2 // VAQ T2, CAR2, T2 // ---------- VSUBCUQ T0, PL, CAR1 // VSCBIQ PL, T0, CAR1 VSUBUQM T0, PL, TT0 // VSQ PL, T0, TT0 VSUBECUQ T1, PH, CAR1, CAR2 // VSBCBIQ T1, PH, CAR1, CAR2 VSUBEUQM T1, PH, CAR1, TT1 // VSBIQ T1, PH, CAR1, TT1 VSUBEUQM...

	SARL $1, (BX)                           // d13b
	SARL $1, (R11)                          // 41d13b
	SARL $1, DX                             // d1fa
	SARL $1, R11                            // 41d1fb
	SARL CL, (BX)                           // d33b
	SARL CL, (R11)                          // 41d33b
	SARL CL, DX                             // d3fa
	SARL CL, R11                            // 41d3fb